xref: /linux/arch/arm64/kvm/vgic/vgic-init.c (revision 9e4e86a604dfd06402933467578c4b79f5412b2c)
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 doesn't depend
38  *   on any sizing information. Private interrupts are allocated if not
39  *   already allocated at vgic-creation time.
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  */
kvm_vgic_early_init(struct kvm * kvm)52 void kvm_vgic_early_init(struct kvm *kvm)
53 {
54 	struct vgic_dist *dist = &kvm->arch.vgic;
55 
56 	xa_init_flags(&dist->lpi_xa, XA_FLAGS_LOCK_IRQ);
57 }
58 
59 /* CREATION */
60 
61 static int vgic_allocate_private_irqs_locked(struct kvm_vcpu *vcpu, u32 type);
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[235]
70  */
kvm_vgic_create(struct kvm * kvm,u32 type)71 int 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 	/*
88 	 * Ensure mutual exclusion with vCPU creation and any vCPU ioctls by:
89 	 *
90 	 *  - Holding kvm->lock to prevent KVM_CREATE_VCPU from reaching
91 	 *    kvm_arch_vcpu_precreate() and ensuring created_vcpus is stable.
92 	 *    This alone is insufficient, as kvm_vm_ioctl_create_vcpu() drops
93 	 *    the kvm->lock before completing the vCPU creation.
94 	 */
95 	lockdep_assert_held(&kvm->lock);
96 
97 	/*
98 	 *  - Acquiring the vCPU mutex for every *online* vCPU to prevent
99 	 *    concurrent vCPU ioctls for vCPUs already visible to userspace.
100 	 */
101 	ret = -EBUSY;
102 	if (kvm_trylock_all_vcpus(kvm))
103 		return ret;
104 
105 	/*
106 	 *  - Taking the config_lock which protects VGIC data structures such
107 	 *    as the per-vCPU arrays of private IRQs (SGIs, PPIs).
108 	 */
109 	mutex_lock(&kvm->arch.config_lock);
110 
111 	/*
112 	 * - Bailing on the entire thing if a vCPU is in the middle of creation,
113 	 *   dropped the kvm->lock, but hasn't reached kvm_arch_vcpu_create().
114 	 *
115 	 * The whole combination of this guarantees that no vCPU can get into
116 	 * KVM with a VGIC configuration inconsistent with the VM's VGIC.
117 	 */
118 	if (kvm->created_vcpus != atomic_read(&kvm->online_vcpus))
119 		goto out_unlock;
120 
121 	if (irqchip_in_kernel(kvm)) {
122 		ret = -EEXIST;
123 		goto out_unlock;
124 	}
125 
126 	kvm_for_each_vcpu(i, vcpu, kvm) {
127 		if (vcpu_has_run_once(vcpu))
128 			goto out_unlock;
129 	}
130 	ret = 0;
131 
132 	if (type == KVM_DEV_TYPE_ARM_VGIC_V2)
133 		kvm->max_vcpus = VGIC_V2_MAX_CPUS;
134 	else if (type == KVM_DEV_TYPE_ARM_VGIC_V3)
135 		kvm->max_vcpus = VGIC_V3_MAX_CPUS;
136 	else if (type == KVM_DEV_TYPE_ARM_VGIC_V5)
137 		kvm->max_vcpus = min(VGIC_V5_MAX_CPUS,
138 				     kvm_vgic_global_state.max_gic_vcpus);
139 
140 	if (atomic_read(&kvm->online_vcpus) > kvm->max_vcpus) {
141 		ret = -E2BIG;
142 		goto out_unlock;
143 	}
144 
145 	kvm->arch.vgic.in_kernel = true;
146 	kvm->arch.vgic.vgic_model = type;
147 	kvm->arch.vgic.implementation_rev = KVM_VGIC_IMP_REV_LATEST;
148 	kvm->arch.vgic.vgic_dist_base = VGIC_ADDR_UNDEF;
149 
150 	switch (type) {
151 	case KVM_DEV_TYPE_ARM_VGIC_V2:
152 		kvm->arch.vgic.vgic_cpu_base = VGIC_ADDR_UNDEF;
153 		break;
154 	case KVM_DEV_TYPE_ARM_VGIC_V3:
155 		INIT_LIST_HEAD(&kvm->arch.vgic.rd_regions);
156 		break;
157 	}
158 
159 	/*
160 	 * We've now created the GIC. Update the system register state
161 	 * to accurately reflect what we've created.
162 	 */
163 	kvm_vgic_finalize_idregs(kvm);
164 
165 	kvm_for_each_vcpu(i, vcpu, kvm) {
166 		ret = vgic_allocate_private_irqs_locked(vcpu, type);
167 		if (ret)
168 			break;
169 	}
170 
171 	if (ret) {
172 		kvm_for_each_vcpu(i, vcpu, kvm) {
173 			struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
174 			kfree(vgic_cpu->private_irqs);
175 			vgic_cpu->private_irqs = NULL;
176 		}
177 
178 		kvm->arch.vgic.vgic_model = 0;
179 		goto out_unlock;
180 	}
181 
182 	if (type == KVM_DEV_TYPE_ARM_VGIC_V3)
183 		kvm->arch.vgic.nassgicap = system_supports_direct_sgis();
184 
185 	/*
186 	 * We now know that we have a GICv5. The Arch Timer PPI interrupts may
187 	 * have been initialised at this stage, but will have done so assuming
188 	 * that we have an older GIC, meaning that the IntIDs won't be
189 	 * correct. We init them again, and this time they will be correct.
190 	 */
191 	if (type == KVM_DEV_TYPE_ARM_VGIC_V5)
192 		kvm_timer_init_vm(kvm);
193 
194 out_unlock:
195 	mutex_unlock(&kvm->arch.config_lock);
196 	kvm_unlock_all_vcpus(kvm);
197 	return ret;
198 }
199 
200 /* INIT/DESTROY */
201 
202 /**
203  * kvm_vgic_dist_init: initialize the dist data structures
204  * @kvm: kvm struct pointer
205  * @nr_spis: number of spis, frozen by caller
206  */
kvm_vgic_dist_init(struct kvm * kvm,unsigned int nr_spis)207 static int kvm_vgic_dist_init(struct kvm *kvm, unsigned int nr_spis)
208 {
209 	struct vgic_dist *dist = &kvm->arch.vgic;
210 	struct kvm_vcpu *vcpu0 = kvm_get_vcpu(kvm, 0);
211 	int i;
212 
213 	dist->active_spis = (atomic_t)ATOMIC_INIT(0);
214 	dist->spis = kzalloc_objs(struct vgic_irq, nr_spis, GFP_KERNEL_ACCOUNT);
215 	if (!dist->spis)
216 		return  -ENOMEM;
217 
218 	/*
219 	 * In the following code we do not take the irq struct lock since
220 	 * no other action on irq structs can happen while the VGIC is
221 	 * not initialized yet:
222 	 * If someone wants to inject an interrupt or does a MMIO access, we
223 	 * require prior initialization in case of a virtual GICv3 or trigger
224 	 * initialization when using a virtual GICv2.
225 	 */
226 	for (i = 0; i < nr_spis; i++) {
227 		struct vgic_irq *irq = &dist->spis[i];
228 
229 		irq->intid = i + VGIC_NR_PRIVATE_IRQS;
230 		INIT_LIST_HEAD(&irq->ap_list);
231 		raw_spin_lock_init(&irq->irq_lock);
232 		irq->vcpu = NULL;
233 		irq->target_vcpu = vcpu0;
234 		refcount_set(&irq->refcount, 0);
235 		switch (dist->vgic_model) {
236 		case KVM_DEV_TYPE_ARM_VGIC_V2:
237 			irq->targets = 0;
238 			irq->group = 0;
239 			break;
240 		case KVM_DEV_TYPE_ARM_VGIC_V3:
241 			irq->mpidr = 0;
242 			irq->group = 1;
243 			break;
244 		default:
245 			kfree(dist->spis);
246 			dist->spis = NULL;
247 			return -EINVAL;
248 		}
249 	}
250 	return 0;
251 }
252 
253 /* Default GICv3 Maintenance Interrupt INTID, as per SBSA */
254 #define DEFAULT_MI_INTID	25
255 
kvm_vgic_vcpu_nv_init(struct kvm_vcpu * vcpu)256 int kvm_vgic_vcpu_nv_init(struct kvm_vcpu *vcpu)
257 {
258 	int ret;
259 
260 	guard(mutex)(&vcpu->kvm->arch.config_lock);
261 
262 	/*
263 	 * Matching the tradition established with the timers, provide
264 	 * a default PPI for the maintenance interrupt. It makes
265 	 * things easier to reason about.
266 	 */
267 	if (vcpu->kvm->arch.vgic.mi_intid == 0)
268 		vcpu->kvm->arch.vgic.mi_intid = DEFAULT_MI_INTID;
269 	ret = kvm_vgic_set_owner(vcpu, vcpu->kvm->arch.vgic.mi_intid, vcpu);
270 
271 	return ret;
272 }
273 
vgic_allocate_private_irq(struct kvm_vcpu * vcpu,int i,u32 type)274 static void vgic_allocate_private_irq(struct kvm_vcpu *vcpu, int i, u32 type)
275 {
276 	struct vgic_irq *irq = &vcpu->arch.vgic_cpu.private_irqs[i];
277 
278 	INIT_LIST_HEAD(&irq->ap_list);
279 	raw_spin_lock_init(&irq->irq_lock);
280 	irq->vcpu = NULL;
281 	irq->target_vcpu = vcpu;
282 	refcount_set(&irq->refcount, 0);
283 
284 	irq->intid = i;
285 	if (vgic_irq_is_sgi(i)) {
286 		/* SGIs */
287 		irq->enabled = 1;
288 		irq->config = VGIC_CONFIG_EDGE;
289 	} else {
290 		/* PPIs */
291 		irq->config = VGIC_CONFIG_LEVEL;
292 	}
293 
294 	switch (type) {
295 	case KVM_DEV_TYPE_ARM_VGIC_V3:
296 		irq->group = 1;
297 		irq->mpidr = kvm_vcpu_get_mpidr_aff(vcpu);
298 		break;
299 	case KVM_DEV_TYPE_ARM_VGIC_V2:
300 		irq->group = 0;
301 		irq->targets = BIT(vcpu->vcpu_id);
302 		break;
303 	}
304 }
305 
vgic_v5_allocate_private_irq(struct kvm_vcpu * vcpu,int i,u32 type)306 static void vgic_v5_allocate_private_irq(struct kvm_vcpu *vcpu, int i, u32 type)
307 {
308 	struct vgic_irq *irq = &vcpu->arch.vgic_cpu.private_irqs[i];
309 	u32 intid = vgic_v5_make_ppi(i);
310 
311 	INIT_LIST_HEAD(&irq->ap_list);
312 	raw_spin_lock_init(&irq->irq_lock);
313 	irq->vcpu = NULL;
314 	irq->target_vcpu = vcpu;
315 	refcount_set(&irq->refcount, 0);
316 
317 	irq->intid = intid;
318 
319 	/* The only Edge architected PPI is the SW_PPI */
320 	if (i == GICV5_ARCH_PPI_SW_PPI)
321 		irq->config = VGIC_CONFIG_EDGE;
322 	else
323 		irq->config = VGIC_CONFIG_LEVEL;
324 
325 	/* Register the GICv5-specific PPI ops */
326 	vgic_v5_set_ppi_ops(vcpu, intid);
327 }
328 
vgic_allocate_private_irqs_locked(struct kvm_vcpu * vcpu,u32 type)329 static int vgic_allocate_private_irqs_locked(struct kvm_vcpu *vcpu, u32 type)
330 {
331 	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
332 	u32 num_private_irqs;
333 	int i;
334 
335 	lockdep_assert_held(&vcpu->kvm->arch.config_lock);
336 
337 	if (vgic_cpu->private_irqs)
338 		return 0;
339 
340 	if (vgic_is_v5(vcpu->kvm))
341 		num_private_irqs = VGIC_V5_NR_PRIVATE_IRQS;
342 	else
343 		num_private_irqs = VGIC_NR_PRIVATE_IRQS;
344 
345 	vgic_cpu->private_irqs = kzalloc_objs(struct vgic_irq,
346 					      num_private_irqs,
347 					      GFP_KERNEL_ACCOUNT);
348 
349 	if (!vgic_cpu->private_irqs)
350 		return -ENOMEM;
351 
352 	/*
353 	 * Enable and configure all SGIs to be edge-triggered and
354 	 * configure all PPIs as level-triggered.
355 	 */
356 	for (i = 0; i < num_private_irqs; i++) {
357 		if (vgic_is_v5(vcpu->kvm))
358 			vgic_v5_allocate_private_irq(vcpu, i, type);
359 		else
360 			vgic_allocate_private_irq(vcpu, i, type);
361 	}
362 
363 	return 0;
364 }
365 
vgic_allocate_private_irqs(struct kvm_vcpu * vcpu,u32 type)366 static int vgic_allocate_private_irqs(struct kvm_vcpu *vcpu, u32 type)
367 {
368 	int ret;
369 
370 	mutex_lock(&vcpu->kvm->arch.config_lock);
371 	ret = vgic_allocate_private_irqs_locked(vcpu, type);
372 	mutex_unlock(&vcpu->kvm->arch.config_lock);
373 
374 	return ret;
375 }
376 
377 /**
378  * kvm_vgic_vcpu_init() - Initialize static VGIC VCPU data
379  * structures and register VCPU-specific KVM iodevs
380  *
381  * @vcpu: pointer to the VCPU being created and initialized
382  *
383  * Only do initialization, but do not actually enable the
384  * VGIC CPU interface
385  */
kvm_vgic_vcpu_init(struct kvm_vcpu * vcpu)386 int kvm_vgic_vcpu_init(struct kvm_vcpu *vcpu)
387 {
388 	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
389 	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
390 	int ret = 0;
391 
392 	vgic_cpu->rd_iodev.base_addr = VGIC_ADDR_UNDEF;
393 
394 	INIT_LIST_HEAD(&vgic_cpu->ap_list_head);
395 	raw_spin_lock_init(&vgic_cpu->ap_list_lock);
396 	atomic_set(&vgic_cpu->vgic_v3.its_vpe.vlpi_count, 0);
397 
398 	if (!irqchip_in_kernel(vcpu->kvm))
399 		return 0;
400 
401 	ret = vgic_allocate_private_irqs(vcpu, dist->vgic_model);
402 	if (ret)
403 		return ret;
404 
405 	/*
406 	 * If we are creating a VCPU with a GICv3 we must also register the
407 	 * KVM io device for the redistributor that belongs to this VCPU.
408 	 */
409 	if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) {
410 		mutex_lock(&vcpu->kvm->slots_lock);
411 		ret = vgic_register_redist_iodev(vcpu);
412 		mutex_unlock(&vcpu->kvm->slots_lock);
413 	}
414 	return ret;
415 }
416 
kvm_vgic_vcpu_reset(struct kvm_vcpu * vcpu)417 static void kvm_vgic_vcpu_reset(struct kvm_vcpu *vcpu)
418 {
419 	const struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
420 
421 	if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V5)
422 		vgic_v5_reset(vcpu);
423 	else if (kvm_vgic_global_state.type == VGIC_V2)
424 		vgic_v2_reset(vcpu);
425 	else
426 		vgic_v3_reset(vcpu);
427 }
428 
429 /*
430  * vgic_init: allocates and initializes dist and vcpu data structures
431  * depending on two dimensioning parameters:
432  * - the number of spis
433  * - the number of vcpus
434  * The function is generally called when nr_spis has been explicitly set
435  * by the guest through the KVM DEVICE API. If not nr_spis is set to 256.
436  * vgic_initialized() returns true when this function has succeeded.
437  */
vgic_init(struct kvm * kvm)438 int vgic_init(struct kvm *kvm)
439 {
440 	struct vgic_dist *dist = &kvm->arch.vgic;
441 	struct kvm_vcpu *vcpu;
442 	int ret = 0;
443 	unsigned long idx;
444 
445 	lockdep_assert_held(&kvm->arch.config_lock);
446 
447 	if (vgic_initialized(kvm))
448 		return 0;
449 
450 	/* Are we also in the middle of creating a VCPU? */
451 	if (kvm->created_vcpus != atomic_read(&kvm->online_vcpus))
452 		return -EBUSY;
453 
454 	if (!vgic_is_v5(kvm)) {
455 		/* freeze the number of spis */
456 		if (!dist->nr_spis)
457 			dist->nr_spis = VGIC_NR_IRQS_LEGACY - VGIC_NR_PRIVATE_IRQS;
458 
459 		ret = kvm_vgic_dist_init(kvm, dist->nr_spis);
460 		if (ret)
461 			return ret;
462 
463 		/*
464 		 * Ensure vPEs are allocated if direct IRQ injection (e.g. vSGIs,
465 		 * vLPIs) is supported.
466 		 */
467 		if (vgic_supports_direct_irqs(kvm)) {
468 			ret = vgic_v4_init(kvm);
469 			if (ret)
470 				return ret;
471 		}
472 	} else {
473 		ret = vgic_v5_init(kvm);
474 		if (ret)
475 			return ret;
476 	}
477 
478 	kvm_for_each_vcpu(idx, vcpu, kvm)
479 		kvm_vgic_vcpu_reset(vcpu);
480 
481 	ret = kvm_vgic_setup_default_irq_routing(kvm);
482 	if (ret)
483 		return ret;
484 
485 	vgic_debug_init(kvm);
486 	dist->initialized = true;
487 
488 	return 0;
489 }
490 
kvm_vgic_dist_destroy(struct kvm * kvm)491 static void kvm_vgic_dist_destroy(struct kvm *kvm)
492 {
493 	struct vgic_dist *dist = &kvm->arch.vgic;
494 	struct vgic_redist_region *rdreg, *next;
495 
496 	dist->ready = false;
497 	dist->initialized = false;
498 
499 	kfree(dist->spis);
500 	dist->spis = NULL;
501 	dist->nr_spis = 0;
502 	dist->vgic_dist_base = VGIC_ADDR_UNDEF;
503 
504 	if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) {
505 		list_for_each_entry_safe(rdreg, next, &dist->rd_regions, list)
506 			vgic_v3_free_redist_region(kvm, rdreg);
507 		INIT_LIST_HEAD(&dist->rd_regions);
508 	} else {
509 		dist->vgic_cpu_base = VGIC_ADDR_UNDEF;
510 	}
511 
512 	if (vgic_supports_direct_irqs(kvm))
513 		vgic_v4_teardown(kvm);
514 
515 	xa_destroy(&dist->lpi_xa);
516 }
517 
__kvm_vgic_vcpu_destroy(struct kvm_vcpu * vcpu)518 static void __kvm_vgic_vcpu_destroy(struct kvm_vcpu *vcpu)
519 {
520 	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
521 
522 	/*
523 	 * Retire all pending LPIs on this vcpu anyway as we're
524 	 * going to destroy it.
525 	 */
526 	vgic_flush_pending_lpis(vcpu);
527 
528 	INIT_LIST_HEAD(&vgic_cpu->ap_list_head);
529 	kfree(vgic_cpu->private_irqs);
530 	vgic_cpu->private_irqs = NULL;
531 
532 	if (vcpu->kvm->arch.vgic.vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) {
533 		/*
534 		 * If this vCPU is being destroyed because of a failed creation
535 		 * then unregister the redistributor to avoid leaving behind a
536 		 * dangling pointer to the vCPU struct.
537 		 *
538 		 * vCPUs that have been successfully created (i.e. added to
539 		 * kvm->vcpu_array) get unregistered in kvm_vgic_destroy(), as
540 		 * this function gets called while holding kvm->arch.config_lock
541 		 * in the VM teardown path and would otherwise introduce a lock
542 		 * inversion w.r.t. kvm->srcu.
543 		 *
544 		 * vCPUs that failed creation are torn down outside of the
545 		 * kvm->arch.config_lock and do not get unregistered in
546 		 * kvm_vgic_destroy(), meaning it is both safe and necessary to
547 		 * do so here.
548 		 */
549 		if (kvm_get_vcpu_by_id(vcpu->kvm, vcpu->vcpu_id) != vcpu)
550 			vgic_unregister_redist_iodev(vcpu);
551 
552 		vgic_cpu->rd_iodev.base_addr = VGIC_ADDR_UNDEF;
553 	}
554 }
555 
kvm_vgic_vcpu_destroy(struct kvm_vcpu * vcpu)556 void kvm_vgic_vcpu_destroy(struct kvm_vcpu *vcpu)
557 {
558 	struct kvm *kvm = vcpu->kvm;
559 
560 	mutex_lock(&kvm->slots_lock);
561 	__kvm_vgic_vcpu_destroy(vcpu);
562 	mutex_unlock(&kvm->slots_lock);
563 }
564 
kvm_vgic_destroy(struct kvm * kvm)565 void kvm_vgic_destroy(struct kvm *kvm)
566 {
567 	struct kvm_vcpu *vcpu;
568 	unsigned long i;
569 
570 	mutex_lock(&kvm->slots_lock);
571 	mutex_lock(&kvm->arch.config_lock);
572 
573 	vgic_debug_destroy(kvm);
574 
575 	kvm_for_each_vcpu(i, vcpu, kvm)
576 		__kvm_vgic_vcpu_destroy(vcpu);
577 
578 	kvm_vgic_dist_destroy(kvm);
579 
580 	mutex_unlock(&kvm->arch.config_lock);
581 
582 	if (kvm->arch.vgic.vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3)
583 		kvm_for_each_vcpu(i, vcpu, kvm)
584 			vgic_unregister_redist_iodev(vcpu);
585 
586 	mutex_unlock(&kvm->slots_lock);
587 }
588 
589 /**
590  * vgic_lazy_init: Lazy init is only allowed if the GIC exposed to the guest
591  * is a GICv2. A GICv3 must be explicitly initialized by userspace using the
592  * KVM_DEV_ARM_VGIC_GRP_CTRL KVM_DEVICE group.
593  * @kvm: kvm struct pointer
594  */
vgic_lazy_init(struct kvm * kvm)595 int vgic_lazy_init(struct kvm *kvm)
596 {
597 	int ret = 0;
598 
599 	if (unlikely(!vgic_initialized(kvm))) {
600 		/*
601 		 * We only provide the automatic initialization of the VGIC
602 		 * for the legacy case of a GICv2. Any other type must
603 		 * be explicitly initialized once setup with the respective
604 		 * KVM device call.
605 		 */
606 		if (kvm->arch.vgic.vgic_model != KVM_DEV_TYPE_ARM_VGIC_V2)
607 			return -EBUSY;
608 
609 		mutex_lock(&kvm->arch.config_lock);
610 		ret = vgic_init(kvm);
611 		mutex_unlock(&kvm->arch.config_lock);
612 	}
613 
614 	return ret;
615 }
616 
617 /* RESOURCE MAPPING */
618 
619 /**
620  * kvm_vgic_map_resources - map the MMIO regions
621  * @kvm: kvm struct pointer
622  *
623  * Map the MMIO regions depending on the VGIC model exposed to the guest
624  * called on the first VCPU run.
625  * Also map the virtual CPU interface into the VM.
626  * v2 calls vgic_init() if not already done.
627  * v3 and derivatives return an error if the VGIC is not initialized.
628  */
kvm_vgic_map_resources(struct kvm * kvm)629 int kvm_vgic_map_resources(struct kvm *kvm)
630 {
631 	struct vgic_dist *dist = &kvm->arch.vgic;
632 	bool needs_dist = true;
633 	enum vgic_type type;
634 	gpa_t dist_base;
635 	int ret = 0;
636 
637 	if (likely(smp_load_acquire(&dist->ready)))
638 		return 0;
639 
640 	mutex_lock(&kvm->slots_lock);
641 	mutex_lock(&kvm->arch.config_lock);
642 	if (dist->ready)
643 		goto out;
644 
645 	if (!irqchip_in_kernel(kvm))
646 		goto out;
647 
648 	if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V2) {
649 		ret = vgic_v2_map_resources(kvm);
650 		type = VGIC_V2;
651 	} else if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) {
652 		ret = vgic_v3_map_resources(kvm);
653 		type = VGIC_V3;
654 	} else {
655 		ret = vgic_v5_map_resources(kvm);
656 		type = VGIC_V5;
657 		needs_dist = false;
658 	}
659 
660 	if (ret)
661 		goto out;
662 
663 	if (needs_dist) {
664 		dist_base = dist->vgic_dist_base;
665 		mutex_unlock(&kvm->arch.config_lock);
666 
667 		ret = vgic_register_dist_iodev(kvm, dist_base, type);
668 		if (ret) {
669 			kvm_err("Unable to register VGIC dist MMIO regions\n");
670 			goto out_slots;
671 		}
672 	} else {
673 		mutex_unlock(&kvm->arch.config_lock);
674 	}
675 
676 	smp_store_release(&dist->ready, true);
677 	goto out_slots;
678 out:
679 	mutex_unlock(&kvm->arch.config_lock);
680 out_slots:
681 	if (ret)
682 		kvm_vm_dead(kvm);
683 
684 	mutex_unlock(&kvm->slots_lock);
685 
686 	return ret;
687 }
688 
kvm_vgic_finalize_idregs(struct kvm * kvm)689 void kvm_vgic_finalize_idregs(struct kvm *kvm)
690 {
691 	u32 type = kvm->arch.vgic.vgic_model;
692 	u64 aa64pfr0, aa64pfr2, pfr1;
693 
694 	aa64pfr0 = kvm_read_vm_id_reg(kvm, SYS_ID_AA64PFR0_EL1) & ~ID_AA64PFR0_EL1_GIC;
695 	aa64pfr2 = kvm_read_vm_id_reg(kvm, SYS_ID_AA64PFR2_EL1) & ~ID_AA64PFR2_EL1_GCIE;
696 	pfr1 = kvm_read_vm_id_reg(kvm, SYS_ID_PFR1_EL1) & ~ID_PFR1_EL1_GIC;
697 
698 	switch (type) {
699 	case KVM_DEV_TYPE_ARM_VGIC_V2:
700 		break;
701 	case KVM_DEV_TYPE_ARM_VGIC_V3:
702 		aa64pfr0 |= SYS_FIELD_PREP_ENUM(ID_AA64PFR0_EL1, GIC, IMP);
703 		if (kvm_supports_32bit_el0())
704 			pfr1 |= SYS_FIELD_PREP_ENUM(ID_PFR1_EL1, GIC, GICv3);
705 		break;
706 	case KVM_DEV_TYPE_ARM_VGIC_V5:
707 		aa64pfr2 |= SYS_FIELD_PREP_ENUM(ID_AA64PFR2_EL1, GCIE, IMP);
708 		break;
709 	default:
710 		WARN_ONCE(1, "Unknown VGIC type!!!\n");
711 	}
712 
713 	kvm_set_vm_id_reg(kvm, SYS_ID_AA64PFR0_EL1, aa64pfr0);
714 	kvm_set_vm_id_reg(kvm, SYS_ID_AA64PFR2_EL1, aa64pfr2);
715 	kvm_set_vm_id_reg(kvm, SYS_ID_PFR1_EL1, pfr1);
716 }
717 
718 /* GENERIC PROBE */
719 
kvm_vgic_cpu_up(void)720 void kvm_vgic_cpu_up(void)
721 {
722 	enable_percpu_irq(kvm_vgic_global_state.maint_irq, 0);
723 }
724 
725 
kvm_vgic_cpu_down(void)726 void kvm_vgic_cpu_down(void)
727 {
728 	disable_percpu_irq(kvm_vgic_global_state.maint_irq);
729 }
730 
vgic_maintenance_handler(int irq,void * data)731 static irqreturn_t vgic_maintenance_handler(int irq, void *data)
732 {
733 	struct kvm_vcpu *vcpu = *(struct kvm_vcpu **)data;
734 
735 	/*
736 	 * We cannot rely on the vgic maintenance interrupt to be
737 	 * delivered synchronously. This means we can only use it to
738 	 * exit the VM, and we perform the handling of EOIed
739 	 * interrupts on the exit path (see vgic_fold_lr_state).
740 	 *
741 	 * Of course, NV throws a wrench in this plan, and needs
742 	 * something special.
743 	 */
744 	if (vcpu && vgic_state_is_nested(vcpu))
745 		vgic_v3_handle_nested_maint_irq(vcpu);
746 
747 	return IRQ_HANDLED;
748 }
749 
750 static struct gic_kvm_info *gic_kvm_info;
751 
vgic_set_kvm_info(const struct gic_kvm_info * info)752 void __init vgic_set_kvm_info(const struct gic_kvm_info *info)
753 {
754 	BUG_ON(gic_kvm_info != NULL);
755 	gic_kvm_info = kmalloc_obj(*gic_kvm_info);
756 	if (gic_kvm_info)
757 		*gic_kvm_info = *info;
758 }
759 
760 /**
761  * kvm_vgic_init_cpu_hardware - initialize the GIC VE hardware
762  *
763  * For a specific CPU, initialize the GIC VE hardware.
764  */
kvm_vgic_init_cpu_hardware(void)765 void kvm_vgic_init_cpu_hardware(void)
766 {
767 	BUG_ON(preemptible());
768 
769 	/*
770 	 * We want to make sure the list registers start out clear so that we
771 	 * only have the program the used registers.
772 	 */
773 	if (kvm_vgic_global_state.type == VGIC_V2) {
774 		vgic_v2_init_lrs();
775 	} else if (kvm_vgic_global_state.type == VGIC_V3 ||
776 		   kvm_vgic_global_state.has_gcie_v3_compat) {
777 		kvm_call_hyp(__vgic_v3_init_lrs);
778 	}
779 }
780 
781 /**
782  * kvm_vgic_hyp_init: populates the kvm_vgic_global_state variable
783  * according to the host GIC model. Accordingly calls either
784  * vgic_v2/v3_probe which registers the KVM_DEVICE that can be
785  * instantiated by a guest later on .
786  */
kvm_vgic_hyp_init(void)787 int kvm_vgic_hyp_init(void)
788 {
789 	bool has_mask;
790 	int ret;
791 
792 	if (!gic_kvm_info)
793 		return -ENODEV;
794 
795 	has_mask = !gic_kvm_info->no_maint_irq_mask;
796 
797 	if (has_mask && !gic_kvm_info->maint_irq) {
798 		kvm_err("No vgic maintenance irq\n");
799 		return -ENXIO;
800 	}
801 
802 	/*
803 	 * If we get one of these oddball non-GICs, taint the kernel,
804 	 * as we have no idea of how they *really* behave.
805 	 */
806 	if (gic_kvm_info->no_hw_deactivation) {
807 		kvm_info("Non-architectural vgic, tainting kernel\n");
808 		add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
809 		kvm_vgic_global_state.no_hw_deactivation = true;
810 	}
811 
812 	switch (gic_kvm_info->type) {
813 	case GIC_V2:
814 		ret = vgic_v2_probe(gic_kvm_info);
815 		break;
816 	case GIC_V3:
817 		ret = vgic_v3_probe(gic_kvm_info);
818 		if (!ret) {
819 			static_branch_enable(&kvm_vgic_global_state.gicv3_cpuif);
820 			kvm_info("GIC system register CPU interface enabled\n");
821 		}
822 		break;
823 	case GIC_V5:
824 		ret = vgic_v5_probe(gic_kvm_info);
825 		break;
826 	default:
827 		ret = -ENODEV;
828 	}
829 
830 	kvm_vgic_global_state.maint_irq = gic_kvm_info->maint_irq;
831 
832 	kfree(gic_kvm_info);
833 	gic_kvm_info = NULL;
834 
835 	if (ret)
836 		return ret;
837 
838 	if (!has_mask && !kvm_vgic_global_state.maint_irq)
839 		return 0;
840 
841 	ret = request_percpu_irq(kvm_vgic_global_state.maint_irq,
842 				 vgic_maintenance_handler,
843 				 "vgic", kvm_get_running_vcpus());
844 	if (ret) {
845 		kvm_err("Cannot register interrupt %d\n",
846 			kvm_vgic_global_state.maint_irq);
847 		return ret;
848 	}
849 
850 	kvm_info("vgic interrupt IRQ%d\n", kvm_vgic_global_state.maint_irq);
851 	return 0;
852 }
853