xref: /freebsd/sys/amd64/vmm/x86.c (revision 0bf56da32d83fbd3b5db8d6c72cd1e7cc26fbc66)
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 | CTLFLAG_MPSAFE, 0,
54     NULL);
55 
56 #define	CPUID_VM_HIGH		0x40000000
57 
58 static const char bhyve_id[12] = "bhyve bhyve ";
59 
60 static uint64_t bhyve_xcpuids;
61 SYSCTL_ULONG(_hw_vmm, OID_AUTO, bhyve_xcpuids, CTLFLAG_RW, &bhyve_xcpuids, 0,
62     "Number of times an unknown cpuid leaf was accessed");
63 
64 #if __FreeBSD_version < 1200060	/* Remove after 11 EOL helps MFCing */
65 extern u_int threads_per_core;
66 SYSCTL_UINT(_hw_vmm_topology, OID_AUTO, threads_per_core, CTLFLAG_RDTUN,
67     &threads_per_core, 0, NULL);
68 
69 extern u_int cores_per_package;
70 SYSCTL_UINT(_hw_vmm_topology, OID_AUTO, cores_per_package, CTLFLAG_RDTUN,
71     &cores_per_package, 0, NULL);
72 #endif
73 
74 static int cpuid_leaf_b = 1;
75 SYSCTL_INT(_hw_vmm_topology, OID_AUTO, cpuid_leaf_b, CTLFLAG_RDTUN,
76     &cpuid_leaf_b, 0, NULL);
77 
78 /*
79  * Round up to the next power of two, if necessary, and then take log2.
80  * Returns -1 if argument is zero.
81  */
82 static __inline int
83 log2(u_int x)
84 {
85 
86 	return (fls(x << (1 - powerof2(x))) - 1);
87 }
88 
89 int
90 x86_emulate_cpuid(struct vm *vm, int vcpu_id,
91 		  uint32_t *eax, uint32_t *ebx, uint32_t *ecx, uint32_t *edx)
92 {
93 	const struct xsave_limits *limits;
94 	uint64_t cr4;
95 	int error, enable_invpcid, level, width, x2apic_id;
96 	unsigned int func, regs[4], logical_cpus;
97 	enum x2apic_state x2apic_state;
98 	uint16_t cores, maxcpus, sockets, threads;
99 
100 	VCPU_CTR2(vm, vcpu_id, "cpuid %#x,%#x", *eax, *ecx);
101 
102 	/*
103 	 * Requests for invalid CPUID levels should map to the highest
104 	 * available level instead.
105 	 */
106 	if (cpu_exthigh != 0 && *eax >= 0x80000000) {
107 		if (*eax > cpu_exthigh)
108 			*eax = cpu_exthigh;
109 	} else if (*eax >= 0x40000000) {
110 		if (*eax > CPUID_VM_HIGH)
111 			*eax = CPUID_VM_HIGH;
112 	} else if (*eax > cpu_high) {
113 		*eax = cpu_high;
114 	}
115 
116 	func = *eax;
117 
118 	/*
119 	 * In general the approach used for CPU topology is to
120 	 * advertise a flat topology where all CPUs are packages with
121 	 * no multi-core or SMT.
122 	 */
123 	switch (func) {
124 		/*
125 		 * Pass these through to the guest
126 		 */
127 		case CPUID_0000_0000:
128 		case CPUID_0000_0002:
129 		case CPUID_0000_0003:
130 		case CPUID_8000_0000:
131 		case CPUID_8000_0002:
132 		case CPUID_8000_0003:
133 		case CPUID_8000_0004:
134 		case CPUID_8000_0006:
135 			cpuid_count(*eax, *ecx, regs);
136 			break;
137 		case CPUID_8000_0008:
138 			cpuid_count(*eax, *ecx, regs);
139 			if (vmm_is_svm()) {
140 				/*
141 				 * As on Intel (0000_0007:0, EDX), mask out
142 				 * unsupported or unsafe AMD extended features
143 				 * (8000_0008 EBX).
144 				 */
145 				regs[1] &= (AMDFEID_CLZERO | AMDFEID_IRPERF |
146 				    AMDFEID_XSAVEERPTR);
147 
148 				vm_get_topology(vm, &sockets, &cores, &threads,
149 				    &maxcpus);
150 				/*
151 				 * Here, width is ApicIdCoreIdSize, present on
152 				 * at least Family 15h and newer.  It
153 				 * represents the "number of bits in the
154 				 * initial apicid that indicate thread id
155 				 * within a package."
156 				 *
157 				 * Our topo_probe_amd() uses it for
158 				 * pkg_id_shift and other OSes may rely on it.
159 				 */
160 				width = MIN(0xF, log2(threads * cores));
161 				if (width < 0x4)
162 					width = 0;
163 				logical_cpus = MIN(0xFF, threads * cores - 1);
164 				regs[2] = (width << AMDID_COREID_SIZE_SHIFT) | logical_cpus;
165 			}
166 			break;
167 
168 		case CPUID_8000_0001:
169 			cpuid_count(*eax, *ecx, regs);
170 
171 			/*
172 			 * Hide SVM from guest.
173 			 */
174 			regs[2] &= ~AMDID2_SVM;
175 
176 			/*
177 			 * Don't advertise extended performance counter MSRs
178 			 * to the guest.
179 			 */
180 			regs[2] &= ~AMDID2_PCXC;
181 			regs[2] &= ~AMDID2_PNXC;
182 			regs[2] &= ~AMDID2_PTSCEL2I;
183 
184 			/*
185 			 * Don't advertise Instruction Based Sampling feature.
186 			 */
187 			regs[2] &= ~AMDID2_IBS;
188 
189 			/* NodeID MSR not available */
190 			regs[2] &= ~AMDID2_NODE_ID;
191 
192 			/* Don't advertise the OS visible workaround feature */
193 			regs[2] &= ~AMDID2_OSVW;
194 
195 			/* Hide mwaitx/monitorx capability from the guest */
196 			regs[2] &= ~AMDID2_MWAITX;
197 
198 			/*
199 			 * Hide rdtscp/ia32_tsc_aux until we know how
200 			 * to deal with them.
201 			 */
202 			regs[3] &= ~AMDID_RDTSCP;
203 			break;
204 
205 		case CPUID_8000_0007:
206 			/*
207 			 * AMD uses this leaf to advertise the processor's
208 			 * power monitoring and RAS capabilities. These
209 			 * features are hardware-specific and exposing
210 			 * them to a guest doesn't make a lot of sense.
211 			 *
212 			 * Intel uses this leaf only to advertise the
213 			 * "Invariant TSC" feature with all other bits
214 			 * being reserved (set to zero).
215 			 */
216 			regs[0] = 0;
217 			regs[1] = 0;
218 			regs[2] = 0;
219 			regs[3] = 0;
220 
221 			/*
222 			 * "Invariant TSC" can be advertised to the guest if:
223 			 * - host TSC frequency is invariant
224 			 * - host TSCs are synchronized across physical cpus
225 			 *
226 			 * XXX This still falls short because the vcpu
227 			 * can observe the TSC moving backwards as it
228 			 * migrates across physical cpus. But at least
229 			 * it should discourage the guest from using the
230 			 * TSC to keep track of time.
231 			 */
232 			if (tsc_is_invariant && smp_tsc)
233 				regs[3] |= AMDPM_TSC_INVARIANT;
234 			break;
235 
236 		case CPUID_8000_001D:
237 			/* AMD Cache topology, like 0000_0004 for Intel. */
238 			if (!vmm_is_svm())
239 				goto default_leaf;
240 
241 			/*
242 			 * Similar to Intel, generate a ficticious cache
243 			 * topology for the guest with L3 shared by the
244 			 * package, and L1 and L2 local to a core.
245 			 */
246 			vm_get_topology(vm, &sockets, &cores, &threads,
247 			    &maxcpus);
248 			switch (*ecx) {
249 			case 0:
250 				logical_cpus = threads;
251 				level = 1;
252 				func = 1;	/* data cache */
253 				break;
254 			case 1:
255 				logical_cpus = threads;
256 				level = 2;
257 				func = 3;	/* unified cache */
258 				break;
259 			case 2:
260 				logical_cpus = threads * cores;
261 				level = 3;
262 				func = 3;	/* unified cache */
263 				break;
264 			default:
265 				logical_cpus = 0;
266 				level = 0;
267 				func = 0;
268 				break;
269 			}
270 
271 			logical_cpus = MIN(0xfff, logical_cpus - 1);
272 			regs[0] = (logical_cpus << 14) | (1 << 8) |
273 			    (level << 5) | func;
274 			regs[1] = (func > 0) ? (CACHE_LINE_SIZE - 1) : 0;
275 			regs[2] = 0;
276 			regs[3] = 0;
277 			break;
278 
279 		case CPUID_8000_001E:
280 			/*
281 			 * AMD Family 16h+ and Hygon Family 18h additional
282 			 * identifiers.
283 			 */
284 			if (!vmm_is_svm() || CPUID_TO_FAMILY(cpu_id) < 0x16)
285 				goto default_leaf;
286 
287 			vm_get_topology(vm, &sockets, &cores, &threads,
288 			    &maxcpus);
289 			regs[0] = vcpu_id;
290 			threads = MIN(0xFF, threads - 1);
291 			regs[1] = (threads << 8) |
292 			    (vcpu_id >> log2(threads + 1));
293 			/*
294 			 * XXX Bhyve topology cannot yet represent >1 node per
295 			 * processor.
296 			 */
297 			regs[2] = 0;
298 			regs[3] = 0;
299 			break;
300 
301 		case CPUID_0000_0001:
302 			do_cpuid(1, regs);
303 
304 			error = vm_get_x2apic_state(vm, vcpu_id, &x2apic_state);
305 			if (error) {
306 				panic("x86_emulate_cpuid: error %d "
307 				      "fetching x2apic state", error);
308 			}
309 
310 			/*
311 			 * Override the APIC ID only in ebx
312 			 */
313 			regs[1] &= ~(CPUID_LOCAL_APIC_ID);
314 			regs[1] |= (vcpu_id << CPUID_0000_0001_APICID_SHIFT);
315 
316 			/*
317 			 * Don't expose VMX, SpeedStep, TME or SMX capability.
318 			 * Advertise x2APIC capability and Hypervisor guest.
319 			 */
320 			regs[2] &= ~(CPUID2_VMX | CPUID2_EST | CPUID2_TM2);
321 			regs[2] &= ~(CPUID2_SMX);
322 
323 			regs[2] |= CPUID2_HV;
324 
325 			if (x2apic_state != X2APIC_DISABLED)
326 				regs[2] |= CPUID2_X2APIC;
327 			else
328 				regs[2] &= ~CPUID2_X2APIC;
329 
330 			/*
331 			 * Only advertise CPUID2_XSAVE in the guest if
332 			 * the host is using XSAVE.
333 			 */
334 			if (!(regs[2] & CPUID2_OSXSAVE))
335 				regs[2] &= ~CPUID2_XSAVE;
336 
337 			/*
338 			 * If CPUID2_XSAVE is being advertised and the
339 			 * guest has set CR4_XSAVE, set
340 			 * CPUID2_OSXSAVE.
341 			 */
342 			regs[2] &= ~CPUID2_OSXSAVE;
343 			if (regs[2] & CPUID2_XSAVE) {
344 				error = vm_get_register(vm, vcpu_id,
345 				    VM_REG_GUEST_CR4, &cr4);
346 				if (error)
347 					panic("x86_emulate_cpuid: error %d "
348 					      "fetching %%cr4", error);
349 				if (cr4 & CR4_XSAVE)
350 					regs[2] |= CPUID2_OSXSAVE;
351 			}
352 
353 			/*
354 			 * Hide monitor/mwait until we know how to deal with
355 			 * these instructions.
356 			 */
357 			regs[2] &= ~CPUID2_MON;
358 
359                         /*
360 			 * Hide the performance and debug features.
361 			 */
362 			regs[2] &= ~CPUID2_PDCM;
363 
364 			/*
365 			 * No TSC deadline support in the APIC yet
366 			 */
367 			regs[2] &= ~CPUID2_TSCDLT;
368 
369 			/*
370 			 * Hide thermal monitoring
371 			 */
372 			regs[3] &= ~(CPUID_ACPI | CPUID_TM);
373 
374 			/*
375 			 * Hide the debug store capability.
376 			 */
377 			regs[3] &= ~CPUID_DS;
378 
379 			/*
380 			 * Advertise the Machine Check and MTRR capability.
381 			 *
382 			 * Some guest OSes (e.g. Windows) will not boot if
383 			 * these features are absent.
384 			 */
385 			regs[3] |= (CPUID_MCA | CPUID_MCE | CPUID_MTRR);
386 
387 			vm_get_topology(vm, &sockets, &cores, &threads,
388 			    &maxcpus);
389 			logical_cpus = threads * cores;
390 			regs[1] &= ~CPUID_HTT_CORES;
391 			regs[1] |= (logical_cpus & 0xff) << 16;
392 			regs[3] |= CPUID_HTT;
393 			break;
394 
395 		case CPUID_0000_0004:
396 			cpuid_count(*eax, *ecx, regs);
397 
398 			if (regs[0] || regs[1] || regs[2] || regs[3]) {
399 				vm_get_topology(vm, &sockets, &cores, &threads,
400 				    &maxcpus);
401 				regs[0] &= 0x3ff;
402 				regs[0] |= (cores - 1) << 26;
403 				/*
404 				 * Cache topology:
405 				 * - L1 and L2 are shared only by the logical
406 				 *   processors in a single core.
407 				 * - L3 and above are shared by all logical
408 				 *   processors in the package.
409 				 */
410 				logical_cpus = threads;
411 				level = (regs[0] >> 5) & 0x7;
412 				if (level >= 3)
413 					logical_cpus *= cores;
414 				regs[0] |= (logical_cpus - 1) << 14;
415 			}
416 			break;
417 
418 		case CPUID_0000_0007:
419 			regs[0] = 0;
420 			regs[1] = 0;
421 			regs[2] = 0;
422 			regs[3] = 0;
423 
424 			/* leaf 0 */
425 			if (*ecx == 0) {
426 				cpuid_count(*eax, *ecx, regs);
427 
428 				/* Only leaf 0 is supported */
429 				regs[0] = 0;
430 
431 				/*
432 				 * Expose known-safe features.
433 				 */
434 				regs[1] &= (CPUID_STDEXT_FSGSBASE |
435 				    CPUID_STDEXT_BMI1 | CPUID_STDEXT_HLE |
436 				    CPUID_STDEXT_AVX2 | CPUID_STDEXT_BMI2 |
437 				    CPUID_STDEXT_ERMS | CPUID_STDEXT_RTM |
438 				    CPUID_STDEXT_AVX512F |
439 				    CPUID_STDEXT_RDSEED |
440 				    CPUID_STDEXT_AVX512PF |
441 				    CPUID_STDEXT_AVX512ER |
442 				    CPUID_STDEXT_AVX512CD | CPUID_STDEXT_SHA);
443 				regs[2] = 0;
444 				regs[3] &= CPUID_STDEXT3_MD_CLEAR;
445 
446 				/* Advertise INVPCID if it is enabled. */
447 				error = vm_get_capability(vm, vcpu_id,
448 				    VM_CAP_ENABLE_INVPCID, &enable_invpcid);
449 				if (error == 0 && enable_invpcid)
450 					regs[1] |= CPUID_STDEXT_INVPCID;
451 			}
452 			break;
453 
454 		case CPUID_0000_0006:
455 			regs[0] = CPUTPM1_ARAT;
456 			regs[1] = 0;
457 			regs[2] = 0;
458 			regs[3] = 0;
459 			break;
460 
461 		case CPUID_0000_000A:
462 			/*
463 			 * Handle the access, but report 0 for
464 			 * all options
465 			 */
466 			regs[0] = 0;
467 			regs[1] = 0;
468 			regs[2] = 0;
469 			regs[3] = 0;
470 			break;
471 
472 		case CPUID_0000_000B:
473 			/*
474 			 * Intel processor topology enumeration
475 			 */
476 			if (vmm_is_intel()) {
477 				vm_get_topology(vm, &sockets, &cores, &threads,
478 				    &maxcpus);
479 				if (*ecx == 0) {
480 					logical_cpus = threads;
481 					width = log2(logical_cpus);
482 					level = CPUID_TYPE_SMT;
483 					x2apic_id = vcpu_id;
484 				}
485 
486 				if (*ecx == 1) {
487 					logical_cpus = threads * cores;
488 					width = log2(logical_cpus);
489 					level = CPUID_TYPE_CORE;
490 					x2apic_id = vcpu_id;
491 				}
492 
493 				if (!cpuid_leaf_b || *ecx >= 2) {
494 					width = 0;
495 					logical_cpus = 0;
496 					level = 0;
497 					x2apic_id = 0;
498 				}
499 
500 				regs[0] = width & 0x1f;
501 				regs[1] = logical_cpus & 0xffff;
502 				regs[2] = (level << 8) | (*ecx & 0xff);
503 				regs[3] = x2apic_id;
504 			} else {
505 				regs[0] = 0;
506 				regs[1] = 0;
507 				regs[2] = 0;
508 				regs[3] = 0;
509 			}
510 			break;
511 
512 		case CPUID_0000_000D:
513 			limits = vmm_get_xsave_limits();
514 			if (!limits->xsave_enabled) {
515 				regs[0] = 0;
516 				regs[1] = 0;
517 				regs[2] = 0;
518 				regs[3] = 0;
519 				break;
520 			}
521 
522 			cpuid_count(*eax, *ecx, regs);
523 			switch (*ecx) {
524 			case 0:
525 				/*
526 				 * Only permit the guest to use bits
527 				 * that are active in the host in
528 				 * %xcr0.  Also, claim that the
529 				 * maximum save area size is
530 				 * equivalent to the host's current
531 				 * save area size.  Since this runs
532 				 * "inside" of vmrun(), it runs with
533 				 * the guest's xcr0, so the current
534 				 * save area size is correct as-is.
535 				 */
536 				regs[0] &= limits->xcr0_allowed;
537 				regs[2] = limits->xsave_max_size;
538 				regs[3] &= (limits->xcr0_allowed >> 32);
539 				break;
540 			case 1:
541 				/* Only permit XSAVEOPT. */
542 				regs[0] &= CPUID_EXTSTATE_XSAVEOPT;
543 				regs[1] = 0;
544 				regs[2] = 0;
545 				regs[3] = 0;
546 				break;
547 			default:
548 				/*
549 				 * If the leaf is for a permitted feature,
550 				 * pass through as-is, otherwise return
551 				 * all zeroes.
552 				 */
553 				if (!(limits->xcr0_allowed & (1ul << *ecx))) {
554 					regs[0] = 0;
555 					regs[1] = 0;
556 					regs[2] = 0;
557 					regs[3] = 0;
558 				}
559 				break;
560 			}
561 			break;
562 
563 		case 0x40000000:
564 			regs[0] = CPUID_VM_HIGH;
565 			bcopy(bhyve_id, &regs[1], 4);
566 			bcopy(bhyve_id + 4, &regs[2], 4);
567 			bcopy(bhyve_id + 8, &regs[3], 4);
568 			break;
569 
570 		default:
571 default_leaf:
572 			/*
573 			 * The leaf value has already been clamped so
574 			 * simply pass this through, keeping count of
575 			 * how many unhandled leaf values have been seen.
576 			 */
577 			atomic_add_long(&bhyve_xcpuids, 1);
578 			cpuid_count(*eax, *ecx, regs);
579 			break;
580 	}
581 
582 	*eax = regs[0];
583 	*ebx = regs[1];
584 	*ecx = regs[2];
585 	*edx = regs[3];
586 
587 	return (1);
588 }
589 
590 bool
591 vm_cpuid_capability(struct vm *vm, int vcpuid, enum vm_cpuid_capability cap)
592 {
593 	bool rv;
594 
595 	KASSERT(cap > 0 && cap < VCC_LAST, ("%s: invalid vm_cpu_capability %d",
596 	    __func__, cap));
597 
598 	/*
599 	 * Simply passthrough the capabilities of the host cpu for now.
600 	 */
601 	rv = false;
602 	switch (cap) {
603 	case VCC_NO_EXECUTE:
604 		if (amd_feature & AMDID_NX)
605 			rv = true;
606 		break;
607 	case VCC_FFXSR:
608 		if (amd_feature & AMDID_FFXSR)
609 			rv = true;
610 		break;
611 	case VCC_TCE:
612 		if (amd_feature2 & AMDID2_TCE)
613 			rv = true;
614 		break;
615 	default:
616 		panic("%s: unknown vm_cpu_capability %d", __func__, cap);
617 	}
618 	return (rv);
619 }
620