xref: /openbmc/linux/arch/arm64/kvm/vgic/vgic-v3.c (revision d9679d00)
1 // SPDX-License-Identifier: GPL-2.0-only
2 
3 #include <linux/irqchip/arm-gic-v3.h>
4 #include <linux/irq.h>
5 #include <linux/irqdomain.h>
6 #include <linux/kvm.h>
7 #include <linux/kvm_host.h>
8 #include <kvm/arm_vgic.h>
9 #include <asm/kvm_hyp.h>
10 #include <asm/kvm_mmu.h>
11 #include <asm/kvm_asm.h>
12 
13 #include "vgic.h"
14 
15 static bool group0_trap;
16 static bool group1_trap;
17 static bool common_trap;
18 static bool dir_trap;
19 static bool gicv4_enable;
20 
21 void vgic_v3_set_underflow(struct kvm_vcpu *vcpu)
22 {
23 	struct vgic_v3_cpu_if *cpuif = &vcpu->arch.vgic_cpu.vgic_v3;
24 
25 	cpuif->vgic_hcr |= ICH_HCR_UIE;
26 }
27 
28 static bool lr_signals_eoi_mi(u64 lr_val)
29 {
30 	return !(lr_val & ICH_LR_STATE) && (lr_val & ICH_LR_EOI) &&
31 	       !(lr_val & ICH_LR_HW);
32 }
33 
34 void vgic_v3_fold_lr_state(struct kvm_vcpu *vcpu)
35 {
36 	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
37 	struct vgic_v3_cpu_if *cpuif = &vgic_cpu->vgic_v3;
38 	u32 model = vcpu->kvm->arch.vgic.vgic_model;
39 	int lr;
40 
41 	DEBUG_SPINLOCK_BUG_ON(!irqs_disabled());
42 
43 	cpuif->vgic_hcr &= ~ICH_HCR_UIE;
44 
45 	for (lr = 0; lr < cpuif->used_lrs; lr++) {
46 		u64 val = cpuif->vgic_lr[lr];
47 		u32 intid, cpuid;
48 		struct vgic_irq *irq;
49 		bool is_v2_sgi = false;
50 		bool deactivated;
51 
52 		cpuid = val & GICH_LR_PHYSID_CPUID;
53 		cpuid >>= GICH_LR_PHYSID_CPUID_SHIFT;
54 
55 		if (model == KVM_DEV_TYPE_ARM_VGIC_V3) {
56 			intid = val & ICH_LR_VIRTUAL_ID_MASK;
57 		} else {
58 			intid = val & GICH_LR_VIRTUALID;
59 			is_v2_sgi = vgic_irq_is_sgi(intid);
60 		}
61 
62 		/* Notify fds when the guest EOI'ed a level-triggered IRQ */
63 		if (lr_signals_eoi_mi(val) && vgic_valid_spi(vcpu->kvm, intid))
64 			kvm_notify_acked_irq(vcpu->kvm, 0,
65 					     intid - VGIC_NR_PRIVATE_IRQS);
66 
67 		irq = vgic_get_irq(vcpu->kvm, vcpu, intid);
68 		if (!irq)	/* An LPI could have been unmapped. */
69 			continue;
70 
71 		raw_spin_lock(&irq->irq_lock);
72 
73 		/* Always preserve the active bit, note deactivation */
74 		deactivated = irq->active && !(val & ICH_LR_ACTIVE_BIT);
75 		irq->active = !!(val & ICH_LR_ACTIVE_BIT);
76 
77 		if (irq->active && is_v2_sgi)
78 			irq->active_source = cpuid;
79 
80 		/* Edge is the only case where we preserve the pending bit */
81 		if (irq->config == VGIC_CONFIG_EDGE &&
82 		    (val & ICH_LR_PENDING_BIT)) {
83 			irq->pending_latch = true;
84 
85 			if (is_v2_sgi)
86 				irq->source |= (1 << cpuid);
87 		}
88 
89 		/*
90 		 * Clear soft pending state when level irqs have been acked.
91 		 */
92 		if (irq->config == VGIC_CONFIG_LEVEL && !(val & ICH_LR_STATE))
93 			irq->pending_latch = false;
94 
95 		/* Handle resampling for mapped interrupts if required */
96 		vgic_irq_handle_resampling(irq, deactivated, val & ICH_LR_PENDING_BIT);
97 
98 		raw_spin_unlock(&irq->irq_lock);
99 		vgic_put_irq(vcpu->kvm, irq);
100 	}
101 
102 	cpuif->used_lrs = 0;
103 }
104 
105 /* Requires the irq to be locked already */
106 void vgic_v3_populate_lr(struct kvm_vcpu *vcpu, struct vgic_irq *irq, int lr)
107 {
108 	u32 model = vcpu->kvm->arch.vgic.vgic_model;
109 	u64 val = irq->intid;
110 	bool allow_pending = true, is_v2_sgi;
111 
112 	is_v2_sgi = (vgic_irq_is_sgi(irq->intid) &&
113 		     model == KVM_DEV_TYPE_ARM_VGIC_V2);
114 
115 	if (irq->active) {
116 		val |= ICH_LR_ACTIVE_BIT;
117 		if (is_v2_sgi)
118 			val |= irq->active_source << GICH_LR_PHYSID_CPUID_SHIFT;
119 		if (vgic_irq_is_multi_sgi(irq)) {
120 			allow_pending = false;
121 			val |= ICH_LR_EOI;
122 		}
123 	}
124 
125 	if (irq->hw && !vgic_irq_needs_resampling(irq)) {
126 		val |= ICH_LR_HW;
127 		val |= ((u64)irq->hwintid) << ICH_LR_PHYS_ID_SHIFT;
128 		/*
129 		 * Never set pending+active on a HW interrupt, as the
130 		 * pending state is kept at the physical distributor
131 		 * level.
132 		 */
133 		if (irq->active)
134 			allow_pending = false;
135 	} else {
136 		if (irq->config == VGIC_CONFIG_LEVEL) {
137 			val |= ICH_LR_EOI;
138 
139 			/*
140 			 * Software resampling doesn't work very well
141 			 * if we allow P+A, so let's not do that.
142 			 */
143 			if (irq->active)
144 				allow_pending = false;
145 		}
146 	}
147 
148 	if (allow_pending && irq_is_pending(irq)) {
149 		val |= ICH_LR_PENDING_BIT;
150 
151 		if (irq->config == VGIC_CONFIG_EDGE)
152 			irq->pending_latch = false;
153 
154 		if (vgic_irq_is_sgi(irq->intid) &&
155 		    model == KVM_DEV_TYPE_ARM_VGIC_V2) {
156 			u32 src = ffs(irq->source);
157 
158 			if (WARN_RATELIMIT(!src, "No SGI source for INTID %d\n",
159 					   irq->intid))
160 				return;
161 
162 			val |= (src - 1) << GICH_LR_PHYSID_CPUID_SHIFT;
163 			irq->source &= ~(1 << (src - 1));
164 			if (irq->source) {
165 				irq->pending_latch = true;
166 				val |= ICH_LR_EOI;
167 			}
168 		}
169 	}
170 
171 	/*
172 	 * Level-triggered mapped IRQs are special because we only observe
173 	 * rising edges as input to the VGIC.  We therefore lower the line
174 	 * level here, so that we can take new virtual IRQs.  See
175 	 * vgic_v3_fold_lr_state for more info.
176 	 */
177 	if (vgic_irq_is_mapped_level(irq) && (val & ICH_LR_PENDING_BIT))
178 		irq->line_level = false;
179 
180 	if (irq->group)
181 		val |= ICH_LR_GROUP;
182 
183 	val |= (u64)irq->priority << ICH_LR_PRIORITY_SHIFT;
184 
185 	vcpu->arch.vgic_cpu.vgic_v3.vgic_lr[lr] = val;
186 }
187 
188 void vgic_v3_clear_lr(struct kvm_vcpu *vcpu, int lr)
189 {
190 	vcpu->arch.vgic_cpu.vgic_v3.vgic_lr[lr] = 0;
191 }
192 
193 void vgic_v3_set_vmcr(struct kvm_vcpu *vcpu, struct vgic_vmcr *vmcrp)
194 {
195 	struct vgic_v3_cpu_if *cpu_if = &vcpu->arch.vgic_cpu.vgic_v3;
196 	u32 model = vcpu->kvm->arch.vgic.vgic_model;
197 	u32 vmcr;
198 
199 	if (model == KVM_DEV_TYPE_ARM_VGIC_V2) {
200 		vmcr = (vmcrp->ackctl << ICH_VMCR_ACK_CTL_SHIFT) &
201 			ICH_VMCR_ACK_CTL_MASK;
202 		vmcr |= (vmcrp->fiqen << ICH_VMCR_FIQ_EN_SHIFT) &
203 			ICH_VMCR_FIQ_EN_MASK;
204 	} else {
205 		/*
206 		 * When emulating GICv3 on GICv3 with SRE=1 on the
207 		 * VFIQEn bit is RES1 and the VAckCtl bit is RES0.
208 		 */
209 		vmcr = ICH_VMCR_FIQ_EN_MASK;
210 	}
211 
212 	vmcr |= (vmcrp->cbpr << ICH_VMCR_CBPR_SHIFT) & ICH_VMCR_CBPR_MASK;
213 	vmcr |= (vmcrp->eoim << ICH_VMCR_EOIM_SHIFT) & ICH_VMCR_EOIM_MASK;
214 	vmcr |= (vmcrp->abpr << ICH_VMCR_BPR1_SHIFT) & ICH_VMCR_BPR1_MASK;
215 	vmcr |= (vmcrp->bpr << ICH_VMCR_BPR0_SHIFT) & ICH_VMCR_BPR0_MASK;
216 	vmcr |= (vmcrp->pmr << ICH_VMCR_PMR_SHIFT) & ICH_VMCR_PMR_MASK;
217 	vmcr |= (vmcrp->grpen0 << ICH_VMCR_ENG0_SHIFT) & ICH_VMCR_ENG0_MASK;
218 	vmcr |= (vmcrp->grpen1 << ICH_VMCR_ENG1_SHIFT) & ICH_VMCR_ENG1_MASK;
219 
220 	cpu_if->vgic_vmcr = vmcr;
221 }
222 
223 void vgic_v3_get_vmcr(struct kvm_vcpu *vcpu, struct vgic_vmcr *vmcrp)
224 {
225 	struct vgic_v3_cpu_if *cpu_if = &vcpu->arch.vgic_cpu.vgic_v3;
226 	u32 model = vcpu->kvm->arch.vgic.vgic_model;
227 	u32 vmcr;
228 
229 	vmcr = cpu_if->vgic_vmcr;
230 
231 	if (model == KVM_DEV_TYPE_ARM_VGIC_V2) {
232 		vmcrp->ackctl = (vmcr & ICH_VMCR_ACK_CTL_MASK) >>
233 			ICH_VMCR_ACK_CTL_SHIFT;
234 		vmcrp->fiqen = (vmcr & ICH_VMCR_FIQ_EN_MASK) >>
235 			ICH_VMCR_FIQ_EN_SHIFT;
236 	} else {
237 		/*
238 		 * When emulating GICv3 on GICv3 with SRE=1 on the
239 		 * VFIQEn bit is RES1 and the VAckCtl bit is RES0.
240 		 */
241 		vmcrp->fiqen = 1;
242 		vmcrp->ackctl = 0;
243 	}
244 
245 	vmcrp->cbpr = (vmcr & ICH_VMCR_CBPR_MASK) >> ICH_VMCR_CBPR_SHIFT;
246 	vmcrp->eoim = (vmcr & ICH_VMCR_EOIM_MASK) >> ICH_VMCR_EOIM_SHIFT;
247 	vmcrp->abpr = (vmcr & ICH_VMCR_BPR1_MASK) >> ICH_VMCR_BPR1_SHIFT;
248 	vmcrp->bpr  = (vmcr & ICH_VMCR_BPR0_MASK) >> ICH_VMCR_BPR0_SHIFT;
249 	vmcrp->pmr  = (vmcr & ICH_VMCR_PMR_MASK) >> ICH_VMCR_PMR_SHIFT;
250 	vmcrp->grpen0 = (vmcr & ICH_VMCR_ENG0_MASK) >> ICH_VMCR_ENG0_SHIFT;
251 	vmcrp->grpen1 = (vmcr & ICH_VMCR_ENG1_MASK) >> ICH_VMCR_ENG1_SHIFT;
252 }
253 
254 #define INITIAL_PENDBASER_VALUE						  \
255 	(GIC_BASER_CACHEABILITY(GICR_PENDBASER, INNER, RaWb)		| \
256 	GIC_BASER_CACHEABILITY(GICR_PENDBASER, OUTER, SameAsInner)	| \
257 	GIC_BASER_SHAREABILITY(GICR_PENDBASER, InnerShareable))
258 
259 void vgic_v3_enable(struct kvm_vcpu *vcpu)
260 {
261 	struct vgic_v3_cpu_if *vgic_v3 = &vcpu->arch.vgic_cpu.vgic_v3;
262 
263 	/*
264 	 * By forcing VMCR to zero, the GIC will restore the binary
265 	 * points to their reset values. Anything else resets to zero
266 	 * anyway.
267 	 */
268 	vgic_v3->vgic_vmcr = 0;
269 
270 	/*
271 	 * If we are emulating a GICv3, we do it in an non-GICv2-compatible
272 	 * way, so we force SRE to 1 to demonstrate this to the guest.
273 	 * Also, we don't support any form of IRQ/FIQ bypass.
274 	 * This goes with the spec allowing the value to be RAO/WI.
275 	 */
276 	if (vcpu->kvm->arch.vgic.vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) {
277 		vgic_v3->vgic_sre = (ICC_SRE_EL1_DIB |
278 				     ICC_SRE_EL1_DFB |
279 				     ICC_SRE_EL1_SRE);
280 		vcpu->arch.vgic_cpu.pendbaser = INITIAL_PENDBASER_VALUE;
281 	} else {
282 		vgic_v3->vgic_sre = 0;
283 	}
284 
285 	vcpu->arch.vgic_cpu.num_id_bits = (kvm_vgic_global_state.ich_vtr_el2 &
286 					   ICH_VTR_ID_BITS_MASK) >>
287 					   ICH_VTR_ID_BITS_SHIFT;
288 	vcpu->arch.vgic_cpu.num_pri_bits = ((kvm_vgic_global_state.ich_vtr_el2 &
289 					    ICH_VTR_PRI_BITS_MASK) >>
290 					    ICH_VTR_PRI_BITS_SHIFT) + 1;
291 
292 	/* Get the show on the road... */
293 	vgic_v3->vgic_hcr = ICH_HCR_EN;
294 	if (group0_trap)
295 		vgic_v3->vgic_hcr |= ICH_HCR_TALL0;
296 	if (group1_trap)
297 		vgic_v3->vgic_hcr |= ICH_HCR_TALL1;
298 	if (common_trap)
299 		vgic_v3->vgic_hcr |= ICH_HCR_TC;
300 	if (dir_trap)
301 		vgic_v3->vgic_hcr |= ICH_HCR_TDIR;
302 }
303 
304 int vgic_v3_lpi_sync_pending_status(struct kvm *kvm, struct vgic_irq *irq)
305 {
306 	struct kvm_vcpu *vcpu;
307 	int byte_offset, bit_nr;
308 	gpa_t pendbase, ptr;
309 	bool status;
310 	u8 val;
311 	int ret;
312 	unsigned long flags;
313 
314 retry:
315 	vcpu = irq->target_vcpu;
316 	if (!vcpu)
317 		return 0;
318 
319 	pendbase = GICR_PENDBASER_ADDRESS(vcpu->arch.vgic_cpu.pendbaser);
320 
321 	byte_offset = irq->intid / BITS_PER_BYTE;
322 	bit_nr = irq->intid % BITS_PER_BYTE;
323 	ptr = pendbase + byte_offset;
324 
325 	ret = kvm_read_guest_lock(kvm, ptr, &val, 1);
326 	if (ret)
327 		return ret;
328 
329 	status = val & (1 << bit_nr);
330 
331 	raw_spin_lock_irqsave(&irq->irq_lock, flags);
332 	if (irq->target_vcpu != vcpu) {
333 		raw_spin_unlock_irqrestore(&irq->irq_lock, flags);
334 		goto retry;
335 	}
336 	irq->pending_latch = status;
337 	vgic_queue_irq_unlock(vcpu->kvm, irq, flags);
338 
339 	if (status) {
340 		/* clear consumed data */
341 		val &= ~(1 << bit_nr);
342 		ret = kvm_write_guest_lock(kvm, ptr, &val, 1);
343 		if (ret)
344 			return ret;
345 	}
346 	return 0;
347 }
348 
349 /*
350  * The deactivation of the doorbell interrupt will trigger the
351  * unmapping of the associated vPE.
352  */
353 static void unmap_all_vpes(struct vgic_dist *dist)
354 {
355 	struct irq_desc *desc;
356 	int i;
357 
358 	for (i = 0; i < dist->its_vm.nr_vpes; i++) {
359 		desc = irq_to_desc(dist->its_vm.vpes[i]->irq);
360 		irq_domain_deactivate_irq(irq_desc_get_irq_data(desc));
361 	}
362 }
363 
364 static void map_all_vpes(struct vgic_dist *dist)
365 {
366 	struct irq_desc *desc;
367 	int i;
368 
369 	for (i = 0; i < dist->its_vm.nr_vpes; i++) {
370 		desc = irq_to_desc(dist->its_vm.vpes[i]->irq);
371 		irq_domain_activate_irq(irq_desc_get_irq_data(desc), false);
372 	}
373 }
374 
375 /**
376  * vgic_v3_save_pending_tables - Save the pending tables into guest RAM
377  * kvm lock and all vcpu lock must be held
378  */
379 int vgic_v3_save_pending_tables(struct kvm *kvm)
380 {
381 	struct vgic_dist *dist = &kvm->arch.vgic;
382 	struct vgic_irq *irq;
383 	gpa_t last_ptr = ~(gpa_t)0;
384 	bool vlpi_avail = false;
385 	int ret = 0;
386 	u8 val;
387 
388 	if (unlikely(!vgic_initialized(kvm)))
389 		return -ENXIO;
390 
391 	/*
392 	 * A preparation for getting any VLPI states.
393 	 * The above vgic initialized check also ensures that the allocation
394 	 * and enabling of the doorbells have already been done.
395 	 */
396 	if (kvm_vgic_global_state.has_gicv4_1) {
397 		unmap_all_vpes(dist);
398 		vlpi_avail = true;
399 	}
400 
401 	list_for_each_entry(irq, &dist->lpi_list_head, lpi_list) {
402 		int byte_offset, bit_nr;
403 		struct kvm_vcpu *vcpu;
404 		gpa_t pendbase, ptr;
405 		bool is_pending;
406 		bool stored;
407 
408 		vcpu = irq->target_vcpu;
409 		if (!vcpu)
410 			continue;
411 
412 		pendbase = GICR_PENDBASER_ADDRESS(vcpu->arch.vgic_cpu.pendbaser);
413 
414 		byte_offset = irq->intid / BITS_PER_BYTE;
415 		bit_nr = irq->intid % BITS_PER_BYTE;
416 		ptr = pendbase + byte_offset;
417 
418 		if (ptr != last_ptr) {
419 			ret = kvm_read_guest_lock(kvm, ptr, &val, 1);
420 			if (ret)
421 				goto out;
422 			last_ptr = ptr;
423 		}
424 
425 		stored = val & (1U << bit_nr);
426 
427 		is_pending = irq->pending_latch;
428 
429 		if (irq->hw && vlpi_avail)
430 			vgic_v4_get_vlpi_state(irq, &is_pending);
431 
432 		if (stored == is_pending)
433 			continue;
434 
435 		if (is_pending)
436 			val |= 1 << bit_nr;
437 		else
438 			val &= ~(1 << bit_nr);
439 
440 		ret = kvm_write_guest_lock(kvm, ptr, &val, 1);
441 		if (ret)
442 			goto out;
443 	}
444 
445 out:
446 	if (vlpi_avail)
447 		map_all_vpes(dist);
448 
449 	return ret;
450 }
451 
452 /**
453  * vgic_v3_rdist_overlap - check if a region overlaps with any
454  * existing redistributor region
455  *
456  * @kvm: kvm handle
457  * @base: base of the region
458  * @size: size of region
459  *
460  * Return: true if there is an overlap
461  */
462 bool vgic_v3_rdist_overlap(struct kvm *kvm, gpa_t base, size_t size)
463 {
464 	struct vgic_dist *d = &kvm->arch.vgic;
465 	struct vgic_redist_region *rdreg;
466 
467 	list_for_each_entry(rdreg, &d->rd_regions, list) {
468 		if ((base + size > rdreg->base) &&
469 			(base < rdreg->base + vgic_v3_rd_region_size(kvm, rdreg)))
470 			return true;
471 	}
472 	return false;
473 }
474 
475 /*
476  * Check for overlapping regions and for regions crossing the end of memory
477  * for base addresses which have already been set.
478  */
479 bool vgic_v3_check_base(struct kvm *kvm)
480 {
481 	struct vgic_dist *d = &kvm->arch.vgic;
482 	struct vgic_redist_region *rdreg;
483 
484 	if (!IS_VGIC_ADDR_UNDEF(d->vgic_dist_base) &&
485 	    d->vgic_dist_base + KVM_VGIC_V3_DIST_SIZE < d->vgic_dist_base)
486 		return false;
487 
488 	list_for_each_entry(rdreg, &d->rd_regions, list) {
489 		size_t sz = vgic_v3_rd_region_size(kvm, rdreg);
490 
491 		if (vgic_check_iorange(kvm, VGIC_ADDR_UNDEF,
492 				       rdreg->base, SZ_64K, sz))
493 			return false;
494 	}
495 
496 	if (IS_VGIC_ADDR_UNDEF(d->vgic_dist_base))
497 		return true;
498 
499 	return !vgic_v3_rdist_overlap(kvm, d->vgic_dist_base,
500 				      KVM_VGIC_V3_DIST_SIZE);
501 }
502 
503 /**
504  * vgic_v3_rdist_free_slot - Look up registered rdist regions and identify one
505  * which has free space to put a new rdist region.
506  *
507  * @rd_regions: redistributor region list head
508  *
509  * A redistributor regions maps n redistributors, n = region size / (2 x 64kB).
510  * Stride between redistributors is 0 and regions are filled in the index order.
511  *
512  * Return: the redist region handle, if any, that has space to map a new rdist
513  * region.
514  */
515 struct vgic_redist_region *vgic_v3_rdist_free_slot(struct list_head *rd_regions)
516 {
517 	struct vgic_redist_region *rdreg;
518 
519 	list_for_each_entry(rdreg, rd_regions, list) {
520 		if (!vgic_v3_redist_region_full(rdreg))
521 			return rdreg;
522 	}
523 	return NULL;
524 }
525 
526 struct vgic_redist_region *vgic_v3_rdist_region_from_index(struct kvm *kvm,
527 							   u32 index)
528 {
529 	struct list_head *rd_regions = &kvm->arch.vgic.rd_regions;
530 	struct vgic_redist_region *rdreg;
531 
532 	list_for_each_entry(rdreg, rd_regions, list) {
533 		if (rdreg->index == index)
534 			return rdreg;
535 	}
536 	return NULL;
537 }
538 
539 
540 int vgic_v3_map_resources(struct kvm *kvm)
541 {
542 	struct vgic_dist *dist = &kvm->arch.vgic;
543 	struct kvm_vcpu *vcpu;
544 	int ret = 0;
545 	int c;
546 
547 	kvm_for_each_vcpu(c, vcpu, kvm) {
548 		struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
549 
550 		if (IS_VGIC_ADDR_UNDEF(vgic_cpu->rd_iodev.base_addr)) {
551 			kvm_debug("vcpu %d redistributor base not set\n", c);
552 			return -ENXIO;
553 		}
554 	}
555 
556 	if (IS_VGIC_ADDR_UNDEF(dist->vgic_dist_base)) {
557 		kvm_err("Need to set vgic distributor addresses first\n");
558 		return -ENXIO;
559 	}
560 
561 	if (!vgic_v3_check_base(kvm)) {
562 		kvm_err("VGIC redist and dist frames overlap\n");
563 		return -EINVAL;
564 	}
565 
566 	/*
567 	 * For a VGICv3 we require the userland to explicitly initialize
568 	 * the VGIC before we need to use it.
569 	 */
570 	if (!vgic_initialized(kvm)) {
571 		return -EBUSY;
572 	}
573 
574 	ret = vgic_register_dist_iodev(kvm, dist->vgic_dist_base, VGIC_V3);
575 	if (ret) {
576 		kvm_err("Unable to register VGICv3 dist MMIO regions\n");
577 		return ret;
578 	}
579 
580 	if (kvm_vgic_global_state.has_gicv4_1)
581 		vgic_v4_configure_vsgis(kvm);
582 
583 	return 0;
584 }
585 
586 DEFINE_STATIC_KEY_FALSE(vgic_v3_cpuif_trap);
587 
588 static int __init early_group0_trap_cfg(char *buf)
589 {
590 	return strtobool(buf, &group0_trap);
591 }
592 early_param("kvm-arm.vgic_v3_group0_trap", early_group0_trap_cfg);
593 
594 static int __init early_group1_trap_cfg(char *buf)
595 {
596 	return strtobool(buf, &group1_trap);
597 }
598 early_param("kvm-arm.vgic_v3_group1_trap", early_group1_trap_cfg);
599 
600 static int __init early_common_trap_cfg(char *buf)
601 {
602 	return strtobool(buf, &common_trap);
603 }
604 early_param("kvm-arm.vgic_v3_common_trap", early_common_trap_cfg);
605 
606 static int __init early_gicv4_enable(char *buf)
607 {
608 	return strtobool(buf, &gicv4_enable);
609 }
610 early_param("kvm-arm.vgic_v4_enable", early_gicv4_enable);
611 
612 /**
613  * vgic_v3_probe - probe for a VGICv3 compatible interrupt controller
614  * @info:	pointer to the GIC description
615  *
616  * Returns 0 if the VGICv3 has been probed successfully, returns an error code
617  * otherwise
618  */
619 int vgic_v3_probe(const struct gic_kvm_info *info)
620 {
621 	u64 ich_vtr_el2 = kvm_call_hyp_ret(__vgic_v3_get_gic_config);
622 	bool has_v2;
623 	int ret;
624 
625 	has_v2 = ich_vtr_el2 >> 63;
626 	ich_vtr_el2 = (u32)ich_vtr_el2;
627 
628 	/*
629 	 * The ListRegs field is 5 bits, but there is an architectural
630 	 * maximum of 16 list registers. Just ignore bit 4...
631 	 */
632 	kvm_vgic_global_state.nr_lr = (ich_vtr_el2 & 0xf) + 1;
633 	kvm_vgic_global_state.can_emulate_gicv2 = false;
634 	kvm_vgic_global_state.ich_vtr_el2 = ich_vtr_el2;
635 
636 	/* GICv4 support? */
637 	if (info->has_v4) {
638 		kvm_vgic_global_state.has_gicv4 = gicv4_enable;
639 		kvm_vgic_global_state.has_gicv4_1 = info->has_v4_1 && gicv4_enable;
640 		kvm_info("GICv4%s support %sabled\n",
641 			 kvm_vgic_global_state.has_gicv4_1 ? ".1" : "",
642 			 gicv4_enable ? "en" : "dis");
643 	}
644 
645 	kvm_vgic_global_state.vcpu_base = 0;
646 
647 	if (!info->vcpu.start) {
648 		kvm_info("GICv3: no GICV resource entry\n");
649 	} else if (!has_v2) {
650 		pr_warn(FW_BUG "CPU interface incapable of MMIO access\n");
651 	} else if (!PAGE_ALIGNED(info->vcpu.start)) {
652 		pr_warn("GICV physical address 0x%llx not page aligned\n",
653 			(unsigned long long)info->vcpu.start);
654 	} else {
655 		kvm_vgic_global_state.vcpu_base = info->vcpu.start;
656 		kvm_vgic_global_state.can_emulate_gicv2 = true;
657 		ret = kvm_register_vgic_device(KVM_DEV_TYPE_ARM_VGIC_V2);
658 		if (ret) {
659 			kvm_err("Cannot register GICv2 KVM device.\n");
660 			return ret;
661 		}
662 		kvm_info("vgic-v2@%llx\n", info->vcpu.start);
663 	}
664 	ret = kvm_register_vgic_device(KVM_DEV_TYPE_ARM_VGIC_V3);
665 	if (ret) {
666 		kvm_err("Cannot register GICv3 KVM device.\n");
667 		kvm_unregister_device_ops(KVM_DEV_TYPE_ARM_VGIC_V2);
668 		return ret;
669 	}
670 
671 	if (kvm_vgic_global_state.vcpu_base == 0)
672 		kvm_info("disabling GICv2 emulation\n");
673 
674 	if (cpus_have_const_cap(ARM64_WORKAROUND_CAVIUM_30115)) {
675 		group0_trap = true;
676 		group1_trap = true;
677 	}
678 
679 	if (kvm_vgic_global_state.ich_vtr_el2 & ICH_VTR_SEIS_MASK) {
680 		kvm_info("GICv3 with locally generated SEI\n");
681 
682 		group0_trap = true;
683 		group1_trap = true;
684 		if (ich_vtr_el2 & ICH_VTR_TDS_MASK)
685 			dir_trap = true;
686 		else
687 			common_trap = true;
688 	}
689 
690 	if (group0_trap || group1_trap || common_trap | dir_trap) {
691 		kvm_info("GICv3 sysreg trapping enabled ([%s%s%s%s], reduced performance)\n",
692 			 group0_trap ? "G0" : "",
693 			 group1_trap ? "G1" : "",
694 			 common_trap ? "C"  : "",
695 			 dir_trap    ? "D"  : "");
696 		static_branch_enable(&vgic_v3_cpuif_trap);
697 	}
698 
699 	kvm_vgic_global_state.vctrl_base = NULL;
700 	kvm_vgic_global_state.type = VGIC_V3;
701 	kvm_vgic_global_state.max_gic_vcpus = VGIC_V3_MAX_CPUS;
702 
703 	return 0;
704 }
705 
706 void vgic_v3_load(struct kvm_vcpu *vcpu)
707 {
708 	struct vgic_v3_cpu_if *cpu_if = &vcpu->arch.vgic_cpu.vgic_v3;
709 
710 	/*
711 	 * If dealing with a GICv2 emulation on GICv3, VMCR_EL2.VFIQen
712 	 * is dependent on ICC_SRE_EL1.SRE, and we have to perform the
713 	 * VMCR_EL2 save/restore in the world switch.
714 	 */
715 	if (likely(cpu_if->vgic_sre))
716 		kvm_call_hyp(__vgic_v3_write_vmcr, cpu_if->vgic_vmcr);
717 
718 	kvm_call_hyp(__vgic_v3_restore_aprs, cpu_if);
719 
720 	if (has_vhe())
721 		__vgic_v3_activate_traps(cpu_if);
722 
723 	WARN_ON(vgic_v4_load(vcpu));
724 }
725 
726 void vgic_v3_vmcr_sync(struct kvm_vcpu *vcpu)
727 {
728 	struct vgic_v3_cpu_if *cpu_if = &vcpu->arch.vgic_cpu.vgic_v3;
729 
730 	if (likely(cpu_if->vgic_sre))
731 		cpu_if->vgic_vmcr = kvm_call_hyp_ret(__vgic_v3_read_vmcr);
732 }
733 
734 void vgic_v3_put(struct kvm_vcpu *vcpu)
735 {
736 	struct vgic_v3_cpu_if *cpu_if = &vcpu->arch.vgic_cpu.vgic_v3;
737 
738 	WARN_ON(vgic_v4_put(vcpu, false));
739 
740 	vgic_v3_vmcr_sync(vcpu);
741 
742 	kvm_call_hyp(__vgic_v3_save_aprs, cpu_if);
743 
744 	if (has_vhe())
745 		__vgic_v3_deactivate_traps(cpu_if);
746 }
747