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 */
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 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 */
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 /* 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
127 out_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 */
kvm_vgic_dist_init(struct kvm * kvm,unsigned int nr_spis)140 static 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 */
kvm_vgic_vcpu_init(struct kvm_vcpu * vcpu)194 int 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
kvm_vgic_vcpu_enable(struct kvm_vcpu * vcpu)245 static 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 */
vgic_init(struct kvm * kvm)262 int 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, unconditionnaly 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
339 out:
340 return ret;
341 }
342
kvm_vgic_dist_destroy(struct kvm * kvm)343 static 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
__kvm_vgic_vcpu_destroy(struct kvm_vcpu * vcpu)371 static void __kvm_vgic_vcpu_destroy(struct kvm_vcpu *vcpu)
372 {
373 struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
374
375 /*
376 * Retire all pending LPIs on this vcpu anyway as we're
377 * going to destroy it.
378 */
379 vgic_flush_pending_lpis(vcpu);
380
381 INIT_LIST_HEAD(&vgic_cpu->ap_list_head);
382 if (vcpu->kvm->arch.vgic.vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) {
383 vgic_unregister_redist_iodev(vcpu);
384 vgic_cpu->rd_iodev.base_addr = VGIC_ADDR_UNDEF;
385 }
386 }
387
kvm_vgic_vcpu_destroy(struct kvm_vcpu * vcpu)388 void kvm_vgic_vcpu_destroy(struct kvm_vcpu *vcpu)
389 {
390 struct kvm *kvm = vcpu->kvm;
391
392 mutex_lock(&kvm->slots_lock);
393 __kvm_vgic_vcpu_destroy(vcpu);
394 mutex_unlock(&kvm->slots_lock);
395 }
396
kvm_vgic_destroy(struct kvm * kvm)397 void kvm_vgic_destroy(struct kvm *kvm)
398 {
399 struct kvm_vcpu *vcpu;
400 unsigned long i;
401
402 mutex_lock(&kvm->slots_lock);
403
404 vgic_debug_destroy(kvm);
405
406 kvm_for_each_vcpu(i, vcpu, kvm)
407 __kvm_vgic_vcpu_destroy(vcpu);
408
409 mutex_lock(&kvm->arch.config_lock);
410
411 kvm_vgic_dist_destroy(kvm);
412
413 mutex_unlock(&kvm->arch.config_lock);
414 mutex_unlock(&kvm->slots_lock);
415 }
416
417 /**
418 * vgic_lazy_init: Lazy init is only allowed if the GIC exposed to the guest
419 * is a GICv2. A GICv3 must be explicitly initialized by userspace using the
420 * KVM_DEV_ARM_VGIC_GRP_CTRL KVM_DEVICE group.
421 * @kvm: kvm struct pointer
422 */
vgic_lazy_init(struct kvm * kvm)423 int vgic_lazy_init(struct kvm *kvm)
424 {
425 int ret = 0;
426
427 if (unlikely(!vgic_initialized(kvm))) {
428 /*
429 * We only provide the automatic initialization of the VGIC
430 * for the legacy case of a GICv2. Any other type must
431 * be explicitly initialized once setup with the respective
432 * KVM device call.
433 */
434 if (kvm->arch.vgic.vgic_model != KVM_DEV_TYPE_ARM_VGIC_V2)
435 return -EBUSY;
436
437 mutex_lock(&kvm->arch.config_lock);
438 ret = vgic_init(kvm);
439 mutex_unlock(&kvm->arch.config_lock);
440 }
441
442 return ret;
443 }
444
445 /* RESOURCE MAPPING */
446
447 /**
448 * Map the MMIO regions depending on the VGIC model exposed to the guest
449 * called on the first VCPU run.
450 * Also map the virtual CPU interface into the VM.
451 * v2 calls vgic_init() if not already done.
452 * v3 and derivatives return an error if the VGIC is not initialized.
453 * vgic_ready() returns true if this function has succeeded.
454 * @kvm: kvm struct pointer
455 */
kvm_vgic_map_resources(struct kvm * kvm)456 int kvm_vgic_map_resources(struct kvm *kvm)
457 {
458 struct vgic_dist *dist = &kvm->arch.vgic;
459 enum vgic_type type;
460 gpa_t dist_base;
461 int ret = 0;
462
463 if (likely(vgic_ready(kvm)))
464 return 0;
465
466 mutex_lock(&kvm->slots_lock);
467 mutex_lock(&kvm->arch.config_lock);
468 if (vgic_ready(kvm))
469 goto out;
470
471 if (!irqchip_in_kernel(kvm))
472 goto out;
473
474 if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V2) {
475 ret = vgic_v2_map_resources(kvm);
476 type = VGIC_V2;
477 } else {
478 ret = vgic_v3_map_resources(kvm);
479 type = VGIC_V3;
480 }
481
482 if (ret)
483 goto out;
484
485 dist->ready = true;
486 dist_base = dist->vgic_dist_base;
487 mutex_unlock(&kvm->arch.config_lock);
488
489 ret = vgic_register_dist_iodev(kvm, dist_base, type);
490 if (ret)
491 kvm_err("Unable to register VGIC dist MMIO regions\n");
492
493 goto out_slots;
494 out:
495 mutex_unlock(&kvm->arch.config_lock);
496 out_slots:
497 mutex_unlock(&kvm->slots_lock);
498
499 if (ret)
500 kvm_vgic_destroy(kvm);
501
502 return ret;
503 }
504
505 /* GENERIC PROBE */
506
kvm_vgic_cpu_up(void)507 void kvm_vgic_cpu_up(void)
508 {
509 enable_percpu_irq(kvm_vgic_global_state.maint_irq, 0);
510 }
511
512
kvm_vgic_cpu_down(void)513 void kvm_vgic_cpu_down(void)
514 {
515 disable_percpu_irq(kvm_vgic_global_state.maint_irq);
516 }
517
vgic_maintenance_handler(int irq,void * data)518 static irqreturn_t vgic_maintenance_handler(int irq, void *data)
519 {
520 /*
521 * We cannot rely on the vgic maintenance interrupt to be
522 * delivered synchronously. This means we can only use it to
523 * exit the VM, and we perform the handling of EOIed
524 * interrupts on the exit path (see vgic_fold_lr_state).
525 */
526 return IRQ_HANDLED;
527 }
528
529 static struct gic_kvm_info *gic_kvm_info;
530
vgic_set_kvm_info(const struct gic_kvm_info * info)531 void __init vgic_set_kvm_info(const struct gic_kvm_info *info)
532 {
533 BUG_ON(gic_kvm_info != NULL);
534 gic_kvm_info = kmalloc(sizeof(*info), GFP_KERNEL);
535 if (gic_kvm_info)
536 *gic_kvm_info = *info;
537 }
538
539 /**
540 * kvm_vgic_init_cpu_hardware - initialize the GIC VE hardware
541 *
542 * For a specific CPU, initialize the GIC VE hardware.
543 */
kvm_vgic_init_cpu_hardware(void)544 void kvm_vgic_init_cpu_hardware(void)
545 {
546 BUG_ON(preemptible());
547
548 /*
549 * We want to make sure the list registers start out clear so that we
550 * only have the program the used registers.
551 */
552 if (kvm_vgic_global_state.type == VGIC_V2)
553 vgic_v2_init_lrs();
554 else
555 kvm_call_hyp(__vgic_v3_init_lrs);
556 }
557
558 /**
559 * kvm_vgic_hyp_init: populates the kvm_vgic_global_state variable
560 * according to the host GIC model. Accordingly calls either
561 * vgic_v2/v3_probe which registers the KVM_DEVICE that can be
562 * instantiated by a guest later on .
563 */
kvm_vgic_hyp_init(void)564 int kvm_vgic_hyp_init(void)
565 {
566 bool has_mask;
567 int ret;
568
569 if (!gic_kvm_info)
570 return -ENODEV;
571
572 has_mask = !gic_kvm_info->no_maint_irq_mask;
573
574 if (has_mask && !gic_kvm_info->maint_irq) {
575 kvm_err("No vgic maintenance irq\n");
576 return -ENXIO;
577 }
578
579 /*
580 * If we get one of these oddball non-GICs, taint the kernel,
581 * as we have no idea of how they *really* behave.
582 */
583 if (gic_kvm_info->no_hw_deactivation) {
584 kvm_info("Non-architectural vgic, tainting kernel\n");
585 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
586 kvm_vgic_global_state.no_hw_deactivation = true;
587 }
588
589 switch (gic_kvm_info->type) {
590 case GIC_V2:
591 ret = vgic_v2_probe(gic_kvm_info);
592 break;
593 case GIC_V3:
594 ret = vgic_v3_probe(gic_kvm_info);
595 if (!ret) {
596 static_branch_enable(&kvm_vgic_global_state.gicv3_cpuif);
597 kvm_info("GIC system register CPU interface enabled\n");
598 }
599 break;
600 default:
601 ret = -ENODEV;
602 }
603
604 kvm_vgic_global_state.maint_irq = gic_kvm_info->maint_irq;
605
606 kfree(gic_kvm_info);
607 gic_kvm_info = NULL;
608
609 if (ret)
610 return ret;
611
612 if (!has_mask && !kvm_vgic_global_state.maint_irq)
613 return 0;
614
615 ret = request_percpu_irq(kvm_vgic_global_state.maint_irq,
616 vgic_maintenance_handler,
617 "vgic", kvm_get_running_vcpus());
618 if (ret) {
619 kvm_err("Cannot register interrupt %d\n",
620 kvm_vgic_global_state.maint_irq);
621 return ret;
622 }
623
624 kvm_info("vgic interrupt IRQ%d\n", kvm_vgic_global_state.maint_irq);
625 return 0;
626 }
627