xref: /openbmc/linux/arch/x86/kernel/cpu/amd.c (revision e2ad626f)
1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/export.h>
3 #include <linux/bitops.h>
4 #include <linux/elf.h>
5 #include <linux/mm.h>
6 
7 #include <linux/io.h>
8 #include <linux/sched.h>
9 #include <linux/sched/clock.h>
10 #include <linux/random.h>
11 #include <linux/topology.h>
12 #include <asm/processor.h>
13 #include <asm/apic.h>
14 #include <asm/cacheinfo.h>
15 #include <asm/cpu.h>
16 #include <asm/spec-ctrl.h>
17 #include <asm/smp.h>
18 #include <asm/numa.h>
19 #include <asm/pci-direct.h>
20 #include <asm/delay.h>
21 #include <asm/debugreg.h>
22 #include <asm/resctrl.h>
23 
24 #ifdef CONFIG_X86_64
25 # include <asm/mmconfig.h>
26 #endif
27 
28 #include "cpu.h"
29 
30 /*
31  * nodes_per_socket: Stores the number of nodes per socket.
32  * Refer to Fam15h Models 00-0fh BKDG - CPUID Fn8000_001E_ECX
33  * Node Identifiers[10:8]
34  */
35 static u32 nodes_per_socket = 1;
36 
37 /*
38  * AMD errata checking
39  *
40  * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
41  * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
42  * have an OSVW id assigned, which it takes as first argument. Both take a
43  * variable number of family-specific model-stepping ranges created by
44  * AMD_MODEL_RANGE().
45  *
46  * Example:
47  *
48  * const int amd_erratum_319[] =
49  *	AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
50  *			   AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
51  *			   AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
52  */
53 
54 #define AMD_LEGACY_ERRATUM(...)		{ -1, __VA_ARGS__, 0 }
55 #define AMD_OSVW_ERRATUM(osvw_id, ...)	{ osvw_id, __VA_ARGS__, 0 }
56 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
57 	((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
58 #define AMD_MODEL_RANGE_FAMILY(range)	(((range) >> 24) & 0xff)
59 #define AMD_MODEL_RANGE_START(range)	(((range) >> 12) & 0xfff)
60 #define AMD_MODEL_RANGE_END(range)	((range) & 0xfff)
61 
62 static const int amd_erratum_400[] =
63 	AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
64 			    AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
65 
66 static const int amd_erratum_383[] =
67 	AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
68 
69 /* #1054: Instructions Retired Performance Counter May Be Inaccurate */
70 static const int amd_erratum_1054[] =
71 	AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x17, 0, 0, 0x2f, 0xf));
72 
73 static const int amd_zenbleed[] =
74 	AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x17, 0x30, 0x0, 0x4f, 0xf),
75 			   AMD_MODEL_RANGE(0x17, 0x60, 0x0, 0x7f, 0xf),
76 			   AMD_MODEL_RANGE(0x17, 0x90, 0x0, 0x91, 0xf),
77 			   AMD_MODEL_RANGE(0x17, 0xa0, 0x0, 0xaf, 0xf));
78 
79 static const int amd_div0[] =
80 	AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x17, 0x00, 0x0, 0x2f, 0xf),
81 			   AMD_MODEL_RANGE(0x17, 0x50, 0x0, 0x5f, 0xf));
82 
83 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
84 {
85 	int osvw_id = *erratum++;
86 	u32 range;
87 	u32 ms;
88 
89 	if (osvw_id >= 0 && osvw_id < 65536 &&
90 	    cpu_has(cpu, X86_FEATURE_OSVW)) {
91 		u64 osvw_len;
92 
93 		rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
94 		if (osvw_id < osvw_len) {
95 			u64 osvw_bits;
96 
97 			rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
98 			    osvw_bits);
99 			return osvw_bits & (1ULL << (osvw_id & 0x3f));
100 		}
101 	}
102 
103 	/* OSVW unavailable or ID unknown, match family-model-stepping range */
104 	ms = (cpu->x86_model << 4) | cpu->x86_stepping;
105 	while ((range = *erratum++))
106 		if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
107 		    (ms >= AMD_MODEL_RANGE_START(range)) &&
108 		    (ms <= AMD_MODEL_RANGE_END(range)))
109 			return true;
110 
111 	return false;
112 }
113 
114 static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
115 {
116 	u32 gprs[8] = { 0 };
117 	int err;
118 
119 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
120 		  "%s should only be used on K8!\n", __func__);
121 
122 	gprs[1] = msr;
123 	gprs[7] = 0x9c5a203a;
124 
125 	err = rdmsr_safe_regs(gprs);
126 
127 	*p = gprs[0] | ((u64)gprs[2] << 32);
128 
129 	return err;
130 }
131 
132 static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
133 {
134 	u32 gprs[8] = { 0 };
135 
136 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
137 		  "%s should only be used on K8!\n", __func__);
138 
139 	gprs[0] = (u32)val;
140 	gprs[1] = msr;
141 	gprs[2] = val >> 32;
142 	gprs[7] = 0x9c5a203a;
143 
144 	return wrmsr_safe_regs(gprs);
145 }
146 
147 /*
148  *	B step AMD K6 before B 9730xxxx have hardware bugs that can cause
149  *	misexecution of code under Linux. Owners of such processors should
150  *	contact AMD for precise details and a CPU swap.
151  *
152  *	See	http://www.multimania.com/poulot/k6bug.html
153  *	and	section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
154  *		(Publication # 21266  Issue Date: August 1998)
155  *
156  *	The following test is erm.. interesting. AMD neglected to up
157  *	the chip setting when fixing the bug but they also tweaked some
158  *	performance at the same time..
159  */
160 
161 #ifdef CONFIG_X86_32
162 extern __visible void vide(void);
163 __asm__(".text\n"
164 	".globl vide\n"
165 	".type vide, @function\n"
166 	".align 4\n"
167 	"vide: ret\n");
168 #endif
169 
170 static void init_amd_k5(struct cpuinfo_x86 *c)
171 {
172 #ifdef CONFIG_X86_32
173 /*
174  * General Systems BIOSen alias the cpu frequency registers
175  * of the Elan at 0x000df000. Unfortunately, one of the Linux
176  * drivers subsequently pokes it, and changes the CPU speed.
177  * Workaround : Remove the unneeded alias.
178  */
179 #define CBAR		(0xfffc) /* Configuration Base Address  (32-bit) */
180 #define CBAR_ENB	(0x80000000)
181 #define CBAR_KEY	(0X000000CB)
182 	if (c->x86_model == 9 || c->x86_model == 10) {
183 		if (inl(CBAR) & CBAR_ENB)
184 			outl(0 | CBAR_KEY, CBAR);
185 	}
186 #endif
187 }
188 
189 static void init_amd_k6(struct cpuinfo_x86 *c)
190 {
191 #ifdef CONFIG_X86_32
192 	u32 l, h;
193 	int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
194 
195 	if (c->x86_model < 6) {
196 		/* Based on AMD doc 20734R - June 2000 */
197 		if (c->x86_model == 0) {
198 			clear_cpu_cap(c, X86_FEATURE_APIC);
199 			set_cpu_cap(c, X86_FEATURE_PGE);
200 		}
201 		return;
202 	}
203 
204 	if (c->x86_model == 6 && c->x86_stepping == 1) {
205 		const int K6_BUG_LOOP = 1000000;
206 		int n;
207 		void (*f_vide)(void);
208 		u64 d, d2;
209 
210 		pr_info("AMD K6 stepping B detected - ");
211 
212 		/*
213 		 * It looks like AMD fixed the 2.6.2 bug and improved indirect
214 		 * calls at the same time.
215 		 */
216 
217 		n = K6_BUG_LOOP;
218 		f_vide = vide;
219 		OPTIMIZER_HIDE_VAR(f_vide);
220 		d = rdtsc();
221 		while (n--)
222 			f_vide();
223 		d2 = rdtsc();
224 		d = d2-d;
225 
226 		if (d > 20*K6_BUG_LOOP)
227 			pr_cont("system stability may be impaired when more than 32 MB are used.\n");
228 		else
229 			pr_cont("probably OK (after B9730xxxx).\n");
230 	}
231 
232 	/* K6 with old style WHCR */
233 	if (c->x86_model < 8 ||
234 	   (c->x86_model == 8 && c->x86_stepping < 8)) {
235 		/* We can only write allocate on the low 508Mb */
236 		if (mbytes > 508)
237 			mbytes = 508;
238 
239 		rdmsr(MSR_K6_WHCR, l, h);
240 		if ((l&0x0000FFFF) == 0) {
241 			unsigned long flags;
242 			l = (1<<0)|((mbytes/4)<<1);
243 			local_irq_save(flags);
244 			wbinvd();
245 			wrmsr(MSR_K6_WHCR, l, h);
246 			local_irq_restore(flags);
247 			pr_info("Enabling old style K6 write allocation for %d Mb\n",
248 				mbytes);
249 		}
250 		return;
251 	}
252 
253 	if ((c->x86_model == 8 && c->x86_stepping > 7) ||
254 	     c->x86_model == 9 || c->x86_model == 13) {
255 		/* The more serious chips .. */
256 
257 		if (mbytes > 4092)
258 			mbytes = 4092;
259 
260 		rdmsr(MSR_K6_WHCR, l, h);
261 		if ((l&0xFFFF0000) == 0) {
262 			unsigned long flags;
263 			l = ((mbytes>>2)<<22)|(1<<16);
264 			local_irq_save(flags);
265 			wbinvd();
266 			wrmsr(MSR_K6_WHCR, l, h);
267 			local_irq_restore(flags);
268 			pr_info("Enabling new style K6 write allocation for %d Mb\n",
269 				mbytes);
270 		}
271 
272 		return;
273 	}
274 
275 	if (c->x86_model == 10) {
276 		/* AMD Geode LX is model 10 */
277 		/* placeholder for any needed mods */
278 		return;
279 	}
280 #endif
281 }
282 
283 static void init_amd_k7(struct cpuinfo_x86 *c)
284 {
285 #ifdef CONFIG_X86_32
286 	u32 l, h;
287 
288 	/*
289 	 * Bit 15 of Athlon specific MSR 15, needs to be 0
290 	 * to enable SSE on Palomino/Morgan/Barton CPU's.
291 	 * If the BIOS didn't enable it already, enable it here.
292 	 */
293 	if (c->x86_model >= 6 && c->x86_model <= 10) {
294 		if (!cpu_has(c, X86_FEATURE_XMM)) {
295 			pr_info("Enabling disabled K7/SSE Support.\n");
296 			msr_clear_bit(MSR_K7_HWCR, 15);
297 			set_cpu_cap(c, X86_FEATURE_XMM);
298 		}
299 	}
300 
301 	/*
302 	 * It's been determined by AMD that Athlons since model 8 stepping 1
303 	 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
304 	 * As per AMD technical note 27212 0.2
305 	 */
306 	if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) {
307 		rdmsr(MSR_K7_CLK_CTL, l, h);
308 		if ((l & 0xfff00000) != 0x20000000) {
309 			pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
310 				l, ((l & 0x000fffff)|0x20000000));
311 			wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
312 		}
313 	}
314 
315 	/* calling is from identify_secondary_cpu() ? */
316 	if (!c->cpu_index)
317 		return;
318 
319 	/*
320 	 * Certain Athlons might work (for various values of 'work') in SMP
321 	 * but they are not certified as MP capable.
322 	 */
323 	/* Athlon 660/661 is valid. */
324 	if ((c->x86_model == 6) && ((c->x86_stepping == 0) ||
325 	    (c->x86_stepping == 1)))
326 		return;
327 
328 	/* Duron 670 is valid */
329 	if ((c->x86_model == 7) && (c->x86_stepping == 0))
330 		return;
331 
332 	/*
333 	 * Athlon 662, Duron 671, and Athlon >model 7 have capability
334 	 * bit. It's worth noting that the A5 stepping (662) of some
335 	 * Athlon XP's have the MP bit set.
336 	 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
337 	 * more.
338 	 */
339 	if (((c->x86_model == 6) && (c->x86_stepping >= 2)) ||
340 	    ((c->x86_model == 7) && (c->x86_stepping >= 1)) ||
341 	     (c->x86_model > 7))
342 		if (cpu_has(c, X86_FEATURE_MP))
343 			return;
344 
345 	/* If we get here, not a certified SMP capable AMD system. */
346 
347 	/*
348 	 * Don't taint if we are running SMP kernel on a single non-MP
349 	 * approved Athlon
350 	 */
351 	WARN_ONCE(1, "WARNING: This combination of AMD"
352 		" processors is not suitable for SMP.\n");
353 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
354 #endif
355 }
356 
357 #ifdef CONFIG_NUMA
358 /*
359  * To workaround broken NUMA config.  Read the comment in
360  * srat_detect_node().
361  */
362 static int nearby_node(int apicid)
363 {
364 	int i, node;
365 
366 	for (i = apicid - 1; i >= 0; i--) {
367 		node = __apicid_to_node[i];
368 		if (node != NUMA_NO_NODE && node_online(node))
369 			return node;
370 	}
371 	for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
372 		node = __apicid_to_node[i];
373 		if (node != NUMA_NO_NODE && node_online(node))
374 			return node;
375 	}
376 	return first_node(node_online_map); /* Shouldn't happen */
377 }
378 #endif
379 
380 /*
381  * Fix up cpu_core_id for pre-F17h systems to be in the
382  * [0 .. cores_per_node - 1] range. Not really needed but
383  * kept so as not to break existing setups.
384  */
385 static void legacy_fixup_core_id(struct cpuinfo_x86 *c)
386 {
387 	u32 cus_per_node;
388 
389 	if (c->x86 >= 0x17)
390 		return;
391 
392 	cus_per_node = c->x86_max_cores / nodes_per_socket;
393 	c->cpu_core_id %= cus_per_node;
394 }
395 
396 /*
397  * Fixup core topology information for
398  * (1) AMD multi-node processors
399  *     Assumption: Number of cores in each internal node is the same.
400  * (2) AMD processors supporting compute units
401  */
402 static void amd_get_topology(struct cpuinfo_x86 *c)
403 {
404 	int cpu = smp_processor_id();
405 
406 	/* get information required for multi-node processors */
407 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
408 		int err;
409 		u32 eax, ebx, ecx, edx;
410 
411 		cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
412 
413 		c->cpu_die_id  = ecx & 0xff;
414 
415 		if (c->x86 == 0x15)
416 			c->cu_id = ebx & 0xff;
417 
418 		if (c->x86 >= 0x17) {
419 			c->cpu_core_id = ebx & 0xff;
420 
421 			if (smp_num_siblings > 1)
422 				c->x86_max_cores /= smp_num_siblings;
423 		}
424 
425 		/*
426 		 * In case leaf B is available, use it to derive
427 		 * topology information.
428 		 */
429 		err = detect_extended_topology(c);
430 		if (!err)
431 			c->x86_coreid_bits = get_count_order(c->x86_max_cores);
432 
433 		cacheinfo_amd_init_llc_id(c, cpu);
434 
435 	} else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
436 		u64 value;
437 
438 		rdmsrl(MSR_FAM10H_NODE_ID, value);
439 		c->cpu_die_id = value & 7;
440 
441 		per_cpu(cpu_llc_id, cpu) = c->cpu_die_id;
442 	} else
443 		return;
444 
445 	if (nodes_per_socket > 1) {
446 		set_cpu_cap(c, X86_FEATURE_AMD_DCM);
447 		legacy_fixup_core_id(c);
448 	}
449 }
450 
451 /*
452  * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
453  * Assumes number of cores is a power of two.
454  */
455 static void amd_detect_cmp(struct cpuinfo_x86 *c)
456 {
457 	unsigned bits;
458 	int cpu = smp_processor_id();
459 
460 	bits = c->x86_coreid_bits;
461 	/* Low order bits define the core id (index of core in socket) */
462 	c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
463 	/* Convert the initial APIC ID into the socket ID */
464 	c->phys_proc_id = c->initial_apicid >> bits;
465 	/* use socket ID also for last level cache */
466 	per_cpu(cpu_llc_id, cpu) = c->cpu_die_id = c->phys_proc_id;
467 }
468 
469 u32 amd_get_nodes_per_socket(void)
470 {
471 	return nodes_per_socket;
472 }
473 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
474 
475 static void srat_detect_node(struct cpuinfo_x86 *c)
476 {
477 #ifdef CONFIG_NUMA
478 	int cpu = smp_processor_id();
479 	int node;
480 	unsigned apicid = c->apicid;
481 
482 	node = numa_cpu_node(cpu);
483 	if (node == NUMA_NO_NODE)
484 		node = get_llc_id(cpu);
485 
486 	/*
487 	 * On multi-fabric platform (e.g. Numascale NumaChip) a
488 	 * platform-specific handler needs to be called to fixup some
489 	 * IDs of the CPU.
490 	 */
491 	if (x86_cpuinit.fixup_cpu_id)
492 		x86_cpuinit.fixup_cpu_id(c, node);
493 
494 	if (!node_online(node)) {
495 		/*
496 		 * Two possibilities here:
497 		 *
498 		 * - The CPU is missing memory and no node was created.  In
499 		 *   that case try picking one from a nearby CPU.
500 		 *
501 		 * - The APIC IDs differ from the HyperTransport node IDs
502 		 *   which the K8 northbridge parsing fills in.  Assume
503 		 *   they are all increased by a constant offset, but in
504 		 *   the same order as the HT nodeids.  If that doesn't
505 		 *   result in a usable node fall back to the path for the
506 		 *   previous case.
507 		 *
508 		 * This workaround operates directly on the mapping between
509 		 * APIC ID and NUMA node, assuming certain relationship
510 		 * between APIC ID, HT node ID and NUMA topology.  As going
511 		 * through CPU mapping may alter the outcome, directly
512 		 * access __apicid_to_node[].
513 		 */
514 		int ht_nodeid = c->initial_apicid;
515 
516 		if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
517 			node = __apicid_to_node[ht_nodeid];
518 		/* Pick a nearby node */
519 		if (!node_online(node))
520 			node = nearby_node(apicid);
521 	}
522 	numa_set_node(cpu, node);
523 #endif
524 }
525 
526 static void early_init_amd_mc(struct cpuinfo_x86 *c)
527 {
528 #ifdef CONFIG_SMP
529 	unsigned bits, ecx;
530 
531 	/* Multi core CPU? */
532 	if (c->extended_cpuid_level < 0x80000008)
533 		return;
534 
535 	ecx = cpuid_ecx(0x80000008);
536 
537 	c->x86_max_cores = (ecx & 0xff) + 1;
538 
539 	/* CPU telling us the core id bits shift? */
540 	bits = (ecx >> 12) & 0xF;
541 
542 	/* Otherwise recompute */
543 	if (bits == 0) {
544 		while ((1 << bits) < c->x86_max_cores)
545 			bits++;
546 	}
547 
548 	c->x86_coreid_bits = bits;
549 #endif
550 }
551 
552 static void bsp_init_amd(struct cpuinfo_x86 *c)
553 {
554 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
555 
556 		if (c->x86 > 0x10 ||
557 		    (c->x86 == 0x10 && c->x86_model >= 0x2)) {
558 			u64 val;
559 
560 			rdmsrl(MSR_K7_HWCR, val);
561 			if (!(val & BIT(24)))
562 				pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
563 		}
564 	}
565 
566 	if (c->x86 == 0x15) {
567 		unsigned long upperbit;
568 		u32 cpuid, assoc;
569 
570 		cpuid	 = cpuid_edx(0x80000005);
571 		assoc	 = cpuid >> 16 & 0xff;
572 		upperbit = ((cpuid >> 24) << 10) / assoc;
573 
574 		va_align.mask	  = (upperbit - 1) & PAGE_MASK;
575 		va_align.flags    = ALIGN_VA_32 | ALIGN_VA_64;
576 
577 		/* A random value per boot for bit slice [12:upper_bit) */
578 		va_align.bits = get_random_u32() & va_align.mask;
579 	}
580 
581 	if (cpu_has(c, X86_FEATURE_MWAITX))
582 		use_mwaitx_delay();
583 
584 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
585 		u32 ecx;
586 
587 		ecx = cpuid_ecx(0x8000001e);
588 		__max_die_per_package = nodes_per_socket = ((ecx >> 8) & 7) + 1;
589 	} else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
590 		u64 value;
591 
592 		rdmsrl(MSR_FAM10H_NODE_ID, value);
593 		__max_die_per_package = nodes_per_socket = ((value >> 3) & 7) + 1;
594 	}
595 
596 	if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
597 	    !boot_cpu_has(X86_FEATURE_VIRT_SSBD) &&
598 	    c->x86 >= 0x15 && c->x86 <= 0x17) {
599 		unsigned int bit;
600 
601 		switch (c->x86) {
602 		case 0x15: bit = 54; break;
603 		case 0x16: bit = 33; break;
604 		case 0x17: bit = 10; break;
605 		default: return;
606 		}
607 		/*
608 		 * Try to cache the base value so further operations can
609 		 * avoid RMW. If that faults, do not enable SSBD.
610 		 */
611 		if (!rdmsrl_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
612 			setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
613 			setup_force_cpu_cap(X86_FEATURE_SSBD);
614 			x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
615 		}
616 	}
617 
618 	resctrl_cpu_detect(c);
619 }
620 
621 static void early_detect_mem_encrypt(struct cpuinfo_x86 *c)
622 {
623 	u64 msr;
624 
625 	/*
626 	 * BIOS support is required for SME and SEV.
627 	 *   For SME: If BIOS has enabled SME then adjust x86_phys_bits by
628 	 *	      the SME physical address space reduction value.
629 	 *	      If BIOS has not enabled SME then don't advertise the
630 	 *	      SME feature (set in scattered.c).
631 	 *	      If the kernel has not enabled SME via any means then
632 	 *	      don't advertise the SME feature.
633 	 *   For SEV: If BIOS has not enabled SEV then don't advertise the
634 	 *            SEV and SEV_ES feature (set in scattered.c).
635 	 *
636 	 *   In all cases, since support for SME and SEV requires long mode,
637 	 *   don't advertise the feature under CONFIG_X86_32.
638 	 */
639 	if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) {
640 		/* Check if memory encryption is enabled */
641 		rdmsrl(MSR_AMD64_SYSCFG, msr);
642 		if (!(msr & MSR_AMD64_SYSCFG_MEM_ENCRYPT))
643 			goto clear_all;
644 
645 		/*
646 		 * Always adjust physical address bits. Even though this
647 		 * will be a value above 32-bits this is still done for
648 		 * CONFIG_X86_32 so that accurate values are reported.
649 		 */
650 		c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
651 
652 		if (IS_ENABLED(CONFIG_X86_32))
653 			goto clear_all;
654 
655 		if (!sme_me_mask)
656 			setup_clear_cpu_cap(X86_FEATURE_SME);
657 
658 		rdmsrl(MSR_K7_HWCR, msr);
659 		if (!(msr & MSR_K7_HWCR_SMMLOCK))
660 			goto clear_sev;
661 
662 		return;
663 
664 clear_all:
665 		setup_clear_cpu_cap(X86_FEATURE_SME);
666 clear_sev:
667 		setup_clear_cpu_cap(X86_FEATURE_SEV);
668 		setup_clear_cpu_cap(X86_FEATURE_SEV_ES);
669 	}
670 }
671 
672 static void early_init_amd(struct cpuinfo_x86 *c)
673 {
674 	u64 value;
675 	u32 dummy;
676 
677 	early_init_amd_mc(c);
678 
679 	if (c->x86 >= 0xf)
680 		set_cpu_cap(c, X86_FEATURE_K8);
681 
682 	rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
683 
684 	/*
685 	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
686 	 * with P/T states and does not stop in deep C-states
687 	 */
688 	if (c->x86_power & (1 << 8)) {
689 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
690 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
691 	}
692 
693 	/* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
694 	if (c->x86_power & BIT(12))
695 		set_cpu_cap(c, X86_FEATURE_ACC_POWER);
696 
697 	/* Bit 14 indicates the Runtime Average Power Limit interface. */
698 	if (c->x86_power & BIT(14))
699 		set_cpu_cap(c, X86_FEATURE_RAPL);
700 
701 #ifdef CONFIG_X86_64
702 	set_cpu_cap(c, X86_FEATURE_SYSCALL32);
703 #else
704 	/*  Set MTRR capability flag if appropriate */
705 	if (c->x86 == 5)
706 		if (c->x86_model == 13 || c->x86_model == 9 ||
707 		    (c->x86_model == 8 && c->x86_stepping >= 8))
708 			set_cpu_cap(c, X86_FEATURE_K6_MTRR);
709 #endif
710 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
711 	/*
712 	 * ApicID can always be treated as an 8-bit value for AMD APIC versions
713 	 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
714 	 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
715 	 * after 16h.
716 	 */
717 	if (boot_cpu_has(X86_FEATURE_APIC)) {
718 		if (c->x86 > 0x16)
719 			set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
720 		else if (c->x86 >= 0xf) {
721 			/* check CPU config space for extended APIC ID */
722 			unsigned int val;
723 
724 			val = read_pci_config(0, 24, 0, 0x68);
725 			if ((val >> 17 & 0x3) == 0x3)
726 				set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
727 		}
728 	}
729 #endif
730 
731 	/*
732 	 * This is only needed to tell the kernel whether to use VMCALL
733 	 * and VMMCALL.  VMMCALL is never executed except under virt, so
734 	 * we can set it unconditionally.
735 	 */
736 	set_cpu_cap(c, X86_FEATURE_VMMCALL);
737 
738 	/* F16h erratum 793, CVE-2013-6885 */
739 	if (c->x86 == 0x16 && c->x86_model <= 0xf)
740 		msr_set_bit(MSR_AMD64_LS_CFG, 15);
741 
742 	/*
743 	 * Check whether the machine is affected by erratum 400. This is
744 	 * used to select the proper idle routine and to enable the check
745 	 * whether the machine is affected in arch_post_acpi_init(), which
746 	 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
747 	 */
748 	if (cpu_has_amd_erratum(c, amd_erratum_400))
749 		set_cpu_bug(c, X86_BUG_AMD_E400);
750 
751 	early_detect_mem_encrypt(c);
752 
753 	/* Re-enable TopologyExtensions if switched off by BIOS */
754 	if (c->x86 == 0x15 &&
755 	    (c->x86_model >= 0x10 && c->x86_model <= 0x6f) &&
756 	    !cpu_has(c, X86_FEATURE_TOPOEXT)) {
757 
758 		if (msr_set_bit(0xc0011005, 54) > 0) {
759 			rdmsrl(0xc0011005, value);
760 			if (value & BIT_64(54)) {
761 				set_cpu_cap(c, X86_FEATURE_TOPOEXT);
762 				pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
763 			}
764 		}
765 	}
766 
767 	if (cpu_has(c, X86_FEATURE_TOPOEXT))
768 		smp_num_siblings = ((cpuid_ebx(0x8000001e) >> 8) & 0xff) + 1;
769 }
770 
771 static void init_amd_k8(struct cpuinfo_x86 *c)
772 {
773 	u32 level;
774 	u64 value;
775 
776 	/* On C+ stepping K8 rep microcode works well for copy/memset */
777 	level = cpuid_eax(1);
778 	if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
779 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
780 
781 	/*
782 	 * Some BIOSes incorrectly force this feature, but only K8 revision D
783 	 * (model = 0x14) and later actually support it.
784 	 * (AMD Erratum #110, docId: 25759).
785 	 */
786 	if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
787 		clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
788 		if (!rdmsrl_amd_safe(0xc001100d, &value)) {
789 			value &= ~BIT_64(32);
790 			wrmsrl_amd_safe(0xc001100d, value);
791 		}
792 	}
793 
794 	if (!c->x86_model_id[0])
795 		strcpy(c->x86_model_id, "Hammer");
796 
797 #ifdef CONFIG_SMP
798 	/*
799 	 * Disable TLB flush filter by setting HWCR.FFDIS on K8
800 	 * bit 6 of msr C001_0015
801 	 *
802 	 * Errata 63 for SH-B3 steppings
803 	 * Errata 122 for all steppings (F+ have it disabled by default)
804 	 */
805 	msr_set_bit(MSR_K7_HWCR, 6);
806 #endif
807 	set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
808 }
809 
810 static void init_amd_gh(struct cpuinfo_x86 *c)
811 {
812 #ifdef CONFIG_MMCONF_FAM10H
813 	/* do this for boot cpu */
814 	if (c == &boot_cpu_data)
815 		check_enable_amd_mmconf_dmi();
816 
817 	fam10h_check_enable_mmcfg();
818 #endif
819 
820 	/*
821 	 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
822 	 * is always needed when GART is enabled, even in a kernel which has no
823 	 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
824 	 * If it doesn't, we do it here as suggested by the BKDG.
825 	 *
826 	 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
827 	 */
828 	msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
829 
830 	/*
831 	 * On family 10h BIOS may not have properly enabled WC+ support, causing
832 	 * it to be converted to CD memtype. This may result in performance
833 	 * degradation for certain nested-paging guests. Prevent this conversion
834 	 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
835 	 *
836 	 * NOTE: we want to use the _safe accessors so as not to #GP kvm
837 	 * guests on older kvm hosts.
838 	 */
839 	msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
840 
841 	if (cpu_has_amd_erratum(c, amd_erratum_383))
842 		set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
843 }
844 
845 static void init_amd_ln(struct cpuinfo_x86 *c)
846 {
847 	/*
848 	 * Apply erratum 665 fix unconditionally so machines without a BIOS
849 	 * fix work.
850 	 */
851 	msr_set_bit(MSR_AMD64_DE_CFG, 31);
852 }
853 
854 static bool rdrand_force;
855 
856 static int __init rdrand_cmdline(char *str)
857 {
858 	if (!str)
859 		return -EINVAL;
860 
861 	if (!strcmp(str, "force"))
862 		rdrand_force = true;
863 	else
864 		return -EINVAL;
865 
866 	return 0;
867 }
868 early_param("rdrand", rdrand_cmdline);
869 
870 static void clear_rdrand_cpuid_bit(struct cpuinfo_x86 *c)
871 {
872 	/*
873 	 * Saving of the MSR used to hide the RDRAND support during
874 	 * suspend/resume is done by arch/x86/power/cpu.c, which is
875 	 * dependent on CONFIG_PM_SLEEP.
876 	 */
877 	if (!IS_ENABLED(CONFIG_PM_SLEEP))
878 		return;
879 
880 	/*
881 	 * The self-test can clear X86_FEATURE_RDRAND, so check for
882 	 * RDRAND support using the CPUID function directly.
883 	 */
884 	if (!(cpuid_ecx(1) & BIT(30)) || rdrand_force)
885 		return;
886 
887 	msr_clear_bit(MSR_AMD64_CPUID_FN_1, 62);
888 
889 	/*
890 	 * Verify that the CPUID change has occurred in case the kernel is
891 	 * running virtualized and the hypervisor doesn't support the MSR.
892 	 */
893 	if (cpuid_ecx(1) & BIT(30)) {
894 		pr_info_once("BIOS may not properly restore RDRAND after suspend, but hypervisor does not support hiding RDRAND via CPUID.\n");
895 		return;
896 	}
897 
898 	clear_cpu_cap(c, X86_FEATURE_RDRAND);
899 	pr_info_once("BIOS may not properly restore RDRAND after suspend, hiding RDRAND via CPUID. Use rdrand=force to reenable.\n");
900 }
901 
902 static void init_amd_jg(struct cpuinfo_x86 *c)
903 {
904 	/*
905 	 * Some BIOS implementations do not restore proper RDRAND support
906 	 * across suspend and resume. Check on whether to hide the RDRAND
907 	 * instruction support via CPUID.
908 	 */
909 	clear_rdrand_cpuid_bit(c);
910 }
911 
912 static void init_amd_bd(struct cpuinfo_x86 *c)
913 {
914 	u64 value;
915 
916 	/*
917 	 * The way access filter has a performance penalty on some workloads.
918 	 * Disable it on the affected CPUs.
919 	 */
920 	if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
921 		if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
922 			value |= 0x1E;
923 			wrmsrl_safe(MSR_F15H_IC_CFG, value);
924 		}
925 	}
926 
927 	/*
928 	 * Some BIOS implementations do not restore proper RDRAND support
929 	 * across suspend and resume. Check on whether to hide the RDRAND
930 	 * instruction support via CPUID.
931 	 */
932 	clear_rdrand_cpuid_bit(c);
933 }
934 
935 void init_spectral_chicken(struct cpuinfo_x86 *c)
936 {
937 #ifdef CONFIG_CPU_UNRET_ENTRY
938 	u64 value;
939 
940 	/*
941 	 * On Zen2 we offer this chicken (bit) on the altar of Speculation.
942 	 *
943 	 * This suppresses speculation from the middle of a basic block, i.e. it
944 	 * suppresses non-branch predictions.
945 	 *
946 	 * We use STIBP as a heuristic to filter out Zen2 from the rest of F17H
947 	 */
948 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR) && cpu_has(c, X86_FEATURE_AMD_STIBP)) {
949 		if (!rdmsrl_safe(MSR_ZEN2_SPECTRAL_CHICKEN, &value)) {
950 			value |= MSR_ZEN2_SPECTRAL_CHICKEN_BIT;
951 			wrmsrl_safe(MSR_ZEN2_SPECTRAL_CHICKEN, value);
952 		}
953 	}
954 #endif
955 	/*
956 	 * Work around Erratum 1386.  The XSAVES instruction malfunctions in
957 	 * certain circumstances on Zen1/2 uarch, and not all parts have had
958 	 * updated microcode at the time of writing (March 2023).
959 	 *
960 	 * Affected parts all have no supervisor XSAVE states, meaning that
961 	 * the XSAVEC instruction (which works fine) is equivalent.
962 	 */
963 	clear_cpu_cap(c, X86_FEATURE_XSAVES);
964 }
965 
966 static void init_amd_zn(struct cpuinfo_x86 *c)
967 {
968 	set_cpu_cap(c, X86_FEATURE_ZEN);
969 
970 #ifdef CONFIG_NUMA
971 	node_reclaim_distance = 32;
972 #endif
973 
974 	/* Fix up CPUID bits, but only if not virtualised. */
975 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR)) {
976 
977 		/* Erratum 1076: CPB feature bit not being set in CPUID. */
978 		if (!cpu_has(c, X86_FEATURE_CPB))
979 			set_cpu_cap(c, X86_FEATURE_CPB);
980 
981 		/*
982 		 * Zen3 (Fam19 model < 0x10) parts are not susceptible to
983 		 * Branch Type Confusion, but predate the allocation of the
984 		 * BTC_NO bit.
985 		 */
986 		if (c->x86 == 0x19 && !cpu_has(c, X86_FEATURE_BTC_NO))
987 			set_cpu_cap(c, X86_FEATURE_BTC_NO);
988 	}
989 }
990 
991 static bool cpu_has_zenbleed_microcode(void)
992 {
993 	u32 good_rev = 0;
994 
995 	switch (boot_cpu_data.x86_model) {
996 	case 0x30 ... 0x3f: good_rev = 0x0830107a; break;
997 	case 0x60 ... 0x67: good_rev = 0x0860010b; break;
998 	case 0x68 ... 0x6f: good_rev = 0x08608105; break;
999 	case 0x70 ... 0x7f: good_rev = 0x08701032; break;
1000 	case 0xa0 ... 0xaf: good_rev = 0x08a00008; break;
1001 
1002 	default:
1003 		return false;
1004 		break;
1005 	}
1006 
1007 	if (boot_cpu_data.microcode < good_rev)
1008 		return false;
1009 
1010 	return true;
1011 }
1012 
1013 static void zenbleed_check(struct cpuinfo_x86 *c)
1014 {
1015 	if (!cpu_has_amd_erratum(c, amd_zenbleed))
1016 		return;
1017 
1018 	if (cpu_has(c, X86_FEATURE_HYPERVISOR))
1019 		return;
1020 
1021 	if (!cpu_has(c, X86_FEATURE_AVX))
1022 		return;
1023 
1024 	if (!cpu_has_zenbleed_microcode()) {
1025 		pr_notice_once("Zenbleed: please update your microcode for the most optimal fix\n");
1026 		msr_set_bit(MSR_AMD64_DE_CFG, MSR_AMD64_DE_CFG_ZEN2_FP_BACKUP_FIX_BIT);
1027 	} else {
1028 		msr_clear_bit(MSR_AMD64_DE_CFG, MSR_AMD64_DE_CFG_ZEN2_FP_BACKUP_FIX_BIT);
1029 	}
1030 }
1031 
1032 static void init_amd(struct cpuinfo_x86 *c)
1033 {
1034 	early_init_amd(c);
1035 
1036 	/*
1037 	 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
1038 	 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
1039 	 */
1040 	clear_cpu_cap(c, 0*32+31);
1041 
1042 	if (c->x86 >= 0x10)
1043 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
1044 
1045 	/* AMD FSRM also implies FSRS */
1046 	if (cpu_has(c, X86_FEATURE_FSRM))
1047 		set_cpu_cap(c, X86_FEATURE_FSRS);
1048 
1049 	/* get apicid instead of initial apic id from cpuid */
1050 	c->apicid = read_apic_id();
1051 
1052 	/* K6s reports MCEs but don't actually have all the MSRs */
1053 	if (c->x86 < 6)
1054 		clear_cpu_cap(c, X86_FEATURE_MCE);
1055 
1056 	switch (c->x86) {
1057 	case 4:    init_amd_k5(c); break;
1058 	case 5:    init_amd_k6(c); break;
1059 	case 6:	   init_amd_k7(c); break;
1060 	case 0xf:  init_amd_k8(c); break;
1061 	case 0x10: init_amd_gh(c); break;
1062 	case 0x12: init_amd_ln(c); break;
1063 	case 0x15: init_amd_bd(c); break;
1064 	case 0x16: init_amd_jg(c); break;
1065 	case 0x17: init_spectral_chicken(c);
1066 		   fallthrough;
1067 	case 0x19: init_amd_zn(c); break;
1068 	}
1069 
1070 	/*
1071 	 * Enable workaround for FXSAVE leak on CPUs
1072 	 * without a XSaveErPtr feature
1073 	 */
1074 	if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
1075 		set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
1076 
1077 	cpu_detect_cache_sizes(c);
1078 
1079 	amd_detect_cmp(c);
1080 	amd_get_topology(c);
1081 	srat_detect_node(c);
1082 
1083 	init_amd_cacheinfo(c);
1084 
1085 	if (!cpu_has(c, X86_FEATURE_LFENCE_RDTSC) && cpu_has(c, X86_FEATURE_XMM2)) {
1086 		/*
1087 		 * Use LFENCE for execution serialization.  On families which
1088 		 * don't have that MSR, LFENCE is already serializing.
1089 		 * msr_set_bit() uses the safe accessors, too, even if the MSR
1090 		 * is not present.
1091 		 */
1092 		msr_set_bit(MSR_AMD64_DE_CFG,
1093 			    MSR_AMD64_DE_CFG_LFENCE_SERIALIZE_BIT);
1094 
1095 		/* A serializing LFENCE stops RDTSC speculation */
1096 		set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
1097 	}
1098 
1099 	/*
1100 	 * Family 0x12 and above processors have APIC timer
1101 	 * running in deep C states.
1102 	 */
1103 	if (c->x86 > 0x11)
1104 		set_cpu_cap(c, X86_FEATURE_ARAT);
1105 
1106 	/* 3DNow or LM implies PREFETCHW */
1107 	if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
1108 		if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
1109 			set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
1110 
1111 	/* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
1112 	if (!cpu_feature_enabled(X86_FEATURE_XENPV))
1113 		set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
1114 
1115 	/*
1116 	 * Turn on the Instructions Retired free counter on machines not
1117 	 * susceptible to erratum #1054 "Instructions Retired Performance
1118 	 * Counter May Be Inaccurate".
1119 	 */
1120 	if (cpu_has(c, X86_FEATURE_IRPERF) &&
1121 	    !cpu_has_amd_erratum(c, amd_erratum_1054))
1122 		msr_set_bit(MSR_K7_HWCR, MSR_K7_HWCR_IRPERF_EN_BIT);
1123 
1124 	check_null_seg_clears_base(c);
1125 
1126 	/*
1127 	 * Make sure EFER[AIBRSE - Automatic IBRS Enable] is set. The APs are brought up
1128 	 * using the trampoline code and as part of it, MSR_EFER gets prepared there in
1129 	 * order to be replicated onto them. Regardless, set it here again, if not set,
1130 	 * to protect against any future refactoring/code reorganization which might
1131 	 * miss setting this important bit.
1132 	 */
1133 	if (spectre_v2_in_eibrs_mode(spectre_v2_enabled) &&
1134 	    cpu_has(c, X86_FEATURE_AUTOIBRS))
1135 		WARN_ON_ONCE(msr_set_bit(MSR_EFER, _EFER_AUTOIBRS));
1136 
1137 	zenbleed_check(c);
1138 
1139 	if (cpu_has_amd_erratum(c, amd_div0)) {
1140 		pr_notice_once("AMD Zen1 DIV0 bug detected. Disable SMT for full protection.\n");
1141 		setup_force_cpu_bug(X86_BUG_DIV0);
1142 	}
1143 }
1144 
1145 #ifdef CONFIG_X86_32
1146 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
1147 {
1148 	/* AMD errata T13 (order #21922) */
1149 	if (c->x86 == 6) {
1150 		/* Duron Rev A0 */
1151 		if (c->x86_model == 3 && c->x86_stepping == 0)
1152 			size = 64;
1153 		/* Tbird rev A1/A2 */
1154 		if (c->x86_model == 4 &&
1155 			(c->x86_stepping == 0 || c->x86_stepping == 1))
1156 			size = 256;
1157 	}
1158 	return size;
1159 }
1160 #endif
1161 
1162 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
1163 {
1164 	u32 ebx, eax, ecx, edx;
1165 	u16 mask = 0xfff;
1166 
1167 	if (c->x86 < 0xf)
1168 		return;
1169 
1170 	if (c->extended_cpuid_level < 0x80000006)
1171 		return;
1172 
1173 	cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
1174 
1175 	tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
1176 	tlb_lli_4k[ENTRIES] = ebx & mask;
1177 
1178 	/*
1179 	 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
1180 	 * characteristics from the CPUID function 0x80000005 instead.
1181 	 */
1182 	if (c->x86 == 0xf) {
1183 		cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1184 		mask = 0xff;
1185 	}
1186 
1187 	/* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1188 	if (!((eax >> 16) & mask))
1189 		tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
1190 	else
1191 		tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
1192 
1193 	/* a 4M entry uses two 2M entries */
1194 	tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
1195 
1196 	/* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1197 	if (!(eax & mask)) {
1198 		/* Erratum 658 */
1199 		if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
1200 			tlb_lli_2m[ENTRIES] = 1024;
1201 		} else {
1202 			cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1203 			tlb_lli_2m[ENTRIES] = eax & 0xff;
1204 		}
1205 	} else
1206 		tlb_lli_2m[ENTRIES] = eax & mask;
1207 
1208 	tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
1209 }
1210 
1211 static const struct cpu_dev amd_cpu_dev = {
1212 	.c_vendor	= "AMD",
1213 	.c_ident	= { "AuthenticAMD" },
1214 #ifdef CONFIG_X86_32
1215 	.legacy_models = {
1216 		{ .family = 4, .model_names =
1217 		  {
1218 			  [3] = "486 DX/2",
1219 			  [7] = "486 DX/2-WB",
1220 			  [8] = "486 DX/4",
1221 			  [9] = "486 DX/4-WB",
1222 			  [14] = "Am5x86-WT",
1223 			  [15] = "Am5x86-WB"
1224 		  }
1225 		},
1226 	},
1227 	.legacy_cache_size = amd_size_cache,
1228 #endif
1229 	.c_early_init   = early_init_amd,
1230 	.c_detect_tlb	= cpu_detect_tlb_amd,
1231 	.c_bsp_init	= bsp_init_amd,
1232 	.c_init		= init_amd,
1233 	.c_x86_vendor	= X86_VENDOR_AMD,
1234 };
1235 
1236 cpu_dev_register(amd_cpu_dev);
1237 
1238 static DEFINE_PER_CPU_READ_MOSTLY(unsigned long[4], amd_dr_addr_mask);
1239 
1240 static unsigned int amd_msr_dr_addr_masks[] = {
1241 	MSR_F16H_DR0_ADDR_MASK,
1242 	MSR_F16H_DR1_ADDR_MASK,
1243 	MSR_F16H_DR1_ADDR_MASK + 1,
1244 	MSR_F16H_DR1_ADDR_MASK + 2
1245 };
1246 
1247 void amd_set_dr_addr_mask(unsigned long mask, unsigned int dr)
1248 {
1249 	int cpu = smp_processor_id();
1250 
1251 	if (!cpu_feature_enabled(X86_FEATURE_BPEXT))
1252 		return;
1253 
1254 	if (WARN_ON_ONCE(dr >= ARRAY_SIZE(amd_msr_dr_addr_masks)))
1255 		return;
1256 
1257 	if (per_cpu(amd_dr_addr_mask, cpu)[dr] == mask)
1258 		return;
1259 
1260 	wrmsr(amd_msr_dr_addr_masks[dr], mask, 0);
1261 	per_cpu(amd_dr_addr_mask, cpu)[dr] = mask;
1262 }
1263 
1264 unsigned long amd_get_dr_addr_mask(unsigned int dr)
1265 {
1266 	if (!cpu_feature_enabled(X86_FEATURE_BPEXT))
1267 		return 0;
1268 
1269 	if (WARN_ON_ONCE(dr >= ARRAY_SIZE(amd_msr_dr_addr_masks)))
1270 		return 0;
1271 
1272 	return per_cpu(amd_dr_addr_mask[dr], smp_processor_id());
1273 }
1274 EXPORT_SYMBOL_GPL(amd_get_dr_addr_mask);
1275 
1276 u32 amd_get_highest_perf(void)
1277 {
1278 	struct cpuinfo_x86 *c = &boot_cpu_data;
1279 
1280 	if (c->x86 == 0x17 && ((c->x86_model >= 0x30 && c->x86_model < 0x40) ||
1281 			       (c->x86_model >= 0x70 && c->x86_model < 0x80)))
1282 		return 166;
1283 
1284 	if (c->x86 == 0x19 && ((c->x86_model >= 0x20 && c->x86_model < 0x30) ||
1285 			       (c->x86_model >= 0x40 && c->x86_model < 0x70)))
1286 		return 166;
1287 
1288 	return 255;
1289 }
1290 EXPORT_SYMBOL_GPL(amd_get_highest_perf);
1291 
1292 static void zenbleed_check_cpu(void *unused)
1293 {
1294 	struct cpuinfo_x86 *c = &cpu_data(smp_processor_id());
1295 
1296 	zenbleed_check(c);
1297 }
1298 
1299 void amd_check_microcode(void)
1300 {
1301 	on_each_cpu(zenbleed_check_cpu, NULL, 1);
1302 }
1303 
1304 bool cpu_has_ibpb_brtype_microcode(void)
1305 {
1306 	switch (boot_cpu_data.x86) {
1307 	/* Zen1/2 IBPB flushes branch type predictions too. */
1308 	case 0x17:
1309 		return boot_cpu_has(X86_FEATURE_AMD_IBPB);
1310 	case 0x19:
1311 		/* Poke the MSR bit on Zen3/4 to check its presence. */
1312 		if (!wrmsrl_safe(MSR_IA32_PRED_CMD, PRED_CMD_SBPB)) {
1313 			setup_force_cpu_cap(X86_FEATURE_SBPB);
1314 			return true;
1315 		} else {
1316 			return false;
1317 		}
1318 	default:
1319 		return false;
1320 	}
1321 }
1322 
1323 /*
1324  * Issue a DIV 0/1 insn to clear any division data from previous DIV
1325  * operations.
1326  */
1327 void noinstr amd_clear_divider(void)
1328 {
1329 	asm volatile(ALTERNATIVE("", "div %2\n\t", X86_BUG_DIV0)
1330 		     :: "a" (0), "d" (0), "r" (1));
1331 }
1332 EXPORT_SYMBOL_GPL(amd_clear_divider);
1333