xref: /freebsd/sys/amd64/vmm/x86.c (revision b9f654b163bce26de79705e77b872427c9f2afa1)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2011 NetApp, Inc.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  * $FreeBSD$
29  */
30 
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33 
34 #include <sys/param.h>
35 #include <sys/pcpu.h>
36 #include <sys/systm.h>
37 #include <sys/sysctl.h>
38 
39 #include <machine/clock.h>
40 #include <machine/cpufunc.h>
41 #include <machine/md_var.h>
42 #include <machine/segments.h>
43 #include <machine/specialreg.h>
44 
45 #include <machine/vmm.h>
46 
47 #include "vmm_host.h"
48 #include "vmm_ktr.h"
49 #include "vmm_util.h"
50 #include "x86.h"
51 
52 SYSCTL_DECL(_hw_vmm);
53 static SYSCTL_NODE(_hw_vmm, OID_AUTO, topology, CTLFLAG_RD, 0, NULL);
54 
55 #define	CPUID_VM_HIGH		0x40000000
56 
57 static const char bhyve_id[12] = "bhyve bhyve ";
58 
59 static uint64_t bhyve_xcpuids;
60 SYSCTL_ULONG(_hw_vmm, OID_AUTO, bhyve_xcpuids, CTLFLAG_RW, &bhyve_xcpuids, 0,
61     "Number of times an unknown cpuid leaf was accessed");
62 
63 #if __FreeBSD_version < 1200060	/* Remove after 11 EOL helps MFCing */
64 extern u_int threads_per_core;
65 SYSCTL_UINT(_hw_vmm_topology, OID_AUTO, threads_per_core, CTLFLAG_RDTUN,
66     &threads_per_core, 0, NULL);
67 
68 extern u_int cores_per_package;
69 SYSCTL_UINT(_hw_vmm_topology, OID_AUTO, cores_per_package, CTLFLAG_RDTUN,
70     &cores_per_package, 0, NULL);
71 #endif
72 
73 static int cpuid_leaf_b = 1;
74 SYSCTL_INT(_hw_vmm_topology, OID_AUTO, cpuid_leaf_b, CTLFLAG_RDTUN,
75     &cpuid_leaf_b, 0, NULL);
76 
77 /*
78  * Round up to the next power of two, if necessary, and then take log2.
79  * Returns -1 if argument is zero.
80  */
81 static __inline int
82 log2(u_int x)
83 {
84 
85 	return (fls(x << (1 - powerof2(x))) - 1);
86 }
87 
88 int
89 x86_emulate_cpuid(struct vm *vm, int vcpu_id,
90 		  uint32_t *eax, uint32_t *ebx, uint32_t *ecx, uint32_t *edx)
91 {
92 	const struct xsave_limits *limits;
93 	uint64_t cr4;
94 	int error, enable_invpcid, level, width, x2apic_id;
95 	unsigned int func, regs[4], logical_cpus;
96 	enum x2apic_state x2apic_state;
97 	uint16_t cores, maxcpus, sockets, threads;
98 
99 	VCPU_CTR2(vm, vcpu_id, "cpuid %#x,%#x", *eax, *ecx);
100 
101 	/*
102 	 * Requests for invalid CPUID levels should map to the highest
103 	 * available level instead.
104 	 */
105 	if (cpu_exthigh != 0 && *eax >= 0x80000000) {
106 		if (*eax > cpu_exthigh)
107 			*eax = cpu_exthigh;
108 	} else if (*eax >= 0x40000000) {
109 		if (*eax > CPUID_VM_HIGH)
110 			*eax = CPUID_VM_HIGH;
111 	} else if (*eax > cpu_high) {
112 		*eax = cpu_high;
113 	}
114 
115 	func = *eax;
116 
117 	/*
118 	 * In general the approach used for CPU topology is to
119 	 * advertise a flat topology where all CPUs are packages with
120 	 * no multi-core or SMT.
121 	 */
122 	switch (func) {
123 		/*
124 		 * Pass these through to the guest
125 		 */
126 		case CPUID_0000_0000:
127 		case CPUID_0000_0002:
128 		case CPUID_0000_0003:
129 		case CPUID_8000_0000:
130 		case CPUID_8000_0002:
131 		case CPUID_8000_0003:
132 		case CPUID_8000_0004:
133 		case CPUID_8000_0006:
134 			cpuid_count(*eax, *ecx, regs);
135 			break;
136 		case CPUID_8000_0008:
137 			cpuid_count(*eax, *ecx, regs);
138 			if (vmm_is_amd()) {
139 				/*
140 				 * As on Intel (0000_0007:0, EDX), mask out
141 				 * unsupported or unsafe AMD extended features
142 				 * (8000_0008 EBX).
143 				 */
144 				regs[1] &= (AMDFEID_CLZERO | AMDFEID_IRPERF |
145 				    AMDFEID_XSAVEERPTR);
146 
147 				vm_get_topology(vm, &sockets, &cores, &threads,
148 				    &maxcpus);
149 				/*
150 				 * Here, width is ApicIdCoreIdSize, present on
151 				 * at least Family 15h and newer.  It
152 				 * represents the "number of bits in the
153 				 * initial apicid that indicate thread id
154 				 * within a package."
155 				 *
156 				 * Our topo_probe_amd() uses it for
157 				 * pkg_id_shift and other OSes may rely on it.
158 				 */
159 				width = MIN(0xF, log2(threads * cores));
160 				if (width < 0x4)
161 					width = 0;
162 				logical_cpus = MIN(0xFF, threads * cores - 1);
163 				regs[2] = (width << AMDID_COREID_SIZE_SHIFT) | logical_cpus;
164 			}
165 			break;
166 
167 		case CPUID_8000_0001:
168 			cpuid_count(*eax, *ecx, regs);
169 
170 			/*
171 			 * Hide SVM from guest.
172 			 */
173 			regs[2] &= ~AMDID2_SVM;
174 
175 			/*
176 			 * Don't advertise extended performance counter MSRs
177 			 * to the guest.
178 			 */
179 			regs[2] &= ~AMDID2_PCXC;
180 			regs[2] &= ~AMDID2_PNXC;
181 			regs[2] &= ~AMDID2_PTSCEL2I;
182 
183 			/*
184 			 * Don't advertise Instruction Based Sampling feature.
185 			 */
186 			regs[2] &= ~AMDID2_IBS;
187 
188 			/* NodeID MSR not available */
189 			regs[2] &= ~AMDID2_NODE_ID;
190 
191 			/* Don't advertise the OS visible workaround feature */
192 			regs[2] &= ~AMDID2_OSVW;
193 
194 			/* Hide mwaitx/monitorx capability from the guest */
195 			regs[2] &= ~AMDID2_MWAITX;
196 
197 			/*
198 			 * Hide rdtscp/ia32_tsc_aux until we know how
199 			 * to deal with them.
200 			 */
201 			regs[3] &= ~AMDID_RDTSCP;
202 			break;
203 
204 		case CPUID_8000_0007:
205 			/*
206 			 * AMD uses this leaf to advertise the processor's
207 			 * power monitoring and RAS capabilities. These
208 			 * features are hardware-specific and exposing
209 			 * them to a guest doesn't make a lot of sense.
210 			 *
211 			 * Intel uses this leaf only to advertise the
212 			 * "Invariant TSC" feature with all other bits
213 			 * being reserved (set to zero).
214 			 */
215 			regs[0] = 0;
216 			regs[1] = 0;
217 			regs[2] = 0;
218 			regs[3] = 0;
219 
220 			/*
221 			 * "Invariant TSC" can be advertised to the guest if:
222 			 * - host TSC frequency is invariant
223 			 * - host TSCs are synchronized across physical cpus
224 			 *
225 			 * XXX This still falls short because the vcpu
226 			 * can observe the TSC moving backwards as it
227 			 * migrates across physical cpus. But at least
228 			 * it should discourage the guest from using the
229 			 * TSC to keep track of time.
230 			 */
231 			if (tsc_is_invariant && smp_tsc)
232 				regs[3] |= AMDPM_TSC_INVARIANT;
233 			break;
234 
235 		case CPUID_8000_001D:
236 			/* AMD Cache topology, like 0000_0004 for Intel. */
237 			if (!vmm_is_amd())
238 				goto default_leaf;
239 
240 			/*
241 			 * Similar to Intel, generate a ficticious cache
242 			 * topology for the guest with L3 shared by the
243 			 * package, and L1 and L2 local to a core.
244 			 */
245 			vm_get_topology(vm, &sockets, &cores, &threads,
246 			    &maxcpus);
247 			switch (*ecx) {
248 			case 0:
249 				logical_cpus = threads;
250 				level = 1;
251 				func = 1;	/* data cache */
252 				break;
253 			case 1:
254 				logical_cpus = threads;
255 				level = 2;
256 				func = 3;	/* unified cache */
257 				break;
258 			case 2:
259 				logical_cpus = threads * cores;
260 				level = 3;
261 				func = 3;	/* unified cache */
262 				break;
263 			default:
264 				logical_cpus = 0;
265 				level = 0;
266 				func = 0;
267 				break;
268 			}
269 
270 			logical_cpus = MIN(0xfff, logical_cpus - 1);
271 			regs[0] = (logical_cpus << 14) | (1 << 8) |
272 			    (level << 5) | func;
273 			regs[1] = (func > 0) ? (CACHE_LINE_SIZE - 1) : 0;
274 			regs[2] = 0;
275 			regs[3] = 0;
276 			break;
277 
278 		case CPUID_8000_001E:
279 			/* AMD Family 16h+ additional identifiers */
280 			if (!vmm_is_amd() || CPUID_TO_FAMILY(cpu_id) < 0x16)
281 				goto default_leaf;
282 
283 			vm_get_topology(vm, &sockets, &cores, &threads,
284 			    &maxcpus);
285 			regs[0] = vcpu_id;
286 			threads = MIN(0xFF, threads - 1);
287 			regs[1] = (threads << 8) |
288 			    (vcpu_id >> log2(threads + 1));
289 			/*
290 			 * XXX Bhyve topology cannot yet represent >1 node per
291 			 * processor.
292 			 */
293 			regs[2] = 0;
294 			regs[3] = 0;
295 			break;
296 
297 		case CPUID_0000_0001:
298 			do_cpuid(1, regs);
299 
300 			error = vm_get_x2apic_state(vm, vcpu_id, &x2apic_state);
301 			if (error) {
302 				panic("x86_emulate_cpuid: error %d "
303 				      "fetching x2apic state", error);
304 			}
305 
306 			/*
307 			 * Override the APIC ID only in ebx
308 			 */
309 			regs[1] &= ~(CPUID_LOCAL_APIC_ID);
310 			regs[1] |= (vcpu_id << CPUID_0000_0001_APICID_SHIFT);
311 
312 			/*
313 			 * Don't expose VMX, SpeedStep, TME or SMX capability.
314 			 * Advertise x2APIC capability and Hypervisor guest.
315 			 */
316 			regs[2] &= ~(CPUID2_VMX | CPUID2_EST | CPUID2_TM2);
317 			regs[2] &= ~(CPUID2_SMX);
318 
319 			regs[2] |= CPUID2_HV;
320 
321 			if (x2apic_state != X2APIC_DISABLED)
322 				regs[2] |= CPUID2_X2APIC;
323 			else
324 				regs[2] &= ~CPUID2_X2APIC;
325 
326 			/*
327 			 * Only advertise CPUID2_XSAVE in the guest if
328 			 * the host is using XSAVE.
329 			 */
330 			if (!(regs[2] & CPUID2_OSXSAVE))
331 				regs[2] &= ~CPUID2_XSAVE;
332 
333 			/*
334 			 * If CPUID2_XSAVE is being advertised and the
335 			 * guest has set CR4_XSAVE, set
336 			 * CPUID2_OSXSAVE.
337 			 */
338 			regs[2] &= ~CPUID2_OSXSAVE;
339 			if (regs[2] & CPUID2_XSAVE) {
340 				error = vm_get_register(vm, vcpu_id,
341 				    VM_REG_GUEST_CR4, &cr4);
342 				if (error)
343 					panic("x86_emulate_cpuid: error %d "
344 					      "fetching %%cr4", error);
345 				if (cr4 & CR4_XSAVE)
346 					regs[2] |= CPUID2_OSXSAVE;
347 			}
348 
349 			/*
350 			 * Hide monitor/mwait until we know how to deal with
351 			 * these instructions.
352 			 */
353 			regs[2] &= ~CPUID2_MON;
354 
355                         /*
356 			 * Hide the performance and debug features.
357 			 */
358 			regs[2] &= ~CPUID2_PDCM;
359 
360 			/*
361 			 * No TSC deadline support in the APIC yet
362 			 */
363 			regs[2] &= ~CPUID2_TSCDLT;
364 
365 			/*
366 			 * Hide thermal monitoring
367 			 */
368 			regs[3] &= ~(CPUID_ACPI | CPUID_TM);
369 
370 			/*
371 			 * Hide the debug store capability.
372 			 */
373 			regs[3] &= ~CPUID_DS;
374 
375 			/*
376 			 * Advertise the Machine Check and MTRR capability.
377 			 *
378 			 * Some guest OSes (e.g. Windows) will not boot if
379 			 * these features are absent.
380 			 */
381 			regs[3] |= (CPUID_MCA | CPUID_MCE | CPUID_MTRR);
382 
383 			vm_get_topology(vm, &sockets, &cores, &threads,
384 			    &maxcpus);
385 			logical_cpus = threads * cores;
386 			regs[1] &= ~CPUID_HTT_CORES;
387 			regs[1] |= (logical_cpus & 0xff) << 16;
388 			regs[3] |= CPUID_HTT;
389 			break;
390 
391 		case CPUID_0000_0004:
392 			cpuid_count(*eax, *ecx, regs);
393 
394 			if (regs[0] || regs[1] || regs[2] || regs[3]) {
395 				vm_get_topology(vm, &sockets, &cores, &threads,
396 				    &maxcpus);
397 				regs[0] &= 0x3ff;
398 				regs[0] |= (cores - 1) << 26;
399 				/*
400 				 * Cache topology:
401 				 * - L1 and L2 are shared only by the logical
402 				 *   processors in a single core.
403 				 * - L3 and above are shared by all logical
404 				 *   processors in the package.
405 				 */
406 				logical_cpus = threads;
407 				level = (regs[0] >> 5) & 0x7;
408 				if (level >= 3)
409 					logical_cpus *= cores;
410 				regs[0] |= (logical_cpus - 1) << 14;
411 			}
412 			break;
413 
414 		case CPUID_0000_0007:
415 			regs[0] = 0;
416 			regs[1] = 0;
417 			regs[2] = 0;
418 			regs[3] = 0;
419 
420 			/* leaf 0 */
421 			if (*ecx == 0) {
422 				cpuid_count(*eax, *ecx, regs);
423 
424 				/* Only leaf 0 is supported */
425 				regs[0] = 0;
426 
427 				/*
428 				 * Expose known-safe features.
429 				 */
430 				regs[1] &= (CPUID_STDEXT_FSGSBASE |
431 				    CPUID_STDEXT_BMI1 | CPUID_STDEXT_HLE |
432 				    CPUID_STDEXT_AVX2 | CPUID_STDEXT_BMI2 |
433 				    CPUID_STDEXT_ERMS | CPUID_STDEXT_RTM |
434 				    CPUID_STDEXT_AVX512F |
435 				    CPUID_STDEXT_RDSEED |
436 				    CPUID_STDEXT_AVX512PF |
437 				    CPUID_STDEXT_AVX512ER |
438 				    CPUID_STDEXT_AVX512CD | CPUID_STDEXT_SHA);
439 				regs[2] = 0;
440 				regs[3] &= CPUID_STDEXT3_MD_CLEAR;
441 
442 				/* Advertise INVPCID if it is enabled. */
443 				error = vm_get_capability(vm, vcpu_id,
444 				    VM_CAP_ENABLE_INVPCID, &enable_invpcid);
445 				if (error == 0 && enable_invpcid)
446 					regs[1] |= CPUID_STDEXT_INVPCID;
447 			}
448 			break;
449 
450 		case CPUID_0000_0006:
451 			regs[0] = CPUTPM1_ARAT;
452 			regs[1] = 0;
453 			regs[2] = 0;
454 			regs[3] = 0;
455 			break;
456 
457 		case CPUID_0000_000A:
458 			/*
459 			 * Handle the access, but report 0 for
460 			 * all options
461 			 */
462 			regs[0] = 0;
463 			regs[1] = 0;
464 			regs[2] = 0;
465 			regs[3] = 0;
466 			break;
467 
468 		case CPUID_0000_000B:
469 			/*
470 			 * Intel processor topology enumeration
471 			 */
472 			if (vmm_is_intel()) {
473 				vm_get_topology(vm, &sockets, &cores, &threads,
474 				    &maxcpus);
475 				if (*ecx == 0) {
476 					logical_cpus = threads;
477 					width = log2(logical_cpus);
478 					level = CPUID_TYPE_SMT;
479 					x2apic_id = vcpu_id;
480 				}
481 
482 				if (*ecx == 1) {
483 					logical_cpus = threads * cores;
484 					width = log2(logical_cpus);
485 					level = CPUID_TYPE_CORE;
486 					x2apic_id = vcpu_id;
487 				}
488 
489 				if (!cpuid_leaf_b || *ecx >= 2) {
490 					width = 0;
491 					logical_cpus = 0;
492 					level = 0;
493 					x2apic_id = 0;
494 				}
495 
496 				regs[0] = width & 0x1f;
497 				regs[1] = logical_cpus & 0xffff;
498 				regs[2] = (level << 8) | (*ecx & 0xff);
499 				regs[3] = x2apic_id;
500 			} else {
501 				regs[0] = 0;
502 				regs[1] = 0;
503 				regs[2] = 0;
504 				regs[3] = 0;
505 			}
506 			break;
507 
508 		case CPUID_0000_000D:
509 			limits = vmm_get_xsave_limits();
510 			if (!limits->xsave_enabled) {
511 				regs[0] = 0;
512 				regs[1] = 0;
513 				regs[2] = 0;
514 				regs[3] = 0;
515 				break;
516 			}
517 
518 			cpuid_count(*eax, *ecx, regs);
519 			switch (*ecx) {
520 			case 0:
521 				/*
522 				 * Only permit the guest to use bits
523 				 * that are active in the host in
524 				 * %xcr0.  Also, claim that the
525 				 * maximum save area size is
526 				 * equivalent to the host's current
527 				 * save area size.  Since this runs
528 				 * "inside" of vmrun(), it runs with
529 				 * the guest's xcr0, so the current
530 				 * save area size is correct as-is.
531 				 */
532 				regs[0] &= limits->xcr0_allowed;
533 				regs[2] = limits->xsave_max_size;
534 				regs[3] &= (limits->xcr0_allowed >> 32);
535 				break;
536 			case 1:
537 				/* Only permit XSAVEOPT. */
538 				regs[0] &= CPUID_EXTSTATE_XSAVEOPT;
539 				regs[1] = 0;
540 				regs[2] = 0;
541 				regs[3] = 0;
542 				break;
543 			default:
544 				/*
545 				 * If the leaf is for a permitted feature,
546 				 * pass through as-is, otherwise return
547 				 * all zeroes.
548 				 */
549 				if (!(limits->xcr0_allowed & (1ul << *ecx))) {
550 					regs[0] = 0;
551 					regs[1] = 0;
552 					regs[2] = 0;
553 					regs[3] = 0;
554 				}
555 				break;
556 			}
557 			break;
558 
559 		case 0x40000000:
560 			regs[0] = CPUID_VM_HIGH;
561 			bcopy(bhyve_id, &regs[1], 4);
562 			bcopy(bhyve_id + 4, &regs[2], 4);
563 			bcopy(bhyve_id + 8, &regs[3], 4);
564 			break;
565 
566 		default:
567 default_leaf:
568 			/*
569 			 * The leaf value has already been clamped so
570 			 * simply pass this through, keeping count of
571 			 * how many unhandled leaf values have been seen.
572 			 */
573 			atomic_add_long(&bhyve_xcpuids, 1);
574 			cpuid_count(*eax, *ecx, regs);
575 			break;
576 	}
577 
578 	*eax = regs[0];
579 	*ebx = regs[1];
580 	*ecx = regs[2];
581 	*edx = regs[3];
582 
583 	return (1);
584 }
585 
586 bool
587 vm_cpuid_capability(struct vm *vm, int vcpuid, enum vm_cpuid_capability cap)
588 {
589 	bool rv;
590 
591 	KASSERT(cap > 0 && cap < VCC_LAST, ("%s: invalid vm_cpu_capability %d",
592 	    __func__, cap));
593 
594 	/*
595 	 * Simply passthrough the capabilities of the host cpu for now.
596 	 */
597 	rv = false;
598 	switch (cap) {
599 	case VCC_NO_EXECUTE:
600 		if (amd_feature & AMDID_NX)
601 			rv = true;
602 		break;
603 	case VCC_FFXSR:
604 		if (amd_feature & AMDID_FFXSR)
605 			rv = true;
606 		break;
607 	case VCC_TCE:
608 		if (amd_feature2 & AMDID2_TCE)
609 			rv = true;
610 		break;
611 	default:
612 		panic("%s: unknown vm_cpu_capability %d", __func__, cap);
613 	}
614 	return (rv);
615 }
616