xref: /freebsd/sys/amd64/vmm/x86.c (revision eb69d1f144a6fcc765d1b9d44a5ae8082353e70b)
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 /*
64  * The default CPU topology is a single thread per package.
65  */
66 static u_int threads_per_core = 1;
67 SYSCTL_UINT(_hw_vmm_topology, OID_AUTO, threads_per_core, CTLFLAG_RDTUN,
68     &threads_per_core, 0, NULL);
69 
70 static u_int cores_per_package = 1;
71 SYSCTL_UINT(_hw_vmm_topology, OID_AUTO, cores_per_package, CTLFLAG_RDTUN,
72     &cores_per_package, 0, NULL);
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 
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 				 * XXX this might appear silly because AMD
141 				 * cpus don't have threads.
142 				 *
143 				 * However this matches the logical cpus as
144 				 * advertised by leaf 0x1 and will work even
145 				 * if the 'threads_per_core' tunable is set
146 				 * incorrectly on an AMD host.
147 				 */
148 				logical_cpus = threads_per_core *
149 				    cores_per_package;
150 				regs[2] = logical_cpus - 1;
151 			}
152 			break;
153 
154 		case CPUID_8000_0001:
155 			cpuid_count(*eax, *ecx, regs);
156 
157 			/*
158 			 * Hide SVM and Topology Extension features from guest.
159 			 */
160 			regs[2] &= ~(AMDID2_SVM | AMDID2_TOPOLOGY);
161 
162 			/*
163 			 * Don't advertise extended performance counter MSRs
164 			 * to the guest.
165 			 */
166 			regs[2] &= ~AMDID2_PCXC;
167 			regs[2] &= ~AMDID2_PNXC;
168 			regs[2] &= ~AMDID2_PTSCEL2I;
169 
170 			/*
171 			 * Don't advertise Instruction Based Sampling feature.
172 			 */
173 			regs[2] &= ~AMDID2_IBS;
174 
175 			/* NodeID MSR not available */
176 			regs[2] &= ~AMDID2_NODE_ID;
177 
178 			/* Don't advertise the OS visible workaround feature */
179 			regs[2] &= ~AMDID2_OSVW;
180 
181 			/* Hide mwaitx/monitorx capability from the guest */
182 			regs[2] &= ~AMDID2_MWAITX;
183 
184 			/*
185 			 * Hide rdtscp/ia32_tsc_aux until we know how
186 			 * to deal with them.
187 			 */
188 			regs[3] &= ~AMDID_RDTSCP;
189 			break;
190 
191 		case CPUID_8000_0007:
192 			/*
193 			 * AMD uses this leaf to advertise the processor's
194 			 * power monitoring and RAS capabilities. These
195 			 * features are hardware-specific and exposing
196 			 * them to a guest doesn't make a lot of sense.
197 			 *
198 			 * Intel uses this leaf only to advertise the
199 			 * "Invariant TSC" feature with all other bits
200 			 * being reserved (set to zero).
201 			 */
202 			regs[0] = 0;
203 			regs[1] = 0;
204 			regs[2] = 0;
205 			regs[3] = 0;
206 
207 			/*
208 			 * "Invariant TSC" can be advertised to the guest if:
209 			 * - host TSC frequency is invariant
210 			 * - host TSCs are synchronized across physical cpus
211 			 *
212 			 * XXX This still falls short because the vcpu
213 			 * can observe the TSC moving backwards as it
214 			 * migrates across physical cpus. But at least
215 			 * it should discourage the guest from using the
216 			 * TSC to keep track of time.
217 			 */
218 			if (tsc_is_invariant && smp_tsc)
219 				regs[3] |= AMDPM_TSC_INVARIANT;
220 			break;
221 
222 		case CPUID_0000_0001:
223 			do_cpuid(1, regs);
224 
225 			error = vm_get_x2apic_state(vm, vcpu_id, &x2apic_state);
226 			if (error) {
227 				panic("x86_emulate_cpuid: error %d "
228 				      "fetching x2apic state", error);
229 			}
230 
231 			/*
232 			 * Override the APIC ID only in ebx
233 			 */
234 			regs[1] &= ~(CPUID_LOCAL_APIC_ID);
235 			regs[1] |= (vcpu_id << CPUID_0000_0001_APICID_SHIFT);
236 
237 			/*
238 			 * Don't expose VMX, SpeedStep, TME or SMX capability.
239 			 * Advertise x2APIC capability and Hypervisor guest.
240 			 */
241 			regs[2] &= ~(CPUID2_VMX | CPUID2_EST | CPUID2_TM2);
242 			regs[2] &= ~(CPUID2_SMX);
243 
244 			regs[2] |= CPUID2_HV;
245 
246 			if (x2apic_state != X2APIC_DISABLED)
247 				regs[2] |= CPUID2_X2APIC;
248 			else
249 				regs[2] &= ~CPUID2_X2APIC;
250 
251 			/*
252 			 * Only advertise CPUID2_XSAVE in the guest if
253 			 * the host is using XSAVE.
254 			 */
255 			if (!(regs[2] & CPUID2_OSXSAVE))
256 				regs[2] &= ~CPUID2_XSAVE;
257 
258 			/*
259 			 * If CPUID2_XSAVE is being advertised and the
260 			 * guest has set CR4_XSAVE, set
261 			 * CPUID2_OSXSAVE.
262 			 */
263 			regs[2] &= ~CPUID2_OSXSAVE;
264 			if (regs[2] & CPUID2_XSAVE) {
265 				error = vm_get_register(vm, vcpu_id,
266 				    VM_REG_GUEST_CR4, &cr4);
267 				if (error)
268 					panic("x86_emulate_cpuid: error %d "
269 					      "fetching %%cr4", error);
270 				if (cr4 & CR4_XSAVE)
271 					regs[2] |= CPUID2_OSXSAVE;
272 			}
273 
274 			/*
275 			 * Hide monitor/mwait until we know how to deal with
276 			 * these instructions.
277 			 */
278 			regs[2] &= ~CPUID2_MON;
279 
280                         /*
281 			 * Hide the performance and debug features.
282 			 */
283 			regs[2] &= ~CPUID2_PDCM;
284 
285 			/*
286 			 * No TSC deadline support in the APIC yet
287 			 */
288 			regs[2] &= ~CPUID2_TSCDLT;
289 
290 			/*
291 			 * Hide thermal monitoring
292 			 */
293 			regs[3] &= ~(CPUID_ACPI | CPUID_TM);
294 
295 			/*
296 			 * Hide the debug store capability.
297 			 */
298 			regs[3] &= ~CPUID_DS;
299 
300 			/*
301 			 * Advertise the Machine Check and MTRR capability.
302 			 *
303 			 * Some guest OSes (e.g. Windows) will not boot if
304 			 * these features are absent.
305 			 */
306 			regs[3] |= (CPUID_MCA | CPUID_MCE | CPUID_MTRR);
307 
308 			logical_cpus = threads_per_core * cores_per_package;
309 			regs[1] &= ~CPUID_HTT_CORES;
310 			regs[1] |= (logical_cpus & 0xff) << 16;
311 			regs[3] |= CPUID_HTT;
312 			break;
313 
314 		case CPUID_0000_0004:
315 			cpuid_count(*eax, *ecx, regs);
316 
317 			if (regs[0] || regs[1] || regs[2] || regs[3]) {
318 				regs[0] &= 0x3ff;
319 				regs[0] |= (cores_per_package - 1) << 26;
320 				/*
321 				 * Cache topology:
322 				 * - L1 and L2 are shared only by the logical
323 				 *   processors in a single core.
324 				 * - L3 and above are shared by all logical
325 				 *   processors in the package.
326 				 */
327 				logical_cpus = threads_per_core;
328 				level = (regs[0] >> 5) & 0x7;
329 				if (level >= 3)
330 					logical_cpus *= cores_per_package;
331 				regs[0] |= (logical_cpus - 1) << 14;
332 			}
333 			break;
334 
335 		case CPUID_0000_0007:
336 			regs[0] = 0;
337 			regs[1] = 0;
338 			regs[2] = 0;
339 			regs[3] = 0;
340 
341 			/* leaf 0 */
342 			if (*ecx == 0) {
343 				cpuid_count(*eax, *ecx, regs);
344 
345 				/* Only leaf 0 is supported */
346 				regs[0] = 0;
347 
348 				/*
349 				 * Expose known-safe features.
350 				 */
351 				regs[1] &= (CPUID_STDEXT_FSGSBASE |
352 				    CPUID_STDEXT_BMI1 | CPUID_STDEXT_HLE |
353 				    CPUID_STDEXT_AVX2 | CPUID_STDEXT_BMI2 |
354 				    CPUID_STDEXT_ERMS | CPUID_STDEXT_RTM |
355 				    CPUID_STDEXT_AVX512F |
356 				    CPUID_STDEXT_AVX512PF |
357 				    CPUID_STDEXT_AVX512ER |
358 				    CPUID_STDEXT_AVX512CD);
359 				regs[2] = 0;
360 				regs[3] = 0;
361 
362 				/* Advertise INVPCID if it is enabled. */
363 				error = vm_get_capability(vm, vcpu_id,
364 				    VM_CAP_ENABLE_INVPCID, &enable_invpcid);
365 				if (error == 0 && enable_invpcid)
366 					regs[1] |= CPUID_STDEXT_INVPCID;
367 			}
368 			break;
369 
370 		case CPUID_0000_0006:
371 			regs[0] = CPUTPM1_ARAT;
372 			regs[1] = 0;
373 			regs[2] = 0;
374 			regs[3] = 0;
375 			break;
376 
377 		case CPUID_0000_000A:
378 			/*
379 			 * Handle the access, but report 0 for
380 			 * all options
381 			 */
382 			regs[0] = 0;
383 			regs[1] = 0;
384 			regs[2] = 0;
385 			regs[3] = 0;
386 			break;
387 
388 		case CPUID_0000_000B:
389 			/*
390 			 * Processor topology enumeration
391 			 */
392 			if (*ecx == 0) {
393 				logical_cpus = threads_per_core;
394 				width = log2(logical_cpus);
395 				level = CPUID_TYPE_SMT;
396 				x2apic_id = vcpu_id;
397 			}
398 
399 			if (*ecx == 1) {
400 				logical_cpus = threads_per_core *
401 				    cores_per_package;
402 				width = log2(logical_cpus);
403 				level = CPUID_TYPE_CORE;
404 				x2apic_id = vcpu_id;
405 			}
406 
407 			if (!cpuid_leaf_b || *ecx >= 2) {
408 				width = 0;
409 				logical_cpus = 0;
410 				level = 0;
411 				x2apic_id = 0;
412 			}
413 
414 			regs[0] = width & 0x1f;
415 			regs[1] = logical_cpus & 0xffff;
416 			regs[2] = (level << 8) | (*ecx & 0xff);
417 			regs[3] = x2apic_id;
418 			break;
419 
420 		case CPUID_0000_000D:
421 			limits = vmm_get_xsave_limits();
422 			if (!limits->xsave_enabled) {
423 				regs[0] = 0;
424 				regs[1] = 0;
425 				regs[2] = 0;
426 				regs[3] = 0;
427 				break;
428 			}
429 
430 			cpuid_count(*eax, *ecx, regs);
431 			switch (*ecx) {
432 			case 0:
433 				/*
434 				 * Only permit the guest to use bits
435 				 * that are active in the host in
436 				 * %xcr0.  Also, claim that the
437 				 * maximum save area size is
438 				 * equivalent to the host's current
439 				 * save area size.  Since this runs
440 				 * "inside" of vmrun(), it runs with
441 				 * the guest's xcr0, so the current
442 				 * save area size is correct as-is.
443 				 */
444 				regs[0] &= limits->xcr0_allowed;
445 				regs[2] = limits->xsave_max_size;
446 				regs[3] &= (limits->xcr0_allowed >> 32);
447 				break;
448 			case 1:
449 				/* Only permit XSAVEOPT. */
450 				regs[0] &= CPUID_EXTSTATE_XSAVEOPT;
451 				regs[1] = 0;
452 				regs[2] = 0;
453 				regs[3] = 0;
454 				break;
455 			default:
456 				/*
457 				 * If the leaf is for a permitted feature,
458 				 * pass through as-is, otherwise return
459 				 * all zeroes.
460 				 */
461 				if (!(limits->xcr0_allowed & (1ul << *ecx))) {
462 					regs[0] = 0;
463 					regs[1] = 0;
464 					regs[2] = 0;
465 					regs[3] = 0;
466 				}
467 				break;
468 			}
469 			break;
470 
471 		case 0x40000000:
472 			regs[0] = CPUID_VM_HIGH;
473 			bcopy(bhyve_id, &regs[1], 4);
474 			bcopy(bhyve_id + 4, &regs[2], 4);
475 			bcopy(bhyve_id + 8, &regs[3], 4);
476 			break;
477 
478 		default:
479 			/*
480 			 * The leaf value has already been clamped so
481 			 * simply pass this through, keeping count of
482 			 * how many unhandled leaf values have been seen.
483 			 */
484 			atomic_add_long(&bhyve_xcpuids, 1);
485 			cpuid_count(*eax, *ecx, regs);
486 			break;
487 	}
488 
489 	*eax = regs[0];
490 	*ebx = regs[1];
491 	*ecx = regs[2];
492 	*edx = regs[3];
493 
494 	return (1);
495 }
496 
497 bool
498 vm_cpuid_capability(struct vm *vm, int vcpuid, enum vm_cpuid_capability cap)
499 {
500 	bool rv;
501 
502 	KASSERT(cap > 0 && cap < VCC_LAST, ("%s: invalid vm_cpu_capability %d",
503 	    __func__, cap));
504 
505 	/*
506 	 * Simply passthrough the capabilities of the host cpu for now.
507 	 */
508 	rv = false;
509 	switch (cap) {
510 	case VCC_NO_EXECUTE:
511 		if (amd_feature & AMDID_NX)
512 			rv = true;
513 		break;
514 	case VCC_FFXSR:
515 		if (amd_feature & AMDID_FFXSR)
516 			rv = true;
517 		break;
518 	case VCC_TCE:
519 		if (amd_feature2 & AMDID2_TCE)
520 			rv = true;
521 		break;
522 	default:
523 		panic("%s: unknown vm_cpu_capability %d", __func__, cap);
524 	}
525 	return (rv);
526 }
527