xref: /openbmc/linux/arch/arm64/kvm/vgic/vgic-init.c (revision 1f012283)
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  */
52 void kvm_vgic_early_init(struct kvm *kvm)
53 {
54 	struct vgic_dist *dist = &kvm->arch.vgic;
55 
56 	INIT_LIST_HEAD(&dist->lpi_list_head);
57 	INIT_LIST_HEAD(&dist->lpi_translation_cache);
58 	raw_spin_lock_init(&dist->lpi_list_lock);
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  */
71 int kvm_vgic_create(struct kvm *kvm, u32 type)
72 {
73 	int i, ret;
74 	struct kvm_vcpu *vcpu;
75 
76 	if (irqchip_in_kernel(kvm))
77 		return -EEXIST;
78 
79 	/*
80 	 * This function is also called by the KVM_CREATE_IRQCHIP handler,
81 	 * which had no chance yet to check the availability of the GICv2
82 	 * emulation. So check this here again. KVM_CREATE_DEVICE does
83 	 * the proper checks already.
84 	 */
85 	if (type == KVM_DEV_TYPE_ARM_VGIC_V2 &&
86 		!kvm_vgic_global_state.can_emulate_gicv2)
87 		return -ENODEV;
88 
89 	ret = -EBUSY;
90 	if (!lock_all_vcpus(kvm))
91 		return ret;
92 
93 	kvm_for_each_vcpu(i, vcpu, kvm) {
94 		if (vcpu->arch.has_run_once)
95 			goto out_unlock;
96 	}
97 	ret = 0;
98 
99 	if (type == KVM_DEV_TYPE_ARM_VGIC_V2)
100 		kvm->arch.max_vcpus = VGIC_V2_MAX_CPUS;
101 	else
102 		kvm->arch.max_vcpus = VGIC_V3_MAX_CPUS;
103 
104 	if (atomic_read(&kvm->online_vcpus) > kvm->arch.max_vcpus) {
105 		ret = -E2BIG;
106 		goto out_unlock;
107 	}
108 
109 	kvm->arch.vgic.in_kernel = true;
110 	kvm->arch.vgic.vgic_model = type;
111 
112 	kvm->arch.vgic.vgic_dist_base = VGIC_ADDR_UNDEF;
113 
114 	if (type == KVM_DEV_TYPE_ARM_VGIC_V2)
115 		kvm->arch.vgic.vgic_cpu_base = VGIC_ADDR_UNDEF;
116 	else
117 		INIT_LIST_HEAD(&kvm->arch.vgic.rd_regions);
118 
119 out_unlock:
120 	unlock_all_vcpus(kvm);
121 	return ret;
122 }
123 
124 /* INIT/DESTROY */
125 
126 /**
127  * kvm_vgic_dist_init: initialize the dist data structures
128  * @kvm: kvm struct pointer
129  * @nr_spis: number of spis, frozen by caller
130  */
131 static int kvm_vgic_dist_init(struct kvm *kvm, unsigned int nr_spis)
132 {
133 	struct vgic_dist *dist = &kvm->arch.vgic;
134 	struct kvm_vcpu *vcpu0 = kvm_get_vcpu(kvm, 0);
135 	int i;
136 
137 	dist->spis = kcalloc(nr_spis, sizeof(struct vgic_irq), GFP_KERNEL_ACCOUNT);
138 	if (!dist->spis)
139 		return  -ENOMEM;
140 
141 	/*
142 	 * In the following code we do not take the irq struct lock since
143 	 * no other action on irq structs can happen while the VGIC is
144 	 * not initialized yet:
145 	 * If someone wants to inject an interrupt or does a MMIO access, we
146 	 * require prior initialization in case of a virtual GICv3 or trigger
147 	 * initialization when using a virtual GICv2.
148 	 */
149 	for (i = 0; i < nr_spis; i++) {
150 		struct vgic_irq *irq = &dist->spis[i];
151 
152 		irq->intid = i + VGIC_NR_PRIVATE_IRQS;
153 		INIT_LIST_HEAD(&irq->ap_list);
154 		raw_spin_lock_init(&irq->irq_lock);
155 		irq->vcpu = NULL;
156 		irq->target_vcpu = vcpu0;
157 		kref_init(&irq->refcount);
158 		switch (dist->vgic_model) {
159 		case KVM_DEV_TYPE_ARM_VGIC_V2:
160 			irq->targets = 0;
161 			irq->group = 0;
162 			break;
163 		case KVM_DEV_TYPE_ARM_VGIC_V3:
164 			irq->mpidr = 0;
165 			irq->group = 1;
166 			break;
167 		default:
168 			kfree(dist->spis);
169 			dist->spis = NULL;
170 			return -EINVAL;
171 		}
172 	}
173 	return 0;
174 }
175 
176 /**
177  * kvm_vgic_vcpu_init() - Initialize static VGIC VCPU data
178  * structures and register VCPU-specific KVM iodevs
179  *
180  * @vcpu: pointer to the VCPU being created and initialized
181  *
182  * Only do initialization, but do not actually enable the
183  * VGIC CPU interface
184  */
185 int kvm_vgic_vcpu_init(struct kvm_vcpu *vcpu)
186 {
187 	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
188 	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
189 	int ret = 0;
190 	int i;
191 
192 	vgic_cpu->rd_iodev.base_addr = VGIC_ADDR_UNDEF;
193 
194 	INIT_LIST_HEAD(&vgic_cpu->ap_list_head);
195 	raw_spin_lock_init(&vgic_cpu->ap_list_lock);
196 	atomic_set(&vgic_cpu->vgic_v3.its_vpe.vlpi_count, 0);
197 
198 	/*
199 	 * Enable and configure all SGIs to be edge-triggered and
200 	 * configure all PPIs as level-triggered.
201 	 */
202 	for (i = 0; i < VGIC_NR_PRIVATE_IRQS; i++) {
203 		struct vgic_irq *irq = &vgic_cpu->private_irqs[i];
204 
205 		INIT_LIST_HEAD(&irq->ap_list);
206 		raw_spin_lock_init(&irq->irq_lock);
207 		irq->intid = i;
208 		irq->vcpu = NULL;
209 		irq->target_vcpu = vcpu;
210 		kref_init(&irq->refcount);
211 		if (vgic_irq_is_sgi(i)) {
212 			/* SGIs */
213 			irq->enabled = 1;
214 			irq->config = VGIC_CONFIG_EDGE;
215 		} else {
216 			/* PPIs */
217 			irq->config = VGIC_CONFIG_LEVEL;
218 		}
219 	}
220 
221 	if (!irqchip_in_kernel(vcpu->kvm))
222 		return 0;
223 
224 	/*
225 	 * If we are creating a VCPU with a GICv3 we must also register the
226 	 * KVM io device for the redistributor that belongs to this VCPU.
227 	 */
228 	if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) {
229 		mutex_lock(&vcpu->kvm->lock);
230 		ret = vgic_register_redist_iodev(vcpu);
231 		mutex_unlock(&vcpu->kvm->lock);
232 	}
233 	return ret;
234 }
235 
236 static void kvm_vgic_vcpu_enable(struct kvm_vcpu *vcpu)
237 {
238 	if (kvm_vgic_global_state.type == VGIC_V2)
239 		vgic_v2_enable(vcpu);
240 	else
241 		vgic_v3_enable(vcpu);
242 }
243 
244 /*
245  * vgic_init: allocates and initializes dist and vcpu data structures
246  * depending on two dimensioning parameters:
247  * - the number of spis
248  * - the number of vcpus
249  * The function is generally called when nr_spis has been explicitly set
250  * by the guest through the KVM DEVICE API. If not nr_spis is set to 256.
251  * vgic_initialized() returns true when this function has succeeded.
252  * Must be called with kvm->lock held!
253  */
254 int vgic_init(struct kvm *kvm)
255 {
256 	struct vgic_dist *dist = &kvm->arch.vgic;
257 	struct kvm_vcpu *vcpu;
258 	int ret = 0, i, idx;
259 
260 	if (vgic_initialized(kvm))
261 		return 0;
262 
263 	/* Are we also in the middle of creating a VCPU? */
264 	if (kvm->created_vcpus != atomic_read(&kvm->online_vcpus))
265 		return -EBUSY;
266 
267 	/* freeze the number of spis */
268 	if (!dist->nr_spis)
269 		dist->nr_spis = VGIC_NR_IRQS_LEGACY - VGIC_NR_PRIVATE_IRQS;
270 
271 	ret = kvm_vgic_dist_init(kvm, dist->nr_spis);
272 	if (ret)
273 		goto out;
274 
275 	/* Initialize groups on CPUs created before the VGIC type was known */
276 	kvm_for_each_vcpu(idx, vcpu, kvm) {
277 		struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
278 
279 		for (i = 0; i < VGIC_NR_PRIVATE_IRQS; i++) {
280 			struct vgic_irq *irq = &vgic_cpu->private_irqs[i];
281 			switch (dist->vgic_model) {
282 			case KVM_DEV_TYPE_ARM_VGIC_V3:
283 				irq->group = 1;
284 				irq->mpidr = kvm_vcpu_get_mpidr_aff(vcpu);
285 				break;
286 			case KVM_DEV_TYPE_ARM_VGIC_V2:
287 				irq->group = 0;
288 				irq->targets = 1U << idx;
289 				break;
290 			default:
291 				ret = -EINVAL;
292 				goto out;
293 			}
294 		}
295 	}
296 
297 	if (vgic_has_its(kvm))
298 		vgic_lpi_translation_cache_init(kvm);
299 
300 	/*
301 	 * If we have GICv4.1 enabled, unconditionnaly request enable the
302 	 * v4 support so that we get HW-accelerated vSGIs. Otherwise, only
303 	 * enable it if we present a virtual ITS to the guest.
304 	 */
305 	if (vgic_supports_direct_msis(kvm)) {
306 		ret = vgic_v4_init(kvm);
307 		if (ret)
308 			goto out;
309 	}
310 
311 	kvm_for_each_vcpu(i, vcpu, kvm)
312 		kvm_vgic_vcpu_enable(vcpu);
313 
314 	ret = kvm_vgic_setup_default_irq_routing(kvm);
315 	if (ret)
316 		goto out;
317 
318 	vgic_debug_init(kvm);
319 
320 	dist->implementation_rev = 2;
321 	dist->initialized = true;
322 
323 out:
324 	return ret;
325 }
326 
327 static void kvm_vgic_dist_destroy(struct kvm *kvm)
328 {
329 	struct vgic_dist *dist = &kvm->arch.vgic;
330 	struct vgic_redist_region *rdreg, *next;
331 
332 	dist->ready = false;
333 	dist->initialized = false;
334 
335 	kfree(dist->spis);
336 	dist->spis = NULL;
337 	dist->nr_spis = 0;
338 	dist->vgic_dist_base = VGIC_ADDR_UNDEF;
339 
340 	if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) {
341 		list_for_each_entry_safe(rdreg, next, &dist->rd_regions, list)
342 			vgic_v3_free_redist_region(rdreg);
343 		INIT_LIST_HEAD(&dist->rd_regions);
344 	} else {
345 		dist->vgic_cpu_base = VGIC_ADDR_UNDEF;
346 	}
347 
348 	if (vgic_has_its(kvm))
349 		vgic_lpi_translation_cache_destroy(kvm);
350 
351 	if (vgic_supports_direct_msis(kvm))
352 		vgic_v4_teardown(kvm);
353 }
354 
355 void kvm_vgic_vcpu_destroy(struct kvm_vcpu *vcpu)
356 {
357 	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
358 
359 	/*
360 	 * Retire all pending LPIs on this vcpu anyway as we're
361 	 * going to destroy it.
362 	 */
363 	vgic_flush_pending_lpis(vcpu);
364 
365 	INIT_LIST_HEAD(&vgic_cpu->ap_list_head);
366 	vgic_cpu->rd_iodev.base_addr = VGIC_ADDR_UNDEF;
367 }
368 
369 /* To be called with kvm->lock held */
370 static void __kvm_vgic_destroy(struct kvm *kvm)
371 {
372 	struct kvm_vcpu *vcpu;
373 	int i;
374 
375 	vgic_debug_destroy(kvm);
376 
377 	kvm_for_each_vcpu(i, vcpu, kvm)
378 		kvm_vgic_vcpu_destroy(vcpu);
379 
380 	kvm_vgic_dist_destroy(kvm);
381 }
382 
383 void kvm_vgic_destroy(struct kvm *kvm)
384 {
385 	mutex_lock(&kvm->lock);
386 	__kvm_vgic_destroy(kvm);
387 	mutex_unlock(&kvm->lock);
388 }
389 
390 /**
391  * vgic_lazy_init: Lazy init is only allowed if the GIC exposed to the guest
392  * is a GICv2. A GICv3 must be explicitly initialized by the guest using the
393  * KVM_DEV_ARM_VGIC_GRP_CTRL KVM_DEVICE group.
394  * @kvm: kvm struct pointer
395  */
396 int vgic_lazy_init(struct kvm *kvm)
397 {
398 	int ret = 0;
399 
400 	if (unlikely(!vgic_initialized(kvm))) {
401 		/*
402 		 * We only provide the automatic initialization of the VGIC
403 		 * for the legacy case of a GICv2. Any other type must
404 		 * be explicitly initialized once setup with the respective
405 		 * KVM device call.
406 		 */
407 		if (kvm->arch.vgic.vgic_model != KVM_DEV_TYPE_ARM_VGIC_V2)
408 			return -EBUSY;
409 
410 		mutex_lock(&kvm->lock);
411 		ret = vgic_init(kvm);
412 		mutex_unlock(&kvm->lock);
413 	}
414 
415 	return ret;
416 }
417 
418 /* RESOURCE MAPPING */
419 
420 /**
421  * Map the MMIO regions depending on the VGIC model exposed to the guest
422  * called on the first VCPU run.
423  * Also map the virtual CPU interface into the VM.
424  * v2 calls vgic_init() if not already done.
425  * v3 and derivatives return an error if the VGIC is not initialized.
426  * vgic_ready() returns true if this function has succeeded.
427  * @kvm: kvm struct pointer
428  */
429 int kvm_vgic_map_resources(struct kvm *kvm)
430 {
431 	struct vgic_dist *dist = &kvm->arch.vgic;
432 	int ret = 0;
433 
434 	if (likely(vgic_ready(kvm)))
435 		return 0;
436 
437 	mutex_lock(&kvm->lock);
438 	if (vgic_ready(kvm))
439 		goto out;
440 
441 	if (!irqchip_in_kernel(kvm))
442 		goto out;
443 
444 	if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V2)
445 		ret = vgic_v2_map_resources(kvm);
446 	else
447 		ret = vgic_v3_map_resources(kvm);
448 
449 	if (ret)
450 		__kvm_vgic_destroy(kvm);
451 	else
452 		dist->ready = true;
453 
454 out:
455 	mutex_unlock(&kvm->lock);
456 	return ret;
457 }
458 
459 /* GENERIC PROBE */
460 
461 static int vgic_init_cpu_starting(unsigned int cpu)
462 {
463 	enable_percpu_irq(kvm_vgic_global_state.maint_irq, 0);
464 	return 0;
465 }
466 
467 
468 static int vgic_init_cpu_dying(unsigned int cpu)
469 {
470 	disable_percpu_irq(kvm_vgic_global_state.maint_irq);
471 	return 0;
472 }
473 
474 static irqreturn_t vgic_maintenance_handler(int irq, void *data)
475 {
476 	/*
477 	 * We cannot rely on the vgic maintenance interrupt to be
478 	 * delivered synchronously. This means we can only use it to
479 	 * exit the VM, and we perform the handling of EOIed
480 	 * interrupts on the exit path (see vgic_fold_lr_state).
481 	 */
482 	return IRQ_HANDLED;
483 }
484 
485 static struct gic_kvm_info *gic_kvm_info;
486 
487 void __init vgic_set_kvm_info(const struct gic_kvm_info *info)
488 {
489 	BUG_ON(gic_kvm_info != NULL);
490 	gic_kvm_info = kmalloc(sizeof(*info), GFP_KERNEL);
491 	if (gic_kvm_info)
492 		*gic_kvm_info = *info;
493 }
494 
495 /**
496  * kvm_vgic_init_cpu_hardware - initialize the GIC VE hardware
497  *
498  * For a specific CPU, initialize the GIC VE hardware.
499  */
500 void kvm_vgic_init_cpu_hardware(void)
501 {
502 	BUG_ON(preemptible());
503 
504 	/*
505 	 * We want to make sure the list registers start out clear so that we
506 	 * only have the program the used registers.
507 	 */
508 	if (kvm_vgic_global_state.type == VGIC_V2)
509 		vgic_v2_init_lrs();
510 	else
511 		kvm_call_hyp(__vgic_v3_init_lrs);
512 }
513 
514 /**
515  * kvm_vgic_hyp_init: populates the kvm_vgic_global_state variable
516  * according to the host GIC model. Accordingly calls either
517  * vgic_v2/v3_probe which registers the KVM_DEVICE that can be
518  * instantiated by a guest later on .
519  */
520 int kvm_vgic_hyp_init(void)
521 {
522 	bool has_mask;
523 	int ret;
524 
525 	if (!gic_kvm_info)
526 		return -ENODEV;
527 
528 	has_mask = !gic_kvm_info->no_maint_irq_mask;
529 
530 	if (has_mask && !gic_kvm_info->maint_irq) {
531 		kvm_err("No vgic maintenance irq\n");
532 		return -ENXIO;
533 	}
534 
535 	/*
536 	 * If we get one of these oddball non-GICs, taint the kernel,
537 	 * as we have no idea of how they *really* behave.
538 	 */
539 	if (gic_kvm_info->no_hw_deactivation) {
540 		kvm_info("Non-architectural vgic, tainting kernel\n");
541 		add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
542 		kvm_vgic_global_state.no_hw_deactivation = true;
543 	}
544 
545 	switch (gic_kvm_info->type) {
546 	case GIC_V2:
547 		ret = vgic_v2_probe(gic_kvm_info);
548 		break;
549 	case GIC_V3:
550 		ret = vgic_v3_probe(gic_kvm_info);
551 		if (!ret) {
552 			static_branch_enable(&kvm_vgic_global_state.gicv3_cpuif);
553 			kvm_info("GIC system register CPU interface enabled\n");
554 		}
555 		break;
556 	default:
557 		ret = -ENODEV;
558 	}
559 
560 	kvm_vgic_global_state.maint_irq = gic_kvm_info->maint_irq;
561 
562 	kfree(gic_kvm_info);
563 	gic_kvm_info = NULL;
564 
565 	if (ret)
566 		return ret;
567 
568 	if (!has_mask)
569 		return 0;
570 
571 	ret = request_percpu_irq(kvm_vgic_global_state.maint_irq,
572 				 vgic_maintenance_handler,
573 				 "vgic", kvm_get_running_vcpus());
574 	if (ret) {
575 		kvm_err("Cannot register interrupt %d\n",
576 			kvm_vgic_global_state.maint_irq);
577 		return ret;
578 	}
579 
580 	ret = cpuhp_setup_state(CPUHP_AP_KVM_ARM_VGIC_INIT_STARTING,
581 				"kvm/arm/vgic:starting",
582 				vgic_init_cpu_starting, vgic_init_cpu_dying);
583 	if (ret) {
584 		kvm_err("Cannot register vgic CPU notifier\n");
585 		goto out_free_irq;
586 	}
587 
588 	kvm_info("vgic interrupt IRQ%d\n", kvm_vgic_global_state.maint_irq);
589 	return 0;
590 
591 out_free_irq:
592 	free_percpu_irq(kvm_vgic_global_state.maint_irq,
593 			kvm_get_running_vcpus());
594 	return ret;
595 }
596