xref: /freebsd/sys/x86/xen/hvm.c (revision 22cf89c938886d14f5796fc49f9f020c23ea8eaf)
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 <contrib/xen/arch-x86/cpuid.h>
60 #include <contrib/xen/hvm/params.h>
61 #include <contrib/xen/vcpu.h>
62 
63 /*--------------------------- Forward Declarations ---------------------------*/
64 static void xen_hvm_cpu_init(void);
65 
66 /*-------------------------------- Global Data -------------------------------*/
67 #ifdef SMP
68 struct cpu_ops xen_hvm_cpu_ops = {
69 	.cpu_init	= xen_hvm_cpu_init,
70 	.cpu_resume	= xen_hvm_cpu_init
71 };
72 #endif
73 
74 static MALLOC_DEFINE(M_XENHVM, "xen_hvm", "Xen HVM PV Support");
75 
76 /**
77  * If non-zero, the hypervisor has been configured to use a direct
78  * IDT event callback for interrupt injection.
79  */
80 int xen_vector_callback_enabled;
81 
82 /**
83  * Signal whether the vector injected for the event channel upcall requires to
84  * be EOI'ed on the local APIC.
85  */
86 bool xen_evtchn_needs_ack;
87 
88 /*------------------------------- Per-CPU Data -------------------------------*/
89 DPCPU_DECLARE(struct vcpu_info *, vcpu_info);
90 
91 /*------------------------------ Sysctl tunables -----------------------------*/
92 int xen_disable_pv_disks = 0;
93 int xen_disable_pv_nics = 0;
94 TUNABLE_INT("hw.xen.disable_pv_disks", &xen_disable_pv_disks);
95 TUNABLE_INT("hw.xen.disable_pv_nics", &xen_disable_pv_nics);
96 
97 /*---------------------- XEN Hypervisor Probe and Setup ----------------------*/
98 
99 uint32_t xen_cpuid_base;
100 
101 static uint32_t
102 xen_hvm_cpuid_base(void)
103 {
104 	uint32_t base, regs[4];
105 
106 	for (base = 0x40000000; base < 0x40010000; base += 0x100) {
107 		do_cpuid(base, regs);
108 		if (!memcmp("XenVMMXenVMM", &regs[1], 12)
109 		    && (regs[0] - base) >= 2)
110 			return (base);
111 	}
112 	return (0);
113 }
114 
115 static void
116 hypervisor_quirks(unsigned int major, unsigned int minor)
117 {
118 #ifdef SMP
119 	if (((major < 4) || (major == 4 && minor <= 5)) &&
120 	    msix_disable_migration == -1) {
121 		/*
122 		 * Xen hypervisors prior to 4.6.0 do not properly
123 		 * handle updates to enabled MSI-X table entries,
124 		 * so disable MSI-X interrupt migration in that
125 		 * case.
126 		 */
127 		if (bootverbose)
128 			printf(
129 "Disabling MSI-X interrupt migration due to Xen hypervisor bug.\n"
130 "Set machdep.msix_disable_migration=0 to forcefully enable it.\n");
131 		msix_disable_migration = 1;
132 	}
133 #endif
134 }
135 
136 static void
137 hypervisor_version(void)
138 {
139 	uint32_t regs[4];
140 	int major, minor;
141 
142 	do_cpuid(xen_cpuid_base + 1, regs);
143 
144 	major = regs[0] >> 16;
145 	minor = regs[0] & 0xffff;
146 	printf("XEN: Hypervisor version %d.%d detected.\n", major, minor);
147 
148 	hypervisor_quirks(major, minor);
149 }
150 
151 /*
152  * Allocate and fill in the hypcall page.
153  */
154 int
155 xen_hvm_init_hypercall_stubs(enum xen_hvm_init_type init_type)
156 {
157 	uint32_t regs[4];
158 
159 	/* Legacy PVH will get here without the cpuid leaf being set. */
160 	if (xen_cpuid_base == 0)
161 		xen_cpuid_base = xen_hvm_cpuid_base();
162 	if (xen_cpuid_base == 0)
163 		return (ENXIO);
164 
165 	if (xen_domain() && init_type == XEN_HVM_INIT_LATE) {
166 		/*
167 		 * If the domain type is already set we can assume that the
168 		 * hypercall page has been populated too, so just print the
169 		 * version (and apply any quirks) and exit.
170 		 */
171 		hypervisor_version();
172 		return 0;
173 	}
174 
175 	if (init_type == XEN_HVM_INIT_LATE)
176 		hypervisor_version();
177 
178 	/*
179 	 * Find the hypercall pages.
180 	 */
181 	do_cpuid(xen_cpuid_base + 2, regs);
182 	if (regs[0] != 1)
183 		return (EINVAL);
184 
185 	wrmsr(regs[1], (init_type == XEN_HVM_INIT_EARLY)
186 	    ? (vm_paddr_t)((uintptr_t)&hypercall_page - KERNBASE)
187 	    : vtophys(&hypercall_page));
188 
189 	return (0);
190 }
191 
192 static void
193 xen_hvm_init_shared_info_page(void)
194 {
195 	struct xen_add_to_physmap xatp;
196 
197 	if (xen_pv_domain()) {
198 		/*
199 		 * Already setup in the PV case, shared_info is passed inside
200 		 * of the start_info struct at start of day.
201 		 */
202 		return;
203 	}
204 
205 	if (HYPERVISOR_shared_info == NULL) {
206 		HYPERVISOR_shared_info = malloc(PAGE_SIZE, M_XENHVM, M_NOWAIT);
207 		if (HYPERVISOR_shared_info == NULL)
208 			panic("Unable to allocate Xen shared info page");
209 	}
210 
211 	xatp.domid = DOMID_SELF;
212 	xatp.idx = 0;
213 	xatp.space = XENMAPSPACE_shared_info;
214 	xatp.gpfn = vtophys(HYPERVISOR_shared_info) >> PAGE_SHIFT;
215 	if (HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp))
216 		panic("HYPERVISOR_memory_op failed");
217 }
218 
219 static int
220 set_percpu_callback(unsigned int vcpu)
221 {
222 	struct xen_hvm_evtchn_upcall_vector vec;
223 	int error;
224 
225 	vec.vcpu = vcpu;
226 	vec.vector = IDT_EVTCHN;
227 	error = HYPERVISOR_hvm_op(HVMOP_set_evtchn_upcall_vector, &vec);
228 
229 	return (error != 0 ? xen_translate_error(error) : 0);
230 }
231 
232 /*
233  * Tell the hypervisor how to contact us for event channel callbacks.
234  */
235 void
236 xen_hvm_set_callback(device_t dev)
237 {
238 	struct xen_hvm_param xhp;
239 	int irq;
240 
241 	if (xen_vector_callback_enabled)
242 		return;
243 
244 	xhp.domid = DOMID_SELF;
245 	xhp.index = HVM_PARAM_CALLBACK_IRQ;
246 	if (xen_feature(XENFEAT_hvm_callback_vector) != 0) {
247 		int error;
248 
249 		error = set_percpu_callback(0);
250 		if (error == 0) {
251 			xen_evtchn_needs_ack = true;
252 			/* Trick toolstack to think we are enlightened */
253 			xhp.value = 1;
254 		} else
255 			xhp.value = HVM_CALLBACK_VECTOR(IDT_EVTCHN);
256 		error = HYPERVISOR_hvm_op(HVMOP_set_param, &xhp);
257 		if (error == 0) {
258 			xen_vector_callback_enabled = 1;
259 			return;
260 		} else if (xen_evtchn_needs_ack)
261 			panic("Unable to setup fake HVM param: %d", error);
262 
263 		printf("Xen HVM callback vector registration failed (%d). "
264 		    "Falling back to emulated device interrupt\n", error);
265 	}
266 	xen_vector_callback_enabled = 0;
267 	if (dev == NULL) {
268 		/*
269 		 * Called from early boot or resume.
270 		 * xenpci will invoke us again later.
271 		 */
272 		return;
273 	}
274 
275 	irq = pci_get_irq(dev);
276 	if (irq < 16) {
277 		xhp.value = HVM_CALLBACK_GSI(irq);
278 	} else {
279 		u_int slot;
280 		u_int pin;
281 
282 		slot = pci_get_slot(dev);
283 		pin = pci_get_intpin(dev) - 1;
284 		xhp.value = HVM_CALLBACK_PCI_INTX(slot, pin);
285 	}
286 
287 	if (HYPERVISOR_hvm_op(HVMOP_set_param, &xhp) != 0)
288 		panic("Can't set evtchn callback");
289 }
290 
291 #define	XEN_MAGIC_IOPORT 0x10
292 enum {
293 	XMI_MAGIC			 = 0x49d2,
294 	XMI_UNPLUG_IDE_DISKS		 = 0x01,
295 	XMI_UNPLUG_NICS			 = 0x02,
296 	XMI_UNPLUG_IDE_EXCEPT_PRI_MASTER = 0x04
297 };
298 
299 static void
300 xen_hvm_disable_emulated_devices(void)
301 {
302 	u_short disable_devs = 0;
303 
304 	if (xen_pv_domain()) {
305 		/*
306 		 * No emulated devices in the PV case, so no need to unplug
307 		 * anything.
308 		 */
309 		if (xen_disable_pv_disks != 0 || xen_disable_pv_nics != 0)
310 			printf("PV devices cannot be disabled in PV guests\n");
311 		return;
312 	}
313 
314 	if (inw(XEN_MAGIC_IOPORT) != XMI_MAGIC)
315 		return;
316 
317 	if (xen_disable_pv_disks == 0) {
318 		if (bootverbose)
319 			printf("XEN: disabling emulated disks\n");
320 		disable_devs |= XMI_UNPLUG_IDE_DISKS;
321 	}
322 	if (xen_disable_pv_nics == 0) {
323 		if (bootverbose)
324 			printf("XEN: disabling emulated nics\n");
325 		disable_devs |= XMI_UNPLUG_NICS;
326 	}
327 
328 	if (disable_devs != 0)
329 		outw(XEN_MAGIC_IOPORT, disable_devs);
330 }
331 
332 static void
333 xen_hvm_init(enum xen_hvm_init_type init_type)
334 {
335 	int error;
336 	int i;
337 
338 	if (init_type == XEN_HVM_INIT_CANCELLED_SUSPEND)
339 		return;
340 
341 	error = xen_hvm_init_hypercall_stubs(init_type);
342 
343 	switch (init_type) {
344 	case XEN_HVM_INIT_LATE:
345 		if (error != 0)
346 			return;
347 
348 		/*
349 		 * If xen_domain_type is not set at this point
350 		 * it means we are inside a (PV)HVM guest, because
351 		 * for PVH the guest type is set much earlier
352 		 * (see hammer_time_xen).
353 		 */
354 		if (!xen_domain()) {
355 			xen_domain_type = XEN_HVM_DOMAIN;
356 			vm_guest = VM_GUEST_XEN;
357 		}
358 
359 		setup_xen_features();
360 #ifdef SMP
361 		cpu_ops = xen_hvm_cpu_ops;
362 #endif
363 		break;
364 	case XEN_HVM_INIT_RESUME:
365 		if (error != 0)
366 			panic("Unable to init Xen hypercall stubs on resume");
367 
368 		/* Clear stale vcpu_info. */
369 		CPU_FOREACH(i)
370 			DPCPU_ID_SET(i, vcpu_info, NULL);
371 		break;
372 	default:
373 		panic("Unsupported HVM initialization type");
374 	}
375 
376 	xen_vector_callback_enabled = 0;
377 	xen_evtchn_needs_ack = false;
378 	xen_hvm_set_callback(NULL);
379 
380 	/*
381 	 * On (PV)HVM domains we need to request the hypervisor to
382 	 * fill the shared info page, for PVH guest the shared_info page
383 	 * is passed inside the start_info struct and is already set, so this
384 	 * functions are no-ops.
385 	 */
386 	xen_hvm_init_shared_info_page();
387 	xen_hvm_disable_emulated_devices();
388 }
389 
390 void
391 xen_hvm_suspend(void)
392 {
393 }
394 
395 void
396 xen_hvm_resume(bool suspend_cancelled)
397 {
398 
399 	xen_hvm_init(suspend_cancelled ?
400 	    XEN_HVM_INIT_CANCELLED_SUSPEND : XEN_HVM_INIT_RESUME);
401 
402 	/* Register vcpu_info area for CPU#0. */
403 	xen_hvm_cpu_init();
404 }
405 
406 static void
407 xen_hvm_sysinit(void *arg __unused)
408 {
409 	xen_hvm_init(XEN_HVM_INIT_LATE);
410 }
411 SYSINIT(xen_hvm_init, SI_SUB_HYPERVISOR, SI_ORDER_FIRST, xen_hvm_sysinit, NULL);
412 
413 static void
414 xen_hvm_cpu_init(void)
415 {
416 	uint32_t regs[4];
417 	int rc;
418 
419 	if (!xen_domain())
420 		return;
421 
422 	if (DPCPU_GET(vcpu_info) != NULL) {
423 		/*
424 		 * vcpu_info is already set.  We're resuming
425 		 * from a failed migration and our pre-suspend
426 		 * configuration is still valid.
427 		 */
428 		return;
429 	}
430 
431 	/*
432 	 * Set vCPU ID. If available fetch the ID from CPUID, if not just use
433 	 * the ACPI ID.
434 	 */
435 	KASSERT(xen_cpuid_base != 0, ("Invalid base Xen CPUID leaf"));
436 	cpuid_count(xen_cpuid_base + 4, 0, regs);
437 	KASSERT((regs[0] & XEN_HVM_CPUID_VCPU_ID_PRESENT) ||
438 	    !xen_pv_domain(),
439 	    ("Xen PV domain without vcpu_id in cpuid"));
440 	PCPU_SET(vcpu_id, (regs[0] & XEN_HVM_CPUID_VCPU_ID_PRESENT) ?
441 	    regs[1] : PCPU_GET(acpi_id));
442 
443 	if (xen_evtchn_needs_ack && !IS_BSP()) {
444 		/*
445 		 * Setup the per-vpcu event channel upcall vector. This is only
446 		 * required when using the new HVMOP_set_evtchn_upcall_vector
447 		 * hypercall, which allows using a different vector for each
448 		 * vCPU. Note that FreeBSD uses the same vector for all vCPUs
449 		 * because it's not dynamically allocated.
450 		 */
451 		rc = set_percpu_callback(PCPU_GET(vcpu_id));
452 		if (rc != 0)
453 			panic("Event channel upcall vector setup failed: %d",
454 			    rc);
455 	}
456 
457 	xen_setup_vcpu_info();
458 }
459 SYSINIT(xen_hvm_cpu_init, SI_SUB_INTR, SI_ORDER_FIRST, xen_hvm_cpu_init, NULL);
460 
461 bool
462 xen_has_iommu_maps(void)
463 {
464 	uint32_t regs[4];
465 
466 	KASSERT(xen_cpuid_base != 0, ("Invalid base Xen CPUID leaf"));
467 	cpuid_count(xen_cpuid_base + 4, 0, regs);
468 
469 	return (regs[0] & XEN_HVM_CPUID_IOMMU_MAPPINGS);
470 }
471