xref: /openbmc/linux/arch/x86/kernel/cpu/amd.c (revision 4d016ae42efb214d4b441b0654771ddf34c72891)
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, 0xa0, 0x0, 0xaf, 0xf));
77 
78 static const int amd_div0[] =
79 	AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x17, 0x00, 0x0, 0x2f, 0xf),
80 			   AMD_MODEL_RANGE(0x17, 0x50, 0x0, 0x5f, 0xf));
81 
82 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
83 {
84 	int osvw_id = *erratum++;
85 	u32 range;
86 	u32 ms;
87 
88 	if (osvw_id >= 0 && osvw_id < 65536 &&
89 	    cpu_has(cpu, X86_FEATURE_OSVW)) {
90 		u64 osvw_len;
91 
92 		rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
93 		if (osvw_id < osvw_len) {
94 			u64 osvw_bits;
95 
96 			rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
97 			    osvw_bits);
98 			return osvw_bits & (1ULL << (osvw_id & 0x3f));
99 		}
100 	}
101 
102 	/* OSVW unavailable or ID unknown, match family-model-stepping range */
103 	ms = (cpu->x86_model << 4) | cpu->x86_stepping;
104 	while ((range = *erratum++))
105 		if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
106 		    (ms >= AMD_MODEL_RANGE_START(range)) &&
107 		    (ms <= AMD_MODEL_RANGE_END(range)))
108 			return true;
109 
110 	return false;
111 }
112 
113 static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
114 {
115 	u32 gprs[8] = { 0 };
116 	int err;
117 
118 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
119 		  "%s should only be used on K8!\n", __func__);
120 
121 	gprs[1] = msr;
122 	gprs[7] = 0x9c5a203a;
123 
124 	err = rdmsr_safe_regs(gprs);
125 
126 	*p = gprs[0] | ((u64)gprs[2] << 32);
127 
128 	return err;
129 }
130 
131 static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
132 {
133 	u32 gprs[8] = { 0 };
134 
135 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
136 		  "%s should only be used on K8!\n", __func__);
137 
138 	gprs[0] = (u32)val;
139 	gprs[1] = msr;
140 	gprs[2] = val >> 32;
141 	gprs[7] = 0x9c5a203a;
142 
143 	return wrmsr_safe_regs(gprs);
144 }
145 
146 /*
147  *	B step AMD K6 before B 9730xxxx have hardware bugs that can cause
148  *	misexecution of code under Linux. Owners of such processors should
149  *	contact AMD for precise details and a CPU swap.
150  *
151  *	See	http://www.multimania.com/poulot/k6bug.html
152  *	and	section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
153  *		(Publication # 21266  Issue Date: August 1998)
154  *
155  *	The following test is erm.. interesting. AMD neglected to up
156  *	the chip setting when fixing the bug but they also tweaked some
157  *	performance at the same time..
158  */
159 
160 #ifdef CONFIG_X86_32
161 extern __visible void vide(void);
162 __asm__(".text\n"
163 	".globl vide\n"
164 	".type vide, @function\n"
165 	".align 4\n"
166 	"vide: ret\n");
167 #endif
168 
169 static void init_amd_k5(struct cpuinfo_x86 *c)
170 {
171 #ifdef CONFIG_X86_32
172 /*
173  * General Systems BIOSen alias the cpu frequency registers
174  * of the Elan at 0x000df000. Unfortunately, one of the Linux
175  * drivers subsequently pokes it, and changes the CPU speed.
176  * Workaround : Remove the unneeded alias.
177  */
178 #define CBAR		(0xfffc) /* Configuration Base Address  (32-bit) */
179 #define CBAR_ENB	(0x80000000)
180 #define CBAR_KEY	(0X000000CB)
181 	if (c->x86_model == 9 || c->x86_model == 10) {
182 		if (inl(CBAR) & CBAR_ENB)
183 			outl(0 | CBAR_KEY, CBAR);
184 	}
185 #endif
186 }
187 
188 static void init_amd_k6(struct cpuinfo_x86 *c)
189 {
190 #ifdef CONFIG_X86_32
191 	u32 l, h;
192 	int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
193 
194 	if (c->x86_model < 6) {
195 		/* Based on AMD doc 20734R - June 2000 */
196 		if (c->x86_model == 0) {
197 			clear_cpu_cap(c, X86_FEATURE_APIC);
198 			set_cpu_cap(c, X86_FEATURE_PGE);
199 		}
200 		return;
201 	}
202 
203 	if (c->x86_model == 6 && c->x86_stepping == 1) {
204 		const int K6_BUG_LOOP = 1000000;
205 		int n;
206 		void (*f_vide)(void);
207 		u64 d, d2;
208 
209 		pr_info("AMD K6 stepping B detected - ");
210 
211 		/*
212 		 * It looks like AMD fixed the 2.6.2 bug and improved indirect
213 		 * calls at the same time.
214 		 */
215 
216 		n = K6_BUG_LOOP;
217 		f_vide = vide;
218 		OPTIMIZER_HIDE_VAR(f_vide);
219 		d = rdtsc();
220 		while (n--)
221 			f_vide();
222 		d2 = rdtsc();
223 		d = d2-d;
224 
225 		if (d > 20*K6_BUG_LOOP)
226 			pr_cont("system stability may be impaired when more than 32 MB are used.\n");
227 		else
228 			pr_cont("probably OK (after B9730xxxx).\n");
229 	}
230 
231 	/* K6 with old style WHCR */
232 	if (c->x86_model < 8 ||
233 	   (c->x86_model == 8 && c->x86_stepping < 8)) {
234 		/* We can only write allocate on the low 508Mb */
235 		if (mbytes > 508)
236 			mbytes = 508;
237 
238 		rdmsr(MSR_K6_WHCR, l, h);
239 		if ((l&0x0000FFFF) == 0) {
240 			unsigned long flags;
241 			l = (1<<0)|((mbytes/4)<<1);
242 			local_irq_save(flags);
243 			wbinvd();
244 			wrmsr(MSR_K6_WHCR, l, h);
245 			local_irq_restore(flags);
246 			pr_info("Enabling old style K6 write allocation for %d Mb\n",
247 				mbytes);
248 		}
249 		return;
250 	}
251 
252 	if ((c->x86_model == 8 && c->x86_stepping > 7) ||
253 	     c->x86_model == 9 || c->x86_model == 13) {
254 		/* The more serious chips .. */
255 
256 		if (mbytes > 4092)
257 			mbytes = 4092;
258 
259 		rdmsr(MSR_K6_WHCR, l, h);
260 		if ((l&0xFFFF0000) == 0) {
261 			unsigned long flags;
262 			l = ((mbytes>>2)<<22)|(1<<16);
263 			local_irq_save(flags);
264 			wbinvd();
265 			wrmsr(MSR_K6_WHCR, l, h);
266 			local_irq_restore(flags);
267 			pr_info("Enabling new style K6 write allocation for %d Mb\n",
268 				mbytes);
269 		}
270 
271 		return;
272 	}
273 
274 	if (c->x86_model == 10) {
275 		/* AMD Geode LX is model 10 */
276 		/* placeholder for any needed mods */
277 		return;
278 	}
279 #endif
280 }
281 
282 static void init_amd_k7(struct cpuinfo_x86 *c)
283 {
284 #ifdef CONFIG_X86_32
285 	u32 l, h;
286 
287 	/*
288 	 * Bit 15 of Athlon specific MSR 15, needs to be 0
289 	 * to enable SSE on Palomino/Morgan/Barton CPU's.
290 	 * If the BIOS didn't enable it already, enable it here.
291 	 */
292 	if (c->x86_model >= 6 && c->x86_model <= 10) {
293 		if (!cpu_has(c, X86_FEATURE_XMM)) {
294 			pr_info("Enabling disabled K7/SSE Support.\n");
295 			msr_clear_bit(MSR_K7_HWCR, 15);
296 			set_cpu_cap(c, X86_FEATURE_XMM);
297 		}
298 	}
299 
300 	/*
301 	 * It's been determined by AMD that Athlons since model 8 stepping 1
302 	 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
303 	 * As per AMD technical note 27212 0.2
304 	 */
305 	if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) {
306 		rdmsr(MSR_K7_CLK_CTL, l, h);
307 		if ((l & 0xfff00000) != 0x20000000) {
308 			pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
309 				l, ((l & 0x000fffff)|0x20000000));
310 			wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
311 		}
312 	}
313 
314 	/* calling is from identify_secondary_cpu() ? */
315 	if (!c->cpu_index)
316 		return;
317 
318 	/*
319 	 * Certain Athlons might work (for various values of 'work') in SMP
320 	 * but they are not certified as MP capable.
321 	 */
322 	/* Athlon 660/661 is valid. */
323 	if ((c->x86_model == 6) && ((c->x86_stepping == 0) ||
324 	    (c->x86_stepping == 1)))
325 		return;
326 
327 	/* Duron 670 is valid */
328 	if ((c->x86_model == 7) && (c->x86_stepping == 0))
329 		return;
330 
331 	/*
332 	 * Athlon 662, Duron 671, and Athlon >model 7 have capability
333 	 * bit. It's worth noting that the A5 stepping (662) of some
334 	 * Athlon XP's have the MP bit set.
335 	 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
336 	 * more.
337 	 */
338 	if (((c->x86_model == 6) && (c->x86_stepping >= 2)) ||
339 	    ((c->x86_model == 7) && (c->x86_stepping >= 1)) ||
340 	     (c->x86_model > 7))
341 		if (cpu_has(c, X86_FEATURE_MP))
342 			return;
343 
344 	/* If we get here, not a certified SMP capable AMD system. */
345 
346 	/*
347 	 * Don't taint if we are running SMP kernel on a single non-MP
348 	 * approved Athlon
349 	 */
350 	WARN_ONCE(1, "WARNING: This combination of AMD"
351 		" processors is not suitable for SMP.\n");
352 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
353 #endif
354 }
355 
356 #ifdef CONFIG_NUMA
357 /*
358  * To workaround broken NUMA config.  Read the comment in
359  * srat_detect_node().
360  */
361 static int nearby_node(int apicid)
362 {
363 	int i, node;
364 
365 	for (i = apicid - 1; i >= 0; i--) {
366 		node = __apicid_to_node[i];
367 		if (node != NUMA_NO_NODE && node_online(node))
368 			return node;
369 	}
370 	for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
371 		node = __apicid_to_node[i];
372 		if (node != NUMA_NO_NODE && node_online(node))
373 			return node;
374 	}
375 	return first_node(node_online_map); /* Shouldn't happen */
376 }
377 #endif
378 
379 /*
380  * Fix up cpu_core_id for pre-F17h systems to be in the
381  * [0 .. cores_per_node - 1] range. Not really needed but
382  * kept so as not to break existing setups.
383  */
384 static void legacy_fixup_core_id(struct cpuinfo_x86 *c)
385 {
386 	u32 cus_per_node;
387 
388 	if (c->x86 >= 0x17)
389 		return;
390 
391 	cus_per_node = c->x86_max_cores / nodes_per_socket;
392 	c->cpu_core_id %= cus_per_node;
393 }
394 
395 /*
396  * Fixup core topology information for
397  * (1) AMD multi-node processors
398  *     Assumption: Number of cores in each internal node is the same.
399  * (2) AMD processors supporting compute units
400  */
401 static void amd_get_topology(struct cpuinfo_x86 *c)
402 {
403 	int cpu = smp_processor_id();
404 
405 	/* get information required for multi-node processors */
406 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
407 		int err;
408 		u32 eax, ebx, ecx, edx;
409 
410 		cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
411 
412 		c->cpu_die_id  = ecx & 0xff;
413 
414 		if (c->x86 == 0x15)
415 			c->cu_id = ebx & 0xff;
416 
417 		if (c->x86 >= 0x17) {
418 			c->cpu_core_id = ebx & 0xff;
419 
420 			if (smp_num_siblings > 1)
421 				c->x86_max_cores /= smp_num_siblings;
422 		}
423 
424 		/*
425 		 * In case leaf B is available, use it to derive
426 		 * topology information.
427 		 */
428 		err = detect_extended_topology(c);
429 		if (!err)
430 			c->x86_coreid_bits = get_count_order(c->x86_max_cores);
431 
432 		cacheinfo_amd_init_llc_id(c, cpu);
433 
434 	} else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
435 		u64 value;
436 
437 		rdmsrl(MSR_FAM10H_NODE_ID, value);
438 		c->cpu_die_id = value & 7;
439 
440 		per_cpu(cpu_llc_id, cpu) = c->cpu_die_id;
441 	} else
442 		return;
443 
444 	if (nodes_per_socket > 1) {
445 		set_cpu_cap(c, X86_FEATURE_AMD_DCM);
446 		legacy_fixup_core_id(c);
447 	}
448 }
449 
450 /*
451  * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
452  * Assumes number of cores is a power of two.
453  */
454 static void amd_detect_cmp(struct cpuinfo_x86 *c)
455 {
456 	unsigned bits;
457 	int cpu = smp_processor_id();
458 
459 	bits = c->x86_coreid_bits;
460 	/* Low order bits define the core id (index of core in socket) */
461 	c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
462 	/* Convert the initial APIC ID into the socket ID */
463 	c->phys_proc_id = c->initial_apicid >> bits;
464 	/* use socket ID also for last level cache */
465 	per_cpu(cpu_llc_id, cpu) = c->cpu_die_id = c->phys_proc_id;
466 }
467 
468 u32 amd_get_nodes_per_socket(void)
469 {
470 	return nodes_per_socket;
471 }
472 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
473 
474 static void srat_detect_node(struct cpuinfo_x86 *c)
475 {
476 #ifdef CONFIG_NUMA
477 	int cpu = smp_processor_id();
478 	int node;
479 	unsigned apicid = c->apicid;
480 
481 	node = numa_cpu_node(cpu);
482 	if (node == NUMA_NO_NODE)
483 		node = get_llc_id(cpu);
484 
485 	/*
486 	 * On multi-fabric platform (e.g. Numascale NumaChip) a
487 	 * platform-specific handler needs to be called to fixup some
488 	 * IDs of the CPU.
489 	 */
490 	if (x86_cpuinit.fixup_cpu_id)
491 		x86_cpuinit.fixup_cpu_id(c, node);
492 
493 	if (!node_online(node)) {
494 		/*
495 		 * Two possibilities here:
496 		 *
497 		 * - The CPU is missing memory and no node was created.  In
498 		 *   that case try picking one from a nearby CPU.
499 		 *
500 		 * - The APIC IDs differ from the HyperTransport node IDs
501 		 *   which the K8 northbridge parsing fills in.  Assume
502 		 *   they are all increased by a constant offset, but in
503 		 *   the same order as the HT nodeids.  If that doesn't
504 		 *   result in a usable node fall back to the path for the
505 		 *   previous case.
506 		 *
507 		 * This workaround operates directly on the mapping between
508 		 * APIC ID and NUMA node, assuming certain relationship
509 		 * between APIC ID, HT node ID and NUMA topology.  As going
510 		 * through CPU mapping may alter the outcome, directly
511 		 * access __apicid_to_node[].
512 		 */
513 		int ht_nodeid = c->initial_apicid;
514 
515 		if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
516 			node = __apicid_to_node[ht_nodeid];
517 		/* Pick a nearby node */
518 		if (!node_online(node))
519 			node = nearby_node(apicid);
520 	}
521 	numa_set_node(cpu, node);
522 #endif
523 }
524 
525 static void early_init_amd_mc(struct cpuinfo_x86 *c)
526 {
527 #ifdef CONFIG_SMP
528 	unsigned bits, ecx;
529 
530 	/* Multi core CPU? */
531 	if (c->extended_cpuid_level < 0x80000008)
532 		return;
533 
534 	ecx = cpuid_ecx(0x80000008);
535 
536 	c->x86_max_cores = (ecx & 0xff) + 1;
537 
538 	/* CPU telling us the core id bits shift? */
539 	bits = (ecx >> 12) & 0xF;
540 
541 	/* Otherwise recompute */
542 	if (bits == 0) {
543 		while ((1 << bits) < c->x86_max_cores)
544 			bits++;
545 	}
546 
547 	c->x86_coreid_bits = bits;
548 #endif
549 }
550 
551 static void bsp_init_amd(struct cpuinfo_x86 *c)
552 {
553 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
554 
555 		if (c->x86 > 0x10 ||
556 		    (c->x86 == 0x10 && c->x86_model >= 0x2)) {
557 			u64 val;
558 
559 			rdmsrl(MSR_K7_HWCR, val);
560 			if (!(val & BIT(24)))
561 				pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
562 		}
563 	}
564 
565 	if (c->x86 == 0x15) {
566 		unsigned long upperbit;
567 		u32 cpuid, assoc;
568 
569 		cpuid	 = cpuid_edx(0x80000005);
570 		assoc	 = cpuid >> 16 & 0xff;
571 		upperbit = ((cpuid >> 24) << 10) / assoc;
572 
573 		va_align.mask	  = (upperbit - 1) & PAGE_MASK;
574 		va_align.flags    = ALIGN_VA_32 | ALIGN_VA_64;
575 
576 		/* A random value per boot for bit slice [12:upper_bit) */
577 		va_align.bits = get_random_u32() & va_align.mask;
578 	}
579 
580 	if (cpu_has(c, X86_FEATURE_MWAITX))
581 		use_mwaitx_delay();
582 
583 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
584 		u32 ecx;
585 
586 		ecx = cpuid_ecx(0x8000001e);
587 		__max_die_per_package = nodes_per_socket = ((ecx >> 8) & 7) + 1;
588 	} else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
589 		u64 value;
590 
591 		rdmsrl(MSR_FAM10H_NODE_ID, value);
592 		__max_die_per_package = nodes_per_socket = ((value >> 3) & 7) + 1;
593 	}
594 
595 	if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
596 	    !boot_cpu_has(X86_FEATURE_VIRT_SSBD) &&
597 	    c->x86 >= 0x15 && c->x86 <= 0x17) {
598 		unsigned int bit;
599 
600 		switch (c->x86) {
601 		case 0x15: bit = 54; break;
602 		case 0x16: bit = 33; break;
603 		case 0x17: bit = 10; break;
604 		default: return;
605 		}
606 		/*
607 		 * Try to cache the base value so further operations can
608 		 * avoid RMW. If that faults, do not enable SSBD.
609 		 */
610 		if (!rdmsrl_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
611 			setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
612 			setup_force_cpu_cap(X86_FEATURE_SSBD);
613 			x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
614 		}
615 	}
616 
617 	resctrl_cpu_detect(c);
618 }
619 
620 static void early_detect_mem_encrypt(struct cpuinfo_x86 *c)
621 {
622 	u64 msr;
623 
624 	/*
625 	 * BIOS support is required for SME and SEV.
626 	 *   For SME: If BIOS has enabled SME then adjust x86_phys_bits by
627 	 *	      the SME physical address space reduction value.
628 	 *	      If BIOS has not enabled SME then don't advertise the
629 	 *	      SME feature (set in scattered.c).
630 	 *	      If the kernel has not enabled SME via any means then
631 	 *	      don't advertise the SME feature.
632 	 *   For SEV: If BIOS has not enabled SEV then don't advertise the
633 	 *            SEV and SEV_ES feature (set in scattered.c).
634 	 *
635 	 *   In all cases, since support for SME and SEV requires long mode,
636 	 *   don't advertise the feature under CONFIG_X86_32.
637 	 */
638 	if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) {
639 		/* Check if memory encryption is enabled */
640 		rdmsrl(MSR_AMD64_SYSCFG, msr);
641 		if (!(msr & MSR_AMD64_SYSCFG_MEM_ENCRYPT))
642 			goto clear_all;
643 
644 		/*
645 		 * Always adjust physical address bits. Even though this
646 		 * will be a value above 32-bits this is still done for
647 		 * CONFIG_X86_32 so that accurate values are reported.
648 		 */
649 		c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
650 
651 		if (IS_ENABLED(CONFIG_X86_32))
652 			goto clear_all;
653 
654 		if (!sme_me_mask)
655 			setup_clear_cpu_cap(X86_FEATURE_SME);
656 
657 		rdmsrl(MSR_K7_HWCR, msr);
658 		if (!(msr & MSR_K7_HWCR_SMMLOCK))
659 			goto clear_sev;
660 
661 		return;
662 
663 clear_all:
664 		setup_clear_cpu_cap(X86_FEATURE_SME);
665 clear_sev:
666 		setup_clear_cpu_cap(X86_FEATURE_SEV);
667 		setup_clear_cpu_cap(X86_FEATURE_SEV_ES);
668 	}
669 }
670 
671 static void early_init_amd(struct cpuinfo_x86 *c)
672 {
673 	u64 value;
674 	u32 dummy;
675 
676 	early_init_amd_mc(c);
677 
678 	if (c->x86 >= 0xf)
679 		set_cpu_cap(c, X86_FEATURE_K8);
680 
681 	rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
682 
683 	/*
684 	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
685 	 * with P/T states and does not stop in deep C-states
686 	 */
687 	if (c->x86_power & (1 << 8)) {
688 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
689 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
690 	}
691 
692 	/* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
693 	if (c->x86_power & BIT(12))
694 		set_cpu_cap(c, X86_FEATURE_ACC_POWER);
695 
696 	/* Bit 14 indicates the Runtime Average Power Limit interface. */
697 	if (c->x86_power & BIT(14))
698 		set_cpu_cap(c, X86_FEATURE_RAPL);
699 
700 #ifdef CONFIG_X86_64
701 	set_cpu_cap(c, X86_FEATURE_SYSCALL32);
702 #else
703 	/*  Set MTRR capability flag if appropriate */
704 	if (c->x86 == 5)
705 		if (c->x86_model == 13 || c->x86_model == 9 ||
706 		    (c->x86_model == 8 && c->x86_stepping >= 8))
707 			set_cpu_cap(c, X86_FEATURE_K6_MTRR);
708 #endif
709 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
710 	/*
711 	 * ApicID can always be treated as an 8-bit value for AMD APIC versions
712 	 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
713 	 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
714 	 * after 16h.
715 	 */
716 	if (boot_cpu_has(X86_FEATURE_APIC)) {
717 		if (c->x86 > 0x16)
718 			set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
719 		else if (c->x86 >= 0xf) {
720 			/* check CPU config space for extended APIC ID */
721 			unsigned int val;
722 
723 			val = read_pci_config(0, 24, 0, 0x68);
724 			if ((val >> 17 & 0x3) == 0x3)
725 				set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
726 		}
727 	}
728 #endif
729 
730 	/*
731 	 * This is only needed to tell the kernel whether to use VMCALL
732 	 * and VMMCALL.  VMMCALL is never executed except under virt, so
733 	 * we can set it unconditionally.
734 	 */
735 	set_cpu_cap(c, X86_FEATURE_VMMCALL);
736 
737 	/* F16h erratum 793, CVE-2013-6885 */
738 	if (c->x86 == 0x16 && c->x86_model <= 0xf)
739 		msr_set_bit(MSR_AMD64_LS_CFG, 15);
740 
741 	/*
742 	 * Check whether the machine is affected by erratum 400. This is
743 	 * used to select the proper idle routine and to enable the check
744 	 * whether the machine is affected in arch_post_acpi_init(), which
745 	 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
746 	 */
747 	if (cpu_has_amd_erratum(c, amd_erratum_400))
748 		set_cpu_bug(c, X86_BUG_AMD_E400);
749 
750 	early_detect_mem_encrypt(c);
751 
752 	/* Re-enable TopologyExtensions if switched off by BIOS */
753 	if (c->x86 == 0x15 &&
754 	    (c->x86_model >= 0x10 && c->x86_model <= 0x6f) &&
755 	    !cpu_has(c, X86_FEATURE_TOPOEXT)) {
756 
757 		if (msr_set_bit(0xc0011005, 54) > 0) {
758 			rdmsrl(0xc0011005, value);
759 			if (value & BIT_64(54)) {
760 				set_cpu_cap(c, X86_FEATURE_TOPOEXT);
761 				pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
762 			}
763 		}
764 	}
765 
766 	if (cpu_has(c, X86_FEATURE_TOPOEXT))
767 		smp_num_siblings = ((cpuid_ebx(0x8000001e) >> 8) & 0xff) + 1;
768 }
769 
770 static void init_amd_k8(struct cpuinfo_x86 *c)
771 {
772 	u32 level;
773 	u64 value;
774 
775 	/* On C+ stepping K8 rep microcode works well for copy/memset */
776 	level = cpuid_eax(1);
777 	if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
778 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
779 
780 	/*
781 	 * Some BIOSes incorrectly force this feature, but only K8 revision D
782 	 * (model = 0x14) and later actually support it.
783 	 * (AMD Erratum #110, docId: 25759).
784 	 */
785 	if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
786 		clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
787 		if (!rdmsrl_amd_safe(0xc001100d, &value)) {
788 			value &= ~BIT_64(32);
789 			wrmsrl_amd_safe(0xc001100d, value);
790 		}
791 	}
792 
793 	if (!c->x86_model_id[0])
794 		strcpy(c->x86_model_id, "Hammer");
795 
796 #ifdef CONFIG_SMP
797 	/*
798 	 * Disable TLB flush filter by setting HWCR.FFDIS on K8
799 	 * bit 6 of msr C001_0015
800 	 *
801 	 * Errata 63 for SH-B3 steppings
802 	 * Errata 122 for all steppings (F+ have it disabled by default)
803 	 */
804 	msr_set_bit(MSR_K7_HWCR, 6);
805 #endif
806 	set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
807 }
808 
809 static void init_amd_gh(struct cpuinfo_x86 *c)
810 {
811 #ifdef CONFIG_MMCONF_FAM10H
812 	/* do this for boot cpu */
813 	if (c == &boot_cpu_data)
814 		check_enable_amd_mmconf_dmi();
815 
816 	fam10h_check_enable_mmcfg();
817 #endif
818 
819 	/*
820 	 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
821 	 * is always needed when GART is enabled, even in a kernel which has no
822 	 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
823 	 * If it doesn't, we do it here as suggested by the BKDG.
824 	 *
825 	 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
826 	 */
827 	msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
828 
829 	/*
830 	 * On family 10h BIOS may not have properly enabled WC+ support, causing
831 	 * it to be converted to CD memtype. This may result in performance
832 	 * degradation for certain nested-paging guests. Prevent this conversion
833 	 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
834 	 *
835 	 * NOTE: we want to use the _safe accessors so as not to #GP kvm
836 	 * guests on older kvm hosts.
837 	 */
838 	msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
839 
840 	if (cpu_has_amd_erratum(c, amd_erratum_383))
841 		set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
842 }
843 
844 static void init_amd_ln(struct cpuinfo_x86 *c)
845 {
846 	/*
847 	 * Apply erratum 665 fix unconditionally so machines without a BIOS
848 	 * fix work.
849 	 */
850 	msr_set_bit(MSR_AMD64_DE_CFG, 31);
851 }
852 
853 static bool rdrand_force;
854 
855 static int __init rdrand_cmdline(char *str)
856 {
857 	if (!str)
858 		return -EINVAL;
859 
860 	if (!strcmp(str, "force"))
861 		rdrand_force = true;
862 	else
863 		return -EINVAL;
864 
865 	return 0;
866 }
867 early_param("rdrand", rdrand_cmdline);
868 
869 static void clear_rdrand_cpuid_bit(struct cpuinfo_x86 *c)
870 {
871 	/*
872 	 * Saving of the MSR used to hide the RDRAND support during
873 	 * suspend/resume is done by arch/x86/power/cpu.c, which is
874 	 * dependent on CONFIG_PM_SLEEP.
875 	 */
876 	if (!IS_ENABLED(CONFIG_PM_SLEEP))
877 		return;
878 
879 	/*
880 	 * The self-test can clear X86_FEATURE_RDRAND, so check for
881 	 * RDRAND support using the CPUID function directly.
882 	 */
883 	if (!(cpuid_ecx(1) & BIT(30)) || rdrand_force)
884 		return;
885 
886 	msr_clear_bit(MSR_AMD64_CPUID_FN_1, 62);
887 
888 	/*
889 	 * Verify that the CPUID change has occurred in case the kernel is
890 	 * running virtualized and the hypervisor doesn't support the MSR.
891 	 */
892 	if (cpuid_ecx(1) & BIT(30)) {
893 		pr_info_once("BIOS may not properly restore RDRAND after suspend, but hypervisor does not support hiding RDRAND via CPUID.\n");
894 		return;
895 	}
896 
897 	clear_cpu_cap(c, X86_FEATURE_RDRAND);
898 	pr_info_once("BIOS may not properly restore RDRAND after suspend, hiding RDRAND via CPUID. Use rdrand=force to reenable.\n");
899 }
900 
901 static void init_amd_jg(struct cpuinfo_x86 *c)
902 {
903 	/*
904 	 * Some BIOS implementations do not restore proper RDRAND support
905 	 * across suspend and resume. Check on whether to hide the RDRAND
906 	 * instruction support via CPUID.
907 	 */
908 	clear_rdrand_cpuid_bit(c);
909 }
910 
911 static void init_amd_bd(struct cpuinfo_x86 *c)
912 {
913 	u64 value;
914 
915 	/*
916 	 * The way access filter has a performance penalty on some workloads.
917 	 * Disable it on the affected CPUs.
918 	 */
919 	if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
920 		if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
921 			value |= 0x1E;
922 			wrmsrl_safe(MSR_F15H_IC_CFG, value);
923 		}
924 	}
925 
926 	/*
927 	 * Some BIOS implementations do not restore proper RDRAND support
928 	 * across suspend and resume. Check on whether to hide the RDRAND
929 	 * instruction support via CPUID.
930 	 */
931 	clear_rdrand_cpuid_bit(c);
932 }
933 
934 void init_spectral_chicken(struct cpuinfo_x86 *c)
935 {
936 #ifdef CONFIG_CPU_UNRET_ENTRY
937 	u64 value;
938 
939 	/*
940 	 * On Zen2 we offer this chicken (bit) on the altar of Speculation.
941 	 *
942 	 * This suppresses speculation from the middle of a basic block, i.e. it
943 	 * suppresses non-branch predictions.
944 	 *
945 	 * We use STIBP as a heuristic to filter out Zen2 from the rest of F17H
946 	 */
947 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR) && cpu_has(c, X86_FEATURE_AMD_STIBP)) {
948 		if (!rdmsrl_safe(MSR_ZEN2_SPECTRAL_CHICKEN, &value)) {
949 			value |= MSR_ZEN2_SPECTRAL_CHICKEN_BIT;
950 			wrmsrl_safe(MSR_ZEN2_SPECTRAL_CHICKEN, value);
951 		}
952 	}
953 #endif
954 	/*
955 	 * Work around Erratum 1386.  The XSAVES instruction malfunctions in
956 	 * certain circumstances on Zen1/2 uarch, and not all parts have had
957 	 * updated microcode at the time of writing (March 2023).
958 	 *
959 	 * Affected parts all have no supervisor XSAVE states, meaning that
960 	 * the XSAVEC instruction (which works fine) is equivalent.
961 	 */
962 	clear_cpu_cap(c, X86_FEATURE_XSAVES);
963 }
964 
965 static void init_amd_zn(struct cpuinfo_x86 *c)
966 {
967 	set_cpu_cap(c, X86_FEATURE_ZEN);
968 
969 #ifdef CONFIG_NUMA
970 	node_reclaim_distance = 32;
971 #endif
972 
973 	/* Fix up CPUID bits, but only if not virtualised. */
974 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR)) {
975 
976 		/* Erratum 1076: CPB feature bit not being set in CPUID. */
977 		if (!cpu_has(c, X86_FEATURE_CPB))
978 			set_cpu_cap(c, X86_FEATURE_CPB);
979 
980 		/*
981 		 * Zen3 (Fam19 model < 0x10) parts are not susceptible to
982 		 * Branch Type Confusion, but predate the allocation of the
983 		 * BTC_NO bit.
984 		 */
985 		if (c->x86 == 0x19 && !cpu_has(c, X86_FEATURE_BTC_NO))
986 			set_cpu_cap(c, X86_FEATURE_BTC_NO);
987 	}
988 }
989 
990 static bool cpu_has_zenbleed_microcode(void)
991 {
992 	u32 good_rev = 0;
993 
994 	switch (boot_cpu_data.x86_model) {
995 	case 0x30 ... 0x3f: good_rev = 0x0830107a; break;
996 	case 0x60 ... 0x67: good_rev = 0x0860010b; break;
997 	case 0x68 ... 0x6f: good_rev = 0x08608105; break;
998 	case 0x70 ... 0x7f: good_rev = 0x08701032; break;
999 	case 0xa0 ... 0xaf: good_rev = 0x08a00008; break;
1000 
1001 	default:
1002 		return false;
1003 		break;
1004 	}
1005 
1006 	if (boot_cpu_data.microcode < good_rev)
1007 		return false;
1008 
1009 	return true;
1010 }
1011 
1012 static void zenbleed_check(struct cpuinfo_x86 *c)
1013 {
1014 	if (!cpu_has_amd_erratum(c, amd_zenbleed))
1015 		return;
1016 
1017 	if (cpu_has(c, X86_FEATURE_HYPERVISOR))
1018 		return;
1019 
1020 	if (!cpu_has(c, X86_FEATURE_AVX))
1021 		return;
1022 
1023 	if (!cpu_has_zenbleed_microcode()) {
1024 		pr_notice_once("Zenbleed: please update your microcode for the most optimal fix\n");
1025 		msr_set_bit(MSR_AMD64_DE_CFG, MSR_AMD64_DE_CFG_ZEN2_FP_BACKUP_FIX_BIT);
1026 	} else {
1027 		msr_clear_bit(MSR_AMD64_DE_CFG, MSR_AMD64_DE_CFG_ZEN2_FP_BACKUP_FIX_BIT);
1028 	}
1029 }
1030 
1031 static void init_amd(struct cpuinfo_x86 *c)
1032 {
1033 	early_init_amd(c);
1034 
1035 	/*
1036 	 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
1037 	 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
1038 	 */
1039 	clear_cpu_cap(c, 0*32+31);
1040 
1041 	if (c->x86 >= 0x10)
1042 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
1043 
1044 	/* AMD FSRM also implies FSRS */
1045 	if (cpu_has(c, X86_FEATURE_FSRM))
1046 		set_cpu_cap(c, X86_FEATURE_FSRS);
1047 
1048 	/* get apicid instead of initial apic id from cpuid */
1049 	c->apicid = hard_smp_processor_id();
1050 
1051 	/* K6s reports MCEs but don't actually have all the MSRs */
1052 	if (c->x86 < 6)
1053 		clear_cpu_cap(c, X86_FEATURE_MCE);
1054 
1055 	switch (c->x86) {
1056 	case 4:    init_amd_k5(c); break;
1057 	case 5:    init_amd_k6(c); break;
1058 	case 6:	   init_amd_k7(c); break;
1059 	case 0xf:  init_amd_k8(c); break;
1060 	case 0x10: init_amd_gh(c); break;
1061 	case 0x12: init_amd_ln(c); break;
1062 	case 0x15: init_amd_bd(c); break;
1063 	case 0x16: init_amd_jg(c); break;
1064 	case 0x17: init_spectral_chicken(c);
1065 		   fallthrough;
1066 	case 0x19: init_amd_zn(c); break;
1067 	}
1068 
1069 	/*
1070 	 * Enable workaround for FXSAVE leak on CPUs
1071 	 * without a XSaveErPtr feature
1072 	 */
1073 	if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
1074 		set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
1075 
1076 	cpu_detect_cache_sizes(c);
1077 
1078 	amd_detect_cmp(c);
1079 	amd_get_topology(c);
1080 	srat_detect_node(c);
1081 
1082 	init_amd_cacheinfo(c);
1083 
1084 	if (!cpu_has(c, X86_FEATURE_LFENCE_RDTSC) && cpu_has(c, X86_FEATURE_XMM2)) {
1085 		/*
1086 		 * Use LFENCE for execution serialization.  On families which
1087 		 * don't have that MSR, LFENCE is already serializing.
1088 		 * msr_set_bit() uses the safe accessors, too, even if the MSR
1089 		 * is not present.
1090 		 */
1091 		msr_set_bit(MSR_AMD64_DE_CFG,
1092 			    MSR_AMD64_DE_CFG_LFENCE_SERIALIZE_BIT);
1093 
1094 		/* A serializing LFENCE stops RDTSC speculation */
1095 		set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
1096 	}
1097 
1098 	/*
1099 	 * Family 0x12 and above processors have APIC timer
1100 	 * running in deep C states.
1101 	 */
1102 	if (c->x86 > 0x11)
1103 		set_cpu_cap(c, X86_FEATURE_ARAT);
1104 
1105 	/* 3DNow or LM implies PREFETCHW */
1106 	if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
1107 		if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
1108 			set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
1109 
1110 	/* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
1111 	if (!cpu_feature_enabled(X86_FEATURE_XENPV))
1112 		set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
1113 
1114 	/*
1115 	 * Turn on the Instructions Retired free counter on machines not
1116 	 * susceptible to erratum #1054 "Instructions Retired Performance
1117 	 * Counter May Be Inaccurate".
1118 	 */
1119 	if (cpu_has(c, X86_FEATURE_IRPERF) &&
1120 	    !cpu_has_amd_erratum(c, amd_erratum_1054))
1121 		msr_set_bit(MSR_K7_HWCR, MSR_K7_HWCR_IRPERF_EN_BIT);
1122 
1123 	check_null_seg_clears_base(c);
1124 
1125 	/*
1126 	 * Make sure EFER[AIBRSE - Automatic IBRS Enable] is set. The APs are brought up
1127 	 * using the trampoline code and as part of it, MSR_EFER gets prepared there in
1128 	 * order to be replicated onto them. Regardless, set it here again, if not set,
1129 	 * to protect against any future refactoring/code reorganization which might
1130 	 * miss setting this important bit.
1131 	 */
1132 	if (spectre_v2_in_eibrs_mode(spectre_v2_enabled) &&
1133 	    cpu_has(c, X86_FEATURE_AUTOIBRS))
1134 		WARN_ON_ONCE(msr_set_bit(MSR_EFER, _EFER_AUTOIBRS));
1135 
1136 	zenbleed_check(c);
1137 
1138 	if (cpu_has_amd_erratum(c, amd_div0)) {
1139 		pr_notice_once("AMD Zen1 DIV0 bug detected. Disable SMT for full protection.\n");
1140 		setup_force_cpu_bug(X86_BUG_DIV0);
1141 	}
1142 }
1143 
1144 #ifdef CONFIG_X86_32
1145 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
1146 {
1147 	/* AMD errata T13 (order #21922) */
1148 	if (c->x86 == 6) {
1149 		/* Duron Rev A0 */
1150 		if (c->x86_model == 3 && c->x86_stepping == 0)
1151 			size = 64;
1152 		/* Tbird rev A1/A2 */
1153 		if (c->x86_model == 4 &&
1154 			(c->x86_stepping == 0 || c->x86_stepping == 1))
1155 			size = 256;
1156 	}
1157 	return size;
1158 }
1159 #endif
1160 
1161 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
1162 {
1163 	u32 ebx, eax, ecx, edx;
1164 	u16 mask = 0xfff;
1165 
1166 	if (c->x86 < 0xf)
1167 		return;
1168 
1169 	if (c->extended_cpuid_level < 0x80000006)
1170 		return;
1171 
1172 	cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
1173 
1174 	tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
1175 	tlb_lli_4k[ENTRIES] = ebx & mask;
1176 
1177 	/*
1178 	 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
1179 	 * characteristics from the CPUID function 0x80000005 instead.
1180 	 */
1181 	if (c->x86 == 0xf) {
1182 		cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1183 		mask = 0xff;
1184 	}
1185 
1186 	/* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1187 	if (!((eax >> 16) & mask))
1188 		tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
1189 	else
1190 		tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
1191 
1192 	/* a 4M entry uses two 2M entries */
1193 	tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
1194 
1195 	/* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1196 	if (!(eax & mask)) {
1197 		/* Erratum 658 */
1198 		if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
1199 			tlb_lli_2m[ENTRIES] = 1024;
1200 		} else {
1201 			cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1202 			tlb_lli_2m[ENTRIES] = eax & 0xff;
1203 		}
1204 	} else
1205 		tlb_lli_2m[ENTRIES] = eax & mask;
1206 
1207 	tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
1208 }
1209 
1210 static const struct cpu_dev amd_cpu_dev = {
1211 	.c_vendor	= "AMD",
1212 	.c_ident	= { "AuthenticAMD" },
1213 #ifdef CONFIG_X86_32
1214 	.legacy_models = {
1215 		{ .family = 4, .model_names =
1216 		  {
1217 			  [3] = "486 DX/2",
1218 			  [7] = "486 DX/2-WB",
1219 			  [8] = "486 DX/4",
1220 			  [9] = "486 DX/4-WB",
1221 			  [14] = "Am5x86-WT",
1222 			  [15] = "Am5x86-WB"
1223 		  }
1224 		},
1225 	},
1226 	.legacy_cache_size = amd_size_cache,
1227 #endif
1228 	.c_early_init   = early_init_amd,
1229 	.c_detect_tlb	= cpu_detect_tlb_amd,
1230 	.c_bsp_init	= bsp_init_amd,
1231 	.c_init		= init_amd,
1232 	.c_x86_vendor	= X86_VENDOR_AMD,
1233 };
1234 
1235 cpu_dev_register(amd_cpu_dev);
1236 
1237 static DEFINE_PER_CPU_READ_MOSTLY(unsigned long[4], amd_dr_addr_mask);
1238 
1239 static unsigned int amd_msr_dr_addr_masks[] = {
1240 	MSR_F16H_DR0_ADDR_MASK,
1241 	MSR_F16H_DR1_ADDR_MASK,
1242 	MSR_F16H_DR1_ADDR_MASK + 1,
1243 	MSR_F16H_DR1_ADDR_MASK + 2
1244 };
1245 
1246 void amd_set_dr_addr_mask(unsigned long mask, unsigned int dr)
1247 {
1248 	int cpu = smp_processor_id();
1249 
1250 	if (!cpu_feature_enabled(X86_FEATURE_BPEXT))
1251 		return;
1252 
1253 	if (WARN_ON_ONCE(dr >= ARRAY_SIZE(amd_msr_dr_addr_masks)))
1254 		return;
1255 
1256 	if (per_cpu(amd_dr_addr_mask, cpu)[dr] == mask)
1257 		return;
1258 
1259 	wrmsr(amd_msr_dr_addr_masks[dr], mask, 0);
1260 	per_cpu(amd_dr_addr_mask, cpu)[dr] = mask;
1261 }
1262 
1263 unsigned long amd_get_dr_addr_mask(unsigned int dr)
1264 {
1265 	if (!cpu_feature_enabled(X86_FEATURE_BPEXT))
1266 		return 0;
1267 
1268 	if (WARN_ON_ONCE(dr >= ARRAY_SIZE(amd_msr_dr_addr_masks)))
1269 		return 0;
1270 
1271 	return per_cpu(amd_dr_addr_mask[dr], smp_processor_id());
1272 }
1273 EXPORT_SYMBOL_GPL(amd_get_dr_addr_mask);
1274 
1275 u32 amd_get_highest_perf(void)
1276 {
1277 	struct cpuinfo_x86 *c = &boot_cpu_data;
1278 
1279 	if (c->x86 == 0x17 && ((c->x86_model >= 0x30 && c->x86_model < 0x40) ||
1280 			       (c->x86_model >= 0x70 && c->x86_model < 0x80)))
1281 		return 166;
1282 
1283 	if (c->x86 == 0x19 && ((c->x86_model >= 0x20 && c->x86_model < 0x30) ||
1284 			       (c->x86_model >= 0x40 && c->x86_model < 0x70)))
1285 		return 166;
1286 
1287 	return 255;
1288 }
1289 EXPORT_SYMBOL_GPL(amd_get_highest_perf);
1290 
1291 static void zenbleed_check_cpu(void *unused)
1292 {
1293 	struct cpuinfo_x86 *c = &cpu_data(smp_processor_id());
1294 
1295 	zenbleed_check(c);
1296 }
1297 
1298 void amd_check_microcode(void)
1299 {
1300 	on_each_cpu(zenbleed_check_cpu, NULL, 1);
1301 }
1302 
1303 bool cpu_has_ibpb_brtype_microcode(void)
1304 {
1305 	switch (boot_cpu_data.x86) {
1306 	/* Zen1/2 IBPB flushes branch type predictions too. */
1307 	case 0x17:
1308 		return boot_cpu_has(X86_FEATURE_AMD_IBPB);
1309 	case 0x19:
1310 		/* Poke the MSR bit on Zen3/4 to check its presence. */
1311 		if (!wrmsrl_safe(MSR_IA32_PRED_CMD, PRED_CMD_SBPB)) {
1312 			setup_force_cpu_cap(X86_FEATURE_SBPB);
1313 			return true;
1314 		} else {
1315 			return false;
1316 		}
1317 	default:
1318 		return false;
1319 	}
1320 }
1321 
1322 /*
1323  * Issue a DIV 0/1 insn to clear any division data from previous DIV
1324  * operations.
1325  */
1326 void noinstr amd_clear_divider(void)
1327 {
1328 	asm volatile(ALTERNATIVE("", "div %2\n\t", X86_BUG_DIV0)
1329 		     :: "a" (0), "d" (0), "r" (1));
1330 }
1331