1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2015, 2016 ARM Ltd.
4 */
5
6#include <linux/uaccess.h>
7#include <linux/interrupt.h>
8#include <linux/cpu.h>
9#include <linux/kvm_host.h>
10#include <kvm/arm_vgic.h>
11#include <asm/kvm_emulate.h>
12#include <asm/kvm_mmu.h>
13#include "vgic.h"
14
15/*
16 * Initialization rules: there are multiple stages to the vgic
17 * initialization, both for the distributor and the CPU interfaces.  The basic
18 * idea is that even though the VGIC is not functional or not requested from
19 * user space, the critical path of the run loop can still call VGIC functions
20 * that just won't do anything, without them having to check additional
21 * initialization flags to ensure they don't look at uninitialized data
22 * structures.
23 *
24 * Distributor:
25 *
26 * - kvm_vgic_early_init(): initialization of static data that doesn't
27 *   depend on any sizing information or emulation type. No allocation
28 *   is allowed there.
29 *
30 * - vgic_init(): allocation and initialization of the generic data
31 *   structures that depend on sizing information (number of CPUs,
32 *   number of interrupts). Also initializes the vcpu specific data
33 *   structures. Can be executed lazily for GICv2.
34 *
35 * CPU Interface:
36 *
37 * - kvm_vgic_vcpu_init(): initialization of static data that
38 *   doesn't depend on any sizing information or emulation type. No
39 *   allocation is allowed there.
40 */
41
42/* EARLY INIT */
43
44/**
45 * kvm_vgic_early_init() - Initialize static VGIC VCPU data structures
46 * @kvm: The VM whose VGIC districutor should be initialized
47 *
48 * Only do initialization of static structures that don't require any
49 * allocation or sizing information from userspace.  vgic_init() called
50 * kvm_vgic_dist_init() which takes care of the rest.
51 */
52void kvm_vgic_early_init(struct kvm *kvm)
53{
54	struct vgic_dist *dist = &kvm->arch.vgic;
55
56	INIT_LIST_HEAD(&dist->lpi_translation_cache);
57	raw_spin_lock_init(&dist->lpi_list_lock);
58	xa_init_flags(&dist->lpi_xa, XA_FLAGS_LOCK_IRQ);
59}
60
61/* CREATION */
62
63/**
64 * kvm_vgic_create: triggered by the instantiation of the VGIC device by
65 * user space, either through the legacy KVM_CREATE_IRQCHIP ioctl (v2 only)
66 * or through the generic KVM_CREATE_DEVICE API ioctl.
67 * irqchip_in_kernel() tells you if this function succeeded or not.
68 * @kvm: kvm struct pointer
69 * @type: KVM_DEV_TYPE_ARM_VGIC_V[23]
70 */
71int kvm_vgic_create(struct kvm *kvm, u32 type)
72{
73	struct kvm_vcpu *vcpu;
74	unsigned long i;
75	int ret;
76
77	/*
78	 * This function is also called by the KVM_CREATE_IRQCHIP handler,
79	 * which had no chance yet to check the availability of the GICv2
80	 * emulation. So check this here again. KVM_CREATE_DEVICE does
81	 * the proper checks already.
82	 */
83	if (type == KVM_DEV_TYPE_ARM_VGIC_V2 &&
84		!kvm_vgic_global_state.can_emulate_gicv2)
85		return -ENODEV;
86
87	/* Must be held to avoid race with vCPU creation */
88	lockdep_assert_held(&kvm->lock);
89
90	ret = -EBUSY;
91	if (!lock_all_vcpus(kvm))
92		return ret;
93
94	mutex_lock(&kvm->arch.config_lock);
95
96	if (irqchip_in_kernel(kvm)) {
97		ret = -EEXIST;
98		goto out_unlock;
99	}
100
101	kvm_for_each_vcpu(i, vcpu, kvm) {
102		if (vcpu_has_run_once(vcpu))
103			goto out_unlock;
104	}
105	ret = 0;
106
107	if (type == KVM_DEV_TYPE_ARM_VGIC_V2)
108		kvm->max_vcpus = VGIC_V2_MAX_CPUS;
109	else
110		kvm->max_vcpus = VGIC_V3_MAX_CPUS;
111
112	if (atomic_read(&kvm->online_vcpus) > kvm->max_vcpus) {
113		ret = -E2BIG;
114		goto out_unlock;
115	}
116
117	kvm->arch.vgic.in_kernel = true;
118	kvm->arch.vgic.vgic_model = type;
119
120	kvm->arch.vgic.vgic_dist_base = VGIC_ADDR_UNDEF;
121
122	if (type == KVM_DEV_TYPE_ARM_VGIC_V2)
123		kvm->arch.vgic.vgic_cpu_base = VGIC_ADDR_UNDEF;
124	else
125		INIT_LIST_HEAD(&kvm->arch.vgic.rd_regions);
126
127out_unlock:
128	mutex_unlock(&kvm->arch.config_lock);
129	unlock_all_vcpus(kvm);
130	return ret;
131}
132
133/* INIT/DESTROY */
134
135/**
136 * kvm_vgic_dist_init: initialize the dist data structures
137 * @kvm: kvm struct pointer
138 * @nr_spis: number of spis, frozen by caller
139 */
140static int kvm_vgic_dist_init(struct kvm *kvm, unsigned int nr_spis)
141{
142	struct vgic_dist *dist = &kvm->arch.vgic;
143	struct kvm_vcpu *vcpu0 = kvm_get_vcpu(kvm, 0);
144	int i;
145
146	dist->spis = kcalloc(nr_spis, sizeof(struct vgic_irq), GFP_KERNEL_ACCOUNT);
147	if (!dist->spis)
148		return  -ENOMEM;
149
150	/*
151	 * In the following code we do not take the irq struct lock since
152	 * no other action on irq structs can happen while the VGIC is
153	 * not initialized yet:
154	 * If someone wants to inject an interrupt or does a MMIO access, we
155	 * require prior initialization in case of a virtual GICv3 or trigger
156	 * initialization when using a virtual GICv2.
157	 */
158	for (i = 0; i < nr_spis; i++) {
159		struct vgic_irq *irq = &dist->spis[i];
160
161		irq->intid = i + VGIC_NR_PRIVATE_IRQS;
162		INIT_LIST_HEAD(&irq->ap_list);
163		raw_spin_lock_init(&irq->irq_lock);
164		irq->vcpu = NULL;
165		irq->target_vcpu = vcpu0;
166		kref_init(&irq->refcount);
167		switch (dist->vgic_model) {
168		case KVM_DEV_TYPE_ARM_VGIC_V2:
169			irq->targets = 0;
170			irq->group = 0;
171			break;
172		case KVM_DEV_TYPE_ARM_VGIC_V3:
173			irq->mpidr = 0;
174			irq->group = 1;
175			break;
176		default:
177			kfree(dist->spis);
178			dist->spis = NULL;
179			return -EINVAL;
180		}
181	}
182	return 0;
183}
184
185/**
186 * kvm_vgic_vcpu_init() - Initialize static VGIC VCPU data
187 * structures and register VCPU-specific KVM iodevs
188 *
189 * @vcpu: pointer to the VCPU being created and initialized
190 *
191 * Only do initialization, but do not actually enable the
192 * VGIC CPU interface
193 */
194int kvm_vgic_vcpu_init(struct kvm_vcpu *vcpu)
195{
196	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
197	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
198	int ret = 0;
199	int i;
200
201	vgic_cpu->rd_iodev.base_addr = VGIC_ADDR_UNDEF;
202
203	INIT_LIST_HEAD(&vgic_cpu->ap_list_head);
204	raw_spin_lock_init(&vgic_cpu->ap_list_lock);
205	atomic_set(&vgic_cpu->vgic_v3.its_vpe.vlpi_count, 0);
206
207	/*
208	 * Enable and configure all SGIs to be edge-triggered and
209	 * configure all PPIs as level-triggered.
210	 */
211	for (i = 0; i < VGIC_NR_PRIVATE_IRQS; i++) {
212		struct vgic_irq *irq = &vgic_cpu->private_irqs[i];
213
214		INIT_LIST_HEAD(&irq->ap_list);
215		raw_spin_lock_init(&irq->irq_lock);
216		irq->intid = i;
217		irq->vcpu = NULL;
218		irq->target_vcpu = vcpu;
219		kref_init(&irq->refcount);
220		if (vgic_irq_is_sgi(i)) {
221			/* SGIs */
222			irq->enabled = 1;
223			irq->config = VGIC_CONFIG_EDGE;
224		} else {
225			/* PPIs */
226			irq->config = VGIC_CONFIG_LEVEL;
227		}
228	}
229
230	if (!irqchip_in_kernel(vcpu->kvm))
231		return 0;
232
233	/*
234	 * If we are creating a VCPU with a GICv3 we must also register the
235	 * KVM io device for the redistributor that belongs to this VCPU.
236	 */
237	if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) {
238		mutex_lock(&vcpu->kvm->slots_lock);
239		ret = vgic_register_redist_iodev(vcpu);
240		mutex_unlock(&vcpu->kvm->slots_lock);
241	}
242	return ret;
243}
244
245static void kvm_vgic_vcpu_enable(struct kvm_vcpu *vcpu)
246{
247	if (kvm_vgic_global_state.type == VGIC_V2)
248		vgic_v2_enable(vcpu);
249	else
250		vgic_v3_enable(vcpu);
251}
252
253/*
254 * vgic_init: allocates and initializes dist and vcpu data structures
255 * depending on two dimensioning parameters:
256 * - the number of spis
257 * - the number of vcpus
258 * The function is generally called when nr_spis has been explicitly set
259 * by the guest through the KVM DEVICE API. If not nr_spis is set to 256.
260 * vgic_initialized() returns true when this function has succeeded.
261 */
262int vgic_init(struct kvm *kvm)
263{
264	struct vgic_dist *dist = &kvm->arch.vgic;
265	struct kvm_vcpu *vcpu;
266	int ret = 0, i;
267	unsigned long idx;
268
269	lockdep_assert_held(&kvm->arch.config_lock);
270
271	if (vgic_initialized(kvm))
272		return 0;
273
274	/* Are we also in the middle of creating a VCPU? */
275	if (kvm->created_vcpus != atomic_read(&kvm->online_vcpus))
276		return -EBUSY;
277
278	/* freeze the number of spis */
279	if (!dist->nr_spis)
280		dist->nr_spis = VGIC_NR_IRQS_LEGACY - VGIC_NR_PRIVATE_IRQS;
281
282	ret = kvm_vgic_dist_init(kvm, dist->nr_spis);
283	if (ret)
284		goto out;
285
286	/* Initialize groups on CPUs created before the VGIC type was known */
287	kvm_for_each_vcpu(idx, vcpu, kvm) {
288		struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
289
290		for (i = 0; i < VGIC_NR_PRIVATE_IRQS; i++) {
291			struct vgic_irq *irq = &vgic_cpu->private_irqs[i];
292			switch (dist->vgic_model) {
293			case KVM_DEV_TYPE_ARM_VGIC_V3:
294				irq->group = 1;
295				irq->mpidr = kvm_vcpu_get_mpidr_aff(vcpu);
296				break;
297			case KVM_DEV_TYPE_ARM_VGIC_V2:
298				irq->group = 0;
299				irq->targets = 1U << idx;
300				break;
301			default:
302				ret = -EINVAL;
303				goto out;
304			}
305		}
306	}
307
308	if (vgic_has_its(kvm))
309		vgic_lpi_translation_cache_init(kvm);
310
311	/*
312	 * If we have GICv4.1 enabled, unconditionally request enable the
313	 * v4 support so that we get HW-accelerated vSGIs. Otherwise, only
314	 * enable it if we present a virtual ITS to the guest.
315	 */
316	if (vgic_supports_direct_msis(kvm)) {
317		ret = vgic_v4_init(kvm);
318		if (ret)
319			goto out;
320	}
321
322	kvm_for_each_vcpu(idx, vcpu, kvm)
323		kvm_vgic_vcpu_enable(vcpu);
324
325	ret = kvm_vgic_setup_default_irq_routing(kvm);
326	if (ret)
327		goto out;
328
329	vgic_debug_init(kvm);
330
331	/*
332	 * If userspace didn't set the GIC implementation revision,
333	 * default to the latest and greatest. You know want it.
334	 */
335	if (!dist->implementation_rev)
336		dist->implementation_rev = KVM_VGIC_IMP_REV_LATEST;
337	dist->initialized = true;
338
339out:
340	return ret;
341}
342
343static void kvm_vgic_dist_destroy(struct kvm *kvm)
344{
345	struct vgic_dist *dist = &kvm->arch.vgic;
346	struct vgic_redist_region *rdreg, *next;
347
348	dist->ready = false;
349	dist->initialized = false;
350
351	kfree(dist->spis);
352	dist->spis = NULL;
353	dist->nr_spis = 0;
354	dist->vgic_dist_base = VGIC_ADDR_UNDEF;
355
356	if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) {
357		list_for_each_entry_safe(rdreg, next, &dist->rd_regions, list)
358			vgic_v3_free_redist_region(rdreg);
359		INIT_LIST_HEAD(&dist->rd_regions);
360	} else {
361		dist->vgic_cpu_base = VGIC_ADDR_UNDEF;
362	}
363
364	if (vgic_has_its(kvm))
365		vgic_lpi_translation_cache_destroy(kvm);
366
367	if (vgic_supports_direct_msis(kvm))
368		vgic_v4_teardown(kvm);
369
370	xa_destroy(&dist->lpi_xa);
371}
372
373static void __kvm_vgic_vcpu_destroy(struct kvm_vcpu *vcpu)
374{
375	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
376
377	/*
378	 * Retire all pending LPIs on this vcpu anyway as we're
379	 * going to destroy it.
380	 */
381	vgic_flush_pending_lpis(vcpu);
382
383	INIT_LIST_HEAD(&vgic_cpu->ap_list_head);
384	if (vcpu->kvm->arch.vgic.vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) {
385		vgic_unregister_redist_iodev(vcpu);
386		vgic_cpu->rd_iodev.base_addr = VGIC_ADDR_UNDEF;
387	}
388}
389
390void kvm_vgic_vcpu_destroy(struct kvm_vcpu *vcpu)
391{
392	struct kvm *kvm = vcpu->kvm;
393
394	mutex_lock(&kvm->slots_lock);
395	__kvm_vgic_vcpu_destroy(vcpu);
396	mutex_unlock(&kvm->slots_lock);
397}
398
399void kvm_vgic_destroy(struct kvm *kvm)
400{
401	struct kvm_vcpu *vcpu;
402	unsigned long i;
403
404	mutex_lock(&kvm->slots_lock);
405
406	vgic_debug_destroy(kvm);
407
408	kvm_for_each_vcpu(i, vcpu, kvm)
409		__kvm_vgic_vcpu_destroy(vcpu);
410
411	mutex_lock(&kvm->arch.config_lock);
412
413	kvm_vgic_dist_destroy(kvm);
414
415	mutex_unlock(&kvm->arch.config_lock);
416	mutex_unlock(&kvm->slots_lock);
417}
418
419/**
420 * vgic_lazy_init: Lazy init is only allowed if the GIC exposed to the guest
421 * is a GICv2. A GICv3 must be explicitly initialized by userspace using the
422 * KVM_DEV_ARM_VGIC_GRP_CTRL KVM_DEVICE group.
423 * @kvm: kvm struct pointer
424 */
425int vgic_lazy_init(struct kvm *kvm)
426{
427	int ret = 0;
428
429	if (unlikely(!vgic_initialized(kvm))) {
430		/*
431		 * We only provide the automatic initialization of the VGIC
432		 * for the legacy case of a GICv2. Any other type must
433		 * be explicitly initialized once setup with the respective
434		 * KVM device call.
435		 */
436		if (kvm->arch.vgic.vgic_model != KVM_DEV_TYPE_ARM_VGIC_V2)
437			return -EBUSY;
438
439		mutex_lock(&kvm->arch.config_lock);
440		ret = vgic_init(kvm);
441		mutex_unlock(&kvm->arch.config_lock);
442	}
443
444	return ret;
445}
446
447/* RESOURCE MAPPING */
448
449/**
450 * kvm_vgic_map_resources - map the MMIO regions
451 * @kvm: kvm struct pointer
452 *
453 * Map the MMIO regions depending on the VGIC model exposed to the guest
454 * called on the first VCPU run.
455 * Also map the virtual CPU interface into the VM.
456 * v2 calls vgic_init() if not already done.
457 * v3 and derivatives return an error if the VGIC is not initialized.
458 * vgic_ready() returns true if this function has succeeded.
459 */
460int kvm_vgic_map_resources(struct kvm *kvm)
461{
462	struct vgic_dist *dist = &kvm->arch.vgic;
463	enum vgic_type type;
464	gpa_t dist_base;
465	int ret = 0;
466
467	if (likely(vgic_ready(kvm)))
468		return 0;
469
470	mutex_lock(&kvm->slots_lock);
471	mutex_lock(&kvm->arch.config_lock);
472	if (vgic_ready(kvm))
473		goto out;
474
475	if (!irqchip_in_kernel(kvm))
476		goto out;
477
478	if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V2) {
479		ret = vgic_v2_map_resources(kvm);
480		type = VGIC_V2;
481	} else {
482		ret = vgic_v3_map_resources(kvm);
483		type = VGIC_V3;
484	}
485
486	if (ret)
487		goto out;
488
489	dist->ready = true;
490	dist_base = dist->vgic_dist_base;
491	mutex_unlock(&kvm->arch.config_lock);
492
493	ret = vgic_register_dist_iodev(kvm, dist_base, type);
494	if (ret)
495		kvm_err("Unable to register VGIC dist MMIO regions\n");
496
497	goto out_slots;
498out:
499	mutex_unlock(&kvm->arch.config_lock);
500out_slots:
501	mutex_unlock(&kvm->slots_lock);
502
503	if (ret)
504		kvm_vgic_destroy(kvm);
505
506	return ret;
507}
508
509/* GENERIC PROBE */
510
511void kvm_vgic_cpu_up(void)
512{
513	enable_percpu_irq(kvm_vgic_global_state.maint_irq, 0);
514}
515
516
517void kvm_vgic_cpu_down(void)
518{
519	disable_percpu_irq(kvm_vgic_global_state.maint_irq);
520}
521
522static irqreturn_t vgic_maintenance_handler(int irq, void *data)
523{
524	/*
525	 * We cannot rely on the vgic maintenance interrupt to be
526	 * delivered synchronously. This means we can only use it to
527	 * exit the VM, and we perform the handling of EOIed
528	 * interrupts on the exit path (see vgic_fold_lr_state).
529	 */
530	return IRQ_HANDLED;
531}
532
533static struct gic_kvm_info *gic_kvm_info;
534
535void __init vgic_set_kvm_info(const struct gic_kvm_info *info)
536{
537	BUG_ON(gic_kvm_info != NULL);
538	gic_kvm_info = kmalloc(sizeof(*info), GFP_KERNEL);
539	if (gic_kvm_info)
540		*gic_kvm_info = *info;
541}
542
543/**
544 * kvm_vgic_init_cpu_hardware - initialize the GIC VE hardware
545 *
546 * For a specific CPU, initialize the GIC VE hardware.
547 */
548void kvm_vgic_init_cpu_hardware(void)
549{
550	BUG_ON(preemptible());
551
552	/*
553	 * We want to make sure the list registers start out clear so that we
554	 * only have the program the used registers.
555	 */
556	if (kvm_vgic_global_state.type == VGIC_V2)
557		vgic_v2_init_lrs();
558	else
559		kvm_call_hyp(__vgic_v3_init_lrs);
560}
561
562/**
563 * kvm_vgic_hyp_init: populates the kvm_vgic_global_state variable
564 * according to the host GIC model. Accordingly calls either
565 * vgic_v2/v3_probe which registers the KVM_DEVICE that can be
566 * instantiated by a guest later on .
567 */
568int kvm_vgic_hyp_init(void)
569{
570	bool has_mask;
571	int ret;
572
573	if (!gic_kvm_info)
574		return -ENODEV;
575
576	has_mask = !gic_kvm_info->no_maint_irq_mask;
577
578	if (has_mask && !gic_kvm_info->maint_irq) {
579		kvm_err("No vgic maintenance irq\n");
580		return -ENXIO;
581	}
582
583	/*
584	 * If we get one of these oddball non-GICs, taint the kernel,
585	 * as we have no idea of how they *really* behave.
586	 */
587	if (gic_kvm_info->no_hw_deactivation) {
588		kvm_info("Non-architectural vgic, tainting kernel\n");
589		add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
590		kvm_vgic_global_state.no_hw_deactivation = true;
591	}
592
593	switch (gic_kvm_info->type) {
594	case GIC_V2:
595		ret = vgic_v2_probe(gic_kvm_info);
596		break;
597	case GIC_V3:
598		ret = vgic_v3_probe(gic_kvm_info);
599		if (!ret) {
600			static_branch_enable(&kvm_vgic_global_state.gicv3_cpuif);
601			kvm_info("GIC system register CPU interface enabled\n");
602		}
603		break;
604	default:
605		ret = -ENODEV;
606	}
607
608	kvm_vgic_global_state.maint_irq = gic_kvm_info->maint_irq;
609
610	kfree(gic_kvm_info);
611	gic_kvm_info = NULL;
612
613	if (ret)
614		return ret;
615
616	if (!has_mask && !kvm_vgic_global_state.maint_irq)
617		return 0;
618
619	ret = request_percpu_irq(kvm_vgic_global_state.maint_irq,
620				 vgic_maintenance_handler,
621				 "vgic", kvm_get_running_vcpus());
622	if (ret) {
623		kvm_err("Cannot register interrupt %d\n",
624			kvm_vgic_global_state.maint_irq);
625		return ret;
626	}
627
628	kvm_info("vgic interrupt IRQ%d\n", kvm_vgic_global_state.maint_irq);
629	return 0;
630}
631