xref: /openbmc/linux/arch/arm64/kvm/vgic/vgic-its.c (revision 7effbd18)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * GICv3 ITS emulation
4  *
5  * Copyright (C) 2015,2016 ARM Ltd.
6  * Author: Andre Przywara <andre.przywara@arm.com>
7  */
8 
9 #include <linux/cpu.h>
10 #include <linux/kvm.h>
11 #include <linux/kvm_host.h>
12 #include <linux/interrupt.h>
13 #include <linux/list.h>
14 #include <linux/uaccess.h>
15 #include <linux/list_sort.h>
16 
17 #include <linux/irqchip/arm-gic-v3.h>
18 
19 #include <asm/kvm_emulate.h>
20 #include <asm/kvm_arm.h>
21 #include <asm/kvm_mmu.h>
22 
23 #include "vgic.h"
24 #include "vgic-mmio.h"
25 
26 static int vgic_its_save_tables_v0(struct vgic_its *its);
27 static int vgic_its_restore_tables_v0(struct vgic_its *its);
28 static int vgic_its_commit_v0(struct vgic_its *its);
29 static int update_lpi_config(struct kvm *kvm, struct vgic_irq *irq,
30 			     struct kvm_vcpu *filter_vcpu, bool needs_inv);
31 
32 /*
33  * Creates a new (reference to a) struct vgic_irq for a given LPI.
34  * If this LPI is already mapped on another ITS, we increase its refcount
35  * and return a pointer to the existing structure.
36  * If this is a "new" LPI, we allocate and initialize a new struct vgic_irq.
37  * This function returns a pointer to the _unlocked_ structure.
38  */
39 static struct vgic_irq *vgic_add_lpi(struct kvm *kvm, u32 intid,
40 				     struct kvm_vcpu *vcpu)
41 {
42 	struct vgic_dist *dist = &kvm->arch.vgic;
43 	struct vgic_irq *irq = vgic_get_irq(kvm, NULL, intid), *oldirq;
44 	unsigned long flags;
45 	int ret;
46 
47 	/* In this case there is no put, since we keep the reference. */
48 	if (irq)
49 		return irq;
50 
51 	irq = kzalloc(sizeof(struct vgic_irq), GFP_KERNEL_ACCOUNT);
52 	if (!irq)
53 		return ERR_PTR(-ENOMEM);
54 
55 	INIT_LIST_HEAD(&irq->lpi_list);
56 	INIT_LIST_HEAD(&irq->ap_list);
57 	raw_spin_lock_init(&irq->irq_lock);
58 
59 	irq->config = VGIC_CONFIG_EDGE;
60 	kref_init(&irq->refcount);
61 	irq->intid = intid;
62 	irq->target_vcpu = vcpu;
63 	irq->group = 1;
64 
65 	raw_spin_lock_irqsave(&dist->lpi_list_lock, flags);
66 
67 	/*
68 	 * There could be a race with another vgic_add_lpi(), so we need to
69 	 * check that we don't add a second list entry with the same LPI.
70 	 */
71 	list_for_each_entry(oldirq, &dist->lpi_list_head, lpi_list) {
72 		if (oldirq->intid != intid)
73 			continue;
74 
75 		/* Someone was faster with adding this LPI, lets use that. */
76 		kfree(irq);
77 		irq = oldirq;
78 
79 		/*
80 		 * This increases the refcount, the caller is expected to
81 		 * call vgic_put_irq() on the returned pointer once it's
82 		 * finished with the IRQ.
83 		 */
84 		vgic_get_irq_kref(irq);
85 
86 		goto out_unlock;
87 	}
88 
89 	list_add_tail(&irq->lpi_list, &dist->lpi_list_head);
90 	dist->lpi_list_count++;
91 
92 out_unlock:
93 	raw_spin_unlock_irqrestore(&dist->lpi_list_lock, flags);
94 
95 	/*
96 	 * We "cache" the configuration table entries in our struct vgic_irq's.
97 	 * However we only have those structs for mapped IRQs, so we read in
98 	 * the respective config data from memory here upon mapping the LPI.
99 	 *
100 	 * Should any of these fail, behave as if we couldn't create the LPI
101 	 * by dropping the refcount and returning the error.
102 	 */
103 	ret = update_lpi_config(kvm, irq, NULL, false);
104 	if (ret) {
105 		vgic_put_irq(kvm, irq);
106 		return ERR_PTR(ret);
107 	}
108 
109 	ret = vgic_v3_lpi_sync_pending_status(kvm, irq);
110 	if (ret) {
111 		vgic_put_irq(kvm, irq);
112 		return ERR_PTR(ret);
113 	}
114 
115 	return irq;
116 }
117 
118 struct its_device {
119 	struct list_head dev_list;
120 
121 	/* the head for the list of ITTEs */
122 	struct list_head itt_head;
123 	u32 num_eventid_bits;
124 	gpa_t itt_addr;
125 	u32 device_id;
126 };
127 
128 #define COLLECTION_NOT_MAPPED ((u32)~0)
129 
130 struct its_collection {
131 	struct list_head coll_list;
132 
133 	u32 collection_id;
134 	u32 target_addr;
135 };
136 
137 #define its_is_collection_mapped(coll) ((coll) && \
138 				((coll)->target_addr != COLLECTION_NOT_MAPPED))
139 
140 struct its_ite {
141 	struct list_head ite_list;
142 
143 	struct vgic_irq *irq;
144 	struct its_collection *collection;
145 	u32 event_id;
146 };
147 
148 struct vgic_translation_cache_entry {
149 	struct list_head	entry;
150 	phys_addr_t		db;
151 	u32			devid;
152 	u32			eventid;
153 	struct vgic_irq		*irq;
154 };
155 
156 /**
157  * struct vgic_its_abi - ITS abi ops and settings
158  * @cte_esz: collection table entry size
159  * @dte_esz: device table entry size
160  * @ite_esz: interrupt translation table entry size
161  * @save tables: save the ITS tables into guest RAM
162  * @restore_tables: restore the ITS internal structs from tables
163  *  stored in guest RAM
164  * @commit: initialize the registers which expose the ABI settings,
165  *  especially the entry sizes
166  */
167 struct vgic_its_abi {
168 	int cte_esz;
169 	int dte_esz;
170 	int ite_esz;
171 	int (*save_tables)(struct vgic_its *its);
172 	int (*restore_tables)(struct vgic_its *its);
173 	int (*commit)(struct vgic_its *its);
174 };
175 
176 #define ABI_0_ESZ	8
177 #define ESZ_MAX		ABI_0_ESZ
178 
179 static const struct vgic_its_abi its_table_abi_versions[] = {
180 	[0] = {
181 	 .cte_esz = ABI_0_ESZ,
182 	 .dte_esz = ABI_0_ESZ,
183 	 .ite_esz = ABI_0_ESZ,
184 	 .save_tables = vgic_its_save_tables_v0,
185 	 .restore_tables = vgic_its_restore_tables_v0,
186 	 .commit = vgic_its_commit_v0,
187 	},
188 };
189 
190 #define NR_ITS_ABIS	ARRAY_SIZE(its_table_abi_versions)
191 
192 inline const struct vgic_its_abi *vgic_its_get_abi(struct vgic_its *its)
193 {
194 	return &its_table_abi_versions[its->abi_rev];
195 }
196 
197 static int vgic_its_set_abi(struct vgic_its *its, u32 rev)
198 {
199 	const struct vgic_its_abi *abi;
200 
201 	its->abi_rev = rev;
202 	abi = vgic_its_get_abi(its);
203 	return abi->commit(its);
204 }
205 
206 /*
207  * Find and returns a device in the device table for an ITS.
208  * Must be called with the its_lock mutex held.
209  */
210 static struct its_device *find_its_device(struct vgic_its *its, u32 device_id)
211 {
212 	struct its_device *device;
213 
214 	list_for_each_entry(device, &its->device_list, dev_list)
215 		if (device_id == device->device_id)
216 			return device;
217 
218 	return NULL;
219 }
220 
221 /*
222  * Find and returns an interrupt translation table entry (ITTE) for a given
223  * Device ID/Event ID pair on an ITS.
224  * Must be called with the its_lock mutex held.
225  */
226 static struct its_ite *find_ite(struct vgic_its *its, u32 device_id,
227 				  u32 event_id)
228 {
229 	struct its_device *device;
230 	struct its_ite *ite;
231 
232 	device = find_its_device(its, device_id);
233 	if (device == NULL)
234 		return NULL;
235 
236 	list_for_each_entry(ite, &device->itt_head, ite_list)
237 		if (ite->event_id == event_id)
238 			return ite;
239 
240 	return NULL;
241 }
242 
243 /* To be used as an iterator this macro misses the enclosing parentheses */
244 #define for_each_lpi_its(dev, ite, its) \
245 	list_for_each_entry(dev, &(its)->device_list, dev_list) \
246 		list_for_each_entry(ite, &(dev)->itt_head, ite_list)
247 
248 #define GIC_LPI_OFFSET 8192
249 
250 #define VITS_TYPER_IDBITS 16
251 #define VITS_TYPER_DEVBITS 16
252 #define VITS_DTE_MAX_DEVID_OFFSET	(BIT(14) - 1)
253 #define VITS_ITE_MAX_EVENTID_OFFSET	(BIT(16) - 1)
254 
255 /*
256  * Finds and returns a collection in the ITS collection table.
257  * Must be called with the its_lock mutex held.
258  */
259 static struct its_collection *find_collection(struct vgic_its *its, int coll_id)
260 {
261 	struct its_collection *collection;
262 
263 	list_for_each_entry(collection, &its->collection_list, coll_list) {
264 		if (coll_id == collection->collection_id)
265 			return collection;
266 	}
267 
268 	return NULL;
269 }
270 
271 #define LPI_PROP_ENABLE_BIT(p)	((p) & LPI_PROP_ENABLED)
272 #define LPI_PROP_PRIORITY(p)	((p) & 0xfc)
273 
274 /*
275  * Reads the configuration data for a given LPI from guest memory and
276  * updates the fields in struct vgic_irq.
277  * If filter_vcpu is not NULL, applies only if the IRQ is targeting this
278  * VCPU. Unconditionally applies if filter_vcpu is NULL.
279  */
280 static int update_lpi_config(struct kvm *kvm, struct vgic_irq *irq,
281 			     struct kvm_vcpu *filter_vcpu, bool needs_inv)
282 {
283 	u64 propbase = GICR_PROPBASER_ADDRESS(kvm->arch.vgic.propbaser);
284 	u8 prop;
285 	int ret;
286 	unsigned long flags;
287 
288 	ret = kvm_read_guest_lock(kvm, propbase + irq->intid - GIC_LPI_OFFSET,
289 				  &prop, 1);
290 
291 	if (ret)
292 		return ret;
293 
294 	raw_spin_lock_irqsave(&irq->irq_lock, flags);
295 
296 	if (!filter_vcpu || filter_vcpu == irq->target_vcpu) {
297 		irq->priority = LPI_PROP_PRIORITY(prop);
298 		irq->enabled = LPI_PROP_ENABLE_BIT(prop);
299 
300 		if (!irq->hw) {
301 			vgic_queue_irq_unlock(kvm, irq, flags);
302 			return 0;
303 		}
304 	}
305 
306 	raw_spin_unlock_irqrestore(&irq->irq_lock, flags);
307 
308 	if (irq->hw)
309 		return its_prop_update_vlpi(irq->host_irq, prop, needs_inv);
310 
311 	return 0;
312 }
313 
314 /*
315  * Create a snapshot of the current LPIs targeting @vcpu, so that we can
316  * enumerate those LPIs without holding any lock.
317  * Returns their number and puts the kmalloc'ed array into intid_ptr.
318  */
319 int vgic_copy_lpi_list(struct kvm *kvm, struct kvm_vcpu *vcpu, u32 **intid_ptr)
320 {
321 	struct vgic_dist *dist = &kvm->arch.vgic;
322 	struct vgic_irq *irq;
323 	unsigned long flags;
324 	u32 *intids;
325 	int irq_count, i = 0;
326 
327 	/*
328 	 * There is an obvious race between allocating the array and LPIs
329 	 * being mapped/unmapped. If we ended up here as a result of a
330 	 * command, we're safe (locks are held, preventing another
331 	 * command). If coming from another path (such as enabling LPIs),
332 	 * we must be careful not to overrun the array.
333 	 */
334 	irq_count = READ_ONCE(dist->lpi_list_count);
335 	intids = kmalloc_array(irq_count, sizeof(intids[0]), GFP_KERNEL_ACCOUNT);
336 	if (!intids)
337 		return -ENOMEM;
338 
339 	raw_spin_lock_irqsave(&dist->lpi_list_lock, flags);
340 	list_for_each_entry(irq, &dist->lpi_list_head, lpi_list) {
341 		if (i == irq_count)
342 			break;
343 		/* We don't need to "get" the IRQ, as we hold the list lock. */
344 		if (vcpu && irq->target_vcpu != vcpu)
345 			continue;
346 		intids[i++] = irq->intid;
347 	}
348 	raw_spin_unlock_irqrestore(&dist->lpi_list_lock, flags);
349 
350 	*intid_ptr = intids;
351 	return i;
352 }
353 
354 static int update_affinity(struct vgic_irq *irq, struct kvm_vcpu *vcpu)
355 {
356 	int ret = 0;
357 	unsigned long flags;
358 
359 	raw_spin_lock_irqsave(&irq->irq_lock, flags);
360 	irq->target_vcpu = vcpu;
361 	raw_spin_unlock_irqrestore(&irq->irq_lock, flags);
362 
363 	if (irq->hw) {
364 		struct its_vlpi_map map;
365 
366 		ret = its_get_vlpi(irq->host_irq, &map);
367 		if (ret)
368 			return ret;
369 
370 		if (map.vpe)
371 			atomic_dec(&map.vpe->vlpi_count);
372 		map.vpe = &vcpu->arch.vgic_cpu.vgic_v3.its_vpe;
373 		atomic_inc(&map.vpe->vlpi_count);
374 
375 		ret = its_map_vlpi(irq->host_irq, &map);
376 	}
377 
378 	return ret;
379 }
380 
381 /*
382  * Promotes the ITS view of affinity of an ITTE (which redistributor this LPI
383  * is targeting) to the VGIC's view, which deals with target VCPUs.
384  * Needs to be called whenever either the collection for a LPIs has
385  * changed or the collection itself got retargeted.
386  */
387 static void update_affinity_ite(struct kvm *kvm, struct its_ite *ite)
388 {
389 	struct kvm_vcpu *vcpu;
390 
391 	if (!its_is_collection_mapped(ite->collection))
392 		return;
393 
394 	vcpu = kvm_get_vcpu(kvm, ite->collection->target_addr);
395 	update_affinity(ite->irq, vcpu);
396 }
397 
398 /*
399  * Updates the target VCPU for every LPI targeting this collection.
400  * Must be called with the its_lock mutex held.
401  */
402 static void update_affinity_collection(struct kvm *kvm, struct vgic_its *its,
403 				       struct its_collection *coll)
404 {
405 	struct its_device *device;
406 	struct its_ite *ite;
407 
408 	for_each_lpi_its(device, ite, its) {
409 		if (ite->collection != coll)
410 			continue;
411 
412 		update_affinity_ite(kvm, ite);
413 	}
414 }
415 
416 static u32 max_lpis_propbaser(u64 propbaser)
417 {
418 	int nr_idbits = (propbaser & 0x1f) + 1;
419 
420 	return 1U << min(nr_idbits, INTERRUPT_ID_BITS_ITS);
421 }
422 
423 /*
424  * Sync the pending table pending bit of LPIs targeting @vcpu
425  * with our own data structures. This relies on the LPI being
426  * mapped before.
427  */
428 static int its_sync_lpi_pending_table(struct kvm_vcpu *vcpu)
429 {
430 	gpa_t pendbase = GICR_PENDBASER_ADDRESS(vcpu->arch.vgic_cpu.pendbaser);
431 	struct vgic_irq *irq;
432 	int last_byte_offset = -1;
433 	int ret = 0;
434 	u32 *intids;
435 	int nr_irqs, i;
436 	unsigned long flags;
437 	u8 pendmask;
438 
439 	nr_irqs = vgic_copy_lpi_list(vcpu->kvm, vcpu, &intids);
440 	if (nr_irqs < 0)
441 		return nr_irqs;
442 
443 	for (i = 0; i < nr_irqs; i++) {
444 		int byte_offset, bit_nr;
445 
446 		byte_offset = intids[i] / BITS_PER_BYTE;
447 		bit_nr = intids[i] % BITS_PER_BYTE;
448 
449 		/*
450 		 * For contiguously allocated LPIs chances are we just read
451 		 * this very same byte in the last iteration. Reuse that.
452 		 */
453 		if (byte_offset != last_byte_offset) {
454 			ret = kvm_read_guest_lock(vcpu->kvm,
455 						  pendbase + byte_offset,
456 						  &pendmask, 1);
457 			if (ret) {
458 				kfree(intids);
459 				return ret;
460 			}
461 			last_byte_offset = byte_offset;
462 		}
463 
464 		irq = vgic_get_irq(vcpu->kvm, NULL, intids[i]);
465 		raw_spin_lock_irqsave(&irq->irq_lock, flags);
466 		irq->pending_latch = pendmask & (1U << bit_nr);
467 		vgic_queue_irq_unlock(vcpu->kvm, irq, flags);
468 		vgic_put_irq(vcpu->kvm, irq);
469 	}
470 
471 	kfree(intids);
472 
473 	return ret;
474 }
475 
476 static unsigned long vgic_mmio_read_its_typer(struct kvm *kvm,
477 					      struct vgic_its *its,
478 					      gpa_t addr, unsigned int len)
479 {
480 	const struct vgic_its_abi *abi = vgic_its_get_abi(its);
481 	u64 reg = GITS_TYPER_PLPIS;
482 
483 	/*
484 	 * We use linear CPU numbers for redistributor addressing,
485 	 * so GITS_TYPER.PTA is 0.
486 	 * Also we force all PROPBASER registers to be the same, so
487 	 * CommonLPIAff is 0 as well.
488 	 * To avoid memory waste in the guest, we keep the number of IDBits and
489 	 * DevBits low - as least for the time being.
490 	 */
491 	reg |= GIC_ENCODE_SZ(VITS_TYPER_DEVBITS, 5) << GITS_TYPER_DEVBITS_SHIFT;
492 	reg |= GIC_ENCODE_SZ(VITS_TYPER_IDBITS, 5) << GITS_TYPER_IDBITS_SHIFT;
493 	reg |= GIC_ENCODE_SZ(abi->ite_esz, 4) << GITS_TYPER_ITT_ENTRY_SIZE_SHIFT;
494 
495 	return extract_bytes(reg, addr & 7, len);
496 }
497 
498 static unsigned long vgic_mmio_read_its_iidr(struct kvm *kvm,
499 					     struct vgic_its *its,
500 					     gpa_t addr, unsigned int len)
501 {
502 	u32 val;
503 
504 	val = (its->abi_rev << GITS_IIDR_REV_SHIFT) & GITS_IIDR_REV_MASK;
505 	val |= (PRODUCT_ID_KVM << GITS_IIDR_PRODUCTID_SHIFT) | IMPLEMENTER_ARM;
506 	return val;
507 }
508 
509 static int vgic_mmio_uaccess_write_its_iidr(struct kvm *kvm,
510 					    struct vgic_its *its,
511 					    gpa_t addr, unsigned int len,
512 					    unsigned long val)
513 {
514 	u32 rev = GITS_IIDR_REV(val);
515 
516 	if (rev >= NR_ITS_ABIS)
517 		return -EINVAL;
518 	return vgic_its_set_abi(its, rev);
519 }
520 
521 static unsigned long vgic_mmio_read_its_idregs(struct kvm *kvm,
522 					       struct vgic_its *its,
523 					       gpa_t addr, unsigned int len)
524 {
525 	switch (addr & 0xffff) {
526 	case GITS_PIDR0:
527 		return 0x92;	/* part number, bits[7:0] */
528 	case GITS_PIDR1:
529 		return 0xb4;	/* part number, bits[11:8] */
530 	case GITS_PIDR2:
531 		return GIC_PIDR2_ARCH_GICv3 | 0x0b;
532 	case GITS_PIDR4:
533 		return 0x40;	/* This is a 64K software visible page */
534 	/* The following are the ID registers for (any) GIC. */
535 	case GITS_CIDR0:
536 		return 0x0d;
537 	case GITS_CIDR1:
538 		return 0xf0;
539 	case GITS_CIDR2:
540 		return 0x05;
541 	case GITS_CIDR3:
542 		return 0xb1;
543 	}
544 
545 	return 0;
546 }
547 
548 static struct vgic_irq *__vgic_its_check_cache(struct vgic_dist *dist,
549 					       phys_addr_t db,
550 					       u32 devid, u32 eventid)
551 {
552 	struct vgic_translation_cache_entry *cte;
553 
554 	list_for_each_entry(cte, &dist->lpi_translation_cache, entry) {
555 		/*
556 		 * If we hit a NULL entry, there is nothing after this
557 		 * point.
558 		 */
559 		if (!cte->irq)
560 			break;
561 
562 		if (cte->db != db || cte->devid != devid ||
563 		    cte->eventid != eventid)
564 			continue;
565 
566 		/*
567 		 * Move this entry to the head, as it is the most
568 		 * recently used.
569 		 */
570 		if (!list_is_first(&cte->entry, &dist->lpi_translation_cache))
571 			list_move(&cte->entry, &dist->lpi_translation_cache);
572 
573 		return cte->irq;
574 	}
575 
576 	return NULL;
577 }
578 
579 static struct vgic_irq *vgic_its_check_cache(struct kvm *kvm, phys_addr_t db,
580 					     u32 devid, u32 eventid)
581 {
582 	struct vgic_dist *dist = &kvm->arch.vgic;
583 	struct vgic_irq *irq;
584 	unsigned long flags;
585 
586 	raw_spin_lock_irqsave(&dist->lpi_list_lock, flags);
587 	irq = __vgic_its_check_cache(dist, db, devid, eventid);
588 	raw_spin_unlock_irqrestore(&dist->lpi_list_lock, flags);
589 
590 	return irq;
591 }
592 
593 static void vgic_its_cache_translation(struct kvm *kvm, struct vgic_its *its,
594 				       u32 devid, u32 eventid,
595 				       struct vgic_irq *irq)
596 {
597 	struct vgic_dist *dist = &kvm->arch.vgic;
598 	struct vgic_translation_cache_entry *cte;
599 	unsigned long flags;
600 	phys_addr_t db;
601 
602 	/* Do not cache a directly injected interrupt */
603 	if (irq->hw)
604 		return;
605 
606 	raw_spin_lock_irqsave(&dist->lpi_list_lock, flags);
607 
608 	if (unlikely(list_empty(&dist->lpi_translation_cache)))
609 		goto out;
610 
611 	/*
612 	 * We could have raced with another CPU caching the same
613 	 * translation behind our back, so let's check it is not in
614 	 * already
615 	 */
616 	db = its->vgic_its_base + GITS_TRANSLATER;
617 	if (__vgic_its_check_cache(dist, db, devid, eventid))
618 		goto out;
619 
620 	/* Always reuse the last entry (LRU policy) */
621 	cte = list_last_entry(&dist->lpi_translation_cache,
622 			      typeof(*cte), entry);
623 
624 	/*
625 	 * Caching the translation implies having an extra reference
626 	 * to the interrupt, so drop the potential reference on what
627 	 * was in the cache, and increment it on the new interrupt.
628 	 */
629 	if (cte->irq)
630 		__vgic_put_lpi_locked(kvm, cte->irq);
631 
632 	vgic_get_irq_kref(irq);
633 
634 	cte->db		= db;
635 	cte->devid	= devid;
636 	cte->eventid	= eventid;
637 	cte->irq	= irq;
638 
639 	/* Move the new translation to the head of the list */
640 	list_move(&cte->entry, &dist->lpi_translation_cache);
641 
642 out:
643 	raw_spin_unlock_irqrestore(&dist->lpi_list_lock, flags);
644 }
645 
646 void vgic_its_invalidate_cache(struct kvm *kvm)
647 {
648 	struct vgic_dist *dist = &kvm->arch.vgic;
649 	struct vgic_translation_cache_entry *cte;
650 	unsigned long flags;
651 
652 	raw_spin_lock_irqsave(&dist->lpi_list_lock, flags);
653 
654 	list_for_each_entry(cte, &dist->lpi_translation_cache, entry) {
655 		/*
656 		 * If we hit a NULL entry, there is nothing after this
657 		 * point.
658 		 */
659 		if (!cte->irq)
660 			break;
661 
662 		__vgic_put_lpi_locked(kvm, cte->irq);
663 		cte->irq = NULL;
664 	}
665 
666 	raw_spin_unlock_irqrestore(&dist->lpi_list_lock, flags);
667 }
668 
669 int vgic_its_resolve_lpi(struct kvm *kvm, struct vgic_its *its,
670 			 u32 devid, u32 eventid, struct vgic_irq **irq)
671 {
672 	struct kvm_vcpu *vcpu;
673 	struct its_ite *ite;
674 
675 	if (!its->enabled)
676 		return -EBUSY;
677 
678 	ite = find_ite(its, devid, eventid);
679 	if (!ite || !its_is_collection_mapped(ite->collection))
680 		return E_ITS_INT_UNMAPPED_INTERRUPT;
681 
682 	vcpu = kvm_get_vcpu(kvm, ite->collection->target_addr);
683 	if (!vcpu)
684 		return E_ITS_INT_UNMAPPED_INTERRUPT;
685 
686 	if (!vgic_lpis_enabled(vcpu))
687 		return -EBUSY;
688 
689 	vgic_its_cache_translation(kvm, its, devid, eventid, ite->irq);
690 
691 	*irq = ite->irq;
692 	return 0;
693 }
694 
695 struct vgic_its *vgic_msi_to_its(struct kvm *kvm, struct kvm_msi *msi)
696 {
697 	u64 address;
698 	struct kvm_io_device *kvm_io_dev;
699 	struct vgic_io_device *iodev;
700 
701 	if (!vgic_has_its(kvm))
702 		return ERR_PTR(-ENODEV);
703 
704 	if (!(msi->flags & KVM_MSI_VALID_DEVID))
705 		return ERR_PTR(-EINVAL);
706 
707 	address = (u64)msi->address_hi << 32 | msi->address_lo;
708 
709 	kvm_io_dev = kvm_io_bus_get_dev(kvm, KVM_MMIO_BUS, address);
710 	if (!kvm_io_dev)
711 		return ERR_PTR(-EINVAL);
712 
713 	if (kvm_io_dev->ops != &kvm_io_gic_ops)
714 		return ERR_PTR(-EINVAL);
715 
716 	iodev = container_of(kvm_io_dev, struct vgic_io_device, dev);
717 	if (iodev->iodev_type != IODEV_ITS)
718 		return ERR_PTR(-EINVAL);
719 
720 	return iodev->its;
721 }
722 
723 /*
724  * Find the target VCPU and the LPI number for a given devid/eventid pair
725  * and make this IRQ pending, possibly injecting it.
726  * Must be called with the its_lock mutex held.
727  * Returns 0 on success, a positive error value for any ITS mapping
728  * related errors and negative error values for generic errors.
729  */
730 static int vgic_its_trigger_msi(struct kvm *kvm, struct vgic_its *its,
731 				u32 devid, u32 eventid)
732 {
733 	struct vgic_irq *irq = NULL;
734 	unsigned long flags;
735 	int err;
736 
737 	err = vgic_its_resolve_lpi(kvm, its, devid, eventid, &irq);
738 	if (err)
739 		return err;
740 
741 	if (irq->hw)
742 		return irq_set_irqchip_state(irq->host_irq,
743 					     IRQCHIP_STATE_PENDING, true);
744 
745 	raw_spin_lock_irqsave(&irq->irq_lock, flags);
746 	irq->pending_latch = true;
747 	vgic_queue_irq_unlock(kvm, irq, flags);
748 
749 	return 0;
750 }
751 
752 int vgic_its_inject_cached_translation(struct kvm *kvm, struct kvm_msi *msi)
753 {
754 	struct vgic_irq *irq;
755 	unsigned long flags;
756 	phys_addr_t db;
757 
758 	db = (u64)msi->address_hi << 32 | msi->address_lo;
759 	irq = vgic_its_check_cache(kvm, db, msi->devid, msi->data);
760 	if (!irq)
761 		return -EWOULDBLOCK;
762 
763 	raw_spin_lock_irqsave(&irq->irq_lock, flags);
764 	irq->pending_latch = true;
765 	vgic_queue_irq_unlock(kvm, irq, flags);
766 
767 	return 0;
768 }
769 
770 /*
771  * Queries the KVM IO bus framework to get the ITS pointer from the given
772  * doorbell address.
773  * We then call vgic_its_trigger_msi() with the decoded data.
774  * According to the KVM_SIGNAL_MSI API description returns 1 on success.
775  */
776 int vgic_its_inject_msi(struct kvm *kvm, struct kvm_msi *msi)
777 {
778 	struct vgic_its *its;
779 	int ret;
780 
781 	if (!vgic_its_inject_cached_translation(kvm, msi))
782 		return 1;
783 
784 	its = vgic_msi_to_its(kvm, msi);
785 	if (IS_ERR(its))
786 		return PTR_ERR(its);
787 
788 	mutex_lock(&its->its_lock);
789 	ret = vgic_its_trigger_msi(kvm, its, msi->devid, msi->data);
790 	mutex_unlock(&its->its_lock);
791 
792 	if (ret < 0)
793 		return ret;
794 
795 	/*
796 	 * KVM_SIGNAL_MSI demands a return value > 0 for success and 0
797 	 * if the guest has blocked the MSI. So we map any LPI mapping
798 	 * related error to that.
799 	 */
800 	if (ret)
801 		return 0;
802 	else
803 		return 1;
804 }
805 
806 /* Requires the its_lock to be held. */
807 static void its_free_ite(struct kvm *kvm, struct its_ite *ite)
808 {
809 	list_del(&ite->ite_list);
810 
811 	/* This put matches the get in vgic_add_lpi. */
812 	if (ite->irq) {
813 		if (ite->irq->hw)
814 			WARN_ON(its_unmap_vlpi(ite->irq->host_irq));
815 
816 		vgic_put_irq(kvm, ite->irq);
817 	}
818 
819 	kfree(ite);
820 }
821 
822 static u64 its_cmd_mask_field(u64 *its_cmd, int word, int shift, int size)
823 {
824 	return (le64_to_cpu(its_cmd[word]) >> shift) & (BIT_ULL(size) - 1);
825 }
826 
827 #define its_cmd_get_command(cmd)	its_cmd_mask_field(cmd, 0,  0,  8)
828 #define its_cmd_get_deviceid(cmd)	its_cmd_mask_field(cmd, 0, 32, 32)
829 #define its_cmd_get_size(cmd)		(its_cmd_mask_field(cmd, 1,  0,  5) + 1)
830 #define its_cmd_get_id(cmd)		its_cmd_mask_field(cmd, 1,  0, 32)
831 #define its_cmd_get_physical_id(cmd)	its_cmd_mask_field(cmd, 1, 32, 32)
832 #define its_cmd_get_collection(cmd)	its_cmd_mask_field(cmd, 2,  0, 16)
833 #define its_cmd_get_ittaddr(cmd)	(its_cmd_mask_field(cmd, 2,  8, 44) << 8)
834 #define its_cmd_get_target_addr(cmd)	its_cmd_mask_field(cmd, 2, 16, 32)
835 #define its_cmd_get_validbit(cmd)	its_cmd_mask_field(cmd, 2, 63,  1)
836 
837 /*
838  * The DISCARD command frees an Interrupt Translation Table Entry (ITTE).
839  * Must be called with the its_lock mutex held.
840  */
841 static int vgic_its_cmd_handle_discard(struct kvm *kvm, struct vgic_its *its,
842 				       u64 *its_cmd)
843 {
844 	u32 device_id = its_cmd_get_deviceid(its_cmd);
845 	u32 event_id = its_cmd_get_id(its_cmd);
846 	struct its_ite *ite;
847 
848 	ite = find_ite(its, device_id, event_id);
849 	if (ite && its_is_collection_mapped(ite->collection)) {
850 		/*
851 		 * Though the spec talks about removing the pending state, we
852 		 * don't bother here since we clear the ITTE anyway and the
853 		 * pending state is a property of the ITTE struct.
854 		 */
855 		vgic_its_invalidate_cache(kvm);
856 
857 		its_free_ite(kvm, ite);
858 		return 0;
859 	}
860 
861 	return E_ITS_DISCARD_UNMAPPED_INTERRUPT;
862 }
863 
864 /*
865  * The MOVI command moves an ITTE to a different collection.
866  * Must be called with the its_lock mutex held.
867  */
868 static int vgic_its_cmd_handle_movi(struct kvm *kvm, struct vgic_its *its,
869 				    u64 *its_cmd)
870 {
871 	u32 device_id = its_cmd_get_deviceid(its_cmd);
872 	u32 event_id = its_cmd_get_id(its_cmd);
873 	u32 coll_id = its_cmd_get_collection(its_cmd);
874 	struct kvm_vcpu *vcpu;
875 	struct its_ite *ite;
876 	struct its_collection *collection;
877 
878 	ite = find_ite(its, device_id, event_id);
879 	if (!ite)
880 		return E_ITS_MOVI_UNMAPPED_INTERRUPT;
881 
882 	if (!its_is_collection_mapped(ite->collection))
883 		return E_ITS_MOVI_UNMAPPED_COLLECTION;
884 
885 	collection = find_collection(its, coll_id);
886 	if (!its_is_collection_mapped(collection))
887 		return E_ITS_MOVI_UNMAPPED_COLLECTION;
888 
889 	ite->collection = collection;
890 	vcpu = kvm_get_vcpu(kvm, collection->target_addr);
891 
892 	vgic_its_invalidate_cache(kvm);
893 
894 	return update_affinity(ite->irq, vcpu);
895 }
896 
897 static bool __is_visible_gfn_locked(struct vgic_its *its, gpa_t gpa)
898 {
899 	gfn_t gfn = gpa >> PAGE_SHIFT;
900 	int idx;
901 	bool ret;
902 
903 	idx = srcu_read_lock(&its->dev->kvm->srcu);
904 	ret = kvm_is_visible_gfn(its->dev->kvm, gfn);
905 	srcu_read_unlock(&its->dev->kvm->srcu, idx);
906 	return ret;
907 }
908 
909 /*
910  * Check whether an ID can be stored into the corresponding guest table.
911  * For a direct table this is pretty easy, but gets a bit nasty for
912  * indirect tables. We check whether the resulting guest physical address
913  * is actually valid (covered by a memslot and guest accessible).
914  * For this we have to read the respective first level entry.
915  */
916 static bool vgic_its_check_id(struct vgic_its *its, u64 baser, u32 id,
917 			      gpa_t *eaddr)
918 {
919 	int l1_tbl_size = GITS_BASER_NR_PAGES(baser) * SZ_64K;
920 	u64 indirect_ptr, type = GITS_BASER_TYPE(baser);
921 	phys_addr_t base = GITS_BASER_ADDR_48_to_52(baser);
922 	int esz = GITS_BASER_ENTRY_SIZE(baser);
923 	int index;
924 
925 	switch (type) {
926 	case GITS_BASER_TYPE_DEVICE:
927 		if (id >= BIT_ULL(VITS_TYPER_DEVBITS))
928 			return false;
929 		break;
930 	case GITS_BASER_TYPE_COLLECTION:
931 		/* as GITS_TYPER.CIL == 0, ITS supports 16-bit collection ID */
932 		if (id >= BIT_ULL(16))
933 			return false;
934 		break;
935 	default:
936 		return false;
937 	}
938 
939 	if (!(baser & GITS_BASER_INDIRECT)) {
940 		phys_addr_t addr;
941 
942 		if (id >= (l1_tbl_size / esz))
943 			return false;
944 
945 		addr = base + id * esz;
946 
947 		if (eaddr)
948 			*eaddr = addr;
949 
950 		return __is_visible_gfn_locked(its, addr);
951 	}
952 
953 	/* calculate and check the index into the 1st level */
954 	index = id / (SZ_64K / esz);
955 	if (index >= (l1_tbl_size / sizeof(u64)))
956 		return false;
957 
958 	/* Each 1st level entry is represented by a 64-bit value. */
959 	if (kvm_read_guest_lock(its->dev->kvm,
960 			   base + index * sizeof(indirect_ptr),
961 			   &indirect_ptr, sizeof(indirect_ptr)))
962 		return false;
963 
964 	indirect_ptr = le64_to_cpu(indirect_ptr);
965 
966 	/* check the valid bit of the first level entry */
967 	if (!(indirect_ptr & BIT_ULL(63)))
968 		return false;
969 
970 	/* Mask the guest physical address and calculate the frame number. */
971 	indirect_ptr &= GENMASK_ULL(51, 16);
972 
973 	/* Find the address of the actual entry */
974 	index = id % (SZ_64K / esz);
975 	indirect_ptr += index * esz;
976 
977 	if (eaddr)
978 		*eaddr = indirect_ptr;
979 
980 	return __is_visible_gfn_locked(its, indirect_ptr);
981 }
982 
983 /*
984  * Check whether an event ID can be stored in the corresponding Interrupt
985  * Translation Table, which starts at device->itt_addr.
986  */
987 static bool vgic_its_check_event_id(struct vgic_its *its, struct its_device *device,
988 		u32 event_id)
989 {
990 	const struct vgic_its_abi *abi = vgic_its_get_abi(its);
991 	int ite_esz = abi->ite_esz;
992 	gpa_t gpa;
993 
994 	/* max table size is: BIT_ULL(device->num_eventid_bits) * ite_esz */
995 	if (event_id >= BIT_ULL(device->num_eventid_bits))
996 		return false;
997 
998 	gpa = device->itt_addr + event_id * ite_esz;
999 	return __is_visible_gfn_locked(its, gpa);
1000 }
1001 
1002 /*
1003  * Add a new collection into the ITS collection table.
1004  * Returns 0 on success, and a negative error value for generic errors.
1005  */
1006 static int vgic_its_alloc_collection(struct vgic_its *its,
1007 				     struct its_collection **colp,
1008 				     u32 coll_id)
1009 {
1010 	struct its_collection *collection;
1011 
1012 	collection = kzalloc(sizeof(*collection), GFP_KERNEL_ACCOUNT);
1013 	if (!collection)
1014 		return -ENOMEM;
1015 
1016 	collection->collection_id = coll_id;
1017 	collection->target_addr = COLLECTION_NOT_MAPPED;
1018 
1019 	list_add_tail(&collection->coll_list, &its->collection_list);
1020 	*colp = collection;
1021 
1022 	return 0;
1023 }
1024 
1025 static void vgic_its_free_collection(struct vgic_its *its, u32 coll_id)
1026 {
1027 	struct its_collection *collection;
1028 	struct its_device *device;
1029 	struct its_ite *ite;
1030 
1031 	/*
1032 	 * Clearing the mapping for that collection ID removes the
1033 	 * entry from the list. If there wasn't any before, we can
1034 	 * go home early.
1035 	 */
1036 	collection = find_collection(its, coll_id);
1037 	if (!collection)
1038 		return;
1039 
1040 	for_each_lpi_its(device, ite, its)
1041 		if (ite->collection &&
1042 		    ite->collection->collection_id == coll_id)
1043 			ite->collection = NULL;
1044 
1045 	list_del(&collection->coll_list);
1046 	kfree(collection);
1047 }
1048 
1049 /* Must be called with its_lock mutex held */
1050 static struct its_ite *vgic_its_alloc_ite(struct its_device *device,
1051 					  struct its_collection *collection,
1052 					  u32 event_id)
1053 {
1054 	struct its_ite *ite;
1055 
1056 	ite = kzalloc(sizeof(*ite), GFP_KERNEL_ACCOUNT);
1057 	if (!ite)
1058 		return ERR_PTR(-ENOMEM);
1059 
1060 	ite->event_id	= event_id;
1061 	ite->collection = collection;
1062 
1063 	list_add_tail(&ite->ite_list, &device->itt_head);
1064 	return ite;
1065 }
1066 
1067 /*
1068  * The MAPTI and MAPI commands map LPIs to ITTEs.
1069  * Must be called with its_lock mutex held.
1070  */
1071 static int vgic_its_cmd_handle_mapi(struct kvm *kvm, struct vgic_its *its,
1072 				    u64 *its_cmd)
1073 {
1074 	u32 device_id = its_cmd_get_deviceid(its_cmd);
1075 	u32 event_id = its_cmd_get_id(its_cmd);
1076 	u32 coll_id = its_cmd_get_collection(its_cmd);
1077 	struct its_ite *ite;
1078 	struct kvm_vcpu *vcpu = NULL;
1079 	struct its_device *device;
1080 	struct its_collection *collection, *new_coll = NULL;
1081 	struct vgic_irq *irq;
1082 	int lpi_nr;
1083 
1084 	device = find_its_device(its, device_id);
1085 	if (!device)
1086 		return E_ITS_MAPTI_UNMAPPED_DEVICE;
1087 
1088 	if (!vgic_its_check_event_id(its, device, event_id))
1089 		return E_ITS_MAPTI_ID_OOR;
1090 
1091 	if (its_cmd_get_command(its_cmd) == GITS_CMD_MAPTI)
1092 		lpi_nr = its_cmd_get_physical_id(its_cmd);
1093 	else
1094 		lpi_nr = event_id;
1095 	if (lpi_nr < GIC_LPI_OFFSET ||
1096 	    lpi_nr >= max_lpis_propbaser(kvm->arch.vgic.propbaser))
1097 		return E_ITS_MAPTI_PHYSICALID_OOR;
1098 
1099 	/* If there is an existing mapping, behavior is UNPREDICTABLE. */
1100 	if (find_ite(its, device_id, event_id))
1101 		return 0;
1102 
1103 	collection = find_collection(its, coll_id);
1104 	if (!collection) {
1105 		int ret;
1106 
1107 		if (!vgic_its_check_id(its, its->baser_coll_table, coll_id, NULL))
1108 			return E_ITS_MAPC_COLLECTION_OOR;
1109 
1110 		ret = vgic_its_alloc_collection(its, &collection, coll_id);
1111 		if (ret)
1112 			return ret;
1113 		new_coll = collection;
1114 	}
1115 
1116 	ite = vgic_its_alloc_ite(device, collection, event_id);
1117 	if (IS_ERR(ite)) {
1118 		if (new_coll)
1119 			vgic_its_free_collection(its, coll_id);
1120 		return PTR_ERR(ite);
1121 	}
1122 
1123 	if (its_is_collection_mapped(collection))
1124 		vcpu = kvm_get_vcpu(kvm, collection->target_addr);
1125 
1126 	irq = vgic_add_lpi(kvm, lpi_nr, vcpu);
1127 	if (IS_ERR(irq)) {
1128 		if (new_coll)
1129 			vgic_its_free_collection(its, coll_id);
1130 		its_free_ite(kvm, ite);
1131 		return PTR_ERR(irq);
1132 	}
1133 	ite->irq = irq;
1134 
1135 	return 0;
1136 }
1137 
1138 /* Requires the its_lock to be held. */
1139 static void vgic_its_free_device(struct kvm *kvm, struct its_device *device)
1140 {
1141 	struct its_ite *ite, *temp;
1142 
1143 	/*
1144 	 * The spec says that unmapping a device with still valid
1145 	 * ITTEs associated is UNPREDICTABLE. We remove all ITTEs,
1146 	 * since we cannot leave the memory unreferenced.
1147 	 */
1148 	list_for_each_entry_safe(ite, temp, &device->itt_head, ite_list)
1149 		its_free_ite(kvm, ite);
1150 
1151 	vgic_its_invalidate_cache(kvm);
1152 
1153 	list_del(&device->dev_list);
1154 	kfree(device);
1155 }
1156 
1157 /* its lock must be held */
1158 static void vgic_its_free_device_list(struct kvm *kvm, struct vgic_its *its)
1159 {
1160 	struct its_device *cur, *temp;
1161 
1162 	list_for_each_entry_safe(cur, temp, &its->device_list, dev_list)
1163 		vgic_its_free_device(kvm, cur);
1164 }
1165 
1166 /* its lock must be held */
1167 static void vgic_its_free_collection_list(struct kvm *kvm, struct vgic_its *its)
1168 {
1169 	struct its_collection *cur, *temp;
1170 
1171 	list_for_each_entry_safe(cur, temp, &its->collection_list, coll_list)
1172 		vgic_its_free_collection(its, cur->collection_id);
1173 }
1174 
1175 /* Must be called with its_lock mutex held */
1176 static struct its_device *vgic_its_alloc_device(struct vgic_its *its,
1177 						u32 device_id, gpa_t itt_addr,
1178 						u8 num_eventid_bits)
1179 {
1180 	struct its_device *device;
1181 
1182 	device = kzalloc(sizeof(*device), GFP_KERNEL_ACCOUNT);
1183 	if (!device)
1184 		return ERR_PTR(-ENOMEM);
1185 
1186 	device->device_id = device_id;
1187 	device->itt_addr = itt_addr;
1188 	device->num_eventid_bits = num_eventid_bits;
1189 	INIT_LIST_HEAD(&device->itt_head);
1190 
1191 	list_add_tail(&device->dev_list, &its->device_list);
1192 	return device;
1193 }
1194 
1195 /*
1196  * MAPD maps or unmaps a device ID to Interrupt Translation Tables (ITTs).
1197  * Must be called with the its_lock mutex held.
1198  */
1199 static int vgic_its_cmd_handle_mapd(struct kvm *kvm, struct vgic_its *its,
1200 				    u64 *its_cmd)
1201 {
1202 	u32 device_id = its_cmd_get_deviceid(its_cmd);
1203 	bool valid = its_cmd_get_validbit(its_cmd);
1204 	u8 num_eventid_bits = its_cmd_get_size(its_cmd);
1205 	gpa_t itt_addr = its_cmd_get_ittaddr(its_cmd);
1206 	struct its_device *device;
1207 
1208 	if (!vgic_its_check_id(its, its->baser_device_table, device_id, NULL))
1209 		return E_ITS_MAPD_DEVICE_OOR;
1210 
1211 	if (valid && num_eventid_bits > VITS_TYPER_IDBITS)
1212 		return E_ITS_MAPD_ITTSIZE_OOR;
1213 
1214 	device = find_its_device(its, device_id);
1215 
1216 	/*
1217 	 * The spec says that calling MAPD on an already mapped device
1218 	 * invalidates all cached data for this device. We implement this
1219 	 * by removing the mapping and re-establishing it.
1220 	 */
1221 	if (device)
1222 		vgic_its_free_device(kvm, device);
1223 
1224 	/*
1225 	 * The spec does not say whether unmapping a not-mapped device
1226 	 * is an error, so we are done in any case.
1227 	 */
1228 	if (!valid)
1229 		return 0;
1230 
1231 	device = vgic_its_alloc_device(its, device_id, itt_addr,
1232 				       num_eventid_bits);
1233 
1234 	return PTR_ERR_OR_ZERO(device);
1235 }
1236 
1237 /*
1238  * The MAPC command maps collection IDs to redistributors.
1239  * Must be called with the its_lock mutex held.
1240  */
1241 static int vgic_its_cmd_handle_mapc(struct kvm *kvm, struct vgic_its *its,
1242 				    u64 *its_cmd)
1243 {
1244 	u16 coll_id;
1245 	u32 target_addr;
1246 	struct its_collection *collection;
1247 	bool valid;
1248 
1249 	valid = its_cmd_get_validbit(its_cmd);
1250 	coll_id = its_cmd_get_collection(its_cmd);
1251 	target_addr = its_cmd_get_target_addr(its_cmd);
1252 
1253 	if (target_addr >= atomic_read(&kvm->online_vcpus))
1254 		return E_ITS_MAPC_PROCNUM_OOR;
1255 
1256 	if (!valid) {
1257 		vgic_its_free_collection(its, coll_id);
1258 		vgic_its_invalidate_cache(kvm);
1259 	} else {
1260 		collection = find_collection(its, coll_id);
1261 
1262 		if (!collection) {
1263 			int ret;
1264 
1265 			if (!vgic_its_check_id(its, its->baser_coll_table,
1266 						coll_id, NULL))
1267 				return E_ITS_MAPC_COLLECTION_OOR;
1268 
1269 			ret = vgic_its_alloc_collection(its, &collection,
1270 							coll_id);
1271 			if (ret)
1272 				return ret;
1273 			collection->target_addr = target_addr;
1274 		} else {
1275 			collection->target_addr = target_addr;
1276 			update_affinity_collection(kvm, its, collection);
1277 		}
1278 	}
1279 
1280 	return 0;
1281 }
1282 
1283 /*
1284  * The CLEAR command removes the pending state for a particular LPI.
1285  * Must be called with the its_lock mutex held.
1286  */
1287 static int vgic_its_cmd_handle_clear(struct kvm *kvm, struct vgic_its *its,
1288 				     u64 *its_cmd)
1289 {
1290 	u32 device_id = its_cmd_get_deviceid(its_cmd);
1291 	u32 event_id = its_cmd_get_id(its_cmd);
1292 	struct its_ite *ite;
1293 
1294 
1295 	ite = find_ite(its, device_id, event_id);
1296 	if (!ite)
1297 		return E_ITS_CLEAR_UNMAPPED_INTERRUPT;
1298 
1299 	ite->irq->pending_latch = false;
1300 
1301 	if (ite->irq->hw)
1302 		return irq_set_irqchip_state(ite->irq->host_irq,
1303 					     IRQCHIP_STATE_PENDING, false);
1304 
1305 	return 0;
1306 }
1307 
1308 int vgic_its_inv_lpi(struct kvm *kvm, struct vgic_irq *irq)
1309 {
1310 	return update_lpi_config(kvm, irq, NULL, true);
1311 }
1312 
1313 /*
1314  * The INV command syncs the configuration bits from the memory table.
1315  * Must be called with the its_lock mutex held.
1316  */
1317 static int vgic_its_cmd_handle_inv(struct kvm *kvm, struct vgic_its *its,
1318 				   u64 *its_cmd)
1319 {
1320 	u32 device_id = its_cmd_get_deviceid(its_cmd);
1321 	u32 event_id = its_cmd_get_id(its_cmd);
1322 	struct its_ite *ite;
1323 
1324 
1325 	ite = find_ite(its, device_id, event_id);
1326 	if (!ite)
1327 		return E_ITS_INV_UNMAPPED_INTERRUPT;
1328 
1329 	return vgic_its_inv_lpi(kvm, ite->irq);
1330 }
1331 
1332 /**
1333  * vgic_its_invall - invalidate all LPIs targetting a given vcpu
1334  * @vcpu: the vcpu for which the RD is targetted by an invalidation
1335  *
1336  * Contrary to the INVALL command, this targets a RD instead of a
1337  * collection, and we don't need to hold the its_lock, since no ITS is
1338  * involved here.
1339  */
1340 int vgic_its_invall(struct kvm_vcpu *vcpu)
1341 {
1342 	struct kvm *kvm = vcpu->kvm;
1343 	int irq_count, i = 0;
1344 	u32 *intids;
1345 
1346 	irq_count = vgic_copy_lpi_list(kvm, vcpu, &intids);
1347 	if (irq_count < 0)
1348 		return irq_count;
1349 
1350 	for (i = 0; i < irq_count; i++) {
1351 		struct vgic_irq *irq = vgic_get_irq(kvm, NULL, intids[i]);
1352 		if (!irq)
1353 			continue;
1354 		update_lpi_config(kvm, irq, vcpu, false);
1355 		vgic_put_irq(kvm, irq);
1356 	}
1357 
1358 	kfree(intids);
1359 
1360 	if (vcpu->arch.vgic_cpu.vgic_v3.its_vpe.its_vm)
1361 		its_invall_vpe(&vcpu->arch.vgic_cpu.vgic_v3.its_vpe);
1362 
1363 	return 0;
1364 }
1365 
1366 /*
1367  * The INVALL command requests flushing of all IRQ data in this collection.
1368  * Find the VCPU mapped to that collection, then iterate over the VM's list
1369  * of mapped LPIs and update the configuration for each IRQ which targets
1370  * the specified vcpu. The configuration will be read from the in-memory
1371  * configuration table.
1372  * Must be called with the its_lock mutex held.
1373  */
1374 static int vgic_its_cmd_handle_invall(struct kvm *kvm, struct vgic_its *its,
1375 				      u64 *its_cmd)
1376 {
1377 	u32 coll_id = its_cmd_get_collection(its_cmd);
1378 	struct its_collection *collection;
1379 	struct kvm_vcpu *vcpu;
1380 
1381 	collection = find_collection(its, coll_id);
1382 	if (!its_is_collection_mapped(collection))
1383 		return E_ITS_INVALL_UNMAPPED_COLLECTION;
1384 
1385 	vcpu = kvm_get_vcpu(kvm, collection->target_addr);
1386 	vgic_its_invall(vcpu);
1387 
1388 	return 0;
1389 }
1390 
1391 /*
1392  * The MOVALL command moves the pending state of all IRQs targeting one
1393  * redistributor to another. We don't hold the pending state in the VCPUs,
1394  * but in the IRQs instead, so there is really not much to do for us here.
1395  * However the spec says that no IRQ must target the old redistributor
1396  * afterwards, so we make sure that no LPI is using the associated target_vcpu.
1397  * This command affects all LPIs in the system that target that redistributor.
1398  */
1399 static int vgic_its_cmd_handle_movall(struct kvm *kvm, struct vgic_its *its,
1400 				      u64 *its_cmd)
1401 {
1402 	u32 target1_addr = its_cmd_get_target_addr(its_cmd);
1403 	u32 target2_addr = its_cmd_mask_field(its_cmd, 3, 16, 32);
1404 	struct kvm_vcpu *vcpu1, *vcpu2;
1405 	struct vgic_irq *irq;
1406 	u32 *intids;
1407 	int irq_count, i;
1408 
1409 	if (target1_addr >= atomic_read(&kvm->online_vcpus) ||
1410 	    target2_addr >= atomic_read(&kvm->online_vcpus))
1411 		return E_ITS_MOVALL_PROCNUM_OOR;
1412 
1413 	if (target1_addr == target2_addr)
1414 		return 0;
1415 
1416 	vcpu1 = kvm_get_vcpu(kvm, target1_addr);
1417 	vcpu2 = kvm_get_vcpu(kvm, target2_addr);
1418 
1419 	irq_count = vgic_copy_lpi_list(kvm, vcpu1, &intids);
1420 	if (irq_count < 0)
1421 		return irq_count;
1422 
1423 	for (i = 0; i < irq_count; i++) {
1424 		irq = vgic_get_irq(kvm, NULL, intids[i]);
1425 
1426 		update_affinity(irq, vcpu2);
1427 
1428 		vgic_put_irq(kvm, irq);
1429 	}
1430 
1431 	vgic_its_invalidate_cache(kvm);
1432 
1433 	kfree(intids);
1434 	return 0;
1435 }
1436 
1437 /*
1438  * The INT command injects the LPI associated with that DevID/EvID pair.
1439  * Must be called with the its_lock mutex held.
1440  */
1441 static int vgic_its_cmd_handle_int(struct kvm *kvm, struct vgic_its *its,
1442 				   u64 *its_cmd)
1443 {
1444 	u32 msi_data = its_cmd_get_id(its_cmd);
1445 	u64 msi_devid = its_cmd_get_deviceid(its_cmd);
1446 
1447 	return vgic_its_trigger_msi(kvm, its, msi_devid, msi_data);
1448 }
1449 
1450 /*
1451  * This function is called with the its_cmd lock held, but the ITS data
1452  * structure lock dropped.
1453  */
1454 static int vgic_its_handle_command(struct kvm *kvm, struct vgic_its *its,
1455 				   u64 *its_cmd)
1456 {
1457 	int ret = -ENODEV;
1458 
1459 	mutex_lock(&its->its_lock);
1460 	switch (its_cmd_get_command(its_cmd)) {
1461 	case GITS_CMD_MAPD:
1462 		ret = vgic_its_cmd_handle_mapd(kvm, its, its_cmd);
1463 		break;
1464 	case GITS_CMD_MAPC:
1465 		ret = vgic_its_cmd_handle_mapc(kvm, its, its_cmd);
1466 		break;
1467 	case GITS_CMD_MAPI:
1468 		ret = vgic_its_cmd_handle_mapi(kvm, its, its_cmd);
1469 		break;
1470 	case GITS_CMD_MAPTI:
1471 		ret = vgic_its_cmd_handle_mapi(kvm, its, its_cmd);
1472 		break;
1473 	case GITS_CMD_MOVI:
1474 		ret = vgic_its_cmd_handle_movi(kvm, its, its_cmd);
1475 		break;
1476 	case GITS_CMD_DISCARD:
1477 		ret = vgic_its_cmd_handle_discard(kvm, its, its_cmd);
1478 		break;
1479 	case GITS_CMD_CLEAR:
1480 		ret = vgic_its_cmd_handle_clear(kvm, its, its_cmd);
1481 		break;
1482 	case GITS_CMD_MOVALL:
1483 		ret = vgic_its_cmd_handle_movall(kvm, its, its_cmd);
1484 		break;
1485 	case GITS_CMD_INT:
1486 		ret = vgic_its_cmd_handle_int(kvm, its, its_cmd);
1487 		break;
1488 	case GITS_CMD_INV:
1489 		ret = vgic_its_cmd_handle_inv(kvm, its, its_cmd);
1490 		break;
1491 	case GITS_CMD_INVALL:
1492 		ret = vgic_its_cmd_handle_invall(kvm, its, its_cmd);
1493 		break;
1494 	case GITS_CMD_SYNC:
1495 		/* we ignore this command: we are in sync all of the time */
1496 		ret = 0;
1497 		break;
1498 	}
1499 	mutex_unlock(&its->its_lock);
1500 
1501 	return ret;
1502 }
1503 
1504 static u64 vgic_sanitise_its_baser(u64 reg)
1505 {
1506 	reg = vgic_sanitise_field(reg, GITS_BASER_SHAREABILITY_MASK,
1507 				  GITS_BASER_SHAREABILITY_SHIFT,
1508 				  vgic_sanitise_shareability);
1509 	reg = vgic_sanitise_field(reg, GITS_BASER_INNER_CACHEABILITY_MASK,
1510 				  GITS_BASER_INNER_CACHEABILITY_SHIFT,
1511 				  vgic_sanitise_inner_cacheability);
1512 	reg = vgic_sanitise_field(reg, GITS_BASER_OUTER_CACHEABILITY_MASK,
1513 				  GITS_BASER_OUTER_CACHEABILITY_SHIFT,
1514 				  vgic_sanitise_outer_cacheability);
1515 
1516 	/* We support only one (ITS) page size: 64K */
1517 	reg = (reg & ~GITS_BASER_PAGE_SIZE_MASK) | GITS_BASER_PAGE_SIZE_64K;
1518 
1519 	return reg;
1520 }
1521 
1522 static u64 vgic_sanitise_its_cbaser(u64 reg)
1523 {
1524 	reg = vgic_sanitise_field(reg, GITS_CBASER_SHAREABILITY_MASK,
1525 				  GITS_CBASER_SHAREABILITY_SHIFT,
1526 				  vgic_sanitise_shareability);
1527 	reg = vgic_sanitise_field(reg, GITS_CBASER_INNER_CACHEABILITY_MASK,
1528 				  GITS_CBASER_INNER_CACHEABILITY_SHIFT,
1529 				  vgic_sanitise_inner_cacheability);
1530 	reg = vgic_sanitise_field(reg, GITS_CBASER_OUTER_CACHEABILITY_MASK,
1531 				  GITS_CBASER_OUTER_CACHEABILITY_SHIFT,
1532 				  vgic_sanitise_outer_cacheability);
1533 
1534 	/* Sanitise the physical address to be 64k aligned. */
1535 	reg &= ~GENMASK_ULL(15, 12);
1536 
1537 	return reg;
1538 }
1539 
1540 static unsigned long vgic_mmio_read_its_cbaser(struct kvm *kvm,
1541 					       struct vgic_its *its,
1542 					       gpa_t addr, unsigned int len)
1543 {
1544 	return extract_bytes(its->cbaser, addr & 7, len);
1545 }
1546 
1547 static void vgic_mmio_write_its_cbaser(struct kvm *kvm, struct vgic_its *its,
1548 				       gpa_t addr, unsigned int len,
1549 				       unsigned long val)
1550 {
1551 	/* When GITS_CTLR.Enable is 1, this register is RO. */
1552 	if (its->enabled)
1553 		return;
1554 
1555 	mutex_lock(&its->cmd_lock);
1556 	its->cbaser = update_64bit_reg(its->cbaser, addr & 7, len, val);
1557 	its->cbaser = vgic_sanitise_its_cbaser(its->cbaser);
1558 	its->creadr = 0;
1559 	/*
1560 	 * CWRITER is architecturally UNKNOWN on reset, but we need to reset
1561 	 * it to CREADR to make sure we start with an empty command buffer.
1562 	 */
1563 	its->cwriter = its->creadr;
1564 	mutex_unlock(&its->cmd_lock);
1565 }
1566 
1567 #define ITS_CMD_BUFFER_SIZE(baser)	((((baser) & 0xff) + 1) << 12)
1568 #define ITS_CMD_SIZE			32
1569 #define ITS_CMD_OFFSET(reg)		((reg) & GENMASK(19, 5))
1570 
1571 /* Must be called with the cmd_lock held. */
1572 static void vgic_its_process_commands(struct kvm *kvm, struct vgic_its *its)
1573 {
1574 	gpa_t cbaser;
1575 	u64 cmd_buf[4];
1576 
1577 	/* Commands are only processed when the ITS is enabled. */
1578 	if (!its->enabled)
1579 		return;
1580 
1581 	cbaser = GITS_CBASER_ADDRESS(its->cbaser);
1582 
1583 	while (its->cwriter != its->creadr) {
1584 		int ret = kvm_read_guest_lock(kvm, cbaser + its->creadr,
1585 					      cmd_buf, ITS_CMD_SIZE);
1586 		/*
1587 		 * If kvm_read_guest() fails, this could be due to the guest
1588 		 * programming a bogus value in CBASER or something else going
1589 		 * wrong from which we cannot easily recover.
1590 		 * According to section 6.3.2 in the GICv3 spec we can just
1591 		 * ignore that command then.
1592 		 */
1593 		if (!ret)
1594 			vgic_its_handle_command(kvm, its, cmd_buf);
1595 
1596 		its->creadr += ITS_CMD_SIZE;
1597 		if (its->creadr == ITS_CMD_BUFFER_SIZE(its->cbaser))
1598 			its->creadr = 0;
1599 	}
1600 }
1601 
1602 /*
1603  * By writing to CWRITER the guest announces new commands to be processed.
1604  * To avoid any races in the first place, we take the its_cmd lock, which
1605  * protects our ring buffer variables, so that there is only one user
1606  * per ITS handling commands at a given time.
1607  */
1608 static void vgic_mmio_write_its_cwriter(struct kvm *kvm, struct vgic_its *its,
1609 					gpa_t addr, unsigned int len,
1610 					unsigned long val)
1611 {
1612 	u64 reg;
1613 
1614 	if (!its)
1615 		return;
1616 
1617 	mutex_lock(&its->cmd_lock);
1618 
1619 	reg = update_64bit_reg(its->cwriter, addr & 7, len, val);
1620 	reg = ITS_CMD_OFFSET(reg);
1621 	if (reg >= ITS_CMD_BUFFER_SIZE(its->cbaser)) {
1622 		mutex_unlock(&its->cmd_lock);
1623 		return;
1624 	}
1625 	its->cwriter = reg;
1626 
1627 	vgic_its_process_commands(kvm, its);
1628 
1629 	mutex_unlock(&its->cmd_lock);
1630 }
1631 
1632 static unsigned long vgic_mmio_read_its_cwriter(struct kvm *kvm,
1633 						struct vgic_its *its,
1634 						gpa_t addr, unsigned int len)
1635 {
1636 	return extract_bytes(its->cwriter, addr & 0x7, len);
1637 }
1638 
1639 static unsigned long vgic_mmio_read_its_creadr(struct kvm *kvm,
1640 					       struct vgic_its *its,
1641 					       gpa_t addr, unsigned int len)
1642 {
1643 	return extract_bytes(its->creadr, addr & 0x7, len);
1644 }
1645 
1646 static int vgic_mmio_uaccess_write_its_creadr(struct kvm *kvm,
1647 					      struct vgic_its *its,
1648 					      gpa_t addr, unsigned int len,
1649 					      unsigned long val)
1650 {
1651 	u32 cmd_offset;
1652 	int ret = 0;
1653 
1654 	mutex_lock(&its->cmd_lock);
1655 
1656 	if (its->enabled) {
1657 		ret = -EBUSY;
1658 		goto out;
1659 	}
1660 
1661 	cmd_offset = ITS_CMD_OFFSET(val);
1662 	if (cmd_offset >= ITS_CMD_BUFFER_SIZE(its->cbaser)) {
1663 		ret = -EINVAL;
1664 		goto out;
1665 	}
1666 
1667 	its->creadr = cmd_offset;
1668 out:
1669 	mutex_unlock(&its->cmd_lock);
1670 	return ret;
1671 }
1672 
1673 #define BASER_INDEX(addr) (((addr) / sizeof(u64)) & 0x7)
1674 static unsigned long vgic_mmio_read_its_baser(struct kvm *kvm,
1675 					      struct vgic_its *its,
1676 					      gpa_t addr, unsigned int len)
1677 {
1678 	u64 reg;
1679 
1680 	switch (BASER_INDEX(addr)) {
1681 	case 0:
1682 		reg = its->baser_device_table;
1683 		break;
1684 	case 1:
1685 		reg = its->baser_coll_table;
1686 		break;
1687 	default:
1688 		reg = 0;
1689 		break;
1690 	}
1691 
1692 	return extract_bytes(reg, addr & 7, len);
1693 }
1694 
1695 #define GITS_BASER_RO_MASK	(GENMASK_ULL(52, 48) | GENMASK_ULL(58, 56))
1696 static void vgic_mmio_write_its_baser(struct kvm *kvm,
1697 				      struct vgic_its *its,
1698 				      gpa_t addr, unsigned int len,
1699 				      unsigned long val)
1700 {
1701 	const struct vgic_its_abi *abi = vgic_its_get_abi(its);
1702 	u64 entry_size, table_type;
1703 	u64 reg, *regptr, clearbits = 0;
1704 
1705 	/* When GITS_CTLR.Enable is 1, we ignore write accesses. */
1706 	if (its->enabled)
1707 		return;
1708 
1709 	switch (BASER_INDEX(addr)) {
1710 	case 0:
1711 		regptr = &its->baser_device_table;
1712 		entry_size = abi->dte_esz;
1713 		table_type = GITS_BASER_TYPE_DEVICE;
1714 		break;
1715 	case 1:
1716 		regptr = &its->baser_coll_table;
1717 		entry_size = abi->cte_esz;
1718 		table_type = GITS_BASER_TYPE_COLLECTION;
1719 		clearbits = GITS_BASER_INDIRECT;
1720 		break;
1721 	default:
1722 		return;
1723 	}
1724 
1725 	reg = update_64bit_reg(*regptr, addr & 7, len, val);
1726 	reg &= ~GITS_BASER_RO_MASK;
1727 	reg &= ~clearbits;
1728 
1729 	reg |= (entry_size - 1) << GITS_BASER_ENTRY_SIZE_SHIFT;
1730 	reg |= table_type << GITS_BASER_TYPE_SHIFT;
1731 	reg = vgic_sanitise_its_baser(reg);
1732 
1733 	*regptr = reg;
1734 
1735 	if (!(reg & GITS_BASER_VALID)) {
1736 		/* Take the its_lock to prevent a race with a save/restore */
1737 		mutex_lock(&its->its_lock);
1738 		switch (table_type) {
1739 		case GITS_BASER_TYPE_DEVICE:
1740 			vgic_its_free_device_list(kvm, its);
1741 			break;
1742 		case GITS_BASER_TYPE_COLLECTION:
1743 			vgic_its_free_collection_list(kvm, its);
1744 			break;
1745 		}
1746 		mutex_unlock(&its->its_lock);
1747 	}
1748 }
1749 
1750 static unsigned long vgic_mmio_read_its_ctlr(struct kvm *vcpu,
1751 					     struct vgic_its *its,
1752 					     gpa_t addr, unsigned int len)
1753 {
1754 	u32 reg = 0;
1755 
1756 	mutex_lock(&its->cmd_lock);
1757 	if (its->creadr == its->cwriter)
1758 		reg |= GITS_CTLR_QUIESCENT;
1759 	if (its->enabled)
1760 		reg |= GITS_CTLR_ENABLE;
1761 	mutex_unlock(&its->cmd_lock);
1762 
1763 	return reg;
1764 }
1765 
1766 static void vgic_mmio_write_its_ctlr(struct kvm *kvm, struct vgic_its *its,
1767 				     gpa_t addr, unsigned int len,
1768 				     unsigned long val)
1769 {
1770 	mutex_lock(&its->cmd_lock);
1771 
1772 	/*
1773 	 * It is UNPREDICTABLE to enable the ITS if any of the CBASER or
1774 	 * device/collection BASER are invalid
1775 	 */
1776 	if (!its->enabled && (val & GITS_CTLR_ENABLE) &&
1777 		(!(its->baser_device_table & GITS_BASER_VALID) ||
1778 		 !(its->baser_coll_table & GITS_BASER_VALID) ||
1779 		 !(its->cbaser & GITS_CBASER_VALID)))
1780 		goto out;
1781 
1782 	its->enabled = !!(val & GITS_CTLR_ENABLE);
1783 	if (!its->enabled)
1784 		vgic_its_invalidate_cache(kvm);
1785 
1786 	/*
1787 	 * Try to process any pending commands. This function bails out early
1788 	 * if the ITS is disabled or no commands have been queued.
1789 	 */
1790 	vgic_its_process_commands(kvm, its);
1791 
1792 out:
1793 	mutex_unlock(&its->cmd_lock);
1794 }
1795 
1796 #define REGISTER_ITS_DESC(off, rd, wr, length, acc)		\
1797 {								\
1798 	.reg_offset = off,					\
1799 	.len = length,						\
1800 	.access_flags = acc,					\
1801 	.its_read = rd,						\
1802 	.its_write = wr,					\
1803 }
1804 
1805 #define REGISTER_ITS_DESC_UACCESS(off, rd, wr, uwr, length, acc)\
1806 {								\
1807 	.reg_offset = off,					\
1808 	.len = length,						\
1809 	.access_flags = acc,					\
1810 	.its_read = rd,						\
1811 	.its_write = wr,					\
1812 	.uaccess_its_write = uwr,				\
1813 }
1814 
1815 static void its_mmio_write_wi(struct kvm *kvm, struct vgic_its *its,
1816 			      gpa_t addr, unsigned int len, unsigned long val)
1817 {
1818 	/* Ignore */
1819 }
1820 
1821 static struct vgic_register_region its_registers[] = {
1822 	REGISTER_ITS_DESC(GITS_CTLR,
1823 		vgic_mmio_read_its_ctlr, vgic_mmio_write_its_ctlr, 4,
1824 		VGIC_ACCESS_32bit),
1825 	REGISTER_ITS_DESC_UACCESS(GITS_IIDR,
1826 		vgic_mmio_read_its_iidr, its_mmio_write_wi,
1827 		vgic_mmio_uaccess_write_its_iidr, 4,
1828 		VGIC_ACCESS_32bit),
1829 	REGISTER_ITS_DESC(GITS_TYPER,
1830 		vgic_mmio_read_its_typer, its_mmio_write_wi, 8,
1831 		VGIC_ACCESS_64bit | VGIC_ACCESS_32bit),
1832 	REGISTER_ITS_DESC(GITS_CBASER,
1833 		vgic_mmio_read_its_cbaser, vgic_mmio_write_its_cbaser, 8,
1834 		VGIC_ACCESS_64bit | VGIC_ACCESS_32bit),
1835 	REGISTER_ITS_DESC(GITS_CWRITER,
1836 		vgic_mmio_read_its_cwriter, vgic_mmio_write_its_cwriter, 8,
1837 		VGIC_ACCESS_64bit | VGIC_ACCESS_32bit),
1838 	REGISTER_ITS_DESC_UACCESS(GITS_CREADR,
1839 		vgic_mmio_read_its_creadr, its_mmio_write_wi,
1840 		vgic_mmio_uaccess_write_its_creadr, 8,
1841 		VGIC_ACCESS_64bit | VGIC_ACCESS_32bit),
1842 	REGISTER_ITS_DESC(GITS_BASER,
1843 		vgic_mmio_read_its_baser, vgic_mmio_write_its_baser, 0x40,
1844 		VGIC_ACCESS_64bit | VGIC_ACCESS_32bit),
1845 	REGISTER_ITS_DESC(GITS_IDREGS_BASE,
1846 		vgic_mmio_read_its_idregs, its_mmio_write_wi, 0x30,
1847 		VGIC_ACCESS_32bit),
1848 };
1849 
1850 /* This is called on setting the LPI enable bit in the redistributor. */
1851 void vgic_enable_lpis(struct kvm_vcpu *vcpu)
1852 {
1853 	if (!(vcpu->arch.vgic_cpu.pendbaser & GICR_PENDBASER_PTZ))
1854 		its_sync_lpi_pending_table(vcpu);
1855 }
1856 
1857 static int vgic_register_its_iodev(struct kvm *kvm, struct vgic_its *its,
1858 				   u64 addr)
1859 {
1860 	struct vgic_io_device *iodev = &its->iodev;
1861 	int ret;
1862 
1863 	mutex_lock(&kvm->slots_lock);
1864 	if (!IS_VGIC_ADDR_UNDEF(its->vgic_its_base)) {
1865 		ret = -EBUSY;
1866 		goto out;
1867 	}
1868 
1869 	its->vgic_its_base = addr;
1870 	iodev->regions = its_registers;
1871 	iodev->nr_regions = ARRAY_SIZE(its_registers);
1872 	kvm_iodevice_init(&iodev->dev, &kvm_io_gic_ops);
1873 
1874 	iodev->base_addr = its->vgic_its_base;
1875 	iodev->iodev_type = IODEV_ITS;
1876 	iodev->its = its;
1877 	ret = kvm_io_bus_register_dev(kvm, KVM_MMIO_BUS, iodev->base_addr,
1878 				      KVM_VGIC_V3_ITS_SIZE, &iodev->dev);
1879 out:
1880 	mutex_unlock(&kvm->slots_lock);
1881 
1882 	return ret;
1883 }
1884 
1885 /* Default is 16 cached LPIs per vcpu */
1886 #define LPI_DEFAULT_PCPU_CACHE_SIZE	16
1887 
1888 void vgic_lpi_translation_cache_init(struct kvm *kvm)
1889 {
1890 	struct vgic_dist *dist = &kvm->arch.vgic;
1891 	unsigned int sz;
1892 	int i;
1893 
1894 	if (!list_empty(&dist->lpi_translation_cache))
1895 		return;
1896 
1897 	sz = atomic_read(&kvm->online_vcpus) * LPI_DEFAULT_PCPU_CACHE_SIZE;
1898 
1899 	for (i = 0; i < sz; i++) {
1900 		struct vgic_translation_cache_entry *cte;
1901 
1902 		/* An allocation failure is not fatal */
1903 		cte = kzalloc(sizeof(*cte), GFP_KERNEL_ACCOUNT);
1904 		if (WARN_ON(!cte))
1905 			break;
1906 
1907 		INIT_LIST_HEAD(&cte->entry);
1908 		list_add(&cte->entry, &dist->lpi_translation_cache);
1909 	}
1910 }
1911 
1912 void vgic_lpi_translation_cache_destroy(struct kvm *kvm)
1913 {
1914 	struct vgic_dist *dist = &kvm->arch.vgic;
1915 	struct vgic_translation_cache_entry *cte, *tmp;
1916 
1917 	vgic_its_invalidate_cache(kvm);
1918 
1919 	list_for_each_entry_safe(cte, tmp,
1920 				 &dist->lpi_translation_cache, entry) {
1921 		list_del(&cte->entry);
1922 		kfree(cte);
1923 	}
1924 }
1925 
1926 #define INITIAL_BASER_VALUE						  \
1927 	(GIC_BASER_CACHEABILITY(GITS_BASER, INNER, RaWb)		| \
1928 	 GIC_BASER_CACHEABILITY(GITS_BASER, OUTER, SameAsInner)		| \
1929 	 GIC_BASER_SHAREABILITY(GITS_BASER, InnerShareable)		| \
1930 	 GITS_BASER_PAGE_SIZE_64K)
1931 
1932 #define INITIAL_PROPBASER_VALUE						  \
1933 	(GIC_BASER_CACHEABILITY(GICR_PROPBASER, INNER, RaWb)		| \
1934 	 GIC_BASER_CACHEABILITY(GICR_PROPBASER, OUTER, SameAsInner)	| \
1935 	 GIC_BASER_SHAREABILITY(GICR_PROPBASER, InnerShareable))
1936 
1937 static int vgic_its_create(struct kvm_device *dev, u32 type)
1938 {
1939 	struct vgic_its *its;
1940 
1941 	if (type != KVM_DEV_TYPE_ARM_VGIC_ITS)
1942 		return -ENODEV;
1943 
1944 	its = kzalloc(sizeof(struct vgic_its), GFP_KERNEL_ACCOUNT);
1945 	if (!its)
1946 		return -ENOMEM;
1947 
1948 	if (vgic_initialized(dev->kvm)) {
1949 		int ret = vgic_v4_init(dev->kvm);
1950 		if (ret < 0) {
1951 			kfree(its);
1952 			return ret;
1953 		}
1954 
1955 		vgic_lpi_translation_cache_init(dev->kvm);
1956 	}
1957 
1958 	mutex_init(&its->its_lock);
1959 	mutex_init(&its->cmd_lock);
1960 
1961 	its->vgic_its_base = VGIC_ADDR_UNDEF;
1962 
1963 	INIT_LIST_HEAD(&its->device_list);
1964 	INIT_LIST_HEAD(&its->collection_list);
1965 
1966 	dev->kvm->arch.vgic.msis_require_devid = true;
1967 	dev->kvm->arch.vgic.has_its = true;
1968 	its->enabled = false;
1969 	its->dev = dev;
1970 
1971 	its->baser_device_table = INITIAL_BASER_VALUE			|
1972 		((u64)GITS_BASER_TYPE_DEVICE << GITS_BASER_TYPE_SHIFT);
1973 	its->baser_coll_table = INITIAL_BASER_VALUE |
1974 		((u64)GITS_BASER_TYPE_COLLECTION << GITS_BASER_TYPE_SHIFT);
1975 	dev->kvm->arch.vgic.propbaser = INITIAL_PROPBASER_VALUE;
1976 
1977 	dev->private = its;
1978 
1979 	return vgic_its_set_abi(its, NR_ITS_ABIS - 1);
1980 }
1981 
1982 static void vgic_its_destroy(struct kvm_device *kvm_dev)
1983 {
1984 	struct kvm *kvm = kvm_dev->kvm;
1985 	struct vgic_its *its = kvm_dev->private;
1986 
1987 	mutex_lock(&its->its_lock);
1988 
1989 	vgic_its_free_device_list(kvm, its);
1990 	vgic_its_free_collection_list(kvm, its);
1991 
1992 	mutex_unlock(&its->its_lock);
1993 	kfree(its);
1994 	kfree(kvm_dev);/* alloc by kvm_ioctl_create_device, free by .destroy */
1995 }
1996 
1997 static int vgic_its_has_attr_regs(struct kvm_device *dev,
1998 				  struct kvm_device_attr *attr)
1999 {
2000 	const struct vgic_register_region *region;
2001 	gpa_t offset = attr->attr;
2002 	int align;
2003 
2004 	align = (offset < GITS_TYPER) || (offset >= GITS_PIDR4) ? 0x3 : 0x7;
2005 
2006 	if (offset & align)
2007 		return -EINVAL;
2008 
2009 	region = vgic_find_mmio_region(its_registers,
2010 				       ARRAY_SIZE(its_registers),
2011 				       offset);
2012 	if (!region)
2013 		return -ENXIO;
2014 
2015 	return 0;
2016 }
2017 
2018 static int vgic_its_attr_regs_access(struct kvm_device *dev,
2019 				     struct kvm_device_attr *attr,
2020 				     u64 *reg, bool is_write)
2021 {
2022 	const struct vgic_register_region *region;
2023 	struct vgic_its *its;
2024 	gpa_t addr, offset;
2025 	unsigned int len;
2026 	int align, ret = 0;
2027 
2028 	its = dev->private;
2029 	offset = attr->attr;
2030 
2031 	/*
2032 	 * Although the spec supports upper/lower 32-bit accesses to
2033 	 * 64-bit ITS registers, the userspace ABI requires 64-bit
2034 	 * accesses to all 64-bit wide registers. We therefore only
2035 	 * support 32-bit accesses to GITS_CTLR, GITS_IIDR and GITS ID
2036 	 * registers
2037 	 */
2038 	if ((offset < GITS_TYPER) || (offset >= GITS_PIDR4))
2039 		align = 0x3;
2040 	else
2041 		align = 0x7;
2042 
2043 	if (offset & align)
2044 		return -EINVAL;
2045 
2046 	mutex_lock(&dev->kvm->lock);
2047 
2048 	if (IS_VGIC_ADDR_UNDEF(its->vgic_its_base)) {
2049 		ret = -ENXIO;
2050 		goto out;
2051 	}
2052 
2053 	region = vgic_find_mmio_region(its_registers,
2054 				       ARRAY_SIZE(its_registers),
2055 				       offset);
2056 	if (!region) {
2057 		ret = -ENXIO;
2058 		goto out;
2059 	}
2060 
2061 	if (!lock_all_vcpus(dev->kvm)) {
2062 		ret = -EBUSY;
2063 		goto out;
2064 	}
2065 
2066 	addr = its->vgic_its_base + offset;
2067 
2068 	len = region->access_flags & VGIC_ACCESS_64bit ? 8 : 4;
2069 
2070 	if (is_write) {
2071 		if (region->uaccess_its_write)
2072 			ret = region->uaccess_its_write(dev->kvm, its, addr,
2073 							len, *reg);
2074 		else
2075 			region->its_write(dev->kvm, its, addr, len, *reg);
2076 	} else {
2077 		*reg = region->its_read(dev->kvm, its, addr, len);
2078 	}
2079 	unlock_all_vcpus(dev->kvm);
2080 out:
2081 	mutex_unlock(&dev->kvm->lock);
2082 	return ret;
2083 }
2084 
2085 static u32 compute_next_devid_offset(struct list_head *h,
2086 				     struct its_device *dev)
2087 {
2088 	struct its_device *next;
2089 	u32 next_offset;
2090 
2091 	if (list_is_last(&dev->dev_list, h))
2092 		return 0;
2093 	next = list_next_entry(dev, dev_list);
2094 	next_offset = next->device_id - dev->device_id;
2095 
2096 	return min_t(u32, next_offset, VITS_DTE_MAX_DEVID_OFFSET);
2097 }
2098 
2099 static u32 compute_next_eventid_offset(struct list_head *h, struct its_ite *ite)
2100 {
2101 	struct its_ite *next;
2102 	u32 next_offset;
2103 
2104 	if (list_is_last(&ite->ite_list, h))
2105 		return 0;
2106 	next = list_next_entry(ite, ite_list);
2107 	next_offset = next->event_id - ite->event_id;
2108 
2109 	return min_t(u32, next_offset, VITS_ITE_MAX_EVENTID_OFFSET);
2110 }
2111 
2112 /**
2113  * entry_fn_t - Callback called on a table entry restore path
2114  * @its: its handle
2115  * @id: id of the entry
2116  * @entry: pointer to the entry
2117  * @opaque: pointer to an opaque data
2118  *
2119  * Return: < 0 on error, 0 if last element was identified, id offset to next
2120  * element otherwise
2121  */
2122 typedef int (*entry_fn_t)(struct vgic_its *its, u32 id, void *entry,
2123 			  void *opaque);
2124 
2125 /**
2126  * scan_its_table - Scan a contiguous table in guest RAM and applies a function
2127  * to each entry
2128  *
2129  * @its: its handle
2130  * @base: base gpa of the table
2131  * @size: size of the table in bytes
2132  * @esz: entry size in bytes
2133  * @start_id: the ID of the first entry in the table
2134  * (non zero for 2d level tables)
2135  * @fn: function to apply on each entry
2136  *
2137  * Return: < 0 on error, 0 if last element was identified, 1 otherwise
2138  * (the last element may not be found on second level tables)
2139  */
2140 static int scan_its_table(struct vgic_its *its, gpa_t base, int size, u32 esz,
2141 			  int start_id, entry_fn_t fn, void *opaque)
2142 {
2143 	struct kvm *kvm = its->dev->kvm;
2144 	unsigned long len = size;
2145 	int id = start_id;
2146 	gpa_t gpa = base;
2147 	char entry[ESZ_MAX];
2148 	int ret;
2149 
2150 	memset(entry, 0, esz);
2151 
2152 	while (true) {
2153 		int next_offset;
2154 		size_t byte_offset;
2155 
2156 		ret = kvm_read_guest_lock(kvm, gpa, entry, esz);
2157 		if (ret)
2158 			return ret;
2159 
2160 		next_offset = fn(its, id, entry, opaque);
2161 		if (next_offset <= 0)
2162 			return next_offset;
2163 
2164 		byte_offset = next_offset * esz;
2165 		if (byte_offset >= len)
2166 			break;
2167 
2168 		id += next_offset;
2169 		gpa += byte_offset;
2170 		len -= byte_offset;
2171 	}
2172 	return 1;
2173 }
2174 
2175 /**
2176  * vgic_its_save_ite - Save an interrupt translation entry at @gpa
2177  */
2178 static int vgic_its_save_ite(struct vgic_its *its, struct its_device *dev,
2179 			      struct its_ite *ite, gpa_t gpa, int ite_esz)
2180 {
2181 	struct kvm *kvm = its->dev->kvm;
2182 	u32 next_offset;
2183 	u64 val;
2184 
2185 	next_offset = compute_next_eventid_offset(&dev->itt_head, ite);
2186 	val = ((u64)next_offset << KVM_ITS_ITE_NEXT_SHIFT) |
2187 	       ((u64)ite->irq->intid << KVM_ITS_ITE_PINTID_SHIFT) |
2188 		ite->collection->collection_id;
2189 	val = cpu_to_le64(val);
2190 	return vgic_write_guest_lock(kvm, gpa, &val, ite_esz);
2191 }
2192 
2193 /**
2194  * vgic_its_restore_ite - restore an interrupt translation entry
2195  * @event_id: id used for indexing
2196  * @ptr: pointer to the ITE entry
2197  * @opaque: pointer to the its_device
2198  */
2199 static int vgic_its_restore_ite(struct vgic_its *its, u32 event_id,
2200 				void *ptr, void *opaque)
2201 {
2202 	struct its_device *dev = opaque;
2203 	struct its_collection *collection;
2204 	struct kvm *kvm = its->dev->kvm;
2205 	struct kvm_vcpu *vcpu = NULL;
2206 	u64 val;
2207 	u64 *p = (u64 *)ptr;
2208 	struct vgic_irq *irq;
2209 	u32 coll_id, lpi_id;
2210 	struct its_ite *ite;
2211 	u32 offset;
2212 
2213 	val = *p;
2214 
2215 	val = le64_to_cpu(val);
2216 
2217 	coll_id = val & KVM_ITS_ITE_ICID_MASK;
2218 	lpi_id = (val & KVM_ITS_ITE_PINTID_MASK) >> KVM_ITS_ITE_PINTID_SHIFT;
2219 
2220 	if (!lpi_id)
2221 		return 1; /* invalid entry, no choice but to scan next entry */
2222 
2223 	if (lpi_id < VGIC_MIN_LPI)
2224 		return -EINVAL;
2225 
2226 	offset = val >> KVM_ITS_ITE_NEXT_SHIFT;
2227 	if (event_id + offset >= BIT_ULL(dev->num_eventid_bits))
2228 		return -EINVAL;
2229 
2230 	collection = find_collection(its, coll_id);
2231 	if (!collection)
2232 		return -EINVAL;
2233 
2234 	if (!vgic_its_check_event_id(its, dev, event_id))
2235 		return -EINVAL;
2236 
2237 	ite = vgic_its_alloc_ite(dev, collection, event_id);
2238 	if (IS_ERR(ite))
2239 		return PTR_ERR(ite);
2240 
2241 	if (its_is_collection_mapped(collection))
2242 		vcpu = kvm_get_vcpu(kvm, collection->target_addr);
2243 
2244 	irq = vgic_add_lpi(kvm, lpi_id, vcpu);
2245 	if (IS_ERR(irq)) {
2246 		its_free_ite(kvm, ite);
2247 		return PTR_ERR(irq);
2248 	}
2249 	ite->irq = irq;
2250 
2251 	return offset;
2252 }
2253 
2254 static int vgic_its_ite_cmp(void *priv, const struct list_head *a,
2255 			    const struct list_head *b)
2256 {
2257 	struct its_ite *itea = container_of(a, struct its_ite, ite_list);
2258 	struct its_ite *iteb = container_of(b, struct its_ite, ite_list);
2259 
2260 	if (itea->event_id < iteb->event_id)
2261 		return -1;
2262 	else
2263 		return 1;
2264 }
2265 
2266 static int vgic_its_save_itt(struct vgic_its *its, struct its_device *device)
2267 {
2268 	const struct vgic_its_abi *abi = vgic_its_get_abi(its);
2269 	gpa_t base = device->itt_addr;
2270 	struct its_ite *ite;
2271 	int ret;
2272 	int ite_esz = abi->ite_esz;
2273 
2274 	list_sort(NULL, &device->itt_head, vgic_its_ite_cmp);
2275 
2276 	list_for_each_entry(ite, &device->itt_head, ite_list) {
2277 		gpa_t gpa = base + ite->event_id * ite_esz;
2278 
2279 		/*
2280 		 * If an LPI carries the HW bit, this means that this
2281 		 * interrupt is controlled by GICv4, and we do not
2282 		 * have direct access to that state without GICv4.1.
2283 		 * Let's simply fail the save operation...
2284 		 */
2285 		if (ite->irq->hw && !kvm_vgic_global_state.has_gicv4_1)
2286 			return -EACCES;
2287 
2288 		ret = vgic_its_save_ite(its, device, ite, gpa, ite_esz);
2289 		if (ret)
2290 			return ret;
2291 	}
2292 	return 0;
2293 }
2294 
2295 /**
2296  * vgic_its_restore_itt - restore the ITT of a device
2297  *
2298  * @its: its handle
2299  * @dev: device handle
2300  *
2301  * Return 0 on success, < 0 on error
2302  */
2303 static int vgic_its_restore_itt(struct vgic_its *its, struct its_device *dev)
2304 {
2305 	const struct vgic_its_abi *abi = vgic_its_get_abi(its);
2306 	gpa_t base = dev->itt_addr;
2307 	int ret;
2308 	int ite_esz = abi->ite_esz;
2309 	size_t max_size = BIT_ULL(dev->num_eventid_bits) * ite_esz;
2310 
2311 	ret = scan_its_table(its, base, max_size, ite_esz, 0,
2312 			     vgic_its_restore_ite, dev);
2313 
2314 	/* scan_its_table returns +1 if all ITEs are invalid */
2315 	if (ret > 0)
2316 		ret = 0;
2317 
2318 	return ret;
2319 }
2320 
2321 /**
2322  * vgic_its_save_dte - Save a device table entry at a given GPA
2323  *
2324  * @its: ITS handle
2325  * @dev: ITS device
2326  * @ptr: GPA
2327  */
2328 static int vgic_its_save_dte(struct vgic_its *its, struct its_device *dev,
2329 			     gpa_t ptr, int dte_esz)
2330 {
2331 	struct kvm *kvm = its->dev->kvm;
2332 	u64 val, itt_addr_field;
2333 	u32 next_offset;
2334 
2335 	itt_addr_field = dev->itt_addr >> 8;
2336 	next_offset = compute_next_devid_offset(&its->device_list, dev);
2337 	val = (1ULL << KVM_ITS_DTE_VALID_SHIFT |
2338 	       ((u64)next_offset << KVM_ITS_DTE_NEXT_SHIFT) |
2339 	       (itt_addr_field << KVM_ITS_DTE_ITTADDR_SHIFT) |
2340 		(dev->num_eventid_bits - 1));
2341 	val = cpu_to_le64(val);
2342 	return vgic_write_guest_lock(kvm, ptr, &val, dte_esz);
2343 }
2344 
2345 /**
2346  * vgic_its_restore_dte - restore a device table entry
2347  *
2348  * @its: its handle
2349  * @id: device id the DTE corresponds to
2350  * @ptr: kernel VA where the 8 byte DTE is located
2351  * @opaque: unused
2352  *
2353  * Return: < 0 on error, 0 if the dte is the last one, id offset to the
2354  * next dte otherwise
2355  */
2356 static int vgic_its_restore_dte(struct vgic_its *its, u32 id,
2357 				void *ptr, void *opaque)
2358 {
2359 	struct its_device *dev;
2360 	u64 baser = its->baser_device_table;
2361 	gpa_t itt_addr;
2362 	u8 num_eventid_bits;
2363 	u64 entry = *(u64 *)ptr;
2364 	bool valid;
2365 	u32 offset;
2366 	int ret;
2367 
2368 	entry = le64_to_cpu(entry);
2369 
2370 	valid = entry >> KVM_ITS_DTE_VALID_SHIFT;
2371 	num_eventid_bits = (entry & KVM_ITS_DTE_SIZE_MASK) + 1;
2372 	itt_addr = ((entry & KVM_ITS_DTE_ITTADDR_MASK)
2373 			>> KVM_ITS_DTE_ITTADDR_SHIFT) << 8;
2374 
2375 	if (!valid)
2376 		return 1;
2377 
2378 	/* dte entry is valid */
2379 	offset = (entry & KVM_ITS_DTE_NEXT_MASK) >> KVM_ITS_DTE_NEXT_SHIFT;
2380 
2381 	if (!vgic_its_check_id(its, baser, id, NULL))
2382 		return -EINVAL;
2383 
2384 	dev = vgic_its_alloc_device(its, id, itt_addr, num_eventid_bits);
2385 	if (IS_ERR(dev))
2386 		return PTR_ERR(dev);
2387 
2388 	ret = vgic_its_restore_itt(its, dev);
2389 	if (ret) {
2390 		vgic_its_free_device(its->dev->kvm, dev);
2391 		return ret;
2392 	}
2393 
2394 	return offset;
2395 }
2396 
2397 static int vgic_its_device_cmp(void *priv, const struct list_head *a,
2398 			       const struct list_head *b)
2399 {
2400 	struct its_device *deva = container_of(a, struct its_device, dev_list);
2401 	struct its_device *devb = container_of(b, struct its_device, dev_list);
2402 
2403 	if (deva->device_id < devb->device_id)
2404 		return -1;
2405 	else
2406 		return 1;
2407 }
2408 
2409 /**
2410  * vgic_its_save_device_tables - Save the device table and all ITT
2411  * into guest RAM
2412  *
2413  * L1/L2 handling is hidden by vgic_its_check_id() helper which directly
2414  * returns the GPA of the device entry
2415  */
2416 static int vgic_its_save_device_tables(struct vgic_its *its)
2417 {
2418 	const struct vgic_its_abi *abi = vgic_its_get_abi(its);
2419 	u64 baser = its->baser_device_table;
2420 	struct its_device *dev;
2421 	int dte_esz = abi->dte_esz;
2422 
2423 	if (!(baser & GITS_BASER_VALID))
2424 		return 0;
2425 
2426 	list_sort(NULL, &its->device_list, vgic_its_device_cmp);
2427 
2428 	list_for_each_entry(dev, &its->device_list, dev_list) {
2429 		int ret;
2430 		gpa_t eaddr;
2431 
2432 		if (!vgic_its_check_id(its, baser,
2433 				       dev->device_id, &eaddr))
2434 			return -EINVAL;
2435 
2436 		ret = vgic_its_save_itt(its, dev);
2437 		if (ret)
2438 			return ret;
2439 
2440 		ret = vgic_its_save_dte(its, dev, eaddr, dte_esz);
2441 		if (ret)
2442 			return ret;
2443 	}
2444 	return 0;
2445 }
2446 
2447 /**
2448  * handle_l1_dte - callback used for L1 device table entries (2 stage case)
2449  *
2450  * @its: its handle
2451  * @id: index of the entry in the L1 table
2452  * @addr: kernel VA
2453  * @opaque: unused
2454  *
2455  * L1 table entries are scanned by steps of 1 entry
2456  * Return < 0 if error, 0 if last dte was found when scanning the L2
2457  * table, +1 otherwise (meaning next L1 entry must be scanned)
2458  */
2459 static int handle_l1_dte(struct vgic_its *its, u32 id, void *addr,
2460 			 void *opaque)
2461 {
2462 	const struct vgic_its_abi *abi = vgic_its_get_abi(its);
2463 	int l2_start_id = id * (SZ_64K / abi->dte_esz);
2464 	u64 entry = *(u64 *)addr;
2465 	int dte_esz = abi->dte_esz;
2466 	gpa_t gpa;
2467 	int ret;
2468 
2469 	entry = le64_to_cpu(entry);
2470 
2471 	if (!(entry & KVM_ITS_L1E_VALID_MASK))
2472 		return 1;
2473 
2474 	gpa = entry & KVM_ITS_L1E_ADDR_MASK;
2475 
2476 	ret = scan_its_table(its, gpa, SZ_64K, dte_esz,
2477 			     l2_start_id, vgic_its_restore_dte, NULL);
2478 
2479 	return ret;
2480 }
2481 
2482 /**
2483  * vgic_its_restore_device_tables - Restore the device table and all ITT
2484  * from guest RAM to internal data structs
2485  */
2486 static int vgic_its_restore_device_tables(struct vgic_its *its)
2487 {
2488 	const struct vgic_its_abi *abi = vgic_its_get_abi(its);
2489 	u64 baser = its->baser_device_table;
2490 	int l1_esz, ret;
2491 	int l1_tbl_size = GITS_BASER_NR_PAGES(baser) * SZ_64K;
2492 	gpa_t l1_gpa;
2493 
2494 	if (!(baser & GITS_BASER_VALID))
2495 		return 0;
2496 
2497 	l1_gpa = GITS_BASER_ADDR_48_to_52(baser);
2498 
2499 	if (baser & GITS_BASER_INDIRECT) {
2500 		l1_esz = GITS_LVL1_ENTRY_SIZE;
2501 		ret = scan_its_table(its, l1_gpa, l1_tbl_size, l1_esz, 0,
2502 				     handle_l1_dte, NULL);
2503 	} else {
2504 		l1_esz = abi->dte_esz;
2505 		ret = scan_its_table(its, l1_gpa, l1_tbl_size, l1_esz, 0,
2506 				     vgic_its_restore_dte, NULL);
2507 	}
2508 
2509 	/* scan_its_table returns +1 if all entries are invalid */
2510 	if (ret > 0)
2511 		ret = 0;
2512 
2513 	if (ret < 0)
2514 		vgic_its_free_device_list(its->dev->kvm, its);
2515 
2516 	return ret;
2517 }
2518 
2519 static int vgic_its_save_cte(struct vgic_its *its,
2520 			     struct its_collection *collection,
2521 			     gpa_t gpa, int esz)
2522 {
2523 	u64 val;
2524 
2525 	val = (1ULL << KVM_ITS_CTE_VALID_SHIFT |
2526 	       ((u64)collection->target_addr << KVM_ITS_CTE_RDBASE_SHIFT) |
2527 	       collection->collection_id);
2528 	val = cpu_to_le64(val);
2529 	return vgic_write_guest_lock(its->dev->kvm, gpa, &val, esz);
2530 }
2531 
2532 /*
2533  * Restore a collection entry into the ITS collection table.
2534  * Return +1 on success, 0 if the entry was invalid (which should be
2535  * interpreted as end-of-table), and a negative error value for generic errors.
2536  */
2537 static int vgic_its_restore_cte(struct vgic_its *its, gpa_t gpa, int esz)
2538 {
2539 	struct its_collection *collection;
2540 	struct kvm *kvm = its->dev->kvm;
2541 	u32 target_addr, coll_id;
2542 	u64 val;
2543 	int ret;
2544 
2545 	BUG_ON(esz > sizeof(val));
2546 	ret = kvm_read_guest_lock(kvm, gpa, &val, esz);
2547 	if (ret)
2548 		return ret;
2549 	val = le64_to_cpu(val);
2550 	if (!(val & KVM_ITS_CTE_VALID_MASK))
2551 		return 0;
2552 
2553 	target_addr = (u32)(val >> KVM_ITS_CTE_RDBASE_SHIFT);
2554 	coll_id = val & KVM_ITS_CTE_ICID_MASK;
2555 
2556 	if (target_addr != COLLECTION_NOT_MAPPED &&
2557 	    target_addr >= atomic_read(&kvm->online_vcpus))
2558 		return -EINVAL;
2559 
2560 	collection = find_collection(its, coll_id);
2561 	if (collection)
2562 		return -EEXIST;
2563 
2564 	if (!vgic_its_check_id(its, its->baser_coll_table, coll_id, NULL))
2565 		return -EINVAL;
2566 
2567 	ret = vgic_its_alloc_collection(its, &collection, coll_id);
2568 	if (ret)
2569 		return ret;
2570 	collection->target_addr = target_addr;
2571 	return 1;
2572 }
2573 
2574 /**
2575  * vgic_its_save_collection_table - Save the collection table into
2576  * guest RAM
2577  */
2578 static int vgic_its_save_collection_table(struct vgic_its *its)
2579 {
2580 	const struct vgic_its_abi *abi = vgic_its_get_abi(its);
2581 	u64 baser = its->baser_coll_table;
2582 	gpa_t gpa = GITS_BASER_ADDR_48_to_52(baser);
2583 	struct its_collection *collection;
2584 	u64 val;
2585 	size_t max_size, filled = 0;
2586 	int ret, cte_esz = abi->cte_esz;
2587 
2588 	if (!(baser & GITS_BASER_VALID))
2589 		return 0;
2590 
2591 	max_size = GITS_BASER_NR_PAGES(baser) * SZ_64K;
2592 
2593 	list_for_each_entry(collection, &its->collection_list, coll_list) {
2594 		ret = vgic_its_save_cte(its, collection, gpa, cte_esz);
2595 		if (ret)
2596 			return ret;
2597 		gpa += cte_esz;
2598 		filled += cte_esz;
2599 	}
2600 
2601 	if (filled == max_size)
2602 		return 0;
2603 
2604 	/*
2605 	 * table is not fully filled, add a last dummy element
2606 	 * with valid bit unset
2607 	 */
2608 	val = 0;
2609 	BUG_ON(cte_esz > sizeof(val));
2610 	ret = vgic_write_guest_lock(its->dev->kvm, gpa, &val, cte_esz);
2611 	return ret;
2612 }
2613 
2614 /**
2615  * vgic_its_restore_collection_table - reads the collection table
2616  * in guest memory and restores the ITS internal state. Requires the
2617  * BASER registers to be restored before.
2618  */
2619 static int vgic_its_restore_collection_table(struct vgic_its *its)
2620 {
2621 	const struct vgic_its_abi *abi = vgic_its_get_abi(its);
2622 	u64 baser = its->baser_coll_table;
2623 	int cte_esz = abi->cte_esz;
2624 	size_t max_size, read = 0;
2625 	gpa_t gpa;
2626 	int ret;
2627 
2628 	if (!(baser & GITS_BASER_VALID))
2629 		return 0;
2630 
2631 	gpa = GITS_BASER_ADDR_48_to_52(baser);
2632 
2633 	max_size = GITS_BASER_NR_PAGES(baser) * SZ_64K;
2634 
2635 	while (read < max_size) {
2636 		ret = vgic_its_restore_cte(its, gpa, cte_esz);
2637 		if (ret <= 0)
2638 			break;
2639 		gpa += cte_esz;
2640 		read += cte_esz;
2641 	}
2642 
2643 	if (ret > 0)
2644 		return 0;
2645 
2646 	if (ret < 0)
2647 		vgic_its_free_collection_list(its->dev->kvm, its);
2648 
2649 	return ret;
2650 }
2651 
2652 /**
2653  * vgic_its_save_tables_v0 - Save the ITS tables into guest ARM
2654  * according to v0 ABI
2655  */
2656 static int vgic_its_save_tables_v0(struct vgic_its *its)
2657 {
2658 	int ret;
2659 
2660 	ret = vgic_its_save_device_tables(its);
2661 	if (ret)
2662 		return ret;
2663 
2664 	return vgic_its_save_collection_table(its);
2665 }
2666 
2667 /**
2668  * vgic_its_restore_tables_v0 - Restore the ITS tables from guest RAM
2669  * to internal data structs according to V0 ABI
2670  *
2671  */
2672 static int vgic_its_restore_tables_v0(struct vgic_its *its)
2673 {
2674 	int ret;
2675 
2676 	ret = vgic_its_restore_collection_table(its);
2677 	if (ret)
2678 		return ret;
2679 
2680 	ret = vgic_its_restore_device_tables(its);
2681 	if (ret)
2682 		vgic_its_free_collection_list(its->dev->kvm, its);
2683 	return ret;
2684 }
2685 
2686 static int vgic_its_commit_v0(struct vgic_its *its)
2687 {
2688 	const struct vgic_its_abi *abi;
2689 
2690 	abi = vgic_its_get_abi(its);
2691 	its->baser_coll_table &= ~GITS_BASER_ENTRY_SIZE_MASK;
2692 	its->baser_device_table &= ~GITS_BASER_ENTRY_SIZE_MASK;
2693 
2694 	its->baser_coll_table |= (GIC_ENCODE_SZ(abi->cte_esz, 5)
2695 					<< GITS_BASER_ENTRY_SIZE_SHIFT);
2696 
2697 	its->baser_device_table |= (GIC_ENCODE_SZ(abi->dte_esz, 5)
2698 					<< GITS_BASER_ENTRY_SIZE_SHIFT);
2699 	return 0;
2700 }
2701 
2702 static void vgic_its_reset(struct kvm *kvm, struct vgic_its *its)
2703 {
2704 	/* We need to keep the ABI specific field values */
2705 	its->baser_coll_table &= ~GITS_BASER_VALID;
2706 	its->baser_device_table &= ~GITS_BASER_VALID;
2707 	its->cbaser = 0;
2708 	its->creadr = 0;
2709 	its->cwriter = 0;
2710 	its->enabled = 0;
2711 	vgic_its_free_device_list(kvm, its);
2712 	vgic_its_free_collection_list(kvm, its);
2713 }
2714 
2715 static int vgic_its_has_attr(struct kvm_device *dev,
2716 			     struct kvm_device_attr *attr)
2717 {
2718 	switch (attr->group) {
2719 	case KVM_DEV_ARM_VGIC_GRP_ADDR:
2720 		switch (attr->attr) {
2721 		case KVM_VGIC_ITS_ADDR_TYPE:
2722 			return 0;
2723 		}
2724 		break;
2725 	case KVM_DEV_ARM_VGIC_GRP_CTRL:
2726 		switch (attr->attr) {
2727 		case KVM_DEV_ARM_VGIC_CTRL_INIT:
2728 			return 0;
2729 		case KVM_DEV_ARM_ITS_CTRL_RESET:
2730 			return 0;
2731 		case KVM_DEV_ARM_ITS_SAVE_TABLES:
2732 			return 0;
2733 		case KVM_DEV_ARM_ITS_RESTORE_TABLES:
2734 			return 0;
2735 		}
2736 		break;
2737 	case KVM_DEV_ARM_VGIC_GRP_ITS_REGS:
2738 		return vgic_its_has_attr_regs(dev, attr);
2739 	}
2740 	return -ENXIO;
2741 }
2742 
2743 static int vgic_its_ctrl(struct kvm *kvm, struct vgic_its *its, u64 attr)
2744 {
2745 	const struct vgic_its_abi *abi = vgic_its_get_abi(its);
2746 	int ret = 0;
2747 
2748 	if (attr == KVM_DEV_ARM_VGIC_CTRL_INIT) /* Nothing to do */
2749 		return 0;
2750 
2751 	mutex_lock(&kvm->lock);
2752 	mutex_lock(&its->its_lock);
2753 
2754 	if (!lock_all_vcpus(kvm)) {
2755 		mutex_unlock(&its->its_lock);
2756 		mutex_unlock(&kvm->lock);
2757 		return -EBUSY;
2758 	}
2759 
2760 	switch (attr) {
2761 	case KVM_DEV_ARM_ITS_CTRL_RESET:
2762 		vgic_its_reset(kvm, its);
2763 		break;
2764 	case KVM_DEV_ARM_ITS_SAVE_TABLES:
2765 		ret = abi->save_tables(its);
2766 		break;
2767 	case KVM_DEV_ARM_ITS_RESTORE_TABLES:
2768 		ret = abi->restore_tables(its);
2769 		break;
2770 	}
2771 
2772 	unlock_all_vcpus(kvm);
2773 	mutex_unlock(&its->its_lock);
2774 	mutex_unlock(&kvm->lock);
2775 	return ret;
2776 }
2777 
2778 /*
2779  * kvm_arch_allow_write_without_running_vcpu - allow writing guest memory
2780  * without the running VCPU when dirty ring is enabled.
2781  *
2782  * The running VCPU is required to track dirty guest pages when dirty ring
2783  * is enabled. Otherwise, the backup bitmap should be used to track the
2784  * dirty guest pages. When vgic/its tables are being saved, the backup
2785  * bitmap is used to track the dirty guest pages due to the missed running
2786  * VCPU in the period.
2787  */
2788 bool kvm_arch_allow_write_without_running_vcpu(struct kvm *kvm)
2789 {
2790 	struct vgic_dist *dist = &kvm->arch.vgic;
2791 
2792 	return dist->table_write_in_progress;
2793 }
2794 
2795 static int vgic_its_set_attr(struct kvm_device *dev,
2796 			     struct kvm_device_attr *attr)
2797 {
2798 	struct vgic_its *its = dev->private;
2799 	int ret;
2800 
2801 	switch (attr->group) {
2802 	case KVM_DEV_ARM_VGIC_GRP_ADDR: {
2803 		u64 __user *uaddr = (u64 __user *)(long)attr->addr;
2804 		unsigned long type = (unsigned long)attr->attr;
2805 		u64 addr;
2806 
2807 		if (type != KVM_VGIC_ITS_ADDR_TYPE)
2808 			return -ENODEV;
2809 
2810 		if (copy_from_user(&addr, uaddr, sizeof(addr)))
2811 			return -EFAULT;
2812 
2813 		ret = vgic_check_iorange(dev->kvm, its->vgic_its_base,
2814 					 addr, SZ_64K, KVM_VGIC_V3_ITS_SIZE);
2815 		if (ret)
2816 			return ret;
2817 
2818 		return vgic_register_its_iodev(dev->kvm, its, addr);
2819 	}
2820 	case KVM_DEV_ARM_VGIC_GRP_CTRL:
2821 		return vgic_its_ctrl(dev->kvm, its, attr->attr);
2822 	case KVM_DEV_ARM_VGIC_GRP_ITS_REGS: {
2823 		u64 __user *uaddr = (u64 __user *)(long)attr->addr;
2824 		u64 reg;
2825 
2826 		if (get_user(reg, uaddr))
2827 			return -EFAULT;
2828 
2829 		return vgic_its_attr_regs_access(dev, attr, &reg, true);
2830 	}
2831 	}
2832 	return -ENXIO;
2833 }
2834 
2835 static int vgic_its_get_attr(struct kvm_device *dev,
2836 			     struct kvm_device_attr *attr)
2837 {
2838 	switch (attr->group) {
2839 	case KVM_DEV_ARM_VGIC_GRP_ADDR: {
2840 		struct vgic_its *its = dev->private;
2841 		u64 addr = its->vgic_its_base;
2842 		u64 __user *uaddr = (u64 __user *)(long)attr->addr;
2843 		unsigned long type = (unsigned long)attr->attr;
2844 
2845 		if (type != KVM_VGIC_ITS_ADDR_TYPE)
2846 			return -ENODEV;
2847 
2848 		if (copy_to_user(uaddr, &addr, sizeof(addr)))
2849 			return -EFAULT;
2850 		break;
2851 	}
2852 	case KVM_DEV_ARM_VGIC_GRP_ITS_REGS: {
2853 		u64 __user *uaddr = (u64 __user *)(long)attr->addr;
2854 		u64 reg;
2855 		int ret;
2856 
2857 		ret = vgic_its_attr_regs_access(dev, attr, &reg, false);
2858 		if (ret)
2859 			return ret;
2860 		return put_user(reg, uaddr);
2861 	}
2862 	default:
2863 		return -ENXIO;
2864 	}
2865 
2866 	return 0;
2867 }
2868 
2869 static struct kvm_device_ops kvm_arm_vgic_its_ops = {
2870 	.name = "kvm-arm-vgic-its",
2871 	.create = vgic_its_create,
2872 	.destroy = vgic_its_destroy,
2873 	.set_attr = vgic_its_set_attr,
2874 	.get_attr = vgic_its_get_attr,
2875 	.has_attr = vgic_its_has_attr,
2876 };
2877 
2878 int kvm_vgic_register_its_device(void)
2879 {
2880 	return kvm_register_device_ops(&kvm_arm_vgic_its_ops,
2881 				       KVM_DEV_TYPE_ARM_VGIC_ITS);
2882 }
2883