xref: /freebsd-13-stable/sys/x86/xen/hvm.c (revision 3bc80996974a61a4223eae4c1ccd47b6ee32a48a)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2008, 2013 Citrix Systems, Inc.
5  * Copyright (c) 2012 Spectra Logic Corporation
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 #include <sys/param.h>
32 #include <sys/bus.h>
33 #include <sys/kernel.h>
34 #include <sys/malloc.h>
35 #include <sys/proc.h>
36 #include <sys/smp.h>
37 #include <sys/systm.h>
38 
39 #include <vm/vm.h>
40 #include <vm/pmap.h>
41 #include <vm/vm_param.h>
42 
43 #include <dev/pci/pcivar.h>
44 
45 #include <machine/cpufunc.h>
46 #include <machine/cpu.h>
47 #include <machine/smp.h>
48 
49 #include <x86/apicreg.h>
50 
51 #include <xen/xen-os.h>
52 #include <xen/error.h>
53 #include <xen/features.h>
54 #include <xen/gnttab.h>
55 #include <xen/hypervisor.h>
56 #include <xen/hvm.h>
57 #include <xen/xen_intr.h>
58 
59 #include <xen/interface/arch-x86/cpuid.h>
60 #include <xen/interface/hvm/params.h>
61 #include <xen/interface/vcpu.h>
62 
63 /*--------------------------- Forward Declarations ---------------------------*/
64 static void xen_hvm_cpu_init(void);
65 
66 /*-------------------------------- Global Data -------------------------------*/
67 enum xen_domain_type xen_domain_type = XEN_NATIVE;
68 
69 #ifdef SMP
70 struct cpu_ops xen_hvm_cpu_ops = {
71 	.cpu_init	= xen_hvm_cpu_init,
72 	.cpu_resume	= xen_hvm_cpu_init
73 };
74 #endif
75 
76 static MALLOC_DEFINE(M_XENHVM, "xen_hvm", "Xen HVM PV Support");
77 
78 /**
79  * If non-zero, the hypervisor has been configured to use a direct
80  * IDT event callback for interrupt injection.
81  */
82 int xen_vector_callback_enabled;
83 
84 /**
85  * Start info flags. ATM this only used to store the initial domain flag for
86  * PVHv2, and it's always empty for HVM guests.
87  */
88 uint32_t hvm_start_flags;
89 
90 /**
91  * Signal whether the vector injected for the event channel upcall requires to
92  * be EOI'ed on the local APIC.
93  */
94 bool xen_evtchn_needs_ack;
95 
96 /*------------------------------- Per-CPU Data -------------------------------*/
97 DPCPU_DEFINE(struct vcpu_info, vcpu_local_info);
98 DPCPU_DEFINE(struct vcpu_info *, vcpu_info);
99 
100 /*------------------ Hypervisor Access Shared Memory Regions -----------------*/
101 shared_info_t *HYPERVISOR_shared_info;
102 
103 /*------------------------------ Sysctl tunables -----------------------------*/
104 int xen_disable_pv_disks = 0;
105 int xen_disable_pv_nics = 0;
106 TUNABLE_INT("hw.xen.disable_pv_disks", &xen_disable_pv_disks);
107 TUNABLE_INT("hw.xen.disable_pv_nics", &xen_disable_pv_nics);
108 
109 /*---------------------- XEN Hypervisor Probe and Setup ----------------------*/
110 
111 static uint32_t cpuid_base;
112 
113 static uint32_t
xen_hvm_cpuid_base(void)114 xen_hvm_cpuid_base(void)
115 {
116 	uint32_t base, regs[4];
117 
118 	for (base = 0x40000000; base < 0x40010000; base += 0x100) {
119 		do_cpuid(base, regs);
120 		if (!memcmp("XenVMMXenVMM", &regs[1], 12)
121 		    && (regs[0] - base) >= 2)
122 			return (base);
123 	}
124 	return (0);
125 }
126 
127 static void
hypervisor_quirks(unsigned int major,unsigned int minor)128 hypervisor_quirks(unsigned int major, unsigned int minor)
129 {
130 #ifdef SMP
131 	if (((major < 4) || (major == 4 && minor <= 5)) &&
132 	    msix_disable_migration == -1) {
133 		/*
134 		 * Xen hypervisors prior to 4.6.0 do not properly
135 		 * handle updates to enabled MSI-X table entries,
136 		 * so disable MSI-X interrupt migration in that
137 		 * case.
138 		 */
139 		if (bootverbose)
140 			printf(
141 "Disabling MSI-X interrupt migration due to Xen hypervisor bug.\n"
142 "Set machdep.msix_disable_migration=0 to forcefully enable it.\n");
143 		msix_disable_migration = 1;
144 	}
145 #endif
146 }
147 
148 static void
hypervisor_version(void)149 hypervisor_version(void)
150 {
151 	uint32_t regs[4];
152 	int major, minor;
153 
154 	do_cpuid(cpuid_base + 1, regs);
155 
156 	major = regs[0] >> 16;
157 	minor = regs[0] & 0xffff;
158 	printf("XEN: Hypervisor version %d.%d detected.\n", major, minor);
159 
160 	hypervisor_quirks(major, minor);
161 }
162 
163 /*
164  * Allocate and fill in the hypcall page.
165  */
166 int
xen_hvm_init_hypercall_stubs(enum xen_hvm_init_type init_type)167 xen_hvm_init_hypercall_stubs(enum xen_hvm_init_type init_type)
168 {
169 	uint32_t regs[4];
170 
171 	/* Legacy PVH will get here without the cpuid leaf being set. */
172 	if (cpuid_base == 0)
173 		cpuid_base = xen_hvm_cpuid_base();
174 	if (cpuid_base == 0)
175 		return (ENXIO);
176 
177 	if (xen_domain() && init_type == XEN_HVM_INIT_LATE) {
178 		/*
179 		 * If the domain type is already set we can assume that the
180 		 * hypercall page has been populated too, so just print the
181 		 * version (and apply any quirks) and exit.
182 		 */
183 		hypervisor_version();
184 		return 0;
185 	}
186 
187 	if (init_type == XEN_HVM_INIT_LATE)
188 		hypervisor_version();
189 
190 	/*
191 	 * Find the hypercall pages.
192 	 */
193 	do_cpuid(cpuid_base + 2, regs);
194 	if (regs[0] != 1)
195 		return (EINVAL);
196 
197 	wrmsr(regs[1], (init_type == XEN_HVM_INIT_EARLY)
198 	    ? (vm_paddr_t)((uintptr_t)&hypercall_page - KERNBASE)
199 	    : vtophys(&hypercall_page));
200 
201 	return (0);
202 }
203 
204 static void
xen_hvm_init_shared_info_page(void)205 xen_hvm_init_shared_info_page(void)
206 {
207 	struct xen_add_to_physmap xatp;
208 
209 	if (xen_pv_domain()) {
210 		/*
211 		 * Already setup in the PV case, shared_info is passed inside
212 		 * of the start_info struct at start of day.
213 		 */
214 		return;
215 	}
216 
217 	if (HYPERVISOR_shared_info == NULL) {
218 		HYPERVISOR_shared_info = malloc(PAGE_SIZE, M_XENHVM, M_NOWAIT);
219 		if (HYPERVISOR_shared_info == NULL)
220 			panic("Unable to allocate Xen shared info page");
221 	}
222 
223 	xatp.domid = DOMID_SELF;
224 	xatp.idx = 0;
225 	xatp.space = XENMAPSPACE_shared_info;
226 	xatp.gpfn = vtophys(HYPERVISOR_shared_info) >> PAGE_SHIFT;
227 	if (HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp))
228 		panic("HYPERVISOR_memory_op failed");
229 }
230 
231 static int
set_percpu_callback(unsigned int vcpu)232 set_percpu_callback(unsigned int vcpu)
233 {
234 	struct xen_hvm_evtchn_upcall_vector vec;
235 	int error;
236 
237 	vec.vcpu = vcpu;
238 	vec.vector = IDT_EVTCHN;
239 	error = HYPERVISOR_hvm_op(HVMOP_set_evtchn_upcall_vector, &vec);
240 
241 	return (error != 0 ? xen_translate_error(error) : 0);
242 }
243 
244 /*
245  * Tell the hypervisor how to contact us for event channel callbacks.
246  */
247 void
xen_hvm_set_callback(device_t dev)248 xen_hvm_set_callback(device_t dev)
249 {
250 	struct xen_hvm_param xhp;
251 	int irq;
252 
253 	if (xen_vector_callback_enabled)
254 		return;
255 
256 	xhp.domid = DOMID_SELF;
257 	xhp.index = HVM_PARAM_CALLBACK_IRQ;
258 	if (xen_feature(XENFEAT_hvm_callback_vector) != 0) {
259 		int error;
260 
261 		error = set_percpu_callback(0);
262 		if (error == 0) {
263 			xen_evtchn_needs_ack = true;
264 			/* Trick toolstack to think we are enlightened */
265 			xhp.value = 1;
266 		} else
267 			xhp.value = HVM_CALLBACK_VECTOR(IDT_EVTCHN);
268 		error = HYPERVISOR_hvm_op(HVMOP_set_param, &xhp);
269 		if (error == 0) {
270 			xen_vector_callback_enabled = 1;
271 			return;
272 		} else if (xen_evtchn_needs_ack)
273 			panic("Unable to setup fake HVM param: %d", error);
274 
275 		printf("Xen HVM callback vector registration failed (%d). "
276 		    "Falling back to emulated device interrupt\n", error);
277 	}
278 	xen_vector_callback_enabled = 0;
279 	if (dev == NULL) {
280 		/*
281 		 * Called from early boot or resume.
282 		 * xenpci will invoke us again later.
283 		 */
284 		return;
285 	}
286 
287 	irq = pci_get_irq(dev);
288 	if (irq < 16) {
289 		xhp.value = HVM_CALLBACK_GSI(irq);
290 	} else {
291 		u_int slot;
292 		u_int pin;
293 
294 		slot = pci_get_slot(dev);
295 		pin = pci_get_intpin(dev) - 1;
296 		xhp.value = HVM_CALLBACK_PCI_INTX(slot, pin);
297 	}
298 
299 	if (HYPERVISOR_hvm_op(HVMOP_set_param, &xhp) != 0)
300 		panic("Can't set evtchn callback");
301 }
302 
303 #define	XEN_MAGIC_IOPORT 0x10
304 enum {
305 	XMI_MAGIC			 = 0x49d2,
306 	XMI_UNPLUG_IDE_DISKS		 = 0x01,
307 	XMI_UNPLUG_NICS			 = 0x02,
308 	XMI_UNPLUG_IDE_EXCEPT_PRI_MASTER = 0x04
309 };
310 
311 static void
xen_hvm_disable_emulated_devices(void)312 xen_hvm_disable_emulated_devices(void)
313 {
314 	u_short disable_devs = 0;
315 
316 	if (xen_pv_domain()) {
317 		/*
318 		 * No emulated devices in the PV case, so no need to unplug
319 		 * anything.
320 		 */
321 		if (xen_disable_pv_disks != 0 || xen_disable_pv_nics != 0)
322 			printf("PV devices cannot be disabled in PV guests\n");
323 		return;
324 	}
325 
326 	if (inw(XEN_MAGIC_IOPORT) != XMI_MAGIC)
327 		return;
328 
329 	if (xen_disable_pv_disks == 0) {
330 		if (bootverbose)
331 			printf("XEN: disabling emulated disks\n");
332 		disable_devs |= XMI_UNPLUG_IDE_DISKS;
333 	}
334 	if (xen_disable_pv_nics == 0) {
335 		if (bootverbose)
336 			printf("XEN: disabling emulated nics\n");
337 		disable_devs |= XMI_UNPLUG_NICS;
338 	}
339 
340 	if (disable_devs != 0)
341 		outw(XEN_MAGIC_IOPORT, disable_devs);
342 }
343 
344 static void
xen_hvm_init(enum xen_hvm_init_type init_type)345 xen_hvm_init(enum xen_hvm_init_type init_type)
346 {
347 	int error;
348 	int i;
349 
350 	if (init_type == XEN_HVM_INIT_CANCELLED_SUSPEND)
351 		return;
352 
353 	error = xen_hvm_init_hypercall_stubs(init_type);
354 
355 	switch (init_type) {
356 	case XEN_HVM_INIT_LATE:
357 		if (error != 0)
358 			return;
359 
360 		/*
361 		 * If xen_domain_type is not set at this point
362 		 * it means we are inside a (PV)HVM guest, because
363 		 * for PVH the guest type is set much earlier
364 		 * (see hammer_time_xen).
365 		 */
366 		if (!xen_domain()) {
367 			xen_domain_type = XEN_HVM_DOMAIN;
368 			vm_guest = VM_GUEST_XEN;
369 		}
370 
371 		setup_xen_features();
372 #ifdef SMP
373 		cpu_ops = xen_hvm_cpu_ops;
374 #endif
375 		break;
376 	case XEN_HVM_INIT_RESUME:
377 		if (error != 0)
378 			panic("Unable to init Xen hypercall stubs on resume");
379 
380 		/* Clear stale vcpu_info. */
381 		CPU_FOREACH(i)
382 			DPCPU_ID_SET(i, vcpu_info, NULL);
383 		break;
384 	default:
385 		panic("Unsupported HVM initialization type");
386 	}
387 
388 	xen_vector_callback_enabled = 0;
389 	xen_evtchn_needs_ack = false;
390 	xen_hvm_set_callback(NULL);
391 
392 	/*
393 	 * On (PV)HVM domains we need to request the hypervisor to
394 	 * fill the shared info page, for PVH guest the shared_info page
395 	 * is passed inside the start_info struct and is already set, so this
396 	 * functions are no-ops.
397 	 */
398 	xen_hvm_init_shared_info_page();
399 	xen_hvm_disable_emulated_devices();
400 }
401 
402 void
xen_hvm_suspend(void)403 xen_hvm_suspend(void)
404 {
405 }
406 
407 void
xen_hvm_resume(bool suspend_cancelled)408 xen_hvm_resume(bool suspend_cancelled)
409 {
410 
411 	xen_hvm_init(suspend_cancelled ?
412 	    XEN_HVM_INIT_CANCELLED_SUSPEND : XEN_HVM_INIT_RESUME);
413 
414 	/* Register vcpu_info area for CPU#0. */
415 	xen_hvm_cpu_init();
416 }
417 
418 static void
xen_hvm_sysinit(void * arg __unused)419 xen_hvm_sysinit(void *arg __unused)
420 {
421 	xen_hvm_init(XEN_HVM_INIT_LATE);
422 }
423 SYSINIT(xen_hvm_init, SI_SUB_HYPERVISOR, SI_ORDER_FIRST, xen_hvm_sysinit, NULL);
424 
425 static void
xen_hvm_cpu_init(void)426 xen_hvm_cpu_init(void)
427 {
428 	struct vcpu_register_vcpu_info info;
429 	struct vcpu_info *vcpu_info;
430 	uint32_t regs[4];
431 	int cpu, rc;
432 
433 	if (!xen_domain())
434 		return;
435 
436 	if (DPCPU_GET(vcpu_info) != NULL) {
437 		/*
438 		 * vcpu_info is already set.  We're resuming
439 		 * from a failed migration and our pre-suspend
440 		 * configuration is still valid.
441 		 */
442 		return;
443 	}
444 
445 	/*
446 	 * Set vCPU ID. If available fetch the ID from CPUID, if not just use
447 	 * the ACPI ID.
448 	 */
449 	KASSERT(cpuid_base != 0, ("Invalid base Xen CPUID leaf"));
450 	cpuid_count(cpuid_base + 4, 0, regs);
451 	KASSERT((regs[0] & XEN_HVM_CPUID_VCPU_ID_PRESENT) ||
452 	    !xen_pv_domain(),
453 	    ("Xen PV domain without vcpu_id in cpuid"));
454 	PCPU_SET(vcpu_id, (regs[0] & XEN_HVM_CPUID_VCPU_ID_PRESENT) ?
455 	    regs[1] : PCPU_GET(acpi_id));
456 
457 	if (xen_evtchn_needs_ack && !IS_BSP()) {
458 		/*
459 		 * Setup the per-vpcu event channel upcall vector. This is only
460 		 * required when using the new HVMOP_set_evtchn_upcall_vector
461 		 * hypercall, which allows using a different vector for each
462 		 * vCPU. Note that FreeBSD uses the same vector for all vCPUs
463 		 * because it's not dynamically allocated.
464 		 */
465 		rc = set_percpu_callback(PCPU_GET(vcpu_id));
466 		if (rc != 0)
467 			panic("Event channel upcall vector setup failed: %d",
468 			    rc);
469 	}
470 
471 	/*
472 	 * Set the vCPU info.
473 	 *
474 	 * NB: the vCPU info for vCPUs < 32 can be fetched from the shared info
475 	 * page, but in order to make sure the mapping code is correct always
476 	 * attempt to map the vCPU info at a custom place.
477 	 */
478 	vcpu_info = DPCPU_PTR(vcpu_local_info);
479 	cpu = PCPU_GET(vcpu_id);
480 	info.mfn = vtophys(vcpu_info) >> PAGE_SHIFT;
481 	info.offset = vtophys(vcpu_info) - trunc_page(vtophys(vcpu_info));
482 
483 	rc = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
484 	if (rc != 0)
485 		DPCPU_SET(vcpu_info, &HYPERVISOR_shared_info->vcpu_info[cpu]);
486 	else
487 		DPCPU_SET(vcpu_info, vcpu_info);
488 }
489 SYSINIT(xen_hvm_cpu_init, SI_SUB_INTR, SI_ORDER_FIRST, xen_hvm_cpu_init, NULL);
490 
491 /* HVM/PVH start_info accessors */
492 static vm_paddr_t
hvm_get_xenstore_mfn(void)493 hvm_get_xenstore_mfn(void)
494 {
495 
496 	return (hvm_get_parameter(HVM_PARAM_STORE_PFN));
497 }
498 
499 static evtchn_port_t
hvm_get_xenstore_evtchn(void)500 hvm_get_xenstore_evtchn(void)
501 {
502 
503 	return (hvm_get_parameter(HVM_PARAM_STORE_EVTCHN));
504 }
505 
506 static vm_paddr_t
hvm_get_console_mfn(void)507 hvm_get_console_mfn(void)
508 {
509 
510 	return (hvm_get_parameter(HVM_PARAM_CONSOLE_PFN));
511 }
512 
513 static evtchn_port_t
hvm_get_console_evtchn(void)514 hvm_get_console_evtchn(void)
515 {
516 
517 	return (hvm_get_parameter(HVM_PARAM_CONSOLE_EVTCHN));
518 }
519 
520 static uint32_t
hvm_get_start_flags(void)521 hvm_get_start_flags(void)
522 {
523 
524 	return (hvm_start_flags);
525 }
526 
527 struct hypervisor_info hypervisor_info = {
528 	.get_xenstore_mfn		= hvm_get_xenstore_mfn,
529 	.get_xenstore_evtchn		= hvm_get_xenstore_evtchn,
530 	.get_console_mfn		= hvm_get_console_mfn,
531 	.get_console_evtchn		= hvm_get_console_evtchn,
532 	.get_start_flags		= hvm_get_start_flags,
533 };
534