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 xa_init_flags(&dist->lpi_xa, XA_FLAGS_LOCK_IRQ);
57 }
58
59 /* CREATION */
60
61 /**
62 * kvm_vgic_create: triggered by the instantiation of the VGIC device by
63 * user space, either through the legacy KVM_CREATE_IRQCHIP ioctl (v2 only)
64 * or through the generic KVM_CREATE_DEVICE API ioctl.
65 * irqchip_in_kernel() tells you if this function succeeded or not.
66 * @kvm: kvm struct pointer
67 * @type: KVM_DEV_TYPE_ARM_VGIC_V[23]
68 */
69 int kvm_vgic_create(struct kvm *kvm, u32 type)
70 {
71 struct kvm_vcpu *vcpu;
72 unsigned long i;
73 int ret;
74
75 /*
76 * This function is also called by the KVM_CREATE_IRQCHIP handler,
77 * which had no chance yet to check the availability of the GICv2
78 * emulation. So check this here again. KVM_CREATE_DEVICE does
79 * the proper checks already.
80 */
81 if (type == KVM_DEV_TYPE_ARM_VGIC_V2 &&
82 !kvm_vgic_global_state.can_emulate_gicv2)
83 return -ENODEV;
84
85 /* Must be held to avoid race with vCPU creation */
86 lockdep_assert_held(&kvm->lock);
87
88 ret = -EBUSY;
89 if (!lock_all_vcpus(kvm))
90 return ret;
91
92 mutex_lock(&kvm->arch.config_lock);
93
94 if (irqchip_in_kernel(kvm)) {
95 ret = -EEXIST;
96 goto out_unlock;
97 }
98
99 kvm_for_each_vcpu(i, vcpu, kvm) {
100 if (vcpu_has_run_once(vcpu))
101 goto out_unlock;
102 }
103 ret = 0;
104
105 if (type == KVM_DEV_TYPE_ARM_VGIC_V2)
106 kvm->max_vcpus = VGIC_V2_MAX_CPUS;
107 else
108 kvm->max_vcpus = VGIC_V3_MAX_CPUS;
109
110 if (atomic_read(&kvm->online_vcpus) > kvm->max_vcpus) {
111 ret = -E2BIG;
112 goto out_unlock;
113 }
114
115 kvm->arch.vgic.in_kernel = true;
116 kvm->arch.vgic.vgic_model = type;
117
118 kvm->arch.vgic.vgic_dist_base = VGIC_ADDR_UNDEF;
119
120 if (type == KVM_DEV_TYPE_ARM_VGIC_V2)
121 kvm->arch.vgic.vgic_cpu_base = VGIC_ADDR_UNDEF;
122 else
123 INIT_LIST_HEAD(&kvm->arch.vgic.rd_regions);
124
125 out_unlock:
126 mutex_unlock(&kvm->arch.config_lock);
127 unlock_all_vcpus(kvm);
128 return ret;
129 }
130
131 /* INIT/DESTROY */
132
133 /**
134 * kvm_vgic_dist_init: initialize the dist data structures
135 * @kvm: kvm struct pointer
136 * @nr_spis: number of spis, frozen by caller
137 */
138 static int kvm_vgic_dist_init(struct kvm *kvm, unsigned int nr_spis)
139 {
140 struct vgic_dist *dist = &kvm->arch.vgic;
141 struct kvm_vcpu *vcpu0 = kvm_get_vcpu(kvm, 0);
142 int i;
143
144 dist->spis = kcalloc(nr_spis, sizeof(struct vgic_irq), GFP_KERNEL_ACCOUNT);
145 if (!dist->spis)
146 return -ENOMEM;
147
148 /*
149 * In the following code we do not take the irq struct lock since
150 * no other action on irq structs can happen while the VGIC is
151 * not initialized yet:
152 * If someone wants to inject an interrupt or does a MMIO access, we
153 * require prior initialization in case of a virtual GICv3 or trigger
154 * initialization when using a virtual GICv2.
155 */
156 for (i = 0; i < nr_spis; i++) {
157 struct vgic_irq *irq = &dist->spis[i];
158
159 irq->intid = i + VGIC_NR_PRIVATE_IRQS;
160 INIT_LIST_HEAD(&irq->ap_list);
161 raw_spin_lock_init(&irq->irq_lock);
162 irq->vcpu = NULL;
163 irq->target_vcpu = vcpu0;
164 kref_init(&irq->refcount);
165 switch (dist->vgic_model) {
166 case KVM_DEV_TYPE_ARM_VGIC_V2:
167 irq->targets = 0;
168 irq->group = 0;
169 break;
170 case KVM_DEV_TYPE_ARM_VGIC_V3:
171 irq->mpidr = 0;
172 irq->group = 1;
173 break;
174 default:
175 kfree(dist->spis);
176 dist->spis = NULL;
177 return -EINVAL;
178 }
179 }
180 return 0;
181 }
182
183 static int vgic_allocate_private_irqs_locked(struct kvm_vcpu *vcpu)
184 {
185 struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
186 int i;
187
188 lockdep_assert_held(&vcpu->kvm->arch.config_lock);
189
190 if (vgic_cpu->private_irqs)
191 return 0;
192
193 vgic_cpu->private_irqs = kcalloc(VGIC_NR_PRIVATE_IRQS,
194 sizeof(struct vgic_irq),
195 GFP_KERNEL_ACCOUNT);
196
197 if (!vgic_cpu->private_irqs)
198 return -ENOMEM;
199
200 /*
201 * Enable and configure all SGIs to be edge-triggered and
202 * configure all PPIs as level-triggered.
203 */
204 for (i = 0; i < VGIC_NR_PRIVATE_IRQS; i++) {
205 struct vgic_irq *irq = &vgic_cpu->private_irqs[i];
206
207 INIT_LIST_HEAD(&irq->ap_list);
208 raw_spin_lock_init(&irq->irq_lock);
209 irq->intid = i;
210 irq->vcpu = NULL;
211 irq->target_vcpu = vcpu;
212 kref_init(&irq->refcount);
213 if (vgic_irq_is_sgi(i)) {
214 /* SGIs */
215 irq->enabled = 1;
216 irq->config = VGIC_CONFIG_EDGE;
217 } else {
218 /* PPIs */
219 irq->config = VGIC_CONFIG_LEVEL;
220 }
221 }
222
223 return 0;
224 }
225
226 static int vgic_allocate_private_irqs(struct kvm_vcpu *vcpu)
227 {
228 int ret;
229
230 mutex_lock(&vcpu->kvm->arch.config_lock);
231 ret = vgic_allocate_private_irqs_locked(vcpu);
232 mutex_unlock(&vcpu->kvm->arch.config_lock);
233
234 return ret;
235 }
236
237 /**
238 * kvm_vgic_vcpu_init() - Initialize static VGIC VCPU data
239 * structures and register VCPU-specific KVM iodevs
240 *
241 * @vcpu: pointer to the VCPU being created and initialized
242 *
243 * Only do initialization, but do not actually enable the
244 * VGIC CPU interface
245 */
246 int kvm_vgic_vcpu_init(struct kvm_vcpu *vcpu)
247 {
248 struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
249 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
250 int ret = 0;
251
252 vgic_cpu->rd_iodev.base_addr = VGIC_ADDR_UNDEF;
253
254 INIT_LIST_HEAD(&vgic_cpu->ap_list_head);
255 raw_spin_lock_init(&vgic_cpu->ap_list_lock);
256 atomic_set(&vgic_cpu->vgic_v3.its_vpe.vlpi_count, 0);
257
258 if (!irqchip_in_kernel(vcpu->kvm))
259 return 0;
260
261 ret = vgic_allocate_private_irqs(vcpu);
262 if (ret)
263 return ret;
264
265 /*
266 * If we are creating a VCPU with a GICv3 we must also register the
267 * KVM io device for the redistributor that belongs to this VCPU.
268 */
269 if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) {
270 mutex_lock(&vcpu->kvm->slots_lock);
271 ret = vgic_register_redist_iodev(vcpu);
272 mutex_unlock(&vcpu->kvm->slots_lock);
273 }
274 return ret;
275 }
276
277 static void kvm_vgic_vcpu_enable(struct kvm_vcpu *vcpu)
278 {
279 if (kvm_vgic_global_state.type == VGIC_V2)
280 vgic_v2_enable(vcpu);
281 else
282 vgic_v3_enable(vcpu);
283 }
284
285 /*
286 * vgic_init: allocates and initializes dist and vcpu data structures
287 * depending on two dimensioning parameters:
288 * - the number of spis
289 * - the number of vcpus
290 * The function is generally called when nr_spis has been explicitly set
291 * by the guest through the KVM DEVICE API. If not nr_spis is set to 256.
292 * vgic_initialized() returns true when this function has succeeded.
293 */
294 int vgic_init(struct kvm *kvm)
295 {
296 struct vgic_dist *dist = &kvm->arch.vgic;
297 struct kvm_vcpu *vcpu;
298 int ret = 0, i;
299 unsigned long idx;
300
301 lockdep_assert_held(&kvm->arch.config_lock);
302
303 if (vgic_initialized(kvm))
304 return 0;
305
306 /* Are we also in the middle of creating a VCPU? */
307 if (kvm->created_vcpus != atomic_read(&kvm->online_vcpus))
308 return -EBUSY;
309
310 /* freeze the number of spis */
311 if (!dist->nr_spis)
312 dist->nr_spis = VGIC_NR_IRQS_LEGACY - VGIC_NR_PRIVATE_IRQS;
313
314 ret = kvm_vgic_dist_init(kvm, dist->nr_spis);
315 if (ret)
316 goto out;
317
318 /* Initialize groups on CPUs created before the VGIC type was known */
319 kvm_for_each_vcpu(idx, vcpu, kvm) {
320 ret = vgic_allocate_private_irqs_locked(vcpu);
321 if (ret)
322 goto out;
323
324 for (i = 0; i < VGIC_NR_PRIVATE_IRQS; i++) {
325 struct vgic_irq *irq = vgic_get_irq(kvm, vcpu, i);
326
327 switch (dist->vgic_model) {
328 case KVM_DEV_TYPE_ARM_VGIC_V3:
329 irq->group = 1;
330 irq->mpidr = kvm_vcpu_get_mpidr_aff(vcpu);
331 break;
332 case KVM_DEV_TYPE_ARM_VGIC_V2:
333 irq->group = 0;
334 irq->targets = 1U << idx;
335 break;
336 default:
337 ret = -EINVAL;
338 }
339
340 vgic_put_irq(kvm, irq);
341
342 if (ret)
343 goto out;
344 }
345 }
346
347 /*
348 * If we have GICv4.1 enabled, unconditionally request enable the
349 * v4 support so that we get HW-accelerated vSGIs. Otherwise, only
350 * enable it if we present a virtual ITS to the guest.
351 */
352 if (vgic_supports_direct_msis(kvm)) {
353 ret = vgic_v4_init(kvm);
354 if (ret)
355 goto out;
356 }
357
358 kvm_for_each_vcpu(idx, vcpu, kvm)
359 kvm_vgic_vcpu_enable(vcpu);
360
361 ret = kvm_vgic_setup_default_irq_routing(kvm);
362 if (ret)
363 goto out;
364
365 vgic_debug_init(kvm);
366
367 /*
368 * If userspace didn't set the GIC implementation revision,
369 * default to the latest and greatest. You know want it.
370 */
371 if (!dist->implementation_rev)
372 dist->implementation_rev = KVM_VGIC_IMP_REV_LATEST;
373 dist->initialized = true;
374
375 out:
376 return ret;
377 }
378
379 static void kvm_vgic_dist_destroy(struct kvm *kvm)
380 {
381 struct vgic_dist *dist = &kvm->arch.vgic;
382 struct vgic_redist_region *rdreg, *next;
383
384 dist->ready = false;
385 dist->initialized = false;
386
387 kfree(dist->spis);
388 dist->spis = NULL;
389 dist->nr_spis = 0;
390 dist->vgic_dist_base = VGIC_ADDR_UNDEF;
391
392 if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) {
393 list_for_each_entry_safe(rdreg, next, &dist->rd_regions, list)
394 vgic_v3_free_redist_region(kvm, rdreg);
395 INIT_LIST_HEAD(&dist->rd_regions);
396 } else {
397 dist->vgic_cpu_base = VGIC_ADDR_UNDEF;
398 }
399
400 if (vgic_supports_direct_msis(kvm))
401 vgic_v4_teardown(kvm);
402
403 xa_destroy(&dist->lpi_xa);
404 }
405
406 static void __kvm_vgic_vcpu_destroy(struct kvm_vcpu *vcpu)
407 {
408 struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
409
410 /*
411 * Retire all pending LPIs on this vcpu anyway as we're
412 * going to destroy it.
413 */
414 vgic_flush_pending_lpis(vcpu);
415
416 INIT_LIST_HEAD(&vgic_cpu->ap_list_head);
417 kfree(vgic_cpu->private_irqs);
418 vgic_cpu->private_irqs = NULL;
419
420 if (vcpu->kvm->arch.vgic.vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3)
421 vgic_cpu->rd_iodev.base_addr = VGIC_ADDR_UNDEF;
422 }
423
424 void kvm_vgic_vcpu_destroy(struct kvm_vcpu *vcpu)
425 {
426 struct kvm *kvm = vcpu->kvm;
427
428 mutex_lock(&kvm->slots_lock);
429 __kvm_vgic_vcpu_destroy(vcpu);
430 mutex_unlock(&kvm->slots_lock);
431 }
432
433 void kvm_vgic_destroy(struct kvm *kvm)
434 {
435 struct kvm_vcpu *vcpu;
436 unsigned long i;
437
438 mutex_lock(&kvm->slots_lock);
439 mutex_lock(&kvm->arch.config_lock);
440
441 vgic_debug_destroy(kvm);
442
443 kvm_for_each_vcpu(i, vcpu, kvm)
444 __kvm_vgic_vcpu_destroy(vcpu);
445
446 kvm_vgic_dist_destroy(kvm);
447
448 mutex_unlock(&kvm->arch.config_lock);
449
450 if (kvm->arch.vgic.vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3)
451 kvm_for_each_vcpu(i, vcpu, kvm)
452 vgic_unregister_redist_iodev(vcpu);
453
454 mutex_unlock(&kvm->slots_lock);
455 }
456
457 /**
458 * vgic_lazy_init: Lazy init is only allowed if the GIC exposed to the guest
459 * is a GICv2. A GICv3 must be explicitly initialized by userspace using the
460 * KVM_DEV_ARM_VGIC_GRP_CTRL KVM_DEVICE group.
461 * @kvm: kvm struct pointer
462 */
463 int vgic_lazy_init(struct kvm *kvm)
464 {
465 int ret = 0;
466
467 if (unlikely(!vgic_initialized(kvm))) {
468 /*
469 * We only provide the automatic initialization of the VGIC
470 * for the legacy case of a GICv2. Any other type must
471 * be explicitly initialized once setup with the respective
472 * KVM device call.
473 */
474 if (kvm->arch.vgic.vgic_model != KVM_DEV_TYPE_ARM_VGIC_V2)
475 return -EBUSY;
476
477 mutex_lock(&kvm->arch.config_lock);
478 ret = vgic_init(kvm);
479 mutex_unlock(&kvm->arch.config_lock);
480 }
481
482 return ret;
483 }
484
485 /* RESOURCE MAPPING */
486
487 /**
488 * kvm_vgic_map_resources - map the MMIO regions
489 * @kvm: kvm struct pointer
490 *
491 * Map the MMIO regions depending on the VGIC model exposed to the guest
492 * called on the first VCPU run.
493 * Also map the virtual CPU interface into the VM.
494 * v2 calls vgic_init() if not already done.
495 * v3 and derivatives return an error if the VGIC is not initialized.
496 * vgic_ready() returns true if this function has succeeded.
497 */
498 int kvm_vgic_map_resources(struct kvm *kvm)
499 {
500 struct vgic_dist *dist = &kvm->arch.vgic;
501 enum vgic_type type;
502 gpa_t dist_base;
503 int ret = 0;
504
505 if (likely(vgic_ready(kvm)))
506 return 0;
507
508 mutex_lock(&kvm->slots_lock);
509 mutex_lock(&kvm->arch.config_lock);
510 if (vgic_ready(kvm))
511 goto out;
512
513 if (!irqchip_in_kernel(kvm))
514 goto out;
515
516 if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V2) {
517 ret = vgic_v2_map_resources(kvm);
518 type = VGIC_V2;
519 } else {
520 ret = vgic_v3_map_resources(kvm);
521 type = VGIC_V3;
522 }
523
524 if (ret)
525 goto out;
526
527 dist->ready = true;
528 dist_base = dist->vgic_dist_base;
529 mutex_unlock(&kvm->arch.config_lock);
530
531 ret = vgic_register_dist_iodev(kvm, dist_base, type);
532 if (ret)
533 kvm_err("Unable to register VGIC dist MMIO regions\n");
534
535 goto out_slots;
536 out:
537 mutex_unlock(&kvm->arch.config_lock);
538 out_slots:
539 mutex_unlock(&kvm->slots_lock);
540
541 if (ret)
542 kvm_vgic_destroy(kvm);
543
544 return ret;
545 }
546
547 /* GENERIC PROBE */
548
549 void kvm_vgic_cpu_up(void)
550 {
551 enable_percpu_irq(kvm_vgic_global_state.maint_irq, 0);
552 }
553
554
555 void kvm_vgic_cpu_down(void)
556 {
557 disable_percpu_irq(kvm_vgic_global_state.maint_irq);
558 }
559
560 static irqreturn_t vgic_maintenance_handler(int irq, void *data)
561 {
562 /*
563 * We cannot rely on the vgic maintenance interrupt to be
564 * delivered synchronously. This means we can only use it to
565 * exit the VM, and we perform the handling of EOIed
566 * interrupts on the exit path (see vgic_fold_lr_state).
567 */
568 return IRQ_HANDLED;
569 }
570
571 static struct gic_kvm_info *gic_kvm_info;
572
573 void __init vgic_set_kvm_info(const struct gic_kvm_info *info)
574 {
575 BUG_ON(gic_kvm_info != NULL);
576 gic_kvm_info = kmalloc(sizeof(*info), GFP_KERNEL);
577 if (gic_kvm_info)
578 *gic_kvm_info = *info;
579 }
580
581 /**
582 * kvm_vgic_init_cpu_hardware - initialize the GIC VE hardware
583 *
584 * For a specific CPU, initialize the GIC VE hardware.
585 */
586 void kvm_vgic_init_cpu_hardware(void)
587 {
588 BUG_ON(preemptible());
589
590 /*
591 * We want to make sure the list registers start out clear so that we
592 * only have the program the used registers.
593 */
594 if (kvm_vgic_global_state.type == VGIC_V2)
595 vgic_v2_init_lrs();
596 else
597 kvm_call_hyp(__vgic_v3_init_lrs);
598 }
599
600 /**
601 * kvm_vgic_hyp_init: populates the kvm_vgic_global_state variable
602 * according to the host GIC model. Accordingly calls either
603 * vgic_v2/v3_probe which registers the KVM_DEVICE that can be
604 * instantiated by a guest later on .
605 */
606 int kvm_vgic_hyp_init(void)
607 {
608 bool has_mask;
609 int ret;
610
611 if (!gic_kvm_info)
612 return -ENODEV;
613
614 has_mask = !gic_kvm_info->no_maint_irq_mask;
615
616 if (has_mask && !gic_kvm_info->maint_irq) {
617 kvm_err("No vgic maintenance irq\n");
618 return -ENXIO;
619 }
620
621 /*
622 * If we get one of these oddball non-GICs, taint the kernel,
623 * as we have no idea of how they *really* behave.
624 */
625 if (gic_kvm_info->no_hw_deactivation) {
626 kvm_info("Non-architectural vgic, tainting kernel\n");
627 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
628 kvm_vgic_global_state.no_hw_deactivation = true;
629 }
630
631 switch (gic_kvm_info->type) {
632 case GIC_V2:
633 ret = vgic_v2_probe(gic_kvm_info);
634 break;
635 case GIC_V3:
636 ret = vgic_v3_probe(gic_kvm_info);
637 if (!ret) {
638 static_branch_enable(&kvm_vgic_global_state.gicv3_cpuif);
639 kvm_info("GIC system register CPU interface enabled\n");
640 }
641 break;
642 default:
643 ret = -ENODEV;
644 }
645
646 kvm_vgic_global_state.maint_irq = gic_kvm_info->maint_irq;
647
648 kfree(gic_kvm_info);
649 gic_kvm_info = NULL;
650
651 if (ret)
652 return ret;
653
654 if (!has_mask && !kvm_vgic_global_state.maint_irq)
655 return 0;
656
657 ret = request_percpu_irq(kvm_vgic_global_state.maint_irq,
658 vgic_maintenance_handler,
659 "vgic", kvm_get_running_vcpus());
660 if (ret) {
661 kvm_err("Cannot register interrupt %d\n",
662 kvm_vgic_global_state.maint_irq);
663 return ret;
664 }
665
666 kvm_info("vgic interrupt IRQ%d\n", kvm_vgic_global_state.maint_irq);
667 return 0;
668 }
669
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.