• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 const int amd_erratum_1485[] =
84 	AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x19, 0x10, 0x0, 0x1f, 0xf),
85 			   AMD_MODEL_RANGE(0x19, 0x60, 0x0, 0xaf, 0xf));
86 
cpu_has_amd_erratum(struct cpuinfo_x86 * cpu,const int * erratum)87 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
88 {
89 	int osvw_id = *erratum++;
90 	u32 range;
91 	u32 ms;
92 
93 	if (osvw_id >= 0 && osvw_id < 65536 &&
94 	    cpu_has(cpu, X86_FEATURE_OSVW)) {
95 		u64 osvw_len;
96 
97 		rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
98 		if (osvw_id < osvw_len) {
99 			u64 osvw_bits;
100 
101 			rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
102 			    osvw_bits);
103 			return osvw_bits & (1ULL << (osvw_id & 0x3f));
104 		}
105 	}
106 
107 	/* OSVW unavailable or ID unknown, match family-model-stepping range */
108 	ms = (cpu->x86_model << 4) | cpu->x86_stepping;
109 	while ((range = *erratum++))
110 		if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
111 		    (ms >= AMD_MODEL_RANGE_START(range)) &&
112 		    (ms <= AMD_MODEL_RANGE_END(range)))
113 			return true;
114 
115 	return false;
116 }
117 
rdmsrl_amd_safe(unsigned msr,unsigned long long * p)118 static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
119 {
120 	u32 gprs[8] = { 0 };
121 	int err;
122 
123 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
124 		  "%s should only be used on K8!\n", __func__);
125 
126 	gprs[1] = msr;
127 	gprs[7] = 0x9c5a203a;
128 
129 	err = rdmsr_safe_regs(gprs);
130 
131 	*p = gprs[0] | ((u64)gprs[2] << 32);
132 
133 	return err;
134 }
135 
wrmsrl_amd_safe(unsigned msr,unsigned long long val)136 static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
137 {
138 	u32 gprs[8] = { 0 };
139 
140 	WARN_ONCE((boot_cpu_data.x86 != 0xf),
141 		  "%s should only be used on K8!\n", __func__);
142 
143 	gprs[0] = (u32)val;
144 	gprs[1] = msr;
145 	gprs[2] = val >> 32;
146 	gprs[7] = 0x9c5a203a;
147 
148 	return wrmsr_safe_regs(gprs);
149 }
150 
151 /*
152  *	B step AMD K6 before B 9730xxxx have hardware bugs that can cause
153  *	misexecution of code under Linux. Owners of such processors should
154  *	contact AMD for precise details and a CPU swap.
155  *
156  *	See	http://www.multimania.com/poulot/k6bug.html
157  *	and	section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
158  *		(Publication # 21266  Issue Date: August 1998)
159  *
160  *	The following test is erm.. interesting. AMD neglected to up
161  *	the chip setting when fixing the bug but they also tweaked some
162  *	performance at the same time..
163  */
164 
165 #ifdef CONFIG_X86_32
166 extern __visible void vide(void);
167 __asm__(".text\n"
168 	".globl vide\n"
169 	".type vide, @function\n"
170 	".align 4\n"
171 	"vide: ret\n");
172 #endif
173 
init_amd_k5(struct cpuinfo_x86 * c)174 static void init_amd_k5(struct cpuinfo_x86 *c)
175 {
176 #ifdef CONFIG_X86_32
177 /*
178  * General Systems BIOSen alias the cpu frequency registers
179  * of the Elan at 0x000df000. Unfortunately, one of the Linux
180  * drivers subsequently pokes it, and changes the CPU speed.
181  * Workaround : Remove the unneeded alias.
182  */
183 #define CBAR		(0xfffc) /* Configuration Base Address  (32-bit) */
184 #define CBAR_ENB	(0x80000000)
185 #define CBAR_KEY	(0X000000CB)
186 	if (c->x86_model == 9 || c->x86_model == 10) {
187 		if (inl(CBAR) & CBAR_ENB)
188 			outl(0 | CBAR_KEY, CBAR);
189 	}
190 #endif
191 }
192 
init_amd_k6(struct cpuinfo_x86 * c)193 static void init_amd_k6(struct cpuinfo_x86 *c)
194 {
195 #ifdef CONFIG_X86_32
196 	u32 l, h;
197 	int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
198 
199 	if (c->x86_model < 6) {
200 		/* Based on AMD doc 20734R - June 2000 */
201 		if (c->x86_model == 0) {
202 			clear_cpu_cap(c, X86_FEATURE_APIC);
203 			set_cpu_cap(c, X86_FEATURE_PGE);
204 		}
205 		return;
206 	}
207 
208 	if (c->x86_model == 6 && c->x86_stepping == 1) {
209 		const int K6_BUG_LOOP = 1000000;
210 		int n;
211 		void (*f_vide)(void);
212 		u64 d, d2;
213 
214 		pr_info("AMD K6 stepping B detected - ");
215 
216 		/*
217 		 * It looks like AMD fixed the 2.6.2 bug and improved indirect
218 		 * calls at the same time.
219 		 */
220 
221 		n = K6_BUG_LOOP;
222 		f_vide = vide;
223 		OPTIMIZER_HIDE_VAR(f_vide);
224 		d = rdtsc();
225 		while (n--)
226 			f_vide();
227 		d2 = rdtsc();
228 		d = d2-d;
229 
230 		if (d > 20*K6_BUG_LOOP)
231 			pr_cont("system stability may be impaired when more than 32 MB are used.\n");
232 		else
233 			pr_cont("probably OK (after B9730xxxx).\n");
234 	}
235 
236 	/* K6 with old style WHCR */
237 	if (c->x86_model < 8 ||
238 	   (c->x86_model == 8 && c->x86_stepping < 8)) {
239 		/* We can only write allocate on the low 508Mb */
240 		if (mbytes > 508)
241 			mbytes = 508;
242 
243 		rdmsr(MSR_K6_WHCR, l, h);
244 		if ((l&0x0000FFFF) == 0) {
245 			unsigned long flags;
246 			l = (1<<0)|((mbytes/4)<<1);
247 			local_irq_save(flags);
248 			wbinvd();
249 			wrmsr(MSR_K6_WHCR, l, h);
250 			local_irq_restore(flags);
251 			pr_info("Enabling old style K6 write allocation for %d Mb\n",
252 				mbytes);
253 		}
254 		return;
255 	}
256 
257 	if ((c->x86_model == 8 && c->x86_stepping > 7) ||
258 	     c->x86_model == 9 || c->x86_model == 13) {
259 		/* The more serious chips .. */
260 
261 		if (mbytes > 4092)
262 			mbytes = 4092;
263 
264 		rdmsr(MSR_K6_WHCR, l, h);
265 		if ((l&0xFFFF0000) == 0) {
266 			unsigned long flags;
267 			l = ((mbytes>>2)<<22)|(1<<16);
268 			local_irq_save(flags);
269 			wbinvd();
270 			wrmsr(MSR_K6_WHCR, l, h);
271 			local_irq_restore(flags);
272 			pr_info("Enabling new style K6 write allocation for %d Mb\n",
273 				mbytes);
274 		}
275 
276 		return;
277 	}
278 
279 	if (c->x86_model == 10) {
280 		/* AMD Geode LX is model 10 */
281 		/* placeholder for any needed mods */
282 		return;
283 	}
284 #endif
285 }
286 
init_amd_k7(struct cpuinfo_x86 * c)287 static void init_amd_k7(struct cpuinfo_x86 *c)
288 {
289 #ifdef CONFIG_X86_32
290 	u32 l, h;
291 
292 	/*
293 	 * Bit 15 of Athlon specific MSR 15, needs to be 0
294 	 * to enable SSE on Palomino/Morgan/Barton CPU's.
295 	 * If the BIOS didn't enable it already, enable it here.
296 	 */
297 	if (c->x86_model >= 6 && c->x86_model <= 10) {
298 		if (!cpu_has(c, X86_FEATURE_XMM)) {
299 			pr_info("Enabling disabled K7/SSE Support.\n");
300 			msr_clear_bit(MSR_K7_HWCR, 15);
301 			set_cpu_cap(c, X86_FEATURE_XMM);
302 		}
303 	}
304 
305 	/*
306 	 * It's been determined by AMD that Athlons since model 8 stepping 1
307 	 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
308 	 * As per AMD technical note 27212 0.2
309 	 */
310 	if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) {
311 		rdmsr(MSR_K7_CLK_CTL, l, h);
312 		if ((l & 0xfff00000) != 0x20000000) {
313 			pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
314 				l, ((l & 0x000fffff)|0x20000000));
315 			wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
316 		}
317 	}
318 
319 	/* calling is from identify_secondary_cpu() ? */
320 	if (!c->cpu_index)
321 		return;
322 
323 	/*
324 	 * Certain Athlons might work (for various values of 'work') in SMP
325 	 * but they are not certified as MP capable.
326 	 */
327 	/* Athlon 660/661 is valid. */
328 	if ((c->x86_model == 6) && ((c->x86_stepping == 0) ||
329 	    (c->x86_stepping == 1)))
330 		return;
331 
332 	/* Duron 670 is valid */
333 	if ((c->x86_model == 7) && (c->x86_stepping == 0))
334 		return;
335 
336 	/*
337 	 * Athlon 662, Duron 671, and Athlon >model 7 have capability
338 	 * bit. It's worth noting that the A5 stepping (662) of some
339 	 * Athlon XP's have the MP bit set.
340 	 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
341 	 * more.
342 	 */
343 	if (((c->x86_model == 6) && (c->x86_stepping >= 2)) ||
344 	    ((c->x86_model == 7) && (c->x86_stepping >= 1)) ||
345 	     (c->x86_model > 7))
346 		if (cpu_has(c, X86_FEATURE_MP))
347 			return;
348 
349 	/* If we get here, not a certified SMP capable AMD system. */
350 
351 	/*
352 	 * Don't taint if we are running SMP kernel on a single non-MP
353 	 * approved Athlon
354 	 */
355 	WARN_ONCE(1, "WARNING: This combination of AMD"
356 		" processors is not suitable for SMP.\n");
357 	add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
358 #endif
359 }
360 
361 #ifdef CONFIG_NUMA
362 /*
363  * To workaround broken NUMA config.  Read the comment in
364  * srat_detect_node().
365  */
nearby_node(int apicid)366 static int nearby_node(int apicid)
367 {
368 	int i, node;
369 
370 	for (i = apicid - 1; i >= 0; i--) {
371 		node = __apicid_to_node[i];
372 		if (node != NUMA_NO_NODE && node_online(node))
373 			return node;
374 	}
375 	for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
376 		node = __apicid_to_node[i];
377 		if (node != NUMA_NO_NODE && node_online(node))
378 			return node;
379 	}
380 	return first_node(node_online_map); /* Shouldn't happen */
381 }
382 #endif
383 
384 /*
385  * Fix up cpu_core_id for pre-F17h systems to be in the
386  * [0 .. cores_per_node - 1] range. Not really needed but
387  * kept so as not to break existing setups.
388  */
legacy_fixup_core_id(struct cpuinfo_x86 * c)389 static void legacy_fixup_core_id(struct cpuinfo_x86 *c)
390 {
391 	u32 cus_per_node;
392 
393 	if (c->x86 >= 0x17)
394 		return;
395 
396 	cus_per_node = c->x86_max_cores / nodes_per_socket;
397 	c->cpu_core_id %= cus_per_node;
398 }
399 
400 /*
401  * Fixup core topology information for
402  * (1) AMD multi-node processors
403  *     Assumption: Number of cores in each internal node is the same.
404  * (2) AMD processors supporting compute units
405  */
amd_get_topology(struct cpuinfo_x86 * c)406 static void amd_get_topology(struct cpuinfo_x86 *c)
407 {
408 	int cpu = smp_processor_id();
409 
410 	/* get information required for multi-node processors */
411 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
412 		int err;
413 		u32 eax, ebx, ecx, edx;
414 
415 		cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
416 
417 		c->cpu_die_id  = ecx & 0xff;
418 
419 		if (c->x86 == 0x15)
420 			c->cu_id = ebx & 0xff;
421 
422 		if (c->x86 >= 0x17) {
423 			c->cpu_core_id = ebx & 0xff;
424 
425 			if (smp_num_siblings > 1)
426 				c->x86_max_cores /= smp_num_siblings;
427 		}
428 
429 		/*
430 		 * In case leaf B is available, use it to derive
431 		 * topology information.
432 		 */
433 		err = detect_extended_topology(c);
434 		if (!err)
435 			c->x86_coreid_bits = get_count_order(c->x86_max_cores);
436 
437 		cacheinfo_amd_init_llc_id(c, cpu);
438 
439 	} else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
440 		u64 value;
441 
442 		rdmsrl(MSR_FAM10H_NODE_ID, value);
443 		c->cpu_die_id = value & 7;
444 
445 		per_cpu(cpu_llc_id, cpu) = c->cpu_die_id;
446 	} else
447 		return;
448 
449 	if (nodes_per_socket > 1) {
450 		set_cpu_cap(c, X86_FEATURE_AMD_DCM);
451 		legacy_fixup_core_id(c);
452 	}
453 }
454 
455 /*
456  * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
457  * Assumes number of cores is a power of two.
458  */
amd_detect_cmp(struct cpuinfo_x86 * c)459 static void amd_detect_cmp(struct cpuinfo_x86 *c)
460 {
461 	unsigned bits;
462 	int cpu = smp_processor_id();
463 
464 	bits = c->x86_coreid_bits;
465 	/* Low order bits define the core id (index of core in socket) */
466 	c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
467 	/* Convert the initial APIC ID into the socket ID */
468 	c->phys_proc_id = c->initial_apicid >> bits;
469 	/* use socket ID also for last level cache */
470 	per_cpu(cpu_llc_id, cpu) = c->cpu_die_id = c->phys_proc_id;
471 }
472 
amd_detect_ppin(struct cpuinfo_x86 * c)473 static void amd_detect_ppin(struct cpuinfo_x86 *c)
474 {
475 	unsigned long long val;
476 
477 	if (!cpu_has(c, X86_FEATURE_AMD_PPIN))
478 		return;
479 
480 	/* When PPIN is defined in CPUID, still need to check PPIN_CTL MSR */
481 	if (rdmsrl_safe(MSR_AMD_PPIN_CTL, &val))
482 		goto clear_ppin;
483 
484 	/* PPIN is locked in disabled mode, clear feature bit */
485 	if ((val & 3UL) == 1UL)
486 		goto clear_ppin;
487 
488 	/* If PPIN is disabled, try to enable it */
489 	if (!(val & 2UL)) {
490 		wrmsrl_safe(MSR_AMD_PPIN_CTL,  val | 2UL);
491 		rdmsrl_safe(MSR_AMD_PPIN_CTL, &val);
492 	}
493 
494 	/* If PPIN_EN bit is 1, return from here; otherwise fall through */
495 	if (val & 2UL)
496 		return;
497 
498 clear_ppin:
499 	clear_cpu_cap(c, X86_FEATURE_AMD_PPIN);
500 }
501 
amd_get_nodes_per_socket(void)502 u32 amd_get_nodes_per_socket(void)
503 {
504 	return nodes_per_socket;
505 }
506 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
507 
srat_detect_node(struct cpuinfo_x86 * c)508 static void srat_detect_node(struct cpuinfo_x86 *c)
509 {
510 #ifdef CONFIG_NUMA
511 	int cpu = smp_processor_id();
512 	int node;
513 	unsigned apicid = c->apicid;
514 
515 	node = numa_cpu_node(cpu);
516 	if (node == NUMA_NO_NODE)
517 		node = get_llc_id(cpu);
518 
519 	/*
520 	 * On multi-fabric platform (e.g. Numascale NumaChip) a
521 	 * platform-specific handler needs to be called to fixup some
522 	 * IDs of the CPU.
523 	 */
524 	if (x86_cpuinit.fixup_cpu_id)
525 		x86_cpuinit.fixup_cpu_id(c, node);
526 
527 	if (!node_online(node)) {
528 		/*
529 		 * Two possibilities here:
530 		 *
531 		 * - The CPU is missing memory and no node was created.  In
532 		 *   that case try picking one from a nearby CPU.
533 		 *
534 		 * - The APIC IDs differ from the HyperTransport node IDs
535 		 *   which the K8 northbridge parsing fills in.  Assume
536 		 *   they are all increased by a constant offset, but in
537 		 *   the same order as the HT nodeids.  If that doesn't
538 		 *   result in a usable node fall back to the path for the
539 		 *   previous case.
540 		 *
541 		 * This workaround operates directly on the mapping between
542 		 * APIC ID and NUMA node, assuming certain relationship
543 		 * between APIC ID, HT node ID and NUMA topology.  As going
544 		 * through CPU mapping may alter the outcome, directly
545 		 * access __apicid_to_node[].
546 		 */
547 		int ht_nodeid = c->initial_apicid;
548 
549 		if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
550 			node = __apicid_to_node[ht_nodeid];
551 		/* Pick a nearby node */
552 		if (!node_online(node))
553 			node = nearby_node(apicid);
554 	}
555 	numa_set_node(cpu, node);
556 #endif
557 }
558 
early_init_amd_mc(struct cpuinfo_x86 * c)559 static void early_init_amd_mc(struct cpuinfo_x86 *c)
560 {
561 #ifdef CONFIG_SMP
562 	unsigned bits, ecx;
563 
564 	/* Multi core CPU? */
565 	if (c->extended_cpuid_level < 0x80000008)
566 		return;
567 
568 	ecx = cpuid_ecx(0x80000008);
569 
570 	c->x86_max_cores = (ecx & 0xff) + 1;
571 
572 	/* CPU telling us the core id bits shift? */
573 	bits = (ecx >> 12) & 0xF;
574 
575 	/* Otherwise recompute */
576 	if (bits == 0) {
577 		while ((1 << bits) < c->x86_max_cores)
578 			bits++;
579 	}
580 
581 	c->x86_coreid_bits = bits;
582 #endif
583 }
584 
bsp_init_amd(struct cpuinfo_x86 * c)585 static void bsp_init_amd(struct cpuinfo_x86 *c)
586 {
587 	if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
588 
589 		if (c->x86 > 0x10 ||
590 		    (c->x86 == 0x10 && c->x86_model >= 0x2)) {
591 			u64 val;
592 
593 			rdmsrl(MSR_K7_HWCR, val);
594 			if (!(val & BIT(24)))
595 				pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
596 		}
597 	}
598 
599 	if (c->x86 == 0x15) {
600 		unsigned long upperbit;
601 		u32 cpuid, assoc;
602 
603 		cpuid	 = cpuid_edx(0x80000005);
604 		assoc	 = cpuid >> 16 & 0xff;
605 		upperbit = ((cpuid >> 24) << 10) / assoc;
606 
607 		va_align.mask	  = (upperbit - 1) & PAGE_MASK;
608 		va_align.flags    = ALIGN_VA_32 | ALIGN_VA_64;
609 
610 		/* A random value per boot for bit slice [12:upper_bit) */
611 		va_align.bits = get_random_int() & va_align.mask;
612 	}
613 
614 	if (cpu_has(c, X86_FEATURE_MWAITX))
615 		use_mwaitx_delay();
616 
617 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
618 		u32 ecx;
619 
620 		ecx = cpuid_ecx(0x8000001e);
621 		__max_die_per_package = nodes_per_socket = ((ecx >> 8) & 7) + 1;
622 	} else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
623 		u64 value;
624 
625 		rdmsrl(MSR_FAM10H_NODE_ID, value);
626 		__max_die_per_package = nodes_per_socket = ((value >> 3) & 7) + 1;
627 	}
628 
629 	if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
630 	    !boot_cpu_has(X86_FEATURE_VIRT_SSBD) &&
631 	    c->x86 >= 0x15 && c->x86 <= 0x17) {
632 		unsigned int bit;
633 
634 		switch (c->x86) {
635 		case 0x15: bit = 54; break;
636 		case 0x16: bit = 33; break;
637 		case 0x17: bit = 10; break;
638 		default: return;
639 		}
640 		/*
641 		 * Try to cache the base value so further operations can
642 		 * avoid RMW. If that faults, do not enable SSBD.
643 		 */
644 		if (!rdmsrl_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
645 			setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
646 			setup_force_cpu_cap(X86_FEATURE_SSBD);
647 			x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
648 		}
649 	}
650 
651 	resctrl_cpu_detect(c);
652 }
653 
early_detect_mem_encrypt(struct cpuinfo_x86 * c)654 static void early_detect_mem_encrypt(struct cpuinfo_x86 *c)
655 {
656 	u64 msr;
657 
658 	/*
659 	 * BIOS support is required for SME and SEV.
660 	 *   For SME: If BIOS has enabled SME then adjust x86_phys_bits by
661 	 *	      the SME physical address space reduction value.
662 	 *	      If BIOS has not enabled SME then don't advertise the
663 	 *	      SME feature (set in scattered.c).
664 	 *   For SEV: If BIOS has not enabled SEV then don't advertise the
665 	 *            SEV and SEV_ES feature (set in scattered.c).
666 	 *
667 	 *   In all cases, since support for SME and SEV requires long mode,
668 	 *   don't advertise the feature under CONFIG_X86_32.
669 	 */
670 	if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) {
671 		/* Check if memory encryption is enabled */
672 		rdmsrl(MSR_AMD64_SYSCFG, msr);
673 		if (!(msr & MSR_AMD64_SYSCFG_MEM_ENCRYPT))
674 			goto clear_all;
675 
676 		/*
677 		 * Always adjust physical address bits. Even though this
678 		 * will be a value above 32-bits this is still done for
679 		 * CONFIG_X86_32 so that accurate values are reported.
680 		 */
681 		c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
682 
683 		if (IS_ENABLED(CONFIG_X86_32))
684 			goto clear_all;
685 
686 		rdmsrl(MSR_K7_HWCR, msr);
687 		if (!(msr & MSR_K7_HWCR_SMMLOCK))
688 			goto clear_sev;
689 
690 		return;
691 
692 clear_all:
693 		setup_clear_cpu_cap(X86_FEATURE_SME);
694 clear_sev:
695 		setup_clear_cpu_cap(X86_FEATURE_SEV);
696 		setup_clear_cpu_cap(X86_FEATURE_SEV_ES);
697 	}
698 }
699 
early_init_amd(struct cpuinfo_x86 * c)700 static void early_init_amd(struct cpuinfo_x86 *c)
701 {
702 	u64 value;
703 	u32 dummy;
704 
705 	early_init_amd_mc(c);
706 
707 	if (c->x86 >= 0xf)
708 		set_cpu_cap(c, X86_FEATURE_K8);
709 
710 	rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
711 
712 	/*
713 	 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
714 	 * with P/T states and does not stop in deep C-states
715 	 */
716 	if (c->x86_power & (1 << 8)) {
717 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
718 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
719 	}
720 
721 	/* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
722 	if (c->x86_power & BIT(12))
723 		set_cpu_cap(c, X86_FEATURE_ACC_POWER);
724 
725 	/* Bit 14 indicates the Runtime Average Power Limit interface. */
726 	if (c->x86_power & BIT(14))
727 		set_cpu_cap(c, X86_FEATURE_RAPL);
728 
729 #ifdef CONFIG_X86_64
730 	set_cpu_cap(c, X86_FEATURE_SYSCALL32);
731 #else
732 	/*  Set MTRR capability flag if appropriate */
733 	if (c->x86 == 5)
734 		if (c->x86_model == 13 || c->x86_model == 9 ||
735 		    (c->x86_model == 8 && c->x86_stepping >= 8))
736 			set_cpu_cap(c, X86_FEATURE_K6_MTRR);
737 #endif
738 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
739 	/*
740 	 * ApicID can always be treated as an 8-bit value for AMD APIC versions
741 	 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
742 	 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
743 	 * after 16h.
744 	 */
745 	if (boot_cpu_has(X86_FEATURE_APIC)) {
746 		if (c->x86 > 0x16)
747 			set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
748 		else if (c->x86 >= 0xf) {
749 			/* check CPU config space for extended APIC ID */
750 			unsigned int val;
751 
752 			val = read_pci_config(0, 24, 0, 0x68);
753 			if ((val >> 17 & 0x3) == 0x3)
754 				set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
755 		}
756 	}
757 #endif
758 
759 	/*
760 	 * This is only needed to tell the kernel whether to use VMCALL
761 	 * and VMMCALL.  VMMCALL is never executed except under virt, so
762 	 * we can set it unconditionally.
763 	 */
764 	set_cpu_cap(c, X86_FEATURE_VMMCALL);
765 
766 	/* F16h erratum 793, CVE-2013-6885 */
767 	if (c->x86 == 0x16 && c->x86_model <= 0xf)
768 		msr_set_bit(MSR_AMD64_LS_CFG, 15);
769 
770 	/*
771 	 * Check whether the machine is affected by erratum 400. This is
772 	 * used to select the proper idle routine and to enable the check
773 	 * whether the machine is affected in arch_post_acpi_init(), which
774 	 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
775 	 */
776 	if (cpu_has_amd_erratum(c, amd_erratum_400))
777 		set_cpu_bug(c, X86_BUG_AMD_E400);
778 
779 	early_detect_mem_encrypt(c);
780 
781 	/* Re-enable TopologyExtensions if switched off by BIOS */
782 	if (c->x86 == 0x15 &&
783 	    (c->x86_model >= 0x10 && c->x86_model <= 0x6f) &&
784 	    !cpu_has(c, X86_FEATURE_TOPOEXT)) {
785 
786 		if (msr_set_bit(0xc0011005, 54) > 0) {
787 			rdmsrl(0xc0011005, value);
788 			if (value & BIT_64(54)) {
789 				set_cpu_cap(c, X86_FEATURE_TOPOEXT);
790 				pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
791 			}
792 		}
793 	}
794 
795 	if (cpu_has(c, X86_FEATURE_TOPOEXT))
796 		smp_num_siblings = ((cpuid_ebx(0x8000001e) >> 8) & 0xff) + 1;
797 }
798 
init_amd_k8(struct cpuinfo_x86 * c)799 static void init_amd_k8(struct cpuinfo_x86 *c)
800 {
801 	u32 level;
802 	u64 value;
803 
804 	/* On C+ stepping K8 rep microcode works well for copy/memset */
805 	level = cpuid_eax(1);
806 	if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
807 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
808 
809 	/*
810 	 * Some BIOSes incorrectly force this feature, but only K8 revision D
811 	 * (model = 0x14) and later actually support it.
812 	 * (AMD Erratum #110, docId: 25759).
813 	 */
814 	if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
815 		clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
816 		if (!rdmsrl_amd_safe(0xc001100d, &value)) {
817 			value &= ~BIT_64(32);
818 			wrmsrl_amd_safe(0xc001100d, value);
819 		}
820 	}
821 
822 	if (!c->x86_model_id[0])
823 		strcpy(c->x86_model_id, "Hammer");
824 
825 #ifdef CONFIG_SMP
826 	/*
827 	 * Disable TLB flush filter by setting HWCR.FFDIS on K8
828 	 * bit 6 of msr C001_0015
829 	 *
830 	 * Errata 63 for SH-B3 steppings
831 	 * Errata 122 for all steppings (F+ have it disabled by default)
832 	 */
833 	msr_set_bit(MSR_K7_HWCR, 6);
834 #endif
835 	set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
836 }
837 
init_amd_gh(struct cpuinfo_x86 * c)838 static void init_amd_gh(struct cpuinfo_x86 *c)
839 {
840 #ifdef CONFIG_MMCONF_FAM10H
841 	/* do this for boot cpu */
842 	if (c == &boot_cpu_data)
843 		check_enable_amd_mmconf_dmi();
844 
845 	fam10h_check_enable_mmcfg();
846 #endif
847 
848 	/*
849 	 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
850 	 * is always needed when GART is enabled, even in a kernel which has no
851 	 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
852 	 * If it doesn't, we do it here as suggested by the BKDG.
853 	 *
854 	 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
855 	 */
856 	msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
857 
858 	/*
859 	 * On family 10h BIOS may not have properly enabled WC+ support, causing
860 	 * it to be converted to CD memtype. This may result in performance
861 	 * degradation for certain nested-paging guests. Prevent this conversion
862 	 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
863 	 *
864 	 * NOTE: we want to use the _safe accessors so as not to #GP kvm
865 	 * guests on older kvm hosts.
866 	 */
867 	msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
868 
869 	if (cpu_has_amd_erratum(c, amd_erratum_383))
870 		set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
871 }
872 
init_amd_ln(struct cpuinfo_x86 * c)873 static void init_amd_ln(struct cpuinfo_x86 *c)
874 {
875 	/*
876 	 * Apply erratum 665 fix unconditionally so machines without a BIOS
877 	 * fix work.
878 	 */
879 	msr_set_bit(MSR_AMD64_DE_CFG, 31);
880 }
881 
882 static bool rdrand_force;
883 
rdrand_cmdline(char * str)884 static int __init rdrand_cmdline(char *str)
885 {
886 	if (!str)
887 		return -EINVAL;
888 
889 	if (!strcmp(str, "force"))
890 		rdrand_force = true;
891 	else
892 		return -EINVAL;
893 
894 	return 0;
895 }
896 early_param("rdrand", rdrand_cmdline);
897 
clear_rdrand_cpuid_bit(struct cpuinfo_x86 * c)898 static void clear_rdrand_cpuid_bit(struct cpuinfo_x86 *c)
899 {
900 	/*
901 	 * Saving of the MSR used to hide the RDRAND support during
902 	 * suspend/resume is done by arch/x86/power/cpu.c, which is
903 	 * dependent on CONFIG_PM_SLEEP.
904 	 */
905 	if (!IS_ENABLED(CONFIG_PM_SLEEP))
906 		return;
907 
908 	/*
909 	 * The nordrand option can clear X86_FEATURE_RDRAND, so check for
910 	 * RDRAND support using the CPUID function directly.
911 	 */
912 	if (!(cpuid_ecx(1) & BIT(30)) || rdrand_force)
913 		return;
914 
915 	msr_clear_bit(MSR_AMD64_CPUID_FN_1, 62);
916 
917 	/*
918 	 * Verify that the CPUID change has occurred in case the kernel is
919 	 * running virtualized and the hypervisor doesn't support the MSR.
920 	 */
921 	if (cpuid_ecx(1) & BIT(30)) {
922 		pr_info_once("BIOS may not properly restore RDRAND after suspend, but hypervisor does not support hiding RDRAND via CPUID.\n");
923 		return;
924 	}
925 
926 	clear_cpu_cap(c, X86_FEATURE_RDRAND);
927 	pr_info_once("BIOS may not properly restore RDRAND after suspend, hiding RDRAND via CPUID. Use rdrand=force to reenable.\n");
928 }
929 
init_amd_jg(struct cpuinfo_x86 * c)930 static void init_amd_jg(struct cpuinfo_x86 *c)
931 {
932 	/*
933 	 * Some BIOS implementations do not restore proper RDRAND support
934 	 * across suspend and resume. Check on whether to hide the RDRAND
935 	 * instruction support via CPUID.
936 	 */
937 	clear_rdrand_cpuid_bit(c);
938 }
939 
init_amd_bd(struct cpuinfo_x86 * c)940 static void init_amd_bd(struct cpuinfo_x86 *c)
941 {
942 	u64 value;
943 
944 	/*
945 	 * The way access filter has a performance penalty on some workloads.
946 	 * Disable it on the affected CPUs.
947 	 */
948 	if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
949 		if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
950 			value |= 0x1E;
951 			wrmsrl_safe(MSR_F15H_IC_CFG, value);
952 		}
953 	}
954 
955 	/*
956 	 * Some BIOS implementations do not restore proper RDRAND support
957 	 * across suspend and resume. Check on whether to hide the RDRAND
958 	 * instruction support via CPUID.
959 	 */
960 	clear_rdrand_cpuid_bit(c);
961 }
962 
init_spectral_chicken(struct cpuinfo_x86 * c)963 void init_spectral_chicken(struct cpuinfo_x86 *c)
964 {
965 #ifdef CONFIG_CPU_UNRET_ENTRY
966 	u64 value;
967 
968 	/*
969 	 * On Zen2 we offer this chicken (bit) on the altar of Speculation.
970 	 *
971 	 * This suppresses speculation from the middle of a basic block, i.e. it
972 	 * suppresses non-branch predictions.
973 	 *
974 	 * We use STIBP as a heuristic to filter out Zen2 from the rest of F17H
975 	 */
976 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR) && cpu_has(c, X86_FEATURE_AMD_STIBP)) {
977 		if (!rdmsrl_safe(MSR_ZEN2_SPECTRAL_CHICKEN, &value)) {
978 			value |= MSR_ZEN2_SPECTRAL_CHICKEN_BIT;
979 			wrmsrl_safe(MSR_ZEN2_SPECTRAL_CHICKEN, value);
980 		}
981 	}
982 #endif
983 	/*
984 	 * Work around Erratum 1386.  The XSAVES instruction malfunctions in
985 	 * certain circumstances on Zen1/2 uarch, and not all parts have had
986 	 * updated microcode at the time of writing (March 2023).
987 	 *
988 	 * Affected parts all have no supervisor XSAVE states, meaning that
989 	 * the XSAVEC instruction (which works fine) is equivalent.
990 	 */
991 	clear_cpu_cap(c, X86_FEATURE_XSAVES);
992 }
993 
init_amd_zn(struct cpuinfo_x86 * c)994 static void init_amd_zn(struct cpuinfo_x86 *c)
995 {
996 	set_cpu_cap(c, X86_FEATURE_ZEN);
997 
998 #ifdef CONFIG_NUMA
999 	node_reclaim_distance = 32;
1000 #endif
1001 
1002 	/* Fix up CPUID bits, but only if not virtualised. */
1003 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR)) {
1004 
1005 		/* Erratum 1076: CPB feature bit not being set in CPUID. */
1006 		if (!cpu_has(c, X86_FEATURE_CPB))
1007 			set_cpu_cap(c, X86_FEATURE_CPB);
1008 
1009 		/*
1010 		 * Zen3 (Fam19 model < 0x10) parts are not susceptible to
1011 		 * Branch Type Confusion, but predate the allocation of the
1012 		 * BTC_NO bit.
1013 		 */
1014 		if (c->x86 == 0x19 && !cpu_has(c, X86_FEATURE_BTC_NO))
1015 			set_cpu_cap(c, X86_FEATURE_BTC_NO);
1016 	}
1017 }
1018 
cpu_has_zenbleed_microcode(void)1019 static bool cpu_has_zenbleed_microcode(void)
1020 {
1021 	u32 good_rev = 0;
1022 
1023 	switch (boot_cpu_data.x86_model) {
1024 	case 0x30 ... 0x3f: good_rev = 0x0830107a; break;
1025 	case 0x60 ... 0x67: good_rev = 0x0860010b; break;
1026 	case 0x68 ... 0x6f: good_rev = 0x08608105; break;
1027 	case 0x70 ... 0x7f: good_rev = 0x08701032; break;
1028 	case 0xa0 ... 0xaf: good_rev = 0x08a00008; break;
1029 
1030 	default:
1031 		return false;
1032 		break;
1033 	}
1034 
1035 	if (boot_cpu_data.microcode < good_rev)
1036 		return false;
1037 
1038 	return true;
1039 }
1040 
zenbleed_check(struct cpuinfo_x86 * c)1041 static void zenbleed_check(struct cpuinfo_x86 *c)
1042 {
1043 	if (!cpu_has_amd_erratum(c, amd_zenbleed))
1044 		return;
1045 
1046 	if (cpu_has(c, X86_FEATURE_HYPERVISOR))
1047 		return;
1048 
1049 	if (!cpu_has(c, X86_FEATURE_AVX))
1050 		return;
1051 
1052 	if (!cpu_has_zenbleed_microcode()) {
1053 		pr_notice_once("Zenbleed: please update your microcode for the most optimal fix\n");
1054 		msr_set_bit(MSR_AMD64_DE_CFG, MSR_AMD64_DE_CFG_ZEN2_FP_BACKUP_FIX_BIT);
1055 	} else {
1056 		msr_clear_bit(MSR_AMD64_DE_CFG, MSR_AMD64_DE_CFG_ZEN2_FP_BACKUP_FIX_BIT);
1057 	}
1058 }
1059 
init_amd(struct cpuinfo_x86 * c)1060 static void init_amd(struct cpuinfo_x86 *c)
1061 {
1062 	early_init_amd(c);
1063 
1064 	/*
1065 	 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
1066 	 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
1067 	 */
1068 	clear_cpu_cap(c, 0*32+31);
1069 
1070 	if (c->x86 >= 0x10)
1071 		set_cpu_cap(c, X86_FEATURE_REP_GOOD);
1072 
1073 	/* get apicid instead of initial apic id from cpuid */
1074 	c->apicid = hard_smp_processor_id();
1075 
1076 	/* K6s reports MCEs but don't actually have all the MSRs */
1077 	if (c->x86 < 6)
1078 		clear_cpu_cap(c, X86_FEATURE_MCE);
1079 
1080 	switch (c->x86) {
1081 	case 4:    init_amd_k5(c); break;
1082 	case 5:    init_amd_k6(c); break;
1083 	case 6:	   init_amd_k7(c); break;
1084 	case 0xf:  init_amd_k8(c); break;
1085 	case 0x10: init_amd_gh(c); break;
1086 	case 0x12: init_amd_ln(c); break;
1087 	case 0x15: init_amd_bd(c); break;
1088 	case 0x16: init_amd_jg(c); break;
1089 	case 0x17: init_spectral_chicken(c);
1090 		   fallthrough;
1091 	case 0x19: init_amd_zn(c); break;
1092 	}
1093 
1094 	/*
1095 	 * Enable workaround for FXSAVE leak on CPUs
1096 	 * without a XSaveErPtr feature
1097 	 */
1098 	if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
1099 		set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
1100 
1101 	cpu_detect_cache_sizes(c);
1102 
1103 	amd_detect_cmp(c);
1104 	amd_get_topology(c);
1105 	srat_detect_node(c);
1106 	amd_detect_ppin(c);
1107 
1108 	init_amd_cacheinfo(c);
1109 
1110 	if (cpu_has(c, X86_FEATURE_XMM2)) {
1111 		/*
1112 		 * Use LFENCE for execution serialization.  On families which
1113 		 * don't have that MSR, LFENCE is already serializing.
1114 		 * msr_set_bit() uses the safe accessors, too, even if the MSR
1115 		 * is not present.
1116 		 */
1117 		msr_set_bit(MSR_AMD64_DE_CFG,
1118 			    MSR_AMD64_DE_CFG_LFENCE_SERIALIZE_BIT);
1119 
1120 		/* A serializing LFENCE stops RDTSC speculation */
1121 		set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
1122 	}
1123 
1124 	/*
1125 	 * Family 0x12 and above processors have APIC timer
1126 	 * running in deep C states.
1127 	 */
1128 	if (c->x86 > 0x11)
1129 		set_cpu_cap(c, X86_FEATURE_ARAT);
1130 
1131 	/* 3DNow or LM implies PREFETCHW */
1132 	if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
1133 		if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
1134 			set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
1135 
1136 	/* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
1137 	if (!cpu_has(c, X86_FEATURE_XENPV))
1138 		set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
1139 
1140 	/*
1141 	 * Turn on the Instructions Retired free counter on machines not
1142 	 * susceptible to erratum #1054 "Instructions Retired Performance
1143 	 * Counter May Be Inaccurate".
1144 	 */
1145 	if (cpu_has(c, X86_FEATURE_IRPERF) &&
1146 	    !cpu_has_amd_erratum(c, amd_erratum_1054))
1147 		msr_set_bit(MSR_K7_HWCR, MSR_K7_HWCR_IRPERF_EN_BIT);
1148 
1149 	check_null_seg_clears_base(c);
1150 
1151 	zenbleed_check(c);
1152 
1153 	if (cpu_has_amd_erratum(c, amd_div0)) {
1154 		pr_notice_once("AMD Zen1 DIV0 bug detected. Disable SMT for full protection.\n");
1155 		setup_force_cpu_bug(X86_BUG_DIV0);
1156 	}
1157 
1158 	if (!cpu_has(c, X86_FEATURE_HYPERVISOR) &&
1159 	     cpu_has_amd_erratum(c, amd_erratum_1485))
1160 		msr_set_bit(MSR_ZEN4_BP_CFG, MSR_ZEN4_BP_CFG_SHARED_BTB_FIX_BIT);
1161 }
1162 
1163 #ifdef CONFIG_X86_32
amd_size_cache(struct cpuinfo_x86 * c,unsigned int size)1164 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
1165 {
1166 	/* AMD errata T13 (order #21922) */
1167 	if (c->x86 == 6) {
1168 		/* Duron Rev A0 */
1169 		if (c->x86_model == 3 && c->x86_stepping == 0)
1170 			size = 64;
1171 		/* Tbird rev A1/A2 */
1172 		if (c->x86_model == 4 &&
1173 			(c->x86_stepping == 0 || c->x86_stepping == 1))
1174 			size = 256;
1175 	}
1176 	return size;
1177 }
1178 #endif
1179 
cpu_detect_tlb_amd(struct cpuinfo_x86 * c)1180 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
1181 {
1182 	u32 ebx, eax, ecx, edx;
1183 	u16 mask = 0xfff;
1184 
1185 	if (c->x86 < 0xf)
1186 		return;
1187 
1188 	if (c->extended_cpuid_level < 0x80000006)
1189 		return;
1190 
1191 	cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
1192 
1193 	tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
1194 	tlb_lli_4k[ENTRIES] = ebx & mask;
1195 
1196 	/*
1197 	 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
1198 	 * characteristics from the CPUID function 0x80000005 instead.
1199 	 */
1200 	if (c->x86 == 0xf) {
1201 		cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1202 		mask = 0xff;
1203 	}
1204 
1205 	/* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1206 	if (!((eax >> 16) & mask))
1207 		tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
1208 	else
1209 		tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
1210 
1211 	/* a 4M entry uses two 2M entries */
1212 	tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
1213 
1214 	/* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1215 	if (!(eax & mask)) {
1216 		/* Erratum 658 */
1217 		if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
1218 			tlb_lli_2m[ENTRIES] = 1024;
1219 		} else {
1220 			cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1221 			tlb_lli_2m[ENTRIES] = eax & 0xff;
1222 		}
1223 	} else
1224 		tlb_lli_2m[ENTRIES] = eax & mask;
1225 
1226 	tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
1227 }
1228 
1229 static const struct cpu_dev amd_cpu_dev = {
1230 	.c_vendor	= "AMD",
1231 	.c_ident	= { "AuthenticAMD" },
1232 #ifdef CONFIG_X86_32
1233 	.legacy_models = {
1234 		{ .family = 4, .model_names =
1235 		  {
1236 			  [3] = "486 DX/2",
1237 			  [7] = "486 DX/2-WB",
1238 			  [8] = "486 DX/4",
1239 			  [9] = "486 DX/4-WB",
1240 			  [14] = "Am5x86-WT",
1241 			  [15] = "Am5x86-WB"
1242 		  }
1243 		},
1244 	},
1245 	.legacy_cache_size = amd_size_cache,
1246 #endif
1247 	.c_early_init   = early_init_amd,
1248 	.c_detect_tlb	= cpu_detect_tlb_amd,
1249 	.c_bsp_init	= bsp_init_amd,
1250 	.c_init		= init_amd,
1251 	.c_x86_vendor	= X86_VENDOR_AMD,
1252 };
1253 
1254 cpu_dev_register(amd_cpu_dev);
1255 
set_dr_addr_mask(unsigned long mask,int dr)1256 void set_dr_addr_mask(unsigned long mask, int dr)
1257 {
1258 	if (!boot_cpu_has(X86_FEATURE_BPEXT))
1259 		return;
1260 
1261 	switch (dr) {
1262 	case 0:
1263 		wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0);
1264 		break;
1265 	case 1:
1266 	case 2:
1267 	case 3:
1268 		wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0);
1269 		break;
1270 	default:
1271 		break;
1272 	}
1273 }
1274 
amd_get_highest_perf(void)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 
cpu_has_ibpb_brtype_microcode(void)1291 bool cpu_has_ibpb_brtype_microcode(void)
1292 {
1293 	switch (boot_cpu_data.x86) {
1294 	/* Zen1/2 IBPB flushes branch type predictions too. */
1295 	case 0x17:
1296 		return boot_cpu_has(X86_FEATURE_AMD_IBPB);
1297 	case 0x19:
1298 		/* Poke the MSR bit on Zen3/4 to check its presence. */
1299 		if (!wrmsrl_safe(MSR_IA32_PRED_CMD, PRED_CMD_SBPB)) {
1300 			setup_force_cpu_cap(X86_FEATURE_SBPB);
1301 			return true;
1302 		} else {
1303 			return false;
1304 		}
1305 	default:
1306 		return false;
1307 	}
1308 }
1309 
zenbleed_check_cpu(void * unused)1310 static void zenbleed_check_cpu(void *unused)
1311 {
1312 	struct cpuinfo_x86 *c = &cpu_data(smp_processor_id());
1313 
1314 	zenbleed_check(c);
1315 }
1316 
amd_check_microcode(void)1317 void amd_check_microcode(void)
1318 {
1319 	if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
1320 		return;
1321 
1322 	on_each_cpu(zenbleed_check_cpu, NULL, 1);
1323 }
1324 
1325 /*
1326  * Issue a DIV 0/1 insn to clear any division data from previous DIV
1327  * operations.
1328  */
amd_clear_divider(void)1329 void noinstr amd_clear_divider(void)
1330 {
1331 	asm volatile(ALTERNATIVE("", "div %2\n\t", X86_BUG_DIV0)
1332 		     :: "a" (0), "d" (0), "r" (1));
1333 }
1334 EXPORT_SYMBOL_GPL(amd_clear_divider);
1335