xref: /freebsd/sys/amd64/vmm/x86.c (revision bd1da0a002e9a43cfb5220835c7a42804d90dc56)
1 /*-
2  * Copyright (c) 2011 NetApp, Inc.
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  *
26  * $FreeBSD$
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include <sys/param.h>
33 #include <sys/pcpu.h>
34 #include <sys/systm.h>
35 #include <sys/sysctl.h>
36 
37 #include <machine/clock.h>
38 #include <machine/cpufunc.h>
39 #include <machine/md_var.h>
40 #include <machine/segments.h>
41 #include <machine/specialreg.h>
42 
43 #include <machine/vmm.h>
44 
45 #include "vmm_host.h"
46 #include "vmm_ktr.h"
47 #include "vmm_util.h"
48 #include "x86.h"
49 
50 SYSCTL_DECL(_hw_vmm);
51 static SYSCTL_NODE(_hw_vmm, OID_AUTO, topology, CTLFLAG_RD, 0, NULL);
52 
53 #define	CPUID_VM_HIGH		0x40000000
54 
55 static const char bhyve_id[12] = "bhyve bhyve ";
56 
57 static uint64_t bhyve_xcpuids;
58 SYSCTL_ULONG(_hw_vmm, OID_AUTO, bhyve_xcpuids, CTLFLAG_RW, &bhyve_xcpuids, 0,
59     "Number of times an unknown cpuid leaf was accessed");
60 
61 /*
62  * The default CPU topology is a single thread per package.
63  */
64 static u_int threads_per_core = 1;
65 SYSCTL_UINT(_hw_vmm_topology, OID_AUTO, threads_per_core, CTLFLAG_RDTUN,
66     &threads_per_core, 0, NULL);
67 
68 static u_int cores_per_package = 1;
69 SYSCTL_UINT(_hw_vmm_topology, OID_AUTO, cores_per_package, CTLFLAG_RDTUN,
70     &cores_per_package, 0, NULL);
71 
72 static int cpuid_leaf_b = 1;
73 SYSCTL_INT(_hw_vmm_topology, OID_AUTO, cpuid_leaf_b, CTLFLAG_RDTUN,
74     &cpuid_leaf_b, 0, NULL);
75 
76 /*
77  * Round up to the next power of two, if necessary, and then take log2.
78  * Returns -1 if argument is zero.
79  */
80 static __inline int
81 log2(u_int x)
82 {
83 
84 	return (fls(x << (1 - powerof2(x))) - 1);
85 }
86 
87 int
88 x86_emulate_cpuid(struct vm *vm, int vcpu_id,
89 		  uint32_t *eax, uint32_t *ebx, uint32_t *ecx, uint32_t *edx)
90 {
91 	const struct xsave_limits *limits;
92 	uint64_t cr4;
93 	int error, enable_invpcid, level, width, x2apic_id;
94 	unsigned int func, regs[4], logical_cpus;
95 	enum x2apic_state x2apic_state;
96 
97 	VCPU_CTR2(vm, vcpu_id, "cpuid %#x,%#x", *eax, *ecx);
98 
99 	/*
100 	 * Requests for invalid CPUID levels should map to the highest
101 	 * available level instead.
102 	 */
103 	if (cpu_exthigh != 0 && *eax >= 0x80000000) {
104 		if (*eax > cpu_exthigh)
105 			*eax = cpu_exthigh;
106 	} else if (*eax >= 0x40000000) {
107 		if (*eax > CPUID_VM_HIGH)
108 			*eax = CPUID_VM_HIGH;
109 	} else if (*eax > cpu_high) {
110 		*eax = cpu_high;
111 	}
112 
113 	func = *eax;
114 
115 	/*
116 	 * In general the approach used for CPU topology is to
117 	 * advertise a flat topology where all CPUs are packages with
118 	 * no multi-core or SMT.
119 	 */
120 	switch (func) {
121 		/*
122 		 * Pass these through to the guest
123 		 */
124 		case CPUID_0000_0000:
125 		case CPUID_0000_0002:
126 		case CPUID_0000_0003:
127 		case CPUID_8000_0000:
128 		case CPUID_8000_0002:
129 		case CPUID_8000_0003:
130 		case CPUID_8000_0004:
131 		case CPUID_8000_0006:
132 			cpuid_count(*eax, *ecx, regs);
133 			break;
134 		case CPUID_8000_0008:
135 			cpuid_count(*eax, *ecx, regs);
136 			if (vmm_is_amd()) {
137 				/*
138 				 * XXX this might appear silly because AMD
139 				 * cpus don't have threads.
140 				 *
141 				 * However this matches the logical cpus as
142 				 * advertised by leaf 0x1 and will work even
143 				 * if the 'threads_per_core' tunable is set
144 				 * incorrectly on an AMD host.
145 				 */
146 				logical_cpus = threads_per_core *
147 				    cores_per_package;
148 				regs[2] = logical_cpus - 1;
149 			}
150 			break;
151 
152 		case CPUID_8000_0001:
153 			cpuid_count(*eax, *ecx, regs);
154 
155 			/*
156 			 * Hide SVM and Topology Extension features from guest.
157 			 */
158 			regs[2] &= ~(AMDID2_SVM | AMDID2_TOPOLOGY);
159 
160 			/*
161 			 * Don't advertise extended performance counter MSRs
162 			 * to the guest.
163 			 */
164 			regs[2] &= ~AMDID2_PCXC;
165 			regs[2] &= ~AMDID2_PNXC;
166 			regs[2] &= ~AMDID2_PTSCEL2I;
167 
168 			/*
169 			 * Don't advertise Instruction Based Sampling feature.
170 			 */
171 			regs[2] &= ~AMDID2_IBS;
172 
173 			/* NodeID MSR not available */
174 			regs[2] &= ~AMDID2_NODE_ID;
175 
176 			/* Don't advertise the OS visible workaround feature */
177 			regs[2] &= ~AMDID2_OSVW;
178 
179 			/*
180 			 * Hide rdtscp/ia32_tsc_aux until we know how
181 			 * to deal with them.
182 			 */
183 			regs[3] &= ~AMDID_RDTSCP;
184 			break;
185 
186 		case CPUID_8000_0007:
187 			/*
188 			 * AMD uses this leaf to advertise the processor's
189 			 * power monitoring and RAS capabilities. These
190 			 * features are hardware-specific and exposing
191 			 * them to a guest doesn't make a lot of sense.
192 			 *
193 			 * Intel uses this leaf only to advertise the
194 			 * "Invariant TSC" feature with all other bits
195 			 * being reserved (set to zero).
196 			 */
197 			regs[0] = 0;
198 			regs[1] = 0;
199 			regs[2] = 0;
200 			regs[3] = 0;
201 
202 			/*
203 			 * "Invariant TSC" can be advertised to the guest if:
204 			 * - host TSC frequency is invariant
205 			 * - host TSCs are synchronized across physical cpus
206 			 *
207 			 * XXX This still falls short because the vcpu
208 			 * can observe the TSC moving backwards as it
209 			 * migrates across physical cpus. But at least
210 			 * it should discourage the guest from using the
211 			 * TSC to keep track of time.
212 			 */
213 			if (tsc_is_invariant && smp_tsc)
214 				regs[3] |= AMDPM_TSC_INVARIANT;
215 			break;
216 
217 		case CPUID_0000_0001:
218 			do_cpuid(1, regs);
219 
220 			error = vm_get_x2apic_state(vm, vcpu_id, &x2apic_state);
221 			if (error) {
222 				panic("x86_emulate_cpuid: error %d "
223 				      "fetching x2apic state", error);
224 			}
225 
226 			/*
227 			 * Override the APIC ID only in ebx
228 			 */
229 			regs[1] &= ~(CPUID_LOCAL_APIC_ID);
230 			regs[1] |= (vcpu_id << CPUID_0000_0001_APICID_SHIFT);
231 
232 			/*
233 			 * Don't expose VMX, SpeedStep or TME capability.
234 			 * Advertise x2APIC capability and Hypervisor guest.
235 			 */
236 			regs[2] &= ~(CPUID2_VMX | CPUID2_EST | CPUID2_TM2);
237 
238 			regs[2] |= CPUID2_HV;
239 
240 			if (x2apic_state != X2APIC_DISABLED)
241 				regs[2] |= CPUID2_X2APIC;
242 			else
243 				regs[2] &= ~CPUID2_X2APIC;
244 
245 			/*
246 			 * Only advertise CPUID2_XSAVE in the guest if
247 			 * the host is using XSAVE.
248 			 */
249 			if (!(regs[2] & CPUID2_OSXSAVE))
250 				regs[2] &= ~CPUID2_XSAVE;
251 
252 			/*
253 			 * If CPUID2_XSAVE is being advertised and the
254 			 * guest has set CR4_XSAVE, set
255 			 * CPUID2_OSXSAVE.
256 			 */
257 			regs[2] &= ~CPUID2_OSXSAVE;
258 			if (regs[2] & CPUID2_XSAVE) {
259 				error = vm_get_register(vm, vcpu_id,
260 				    VM_REG_GUEST_CR4, &cr4);
261 				if (error)
262 					panic("x86_emulate_cpuid: error %d "
263 					      "fetching %%cr4", error);
264 				if (cr4 & CR4_XSAVE)
265 					regs[2] |= CPUID2_OSXSAVE;
266 			}
267 
268 			/*
269 			 * Hide monitor/mwait until we know how to deal with
270 			 * these instructions.
271 			 */
272 			regs[2] &= ~CPUID2_MON;
273 
274                         /*
275 			 * Hide the performance and debug features.
276 			 */
277 			regs[2] &= ~CPUID2_PDCM;
278 
279 			/*
280 			 * No TSC deadline support in the APIC yet
281 			 */
282 			regs[2] &= ~CPUID2_TSCDLT;
283 
284 			/*
285 			 * Hide thermal monitoring
286 			 */
287 			regs[3] &= ~(CPUID_ACPI | CPUID_TM);
288 
289 			/*
290 			 * Hide the debug store capability.
291 			 */
292 			regs[3] &= ~CPUID_DS;
293 
294 			/*
295 			 * Advertise the Machine Check and MTRR capability.
296 			 *
297 			 * Some guest OSes (e.g. Windows) will not boot if
298 			 * these features are absent.
299 			 */
300 			regs[3] |= (CPUID_MCA | CPUID_MCE | CPUID_MTRR);
301 
302 			logical_cpus = threads_per_core * cores_per_package;
303 			regs[1] &= ~CPUID_HTT_CORES;
304 			regs[1] |= (logical_cpus & 0xff) << 16;
305 			regs[3] |= CPUID_HTT;
306 			break;
307 
308 		case CPUID_0000_0004:
309 			cpuid_count(*eax, *ecx, regs);
310 
311 			if (regs[0] || regs[1] || regs[2] || regs[3]) {
312 				regs[0] &= 0x3ff;
313 				regs[0] |= (cores_per_package - 1) << 26;
314 				/*
315 				 * Cache topology:
316 				 * - L1 and L2 are shared only by the logical
317 				 *   processors in a single core.
318 				 * - L3 and above are shared by all logical
319 				 *   processors in the package.
320 				 */
321 				logical_cpus = threads_per_core;
322 				level = (regs[0] >> 5) & 0x7;
323 				if (level >= 3)
324 					logical_cpus *= cores_per_package;
325 				regs[0] |= (logical_cpus - 1) << 14;
326 			}
327 			break;
328 
329 		case CPUID_0000_0007:
330 			regs[0] = 0;
331 			regs[1] = 0;
332 			regs[2] = 0;
333 			regs[3] = 0;
334 
335 			/* leaf 0 */
336 			if (*ecx == 0) {
337 				cpuid_count(*eax, *ecx, regs);
338 
339 				/* Only leaf 0 is supported */
340 				regs[0] = 0;
341 
342 				/*
343 				 * Expose known-safe features.
344 				 */
345 				regs[1] &= (CPUID_STDEXT_FSGSBASE |
346 				    CPUID_STDEXT_BMI1 | CPUID_STDEXT_HLE |
347 				    CPUID_STDEXT_AVX2 | CPUID_STDEXT_BMI2 |
348 				    CPUID_STDEXT_ERMS | CPUID_STDEXT_RTM |
349 				    CPUID_STDEXT_AVX512F |
350 				    CPUID_STDEXT_AVX512PF |
351 				    CPUID_STDEXT_AVX512ER |
352 				    CPUID_STDEXT_AVX512CD);
353 				regs[2] = 0;
354 				regs[3] = 0;
355 
356 				/* Advertise INVPCID if it is enabled. */
357 				error = vm_get_capability(vm, vcpu_id,
358 				    VM_CAP_ENABLE_INVPCID, &enable_invpcid);
359 				if (error == 0 && enable_invpcid)
360 					regs[1] |= CPUID_STDEXT_INVPCID;
361 			}
362 			break;
363 
364 		case CPUID_0000_0006:
365 			regs[0] = CPUTPM1_ARAT;
366 			regs[1] = 0;
367 			regs[2] = 0;
368 			regs[3] = 0;
369 			break;
370 
371 		case CPUID_0000_000A:
372 			/*
373 			 * Handle the access, but report 0 for
374 			 * all options
375 			 */
376 			regs[0] = 0;
377 			regs[1] = 0;
378 			regs[2] = 0;
379 			regs[3] = 0;
380 			break;
381 
382 		case CPUID_0000_000B:
383 			/*
384 			 * Processor topology enumeration
385 			 */
386 			if (*ecx == 0) {
387 				logical_cpus = threads_per_core;
388 				width = log2(logical_cpus);
389 				level = CPUID_TYPE_SMT;
390 				x2apic_id = vcpu_id;
391 			}
392 
393 			if (*ecx == 1) {
394 				logical_cpus = threads_per_core *
395 				    cores_per_package;
396 				width = log2(logical_cpus);
397 				level = CPUID_TYPE_CORE;
398 				x2apic_id = vcpu_id;
399 			}
400 
401 			if (!cpuid_leaf_b || *ecx >= 2) {
402 				width = 0;
403 				logical_cpus = 0;
404 				level = 0;
405 				x2apic_id = 0;
406 			}
407 
408 			regs[0] = width & 0x1f;
409 			regs[1] = logical_cpus & 0xffff;
410 			regs[2] = (level << 8) | (*ecx & 0xff);
411 			regs[3] = x2apic_id;
412 			break;
413 
414 		case CPUID_0000_000D:
415 			limits = vmm_get_xsave_limits();
416 			if (!limits->xsave_enabled) {
417 				regs[0] = 0;
418 				regs[1] = 0;
419 				regs[2] = 0;
420 				regs[3] = 0;
421 				break;
422 			}
423 
424 			cpuid_count(*eax, *ecx, regs);
425 			switch (*ecx) {
426 			case 0:
427 				/*
428 				 * Only permit the guest to use bits
429 				 * that are active in the host in
430 				 * %xcr0.  Also, claim that the
431 				 * maximum save area size is
432 				 * equivalent to the host's current
433 				 * save area size.  Since this runs
434 				 * "inside" of vmrun(), it runs with
435 				 * the guest's xcr0, so the current
436 				 * save area size is correct as-is.
437 				 */
438 				regs[0] &= limits->xcr0_allowed;
439 				regs[2] = limits->xsave_max_size;
440 				regs[3] &= (limits->xcr0_allowed >> 32);
441 				break;
442 			case 1:
443 				/* Only permit XSAVEOPT. */
444 				regs[0] &= CPUID_EXTSTATE_XSAVEOPT;
445 				regs[1] = 0;
446 				regs[2] = 0;
447 				regs[3] = 0;
448 				break;
449 			default:
450 				/*
451 				 * If the leaf is for a permitted feature,
452 				 * pass through as-is, otherwise return
453 				 * all zeroes.
454 				 */
455 				if (!(limits->xcr0_allowed & (1ul << *ecx))) {
456 					regs[0] = 0;
457 					regs[1] = 0;
458 					regs[2] = 0;
459 					regs[3] = 0;
460 				}
461 				break;
462 			}
463 			break;
464 
465 		case 0x40000000:
466 			regs[0] = CPUID_VM_HIGH;
467 			bcopy(bhyve_id, &regs[1], 4);
468 			bcopy(bhyve_id + 4, &regs[2], 4);
469 			bcopy(bhyve_id + 8, &regs[3], 4);
470 			break;
471 
472 		default:
473 			/*
474 			 * The leaf value has already been clamped so
475 			 * simply pass this through, keeping count of
476 			 * how many unhandled leaf values have been seen.
477 			 */
478 			atomic_add_long(&bhyve_xcpuids, 1);
479 			cpuid_count(*eax, *ecx, regs);
480 			break;
481 	}
482 
483 	*eax = regs[0];
484 	*ebx = regs[1];
485 	*ecx = regs[2];
486 	*edx = regs[3];
487 
488 	return (1);
489 }
490