1/*
2 * Copyright (C) 2012 ARM Ltd.
3 * Author: Marc Zyngier <marc.zyngier@arm.com>
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
17 */
18
19#include <linux/cpu.h>
20#include <linux/kvm.h>
21#include <linux/kvm_host.h>
22#include <linux/interrupt.h>
23#include <linux/io.h>
24#include <linux/of.h>
25#include <linux/of_address.h>
26#include <linux/of_irq.h>
27#include <linux/rculist.h>
28#include <linux/uaccess.h>
29
30#include <asm/kvm_emulate.h>
31#include <asm/kvm_arm.h>
32#include <asm/kvm_mmu.h>
33#include <trace/events/kvm.h>
34#include <asm/kvm.h>
35#include <kvm/iodev.h>
36
37#define CREATE_TRACE_POINTS
38#include "trace.h"
39
40/*
41 * How the whole thing works (courtesy of Christoffer Dall):
42 *
43 * - At any time, the dist->irq_pending_on_cpu is the oracle that knows if
44 *   something is pending on the CPU interface.
45 * - Interrupts that are pending on the distributor are stored on the
46 *   vgic.irq_pending vgic bitmap (this bitmap is updated by both user land
47 *   ioctls and guest mmio ops, and other in-kernel peripherals such as the
48 *   arch. timers).
49 * - Every time the bitmap changes, the irq_pending_on_cpu oracle is
50 *   recalculated
51 * - To calculate the oracle, we need info for each cpu from
52 *   compute_pending_for_cpu, which considers:
53 *   - PPI: dist->irq_pending & dist->irq_enable
54 *   - SPI: dist->irq_pending & dist->irq_enable & dist->irq_spi_target
55 *   - irq_spi_target is a 'formatted' version of the GICD_ITARGETSRn
56 *     registers, stored on each vcpu. We only keep one bit of
57 *     information per interrupt, making sure that only one vcpu can
58 *     accept the interrupt.
59 * - If any of the above state changes, we must recalculate the oracle.
60 * - The same is true when injecting an interrupt, except that we only
61 *   consider a single interrupt at a time. The irq_spi_cpu array
62 *   contains the target CPU for each SPI.
63 *
64 * The handling of level interrupts adds some extra complexity. We
65 * need to track when the interrupt has been EOIed, so we can sample
66 * the 'line' again. This is achieved as such:
67 *
68 * - When a level interrupt is moved onto a vcpu, the corresponding
69 *   bit in irq_queued is set. As long as this bit is set, the line
70 *   will be ignored for further interrupts. The interrupt is injected
71 *   into the vcpu with the GICH_LR_EOI bit set (generate a
72 *   maintenance interrupt on EOI).
73 * - When the interrupt is EOIed, the maintenance interrupt fires,
74 *   and clears the corresponding bit in irq_queued. This allows the
75 *   interrupt line to be sampled again.
76 * - Note that level-triggered interrupts can also be set to pending from
77 *   writes to GICD_ISPENDRn and lowering the external input line does not
78 *   cause the interrupt to become inactive in such a situation.
79 *   Conversely, writes to GICD_ICPENDRn do not cause the interrupt to become
80 *   inactive as long as the external input line is held high.
81 *
82 *
83 * Initialization rules: there are multiple stages to the vgic
84 * initialization, both for the distributor and the CPU interfaces.
85 *
86 * Distributor:
87 *
88 * - kvm_vgic_early_init(): initialization of static data that doesn't
89 *   depend on any sizing information or emulation type. No allocation
90 *   is allowed there.
91 *
92 * - vgic_init(): allocation and initialization of the generic data
93 *   structures that depend on sizing information (number of CPUs,
94 *   number of interrupts). Also initializes the vcpu specific data
95 *   structures. Can be executed lazily for GICv2.
96 *   [to be renamed to kvm_vgic_init??]
97 *
98 * CPU Interface:
99 *
100 * - kvm_vgic_cpu_early_init(): initialization of static data that
101 *   doesn't depend on any sizing information or emulation type. No
102 *   allocation is allowed there.
103 */
104
105#include "vgic.h"
106
107static void vgic_retire_disabled_irqs(struct kvm_vcpu *vcpu);
108static void vgic_retire_lr(int lr_nr, struct kvm_vcpu *vcpu);
109static struct vgic_lr vgic_get_lr(const struct kvm_vcpu *vcpu, int lr);
110static void vgic_set_lr(struct kvm_vcpu *vcpu, int lr, struct vgic_lr lr_desc);
111static u64 vgic_get_elrsr(struct kvm_vcpu *vcpu);
112static struct irq_phys_map *vgic_irq_map_search(struct kvm_vcpu *vcpu,
113						int virt_irq);
114static int compute_pending_for_cpu(struct kvm_vcpu *vcpu);
115
116static const struct vgic_ops *vgic_ops;
117static const struct vgic_params *vgic;
118
119static void add_sgi_source(struct kvm_vcpu *vcpu, int irq, int source)
120{
121	vcpu->kvm->arch.vgic.vm_ops.add_sgi_source(vcpu, irq, source);
122}
123
124static bool queue_sgi(struct kvm_vcpu *vcpu, int irq)
125{
126	return vcpu->kvm->arch.vgic.vm_ops.queue_sgi(vcpu, irq);
127}
128
129int kvm_vgic_map_resources(struct kvm *kvm)
130{
131	return kvm->arch.vgic.vm_ops.map_resources(kvm, vgic);
132}
133
134/*
135 * struct vgic_bitmap contains a bitmap made of unsigned longs, but
136 * extracts u32s out of them.
137 *
138 * This does not work on 64-bit BE systems, because the bitmap access
139 * will store two consecutive 32-bit words with the higher-addressed
140 * register's bits at the lower index and the lower-addressed register's
141 * bits at the higher index.
142 *
143 * Therefore, swizzle the register index when accessing the 32-bit word
144 * registers to access the right register's value.
145 */
146#if defined(CONFIG_CPU_BIG_ENDIAN) && BITS_PER_LONG == 64
147#define REG_OFFSET_SWIZZLE	1
148#else
149#define REG_OFFSET_SWIZZLE	0
150#endif
151
152static int vgic_init_bitmap(struct vgic_bitmap *b, int nr_cpus, int nr_irqs)
153{
154	int nr_longs;
155
156	nr_longs = nr_cpus + BITS_TO_LONGS(nr_irqs - VGIC_NR_PRIVATE_IRQS);
157
158	b->private = kzalloc(sizeof(unsigned long) * nr_longs, GFP_KERNEL);
159	if (!b->private)
160		return -ENOMEM;
161
162	b->shared = b->private + nr_cpus;
163
164	return 0;
165}
166
167static void vgic_free_bitmap(struct vgic_bitmap *b)
168{
169	kfree(b->private);
170	b->private = NULL;
171	b->shared = NULL;
172}
173
174/*
175 * Call this function to convert a u64 value to an unsigned long * bitmask
176 * in a way that works on both 32-bit and 64-bit LE and BE platforms.
177 *
178 * Warning: Calling this function may modify *val.
179 */
180static unsigned long *u64_to_bitmask(u64 *val)
181{
182#if defined(CONFIG_CPU_BIG_ENDIAN) && BITS_PER_LONG == 32
183	*val = (*val >> 32) | (*val << 32);
184#endif
185	return (unsigned long *)val;
186}
187
188u32 *vgic_bitmap_get_reg(struct vgic_bitmap *x, int cpuid, u32 offset)
189{
190	offset >>= 2;
191	if (!offset)
192		return (u32 *)(x->private + cpuid) + REG_OFFSET_SWIZZLE;
193	else
194		return (u32 *)(x->shared) + ((offset - 1) ^ REG_OFFSET_SWIZZLE);
195}
196
197static int vgic_bitmap_get_irq_val(struct vgic_bitmap *x,
198				   int cpuid, int irq)
199{
200	if (irq < VGIC_NR_PRIVATE_IRQS)
201		return test_bit(irq, x->private + cpuid);
202
203	return test_bit(irq - VGIC_NR_PRIVATE_IRQS, x->shared);
204}
205
206void vgic_bitmap_set_irq_val(struct vgic_bitmap *x, int cpuid,
207			     int irq, int val)
208{
209	unsigned long *reg;
210
211	if (irq < VGIC_NR_PRIVATE_IRQS) {
212		reg = x->private + cpuid;
213	} else {
214		reg = x->shared;
215		irq -= VGIC_NR_PRIVATE_IRQS;
216	}
217
218	if (val)
219		set_bit(irq, reg);
220	else
221		clear_bit(irq, reg);
222}
223
224static unsigned long *vgic_bitmap_get_cpu_map(struct vgic_bitmap *x, int cpuid)
225{
226	return x->private + cpuid;
227}
228
229unsigned long *vgic_bitmap_get_shared_map(struct vgic_bitmap *x)
230{
231	return x->shared;
232}
233
234static int vgic_init_bytemap(struct vgic_bytemap *x, int nr_cpus, int nr_irqs)
235{
236	int size;
237
238	size  = nr_cpus * VGIC_NR_PRIVATE_IRQS;
239	size += nr_irqs - VGIC_NR_PRIVATE_IRQS;
240
241	x->private = kzalloc(size, GFP_KERNEL);
242	if (!x->private)
243		return -ENOMEM;
244
245	x->shared = x->private + nr_cpus * VGIC_NR_PRIVATE_IRQS / sizeof(u32);
246	return 0;
247}
248
249static void vgic_free_bytemap(struct vgic_bytemap *b)
250{
251	kfree(b->private);
252	b->private = NULL;
253	b->shared = NULL;
254}
255
256u32 *vgic_bytemap_get_reg(struct vgic_bytemap *x, int cpuid, u32 offset)
257{
258	u32 *reg;
259
260	if (offset < VGIC_NR_PRIVATE_IRQS) {
261		reg = x->private;
262		offset += cpuid * VGIC_NR_PRIVATE_IRQS;
263	} else {
264		reg = x->shared;
265		offset -= VGIC_NR_PRIVATE_IRQS;
266	}
267
268	return reg + (offset / sizeof(u32));
269}
270
271#define VGIC_CFG_LEVEL	0
272#define VGIC_CFG_EDGE	1
273
274static bool vgic_irq_is_edge(struct kvm_vcpu *vcpu, int irq)
275{
276	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
277	int irq_val;
278
279	irq_val = vgic_bitmap_get_irq_val(&dist->irq_cfg, vcpu->vcpu_id, irq);
280	return irq_val == VGIC_CFG_EDGE;
281}
282
283static int vgic_irq_is_enabled(struct kvm_vcpu *vcpu, int irq)
284{
285	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
286
287	return vgic_bitmap_get_irq_val(&dist->irq_enabled, vcpu->vcpu_id, irq);
288}
289
290static int vgic_irq_is_queued(struct kvm_vcpu *vcpu, int irq)
291{
292	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
293
294	return vgic_bitmap_get_irq_val(&dist->irq_queued, vcpu->vcpu_id, irq);
295}
296
297static int vgic_irq_is_active(struct kvm_vcpu *vcpu, int irq)
298{
299	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
300
301	return vgic_bitmap_get_irq_val(&dist->irq_active, vcpu->vcpu_id, irq);
302}
303
304static void vgic_irq_set_queued(struct kvm_vcpu *vcpu, int irq)
305{
306	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
307
308	vgic_bitmap_set_irq_val(&dist->irq_queued, vcpu->vcpu_id, irq, 1);
309}
310
311static void vgic_irq_clear_queued(struct kvm_vcpu *vcpu, int irq)
312{
313	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
314
315	vgic_bitmap_set_irq_val(&dist->irq_queued, vcpu->vcpu_id, irq, 0);
316}
317
318static void vgic_irq_set_active(struct kvm_vcpu *vcpu, int irq)
319{
320	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
321
322	vgic_bitmap_set_irq_val(&dist->irq_active, vcpu->vcpu_id, irq, 1);
323}
324
325static void vgic_irq_clear_active(struct kvm_vcpu *vcpu, int irq)
326{
327	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
328
329	vgic_bitmap_set_irq_val(&dist->irq_active, vcpu->vcpu_id, irq, 0);
330}
331
332static int vgic_dist_irq_get_level(struct kvm_vcpu *vcpu, int irq)
333{
334	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
335
336	return vgic_bitmap_get_irq_val(&dist->irq_level, vcpu->vcpu_id, irq);
337}
338
339static void vgic_dist_irq_set_level(struct kvm_vcpu *vcpu, int irq)
340{
341	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
342
343	vgic_bitmap_set_irq_val(&dist->irq_level, vcpu->vcpu_id, irq, 1);
344}
345
346static void vgic_dist_irq_clear_level(struct kvm_vcpu *vcpu, int irq)
347{
348	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
349
350	vgic_bitmap_set_irq_val(&dist->irq_level, vcpu->vcpu_id, irq, 0);
351}
352
353static int vgic_dist_irq_soft_pend(struct kvm_vcpu *vcpu, int irq)
354{
355	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
356
357	return vgic_bitmap_get_irq_val(&dist->irq_soft_pend, vcpu->vcpu_id, irq);
358}
359
360static void vgic_dist_irq_clear_soft_pend(struct kvm_vcpu *vcpu, int irq)
361{
362	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
363
364	vgic_bitmap_set_irq_val(&dist->irq_soft_pend, vcpu->vcpu_id, irq, 0);
365	if (!vgic_dist_irq_get_level(vcpu, irq)) {
366		vgic_dist_irq_clear_pending(vcpu, irq);
367		if (!compute_pending_for_cpu(vcpu))
368			clear_bit(vcpu->vcpu_id, dist->irq_pending_on_cpu);
369	}
370}
371
372static int vgic_dist_irq_is_pending(struct kvm_vcpu *vcpu, int irq)
373{
374	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
375
376	return vgic_bitmap_get_irq_val(&dist->irq_pending, vcpu->vcpu_id, irq);
377}
378
379void vgic_dist_irq_set_pending(struct kvm_vcpu *vcpu, int irq)
380{
381	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
382
383	vgic_bitmap_set_irq_val(&dist->irq_pending, vcpu->vcpu_id, irq, 1);
384}
385
386void vgic_dist_irq_clear_pending(struct kvm_vcpu *vcpu, int irq)
387{
388	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
389
390	vgic_bitmap_set_irq_val(&dist->irq_pending, vcpu->vcpu_id, irq, 0);
391}
392
393static void vgic_cpu_irq_set(struct kvm_vcpu *vcpu, int irq)
394{
395	if (irq < VGIC_NR_PRIVATE_IRQS)
396		set_bit(irq, vcpu->arch.vgic_cpu.pending_percpu);
397	else
398		set_bit(irq - VGIC_NR_PRIVATE_IRQS,
399			vcpu->arch.vgic_cpu.pending_shared);
400}
401
402void vgic_cpu_irq_clear(struct kvm_vcpu *vcpu, int irq)
403{
404	if (irq < VGIC_NR_PRIVATE_IRQS)
405		clear_bit(irq, vcpu->arch.vgic_cpu.pending_percpu);
406	else
407		clear_bit(irq - VGIC_NR_PRIVATE_IRQS,
408			  vcpu->arch.vgic_cpu.pending_shared);
409}
410
411static bool vgic_can_sample_irq(struct kvm_vcpu *vcpu, int irq)
412{
413	return !vgic_irq_is_queued(vcpu, irq);
414}
415
416/**
417 * vgic_reg_access - access vgic register
418 * @mmio:   pointer to the data describing the mmio access
419 * @reg:    pointer to the virtual backing of vgic distributor data
420 * @offset: least significant 2 bits used for word offset
421 * @mode:   ACCESS_ mode (see defines above)
422 *
423 * Helper to make vgic register access easier using one of the access
424 * modes defined for vgic register access
425 * (read,raz,write-ignored,setbit,clearbit,write)
426 */
427void vgic_reg_access(struct kvm_exit_mmio *mmio, u32 *reg,
428		     phys_addr_t offset, int mode)
429{
430	int word_offset = (offset & 3) * 8;
431	u32 mask = (1UL << (mmio->len * 8)) - 1;
432	u32 regval;
433
434	/*
435	 * Any alignment fault should have been delivered to the guest
436	 * directly (ARM ARM B3.12.7 "Prioritization of aborts").
437	 */
438
439	if (reg) {
440		regval = *reg;
441	} else {
442		BUG_ON(mode != (ACCESS_READ_RAZ | ACCESS_WRITE_IGNORED));
443		regval = 0;
444	}
445
446	if (mmio->is_write) {
447		u32 data = mmio_data_read(mmio, mask) << word_offset;
448		switch (ACCESS_WRITE_MASK(mode)) {
449		case ACCESS_WRITE_IGNORED:
450			return;
451
452		case ACCESS_WRITE_SETBIT:
453			regval |= data;
454			break;
455
456		case ACCESS_WRITE_CLEARBIT:
457			regval &= ~data;
458			break;
459
460		case ACCESS_WRITE_VALUE:
461			regval = (regval & ~(mask << word_offset)) | data;
462			break;
463		}
464		*reg = regval;
465	} else {
466		switch (ACCESS_READ_MASK(mode)) {
467		case ACCESS_READ_RAZ:
468			regval = 0;
469			/* fall through */
470
471		case ACCESS_READ_VALUE:
472			mmio_data_write(mmio, mask, regval >> word_offset);
473		}
474	}
475}
476
477bool handle_mmio_raz_wi(struct kvm_vcpu *vcpu, struct kvm_exit_mmio *mmio,
478			phys_addr_t offset)
479{
480	vgic_reg_access(mmio, NULL, offset,
481			ACCESS_READ_RAZ | ACCESS_WRITE_IGNORED);
482	return false;
483}
484
485bool vgic_handle_enable_reg(struct kvm *kvm, struct kvm_exit_mmio *mmio,
486			    phys_addr_t offset, int vcpu_id, int access)
487{
488	u32 *reg;
489	int mode = ACCESS_READ_VALUE | access;
490	struct kvm_vcpu *target_vcpu = kvm_get_vcpu(kvm, vcpu_id);
491
492	reg = vgic_bitmap_get_reg(&kvm->arch.vgic.irq_enabled, vcpu_id, offset);
493	vgic_reg_access(mmio, reg, offset, mode);
494	if (mmio->is_write) {
495		if (access & ACCESS_WRITE_CLEARBIT) {
496			if (offset < 4) /* Force SGI enabled */
497				*reg |= 0xffff;
498			vgic_retire_disabled_irqs(target_vcpu);
499		}
500		vgic_update_state(kvm);
501		return true;
502	}
503
504	return false;
505}
506
507bool vgic_handle_set_pending_reg(struct kvm *kvm,
508				 struct kvm_exit_mmio *mmio,
509				 phys_addr_t offset, int vcpu_id)
510{
511	u32 *reg, orig;
512	u32 level_mask;
513	int mode = ACCESS_READ_VALUE | ACCESS_WRITE_SETBIT;
514	struct vgic_dist *dist = &kvm->arch.vgic;
515
516	reg = vgic_bitmap_get_reg(&dist->irq_cfg, vcpu_id, offset);
517	level_mask = (~(*reg));
518
519	/* Mark both level and edge triggered irqs as pending */
520	reg = vgic_bitmap_get_reg(&dist->irq_pending, vcpu_id, offset);
521	orig = *reg;
522	vgic_reg_access(mmio, reg, offset, mode);
523
524	if (mmio->is_write) {
525		/* Set the soft-pending flag only for level-triggered irqs */
526		reg = vgic_bitmap_get_reg(&dist->irq_soft_pend,
527					  vcpu_id, offset);
528		vgic_reg_access(mmio, reg, offset, mode);
529		*reg &= level_mask;
530
531		/* Ignore writes to SGIs */
532		if (offset < 2) {
533			*reg &= ~0xffff;
534			*reg |= orig & 0xffff;
535		}
536
537		vgic_update_state(kvm);
538		return true;
539	}
540
541	return false;
542}
543
544bool vgic_handle_clear_pending_reg(struct kvm *kvm,
545				   struct kvm_exit_mmio *mmio,
546				   phys_addr_t offset, int vcpu_id)
547{
548	u32 *level_active;
549	u32 *reg, orig;
550	int mode = ACCESS_READ_VALUE | ACCESS_WRITE_CLEARBIT;
551	struct vgic_dist *dist = &kvm->arch.vgic;
552
553	reg = vgic_bitmap_get_reg(&dist->irq_pending, vcpu_id, offset);
554	orig = *reg;
555	vgic_reg_access(mmio, reg, offset, mode);
556	if (mmio->is_write) {
557		/* Re-set level triggered level-active interrupts */
558		level_active = vgic_bitmap_get_reg(&dist->irq_level,
559					  vcpu_id, offset);
560		reg = vgic_bitmap_get_reg(&dist->irq_pending, vcpu_id, offset);
561		*reg |= *level_active;
562
563		/* Ignore writes to SGIs */
564		if (offset < 2) {
565			*reg &= ~0xffff;
566			*reg |= orig & 0xffff;
567		}
568
569		/* Clear soft-pending flags */
570		reg = vgic_bitmap_get_reg(&dist->irq_soft_pend,
571					  vcpu_id, offset);
572		vgic_reg_access(mmio, reg, offset, mode);
573
574		vgic_update_state(kvm);
575		return true;
576	}
577	return false;
578}
579
580bool vgic_handle_set_active_reg(struct kvm *kvm,
581				struct kvm_exit_mmio *mmio,
582				phys_addr_t offset, int vcpu_id)
583{
584	u32 *reg;
585	struct vgic_dist *dist = &kvm->arch.vgic;
586
587	reg = vgic_bitmap_get_reg(&dist->irq_active, vcpu_id, offset);
588	vgic_reg_access(mmio, reg, offset,
589			ACCESS_READ_VALUE | ACCESS_WRITE_SETBIT);
590
591	if (mmio->is_write) {
592		vgic_update_state(kvm);
593		return true;
594	}
595
596	return false;
597}
598
599bool vgic_handle_clear_active_reg(struct kvm *kvm,
600				  struct kvm_exit_mmio *mmio,
601				  phys_addr_t offset, int vcpu_id)
602{
603	u32 *reg;
604	struct vgic_dist *dist = &kvm->arch.vgic;
605
606	reg = vgic_bitmap_get_reg(&dist->irq_active, vcpu_id, offset);
607	vgic_reg_access(mmio, reg, offset,
608			ACCESS_READ_VALUE | ACCESS_WRITE_CLEARBIT);
609
610	if (mmio->is_write) {
611		vgic_update_state(kvm);
612		return true;
613	}
614
615	return false;
616}
617
618static u32 vgic_cfg_expand(u16 val)
619{
620	u32 res = 0;
621	int i;
622
623	/*
624	 * Turn a 16bit value like abcd...mnop into a 32bit word
625	 * a0b0c0d0...m0n0o0p0, which is what the HW cfg register is.
626	 */
627	for (i = 0; i < 16; i++)
628		res |= ((val >> i) & VGIC_CFG_EDGE) << (2 * i + 1);
629
630	return res;
631}
632
633static u16 vgic_cfg_compress(u32 val)
634{
635	u16 res = 0;
636	int i;
637
638	/*
639	 * Turn a 32bit word a0b0c0d0...m0n0o0p0 into 16bit value like
640	 * abcd...mnop which is what we really care about.
641	 */
642	for (i = 0; i < 16; i++)
643		res |= ((val >> (i * 2 + 1)) & VGIC_CFG_EDGE) << i;
644
645	return res;
646}
647
648/*
649 * The distributor uses 2 bits per IRQ for the CFG register, but the
650 * LSB is always 0. As such, we only keep the upper bit, and use the
651 * two above functions to compress/expand the bits
652 */
653bool vgic_handle_cfg_reg(u32 *reg, struct kvm_exit_mmio *mmio,
654			 phys_addr_t offset)
655{
656	u32 val;
657
658	if (offset & 4)
659		val = *reg >> 16;
660	else
661		val = *reg & 0xffff;
662
663	val = vgic_cfg_expand(val);
664	vgic_reg_access(mmio, &val, offset,
665			ACCESS_READ_VALUE | ACCESS_WRITE_VALUE);
666	if (mmio->is_write) {
667		/* Ignore writes to read-only SGI and PPI bits */
668		if (offset < 8)
669			return false;
670
671		val = vgic_cfg_compress(val);
672		if (offset & 4) {
673			*reg &= 0xffff;
674			*reg |= val << 16;
675		} else {
676			*reg &= 0xffff << 16;
677			*reg |= val;
678		}
679	}
680
681	return false;
682}
683
684/**
685 * vgic_unqueue_irqs - move pending/active IRQs from LRs to the distributor
686 * @vgic_cpu: Pointer to the vgic_cpu struct holding the LRs
687 *
688 * Move any IRQs that have already been assigned to LRs back to the
689 * emulated distributor state so that the complete emulated state can be read
690 * from the main emulation structures without investigating the LRs.
691 */
692void vgic_unqueue_irqs(struct kvm_vcpu *vcpu)
693{
694	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
695	u64 elrsr = vgic_get_elrsr(vcpu);
696	unsigned long *elrsr_ptr = u64_to_bitmask(&elrsr);
697	int i;
698
699	for_each_clear_bit(i, elrsr_ptr, vgic_cpu->nr_lr) {
700		struct vgic_lr lr = vgic_get_lr(vcpu, i);
701
702		/*
703		 * There are three options for the state bits:
704		 *
705		 * 01: pending
706		 * 10: active
707		 * 11: pending and active
708		 */
709		BUG_ON(!(lr.state & LR_STATE_MASK));
710
711		/* Reestablish SGI source for pending and active IRQs */
712		if (lr.irq < VGIC_NR_SGIS)
713			add_sgi_source(vcpu, lr.irq, lr.source);
714
715		/*
716		 * If the LR holds an active (10) or a pending and active (11)
717		 * interrupt then move the active state to the
718		 * distributor tracking bit.
719		 */
720		if (lr.state & LR_STATE_ACTIVE)
721			vgic_irq_set_active(vcpu, lr.irq);
722
723		/*
724		 * Reestablish the pending state on the distributor and the
725		 * CPU interface and mark the LR as free for other use.
726		 */
727		vgic_retire_lr(i, vcpu);
728
729		/* Finally update the VGIC state. */
730		vgic_update_state(vcpu->kvm);
731	}
732}
733
734const
735struct vgic_io_range *vgic_find_range(const struct vgic_io_range *ranges,
736				      int len, gpa_t offset)
737{
738	while (ranges->len) {
739		if (offset >= ranges->base &&
740		    (offset + len) <= (ranges->base + ranges->len))
741			return ranges;
742		ranges++;
743	}
744
745	return NULL;
746}
747
748static bool vgic_validate_access(const struct vgic_dist *dist,
749				 const struct vgic_io_range *range,
750				 unsigned long offset)
751{
752	int irq;
753
754	if (!range->bits_per_irq)
755		return true;	/* Not an irq-based access */
756
757	irq = offset * 8 / range->bits_per_irq;
758	if (irq >= dist->nr_irqs)
759		return false;
760
761	return true;
762}
763
764/*
765 * Call the respective handler function for the given range.
766 * We split up any 64 bit accesses into two consecutive 32 bit
767 * handler calls and merge the result afterwards.
768 * We do this in a little endian fashion regardless of the host's
769 * or guest's endianness, because the GIC is always LE and the rest of
770 * the code (vgic_reg_access) also puts it in a LE fashion already.
771 * At this point we have already identified the handle function, so
772 * range points to that one entry and offset is relative to this.
773 */
774static bool call_range_handler(struct kvm_vcpu *vcpu,
775			       struct kvm_exit_mmio *mmio,
776			       unsigned long offset,
777			       const struct vgic_io_range *range)
778{
779	struct kvm_exit_mmio mmio32;
780	bool ret;
781
782	if (likely(mmio->len <= 4))
783		return range->handle_mmio(vcpu, mmio, offset);
784
785	/*
786	 * Any access bigger than 4 bytes (that we currently handle in KVM)
787	 * is actually 8 bytes long, caused by a 64-bit access
788	 */
789
790	mmio32.len = 4;
791	mmio32.is_write = mmio->is_write;
792	mmio32.private = mmio->private;
793
794	mmio32.phys_addr = mmio->phys_addr + 4;
795	mmio32.data = &((u32 *)mmio->data)[1];
796	ret = range->handle_mmio(vcpu, &mmio32, offset + 4);
797
798	mmio32.phys_addr = mmio->phys_addr;
799	mmio32.data = &((u32 *)mmio->data)[0];
800	ret |= range->handle_mmio(vcpu, &mmio32, offset);
801
802	return ret;
803}
804
805/**
806 * vgic_handle_mmio_access - handle an in-kernel MMIO access
807 * This is called by the read/write KVM IO device wrappers below.
808 * @vcpu:	pointer to the vcpu performing the access
809 * @this:	pointer to the KVM IO device in charge
810 * @addr:	guest physical address of the access
811 * @len:	size of the access
812 * @val:	pointer to the data region
813 * @is_write:	read or write access
814 *
815 * returns true if the MMIO access could be performed
816 */
817static int vgic_handle_mmio_access(struct kvm_vcpu *vcpu,
818				   struct kvm_io_device *this, gpa_t addr,
819				   int len, void *val, bool is_write)
820{
821	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
822	struct vgic_io_device *iodev = container_of(this,
823						    struct vgic_io_device, dev);
824	struct kvm_run *run = vcpu->run;
825	const struct vgic_io_range *range;
826	struct kvm_exit_mmio mmio;
827	bool updated_state;
828	gpa_t offset;
829
830	offset = addr - iodev->addr;
831	range = vgic_find_range(iodev->reg_ranges, len, offset);
832	if (unlikely(!range || !range->handle_mmio)) {
833		pr_warn("Unhandled access %d %08llx %d\n", is_write, addr, len);
834		return -ENXIO;
835	}
836
837	mmio.phys_addr = addr;
838	mmio.len = len;
839	mmio.is_write = is_write;
840	mmio.data = val;
841	mmio.private = iodev->redist_vcpu;
842
843	spin_lock(&dist->lock);
844	offset -= range->base;
845	if (vgic_validate_access(dist, range, offset)) {
846		updated_state = call_range_handler(vcpu, &mmio, offset, range);
847	} else {
848		if (!is_write)
849			memset(val, 0, len);
850		updated_state = false;
851	}
852	spin_unlock(&dist->lock);
853	run->mmio.is_write	= is_write;
854	run->mmio.len		= len;
855	run->mmio.phys_addr	= addr;
856	memcpy(run->mmio.data, val, len);
857
858	kvm_handle_mmio_return(vcpu, run);
859
860	if (updated_state)
861		vgic_kick_vcpus(vcpu->kvm);
862
863	return 0;
864}
865
866static int vgic_handle_mmio_read(struct kvm_vcpu *vcpu,
867				 struct kvm_io_device *this,
868				 gpa_t addr, int len, void *val)
869{
870	return vgic_handle_mmio_access(vcpu, this, addr, len, val, false);
871}
872
873static int vgic_handle_mmio_write(struct kvm_vcpu *vcpu,
874				  struct kvm_io_device *this,
875				  gpa_t addr, int len, const void *val)
876{
877	return vgic_handle_mmio_access(vcpu, this, addr, len, (void *)val,
878				       true);
879}
880
881struct kvm_io_device_ops vgic_io_ops = {
882	.read	= vgic_handle_mmio_read,
883	.write	= vgic_handle_mmio_write,
884};
885
886/**
887 * vgic_register_kvm_io_dev - register VGIC register frame on the KVM I/O bus
888 * @kvm:            The VM structure pointer
889 * @base:           The (guest) base address for the register frame
890 * @len:            Length of the register frame window
891 * @ranges:         Describing the handler functions for each register
892 * @redist_vcpu_id: The VCPU ID to pass on to the handlers on call
893 * @iodev:          Points to memory to be passed on to the handler
894 *
895 * @iodev stores the parameters of this function to be usable by the handler
896 * respectively the dispatcher function (since the KVM I/O bus framework lacks
897 * an opaque parameter). Initialization is done in this function, but the
898 * reference should be valid and unique for the whole VGIC lifetime.
899 * If the register frame is not mapped for a specific VCPU, pass -1 to
900 * @redist_vcpu_id.
901 */
902int vgic_register_kvm_io_dev(struct kvm *kvm, gpa_t base, int len,
903			     const struct vgic_io_range *ranges,
904			     int redist_vcpu_id,
905			     struct vgic_io_device *iodev)
906{
907	struct kvm_vcpu *vcpu = NULL;
908	int ret;
909
910	if (redist_vcpu_id >= 0)
911		vcpu = kvm_get_vcpu(kvm, redist_vcpu_id);
912
913	iodev->addr		= base;
914	iodev->len		= len;
915	iodev->reg_ranges	= ranges;
916	iodev->redist_vcpu	= vcpu;
917
918	kvm_iodevice_init(&iodev->dev, &vgic_io_ops);
919
920	mutex_lock(&kvm->slots_lock);
921
922	ret = kvm_io_bus_register_dev(kvm, KVM_MMIO_BUS, base, len,
923				      &iodev->dev);
924	mutex_unlock(&kvm->slots_lock);
925
926	/* Mark the iodev as invalid if registration fails. */
927	if (ret)
928		iodev->dev.ops = NULL;
929
930	return ret;
931}
932
933static int vgic_nr_shared_irqs(struct vgic_dist *dist)
934{
935	return dist->nr_irqs - VGIC_NR_PRIVATE_IRQS;
936}
937
938static int compute_active_for_cpu(struct kvm_vcpu *vcpu)
939{
940	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
941	unsigned long *active, *enabled, *act_percpu, *act_shared;
942	unsigned long active_private, active_shared;
943	int nr_shared = vgic_nr_shared_irqs(dist);
944	int vcpu_id;
945
946	vcpu_id = vcpu->vcpu_id;
947	act_percpu = vcpu->arch.vgic_cpu.active_percpu;
948	act_shared = vcpu->arch.vgic_cpu.active_shared;
949
950	active = vgic_bitmap_get_cpu_map(&dist->irq_active, vcpu_id);
951	enabled = vgic_bitmap_get_cpu_map(&dist->irq_enabled, vcpu_id);
952	bitmap_and(act_percpu, active, enabled, VGIC_NR_PRIVATE_IRQS);
953
954	active = vgic_bitmap_get_shared_map(&dist->irq_active);
955	enabled = vgic_bitmap_get_shared_map(&dist->irq_enabled);
956	bitmap_and(act_shared, active, enabled, nr_shared);
957	bitmap_and(act_shared, act_shared,
958		   vgic_bitmap_get_shared_map(&dist->irq_spi_target[vcpu_id]),
959		   nr_shared);
960
961	active_private = find_first_bit(act_percpu, VGIC_NR_PRIVATE_IRQS);
962	active_shared = find_first_bit(act_shared, nr_shared);
963
964	return (active_private < VGIC_NR_PRIVATE_IRQS ||
965		active_shared < nr_shared);
966}
967
968static int compute_pending_for_cpu(struct kvm_vcpu *vcpu)
969{
970	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
971	unsigned long *pending, *enabled, *pend_percpu, *pend_shared;
972	unsigned long pending_private, pending_shared;
973	int nr_shared = vgic_nr_shared_irqs(dist);
974	int vcpu_id;
975
976	vcpu_id = vcpu->vcpu_id;
977	pend_percpu = vcpu->arch.vgic_cpu.pending_percpu;
978	pend_shared = vcpu->arch.vgic_cpu.pending_shared;
979
980	if (!dist->enabled) {
981		bitmap_zero(pend_percpu, VGIC_NR_PRIVATE_IRQS);
982		bitmap_zero(pend_shared, nr_shared);
983		return 0;
984	}
985
986	pending = vgic_bitmap_get_cpu_map(&dist->irq_pending, vcpu_id);
987	enabled = vgic_bitmap_get_cpu_map(&dist->irq_enabled, vcpu_id);
988	bitmap_and(pend_percpu, pending, enabled, VGIC_NR_PRIVATE_IRQS);
989
990	pending = vgic_bitmap_get_shared_map(&dist->irq_pending);
991	enabled = vgic_bitmap_get_shared_map(&dist->irq_enabled);
992	bitmap_and(pend_shared, pending, enabled, nr_shared);
993	bitmap_and(pend_shared, pend_shared,
994		   vgic_bitmap_get_shared_map(&dist->irq_spi_target[vcpu_id]),
995		   nr_shared);
996
997	pending_private = find_first_bit(pend_percpu, VGIC_NR_PRIVATE_IRQS);
998	pending_shared = find_first_bit(pend_shared, nr_shared);
999	return (pending_private < VGIC_NR_PRIVATE_IRQS ||
1000		pending_shared < vgic_nr_shared_irqs(dist));
1001}
1002
1003/*
1004 * Update the interrupt state and determine which CPUs have pending
1005 * or active interrupts. Must be called with distributor lock held.
1006 */
1007void vgic_update_state(struct kvm *kvm)
1008{
1009	struct vgic_dist *dist = &kvm->arch.vgic;
1010	struct kvm_vcpu *vcpu;
1011	int c;
1012
1013	kvm_for_each_vcpu(c, vcpu, kvm) {
1014		if (compute_pending_for_cpu(vcpu))
1015			set_bit(c, dist->irq_pending_on_cpu);
1016
1017		if (compute_active_for_cpu(vcpu))
1018			set_bit(c, dist->irq_active_on_cpu);
1019		else
1020			clear_bit(c, dist->irq_active_on_cpu);
1021	}
1022}
1023
1024static struct vgic_lr vgic_get_lr(const struct kvm_vcpu *vcpu, int lr)
1025{
1026	return vgic_ops->get_lr(vcpu, lr);
1027}
1028
1029static void vgic_set_lr(struct kvm_vcpu *vcpu, int lr,
1030			       struct vgic_lr vlr)
1031{
1032	vgic_ops->set_lr(vcpu, lr, vlr);
1033}
1034
1035static inline u64 vgic_get_elrsr(struct kvm_vcpu *vcpu)
1036{
1037	return vgic_ops->get_elrsr(vcpu);
1038}
1039
1040static inline u64 vgic_get_eisr(struct kvm_vcpu *vcpu)
1041{
1042	return vgic_ops->get_eisr(vcpu);
1043}
1044
1045static inline void vgic_clear_eisr(struct kvm_vcpu *vcpu)
1046{
1047	vgic_ops->clear_eisr(vcpu);
1048}
1049
1050static inline u32 vgic_get_interrupt_status(struct kvm_vcpu *vcpu)
1051{
1052	return vgic_ops->get_interrupt_status(vcpu);
1053}
1054
1055static inline void vgic_enable_underflow(struct kvm_vcpu *vcpu)
1056{
1057	vgic_ops->enable_underflow(vcpu);
1058}
1059
1060static inline void vgic_disable_underflow(struct kvm_vcpu *vcpu)
1061{
1062	vgic_ops->disable_underflow(vcpu);
1063}
1064
1065void vgic_get_vmcr(struct kvm_vcpu *vcpu, struct vgic_vmcr *vmcr)
1066{
1067	vgic_ops->get_vmcr(vcpu, vmcr);
1068}
1069
1070void vgic_set_vmcr(struct kvm_vcpu *vcpu, struct vgic_vmcr *vmcr)
1071{
1072	vgic_ops->set_vmcr(vcpu, vmcr);
1073}
1074
1075static inline void vgic_enable(struct kvm_vcpu *vcpu)
1076{
1077	vgic_ops->enable(vcpu);
1078}
1079
1080static void vgic_retire_lr(int lr_nr, struct kvm_vcpu *vcpu)
1081{
1082	struct vgic_lr vlr = vgic_get_lr(vcpu, lr_nr);
1083
1084	vgic_irq_clear_queued(vcpu, vlr.irq);
1085
1086	/*
1087	 * We must transfer the pending state back to the distributor before
1088	 * retiring the LR, otherwise we may loose edge-triggered interrupts.
1089	 */
1090	if (vlr.state & LR_STATE_PENDING) {
1091		vgic_dist_irq_set_pending(vcpu, vlr.irq);
1092		vlr.hwirq = 0;
1093	}
1094
1095	vlr.state = 0;
1096	vgic_set_lr(vcpu, lr_nr, vlr);
1097}
1098
1099static bool dist_active_irq(struct kvm_vcpu *vcpu)
1100{
1101	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1102
1103	return test_bit(vcpu->vcpu_id, dist->irq_active_on_cpu);
1104}
1105
1106bool kvm_vgic_map_is_active(struct kvm_vcpu *vcpu, struct irq_phys_map *map)
1107{
1108	int i;
1109
1110	for (i = 0; i < vcpu->arch.vgic_cpu.nr_lr; i++) {
1111		struct vgic_lr vlr = vgic_get_lr(vcpu, i);
1112
1113		if (vlr.irq == map->virt_irq && vlr.state & LR_STATE_ACTIVE)
1114			return true;
1115	}
1116
1117	return vgic_irq_is_active(vcpu, map->virt_irq);
1118}
1119
1120/*
1121 * An interrupt may have been disabled after being made pending on the
1122 * CPU interface (the classic case is a timer running while we're
1123 * rebooting the guest - the interrupt would kick as soon as the CPU
1124 * interface gets enabled, with deadly consequences).
1125 *
1126 * The solution is to examine already active LRs, and check the
1127 * interrupt is still enabled. If not, just retire it.
1128 */
1129static void vgic_retire_disabled_irqs(struct kvm_vcpu *vcpu)
1130{
1131	u64 elrsr = vgic_get_elrsr(vcpu);
1132	unsigned long *elrsr_ptr = u64_to_bitmask(&elrsr);
1133	int lr;
1134
1135	for_each_clear_bit(lr, elrsr_ptr, vgic->nr_lr) {
1136		struct vgic_lr vlr = vgic_get_lr(vcpu, lr);
1137
1138		if (!vgic_irq_is_enabled(vcpu, vlr.irq))
1139			vgic_retire_lr(lr, vcpu);
1140	}
1141}
1142
1143static void vgic_queue_irq_to_lr(struct kvm_vcpu *vcpu, int irq,
1144				 int lr_nr, struct vgic_lr vlr)
1145{
1146	if (vgic_irq_is_active(vcpu, irq)) {
1147		vlr.state |= LR_STATE_ACTIVE;
1148		kvm_debug("Set active, clear distributor: 0x%x\n", vlr.state);
1149		vgic_irq_clear_active(vcpu, irq);
1150		vgic_update_state(vcpu->kvm);
1151	} else {
1152		WARN_ON(!vgic_dist_irq_is_pending(vcpu, irq));
1153		vlr.state |= LR_STATE_PENDING;
1154		kvm_debug("Set pending: 0x%x\n", vlr.state);
1155	}
1156
1157	if (!vgic_irq_is_edge(vcpu, irq))
1158		vlr.state |= LR_EOI_INT;
1159
1160	if (vlr.irq >= VGIC_NR_SGIS) {
1161		struct irq_phys_map *map;
1162		map = vgic_irq_map_search(vcpu, irq);
1163
1164		if (map) {
1165			vlr.hwirq = map->phys_irq;
1166			vlr.state |= LR_HW;
1167			vlr.state &= ~LR_EOI_INT;
1168
1169			/*
1170			 * Make sure we're not going to sample this
1171			 * again, as a HW-backed interrupt cannot be
1172			 * in the PENDING_ACTIVE stage.
1173			 */
1174			vgic_irq_set_queued(vcpu, irq);
1175		}
1176	}
1177
1178	vgic_set_lr(vcpu, lr_nr, vlr);
1179}
1180
1181/*
1182 * Queue an interrupt to a CPU virtual interface. Return true on success,
1183 * or false if it wasn't possible to queue it.
1184 * sgi_source must be zero for any non-SGI interrupts.
1185 */
1186bool vgic_queue_irq(struct kvm_vcpu *vcpu, u8 sgi_source_id, int irq)
1187{
1188	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1189	u64 elrsr = vgic_get_elrsr(vcpu);
1190	unsigned long *elrsr_ptr = u64_to_bitmask(&elrsr);
1191	struct vgic_lr vlr;
1192	int lr;
1193
1194	/* Sanitize the input... */
1195	BUG_ON(sgi_source_id & ~7);
1196	BUG_ON(sgi_source_id && irq >= VGIC_NR_SGIS);
1197	BUG_ON(irq >= dist->nr_irqs);
1198
1199	kvm_debug("Queue IRQ%d\n", irq);
1200
1201	/* Do we have an active interrupt for the same CPUID? */
1202	for_each_clear_bit(lr, elrsr_ptr, vgic->nr_lr) {
1203		vlr = vgic_get_lr(vcpu, lr);
1204		if (vlr.irq == irq && vlr.source == sgi_source_id) {
1205			kvm_debug("LR%d piggyback for IRQ%d\n", lr, vlr.irq);
1206			vgic_queue_irq_to_lr(vcpu, irq, lr, vlr);
1207			return true;
1208		}
1209	}
1210
1211	/* Try to use another LR for this interrupt */
1212	lr = find_first_bit(elrsr_ptr, vgic->nr_lr);
1213	if (lr >= vgic->nr_lr)
1214		return false;
1215
1216	kvm_debug("LR%d allocated for IRQ%d %x\n", lr, irq, sgi_source_id);
1217
1218	vlr.irq = irq;
1219	vlr.source = sgi_source_id;
1220	vlr.state = 0;
1221	vgic_queue_irq_to_lr(vcpu, irq, lr, vlr);
1222
1223	return true;
1224}
1225
1226static bool vgic_queue_hwirq(struct kvm_vcpu *vcpu, int irq)
1227{
1228	if (!vgic_can_sample_irq(vcpu, irq))
1229		return true; /* level interrupt, already queued */
1230
1231	if (vgic_queue_irq(vcpu, 0, irq)) {
1232		if (vgic_irq_is_edge(vcpu, irq)) {
1233			vgic_dist_irq_clear_pending(vcpu, irq);
1234			vgic_cpu_irq_clear(vcpu, irq);
1235		} else {
1236			vgic_irq_set_queued(vcpu, irq);
1237		}
1238
1239		return true;
1240	}
1241
1242	return false;
1243}
1244
1245/*
1246 * Fill the list registers with pending interrupts before running the
1247 * guest.
1248 */
1249static void __kvm_vgic_flush_hwstate(struct kvm_vcpu *vcpu)
1250{
1251	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
1252	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1253	unsigned long *pa_percpu, *pa_shared;
1254	int i, vcpu_id;
1255	int overflow = 0;
1256	int nr_shared = vgic_nr_shared_irqs(dist);
1257
1258	vcpu_id = vcpu->vcpu_id;
1259
1260	pa_percpu = vcpu->arch.vgic_cpu.pend_act_percpu;
1261	pa_shared = vcpu->arch.vgic_cpu.pend_act_shared;
1262
1263	bitmap_or(pa_percpu, vgic_cpu->pending_percpu, vgic_cpu->active_percpu,
1264		  VGIC_NR_PRIVATE_IRQS);
1265	bitmap_or(pa_shared, vgic_cpu->pending_shared, vgic_cpu->active_shared,
1266		  nr_shared);
1267	/*
1268	 * We may not have any pending interrupt, or the interrupts
1269	 * may have been serviced from another vcpu. In all cases,
1270	 * move along.
1271	 */
1272	if (!kvm_vgic_vcpu_pending_irq(vcpu) && !dist_active_irq(vcpu))
1273		goto epilog;
1274
1275	/* SGIs */
1276	for_each_set_bit(i, pa_percpu, VGIC_NR_SGIS) {
1277		if (!queue_sgi(vcpu, i))
1278			overflow = 1;
1279	}
1280
1281	/* PPIs */
1282	for_each_set_bit_from(i, pa_percpu, VGIC_NR_PRIVATE_IRQS) {
1283		if (!vgic_queue_hwirq(vcpu, i))
1284			overflow = 1;
1285	}
1286
1287	/* SPIs */
1288	for_each_set_bit(i, pa_shared, nr_shared) {
1289		if (!vgic_queue_hwirq(vcpu, i + VGIC_NR_PRIVATE_IRQS))
1290			overflow = 1;
1291	}
1292
1293
1294
1295
1296epilog:
1297	if (overflow) {
1298		vgic_enable_underflow(vcpu);
1299	} else {
1300		vgic_disable_underflow(vcpu);
1301		/*
1302		 * We're about to run this VCPU, and we've consumed
1303		 * everything the distributor had in store for
1304		 * us. Claim we don't have anything pending. We'll
1305		 * adjust that if needed while exiting.
1306		 */
1307		clear_bit(vcpu_id, dist->irq_pending_on_cpu);
1308	}
1309}
1310
1311static int process_queued_irq(struct kvm_vcpu *vcpu,
1312				   int lr, struct vgic_lr vlr)
1313{
1314	int pending = 0;
1315
1316	/*
1317	 * If the IRQ was EOIed (called from vgic_process_maintenance) or it
1318	 * went from active to non-active (called from vgic_sync_hwirq) it was
1319	 * also ACKed and we we therefore assume we can clear the soft pending
1320	 * state (should it had been set) for this interrupt.
1321	 *
1322	 * Note: if the IRQ soft pending state was set after the IRQ was
1323	 * acked, it actually shouldn't be cleared, but we have no way of
1324	 * knowing that unless we start trapping ACKs when the soft-pending
1325	 * state is set.
1326	 */
1327	vgic_dist_irq_clear_soft_pend(vcpu, vlr.irq);
1328
1329	/*
1330	 * Tell the gic to start sampling this interrupt again.
1331	 */
1332	vgic_irq_clear_queued(vcpu, vlr.irq);
1333
1334	/* Any additional pending interrupt? */
1335	if (vgic_irq_is_edge(vcpu, vlr.irq)) {
1336		BUG_ON(!(vlr.state & LR_HW));
1337		pending = vgic_dist_irq_is_pending(vcpu, vlr.irq);
1338	} else {
1339		if (vgic_dist_irq_get_level(vcpu, vlr.irq)) {
1340			vgic_cpu_irq_set(vcpu, vlr.irq);
1341			pending = 1;
1342		} else {
1343			vgic_dist_irq_clear_pending(vcpu, vlr.irq);
1344			vgic_cpu_irq_clear(vcpu, vlr.irq);
1345		}
1346	}
1347
1348	/*
1349	 * Despite being EOIed, the LR may not have
1350	 * been marked as empty.
1351	 */
1352	vlr.state = 0;
1353	vlr.hwirq = 0;
1354	vgic_set_lr(vcpu, lr, vlr);
1355
1356	return pending;
1357}
1358
1359static bool vgic_process_maintenance(struct kvm_vcpu *vcpu)
1360{
1361	u32 status = vgic_get_interrupt_status(vcpu);
1362	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1363	struct kvm *kvm = vcpu->kvm;
1364	int level_pending = 0;
1365
1366	kvm_debug("STATUS = %08x\n", status);
1367
1368	if (status & INT_STATUS_EOI) {
1369		/*
1370		 * Some level interrupts have been EOIed. Clear their
1371		 * active bit.
1372		 */
1373		u64 eisr = vgic_get_eisr(vcpu);
1374		unsigned long *eisr_ptr = u64_to_bitmask(&eisr);
1375		int lr;
1376
1377		for_each_set_bit(lr, eisr_ptr, vgic->nr_lr) {
1378			struct vgic_lr vlr = vgic_get_lr(vcpu, lr);
1379
1380			WARN_ON(vgic_irq_is_edge(vcpu, vlr.irq));
1381			WARN_ON(vlr.state & LR_STATE_MASK);
1382
1383
1384			/*
1385			 * kvm_notify_acked_irq calls kvm_set_irq()
1386			 * to reset the IRQ level, which grabs the dist->lock
1387			 * so we call this before taking the dist->lock.
1388			 */
1389			kvm_notify_acked_irq(kvm, 0,
1390					     vlr.irq - VGIC_NR_PRIVATE_IRQS);
1391
1392			spin_lock(&dist->lock);
1393			level_pending |= process_queued_irq(vcpu, lr, vlr);
1394			spin_unlock(&dist->lock);
1395		}
1396	}
1397
1398	if (status & INT_STATUS_UNDERFLOW)
1399		vgic_disable_underflow(vcpu);
1400
1401	/*
1402	 * In the next iterations of the vcpu loop, if we sync the vgic state
1403	 * after flushing it, but before entering the guest (this happens for
1404	 * pending signals and vmid rollovers), then make sure we don't pick
1405	 * up any old maintenance interrupts here.
1406	 */
1407	vgic_clear_eisr(vcpu);
1408
1409	return level_pending;
1410}
1411
1412/*
1413 * Save the physical active state, and reset it to inactive.
1414 *
1415 * Return true if there's a pending forwarded interrupt to queue.
1416 */
1417static bool vgic_sync_hwirq(struct kvm_vcpu *vcpu, int lr, struct vgic_lr vlr)
1418{
1419	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1420	bool level_pending;
1421
1422	if (!(vlr.state & LR_HW))
1423		return false;
1424
1425	if (vlr.state & LR_STATE_ACTIVE)
1426		return false;
1427
1428	spin_lock(&dist->lock);
1429	level_pending = process_queued_irq(vcpu, lr, vlr);
1430	spin_unlock(&dist->lock);
1431	return level_pending;
1432}
1433
1434/* Sync back the VGIC state after a guest run */
1435static void __kvm_vgic_sync_hwstate(struct kvm_vcpu *vcpu)
1436{
1437	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1438	u64 elrsr;
1439	unsigned long *elrsr_ptr;
1440	int lr, pending;
1441	bool level_pending;
1442
1443	level_pending = vgic_process_maintenance(vcpu);
1444
1445	/* Deal with HW interrupts, and clear mappings for empty LRs */
1446	for (lr = 0; lr < vgic->nr_lr; lr++) {
1447		struct vgic_lr vlr = vgic_get_lr(vcpu, lr);
1448
1449		level_pending |= vgic_sync_hwirq(vcpu, lr, vlr);
1450		BUG_ON(vlr.irq >= dist->nr_irqs);
1451	}
1452
1453	/* Check if we still have something up our sleeve... */
1454	elrsr = vgic_get_elrsr(vcpu);
1455	elrsr_ptr = u64_to_bitmask(&elrsr);
1456	pending = find_first_zero_bit(elrsr_ptr, vgic->nr_lr);
1457	if (level_pending || pending < vgic->nr_lr)
1458		set_bit(vcpu->vcpu_id, dist->irq_pending_on_cpu);
1459}
1460
1461void kvm_vgic_flush_hwstate(struct kvm_vcpu *vcpu)
1462{
1463	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1464
1465	if (!irqchip_in_kernel(vcpu->kvm))
1466		return;
1467
1468	spin_lock(&dist->lock);
1469	__kvm_vgic_flush_hwstate(vcpu);
1470	spin_unlock(&dist->lock);
1471}
1472
1473void kvm_vgic_sync_hwstate(struct kvm_vcpu *vcpu)
1474{
1475	if (!irqchip_in_kernel(vcpu->kvm))
1476		return;
1477
1478	__kvm_vgic_sync_hwstate(vcpu);
1479}
1480
1481int kvm_vgic_vcpu_pending_irq(struct kvm_vcpu *vcpu)
1482{
1483	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1484
1485	if (!irqchip_in_kernel(vcpu->kvm))
1486		return 0;
1487
1488	return test_bit(vcpu->vcpu_id, dist->irq_pending_on_cpu);
1489}
1490
1491void vgic_kick_vcpus(struct kvm *kvm)
1492{
1493	struct kvm_vcpu *vcpu;
1494	int c;
1495
1496	/*
1497	 * We've injected an interrupt, time to find out who deserves
1498	 * a good kick...
1499	 */
1500	kvm_for_each_vcpu(c, vcpu, kvm) {
1501		if (kvm_vgic_vcpu_pending_irq(vcpu))
1502			kvm_vcpu_kick(vcpu);
1503	}
1504}
1505
1506static int vgic_validate_injection(struct kvm_vcpu *vcpu, int irq, int level)
1507{
1508	int edge_triggered = vgic_irq_is_edge(vcpu, irq);
1509
1510	/*
1511	 * Only inject an interrupt if:
1512	 * - edge triggered and we have a rising edge
1513	 * - level triggered and we change level
1514	 */
1515	if (edge_triggered) {
1516		int state = vgic_dist_irq_is_pending(vcpu, irq);
1517		return level > state;
1518	} else {
1519		int state = vgic_dist_irq_get_level(vcpu, irq);
1520		return level != state;
1521	}
1522}
1523
1524static int vgic_update_irq_pending(struct kvm *kvm, int cpuid,
1525				   struct irq_phys_map *map,
1526				   unsigned int irq_num, bool level)
1527{
1528	struct vgic_dist *dist = &kvm->arch.vgic;
1529	struct kvm_vcpu *vcpu;
1530	int edge_triggered, level_triggered;
1531	int enabled;
1532	bool ret = true, can_inject = true;
1533
1534	trace_vgic_update_irq_pending(cpuid, irq_num, level);
1535
1536	if (irq_num >= min(kvm->arch.vgic.nr_irqs, 1020))
1537		return -EINVAL;
1538
1539	spin_lock(&dist->lock);
1540
1541	vcpu = kvm_get_vcpu(kvm, cpuid);
1542	edge_triggered = vgic_irq_is_edge(vcpu, irq_num);
1543	level_triggered = !edge_triggered;
1544
1545	if (!vgic_validate_injection(vcpu, irq_num, level)) {
1546		ret = false;
1547		goto out;
1548	}
1549
1550	if (irq_num >= VGIC_NR_PRIVATE_IRQS) {
1551		cpuid = dist->irq_spi_cpu[irq_num - VGIC_NR_PRIVATE_IRQS];
1552		if (cpuid == VCPU_NOT_ALLOCATED) {
1553			/* Pretend we use CPU0, and prevent injection */
1554			cpuid = 0;
1555			can_inject = false;
1556		}
1557		vcpu = kvm_get_vcpu(kvm, cpuid);
1558	}
1559
1560	kvm_debug("Inject IRQ%d level %d CPU%d\n", irq_num, level, cpuid);
1561
1562	if (level) {
1563		if (level_triggered)
1564			vgic_dist_irq_set_level(vcpu, irq_num);
1565		vgic_dist_irq_set_pending(vcpu, irq_num);
1566	} else {
1567		if (level_triggered) {
1568			vgic_dist_irq_clear_level(vcpu, irq_num);
1569			if (!vgic_dist_irq_soft_pend(vcpu, irq_num)) {
1570				vgic_dist_irq_clear_pending(vcpu, irq_num);
1571				vgic_cpu_irq_clear(vcpu, irq_num);
1572				if (!compute_pending_for_cpu(vcpu))
1573					clear_bit(cpuid, dist->irq_pending_on_cpu);
1574			}
1575		}
1576
1577		ret = false;
1578		goto out;
1579	}
1580
1581	enabled = vgic_irq_is_enabled(vcpu, irq_num);
1582
1583	if (!enabled || !can_inject) {
1584		ret = false;
1585		goto out;
1586	}
1587
1588	if (!vgic_can_sample_irq(vcpu, irq_num)) {
1589		/*
1590		 * Level interrupt in progress, will be picked up
1591		 * when EOId.
1592		 */
1593		ret = false;
1594		goto out;
1595	}
1596
1597	if (level) {
1598		vgic_cpu_irq_set(vcpu, irq_num);
1599		set_bit(cpuid, dist->irq_pending_on_cpu);
1600	}
1601
1602out:
1603	spin_unlock(&dist->lock);
1604
1605	if (ret) {
1606		/* kick the specified vcpu */
1607		kvm_vcpu_kick(kvm_get_vcpu(kvm, cpuid));
1608	}
1609
1610	return 0;
1611}
1612
1613static int vgic_lazy_init(struct kvm *kvm)
1614{
1615	int ret = 0;
1616
1617	if (unlikely(!vgic_initialized(kvm))) {
1618		/*
1619		 * We only provide the automatic initialization of the VGIC
1620		 * for the legacy case of a GICv2. Any other type must
1621		 * be explicitly initialized once setup with the respective
1622		 * KVM device call.
1623		 */
1624		if (kvm->arch.vgic.vgic_model != KVM_DEV_TYPE_ARM_VGIC_V2)
1625			return -EBUSY;
1626
1627		mutex_lock(&kvm->lock);
1628		ret = vgic_init(kvm);
1629		mutex_unlock(&kvm->lock);
1630	}
1631
1632	return ret;
1633}
1634
1635/**
1636 * kvm_vgic_inject_irq - Inject an IRQ from a device to the vgic
1637 * @kvm:     The VM structure pointer
1638 * @cpuid:   The CPU for PPIs
1639 * @irq_num: The IRQ number that is assigned to the device. This IRQ
1640 *           must not be mapped to a HW interrupt.
1641 * @level:   Edge-triggered:  true:  to trigger the interrupt
1642 *			      false: to ignore the call
1643 *	     Level-sensitive  true:  raise the input signal
1644 *			      false: lower the input signal
1645 *
1646 * The GIC is not concerned with devices being active-LOW or active-HIGH for
1647 * level-sensitive interrupts.  You can think of the level parameter as 1
1648 * being HIGH and 0 being LOW and all devices being active-HIGH.
1649 */
1650int kvm_vgic_inject_irq(struct kvm *kvm, int cpuid, unsigned int irq_num,
1651			bool level)
1652{
1653	struct irq_phys_map *map;
1654	int ret;
1655
1656	ret = vgic_lazy_init(kvm);
1657	if (ret)
1658		return ret;
1659
1660	map = vgic_irq_map_search(kvm_get_vcpu(kvm, cpuid), irq_num);
1661	if (map)
1662		return -EINVAL;
1663
1664	return vgic_update_irq_pending(kvm, cpuid, NULL, irq_num, level);
1665}
1666
1667/**
1668 * kvm_vgic_inject_mapped_irq - Inject a physically mapped IRQ to the vgic
1669 * @kvm:     The VM structure pointer
1670 * @cpuid:   The CPU for PPIs
1671 * @map:     Pointer to a irq_phys_map structure describing the mapping
1672 * @level:   Edge-triggered:  true:  to trigger the interrupt
1673 *			      false: to ignore the call
1674 *	     Level-sensitive  true:  raise the input signal
1675 *			      false: lower the input signal
1676 *
1677 * The GIC is not concerned with devices being active-LOW or active-HIGH for
1678 * level-sensitive interrupts.  You can think of the level parameter as 1
1679 * being HIGH and 0 being LOW and all devices being active-HIGH.
1680 */
1681int kvm_vgic_inject_mapped_irq(struct kvm *kvm, int cpuid,
1682			       struct irq_phys_map *map, bool level)
1683{
1684	int ret;
1685
1686	ret = vgic_lazy_init(kvm);
1687	if (ret)
1688		return ret;
1689
1690	return vgic_update_irq_pending(kvm, cpuid, map, map->virt_irq, level);
1691}
1692
1693static irqreturn_t vgic_maintenance_handler(int irq, void *data)
1694{
1695	/*
1696	 * We cannot rely on the vgic maintenance interrupt to be
1697	 * delivered synchronously. This means we can only use it to
1698	 * exit the VM, and we perform the handling of EOIed
1699	 * interrupts on the exit path (see vgic_process_maintenance).
1700	 */
1701	return IRQ_HANDLED;
1702}
1703
1704static struct list_head *vgic_get_irq_phys_map_list(struct kvm_vcpu *vcpu,
1705						    int virt_irq)
1706{
1707	if (virt_irq < VGIC_NR_PRIVATE_IRQS)
1708		return &vcpu->arch.vgic_cpu.irq_phys_map_list;
1709	else
1710		return &vcpu->kvm->arch.vgic.irq_phys_map_list;
1711}
1712
1713/**
1714 * kvm_vgic_map_phys_irq - map a virtual IRQ to a physical IRQ
1715 * @vcpu: The VCPU pointer
1716 * @virt_irq: The virtual irq number
1717 * @irq: The Linux IRQ number
1718 *
1719 * Establish a mapping between a guest visible irq (@virt_irq) and a
1720 * Linux irq (@irq). On injection, @virt_irq will be associated with
1721 * the physical interrupt represented by @irq. This mapping can be
1722 * established multiple times as long as the parameters are the same.
1723 *
1724 * Returns a valid pointer on success, and an error pointer otherwise
1725 */
1726struct irq_phys_map *kvm_vgic_map_phys_irq(struct kvm_vcpu *vcpu,
1727					   int virt_irq, int irq)
1728{
1729	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1730	struct list_head *root = vgic_get_irq_phys_map_list(vcpu, virt_irq);
1731	struct irq_phys_map *map;
1732	struct irq_phys_map_entry *entry;
1733	struct irq_desc *desc;
1734	struct irq_data *data;
1735	int phys_irq;
1736
1737	desc = irq_to_desc(irq);
1738	if (!desc) {
1739		kvm_err("%s: no interrupt descriptor\n", __func__);
1740		return ERR_PTR(-EINVAL);
1741	}
1742
1743	data = irq_desc_get_irq_data(desc);
1744	while (data->parent_data)
1745		data = data->parent_data;
1746
1747	phys_irq = data->hwirq;
1748
1749	/* Create a new mapping */
1750	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
1751	if (!entry)
1752		return ERR_PTR(-ENOMEM);
1753
1754	spin_lock(&dist->irq_phys_map_lock);
1755
1756	/* Try to match an existing mapping */
1757	map = vgic_irq_map_search(vcpu, virt_irq);
1758	if (map) {
1759		/* Make sure this mapping matches */
1760		if (map->phys_irq != phys_irq	||
1761		    map->irq      != irq)
1762			map = ERR_PTR(-EINVAL);
1763
1764		/* Found an existing, valid mapping */
1765		goto out;
1766	}
1767
1768	map           = &entry->map;
1769	map->virt_irq = virt_irq;
1770	map->phys_irq = phys_irq;
1771	map->irq      = irq;
1772
1773	list_add_tail_rcu(&entry->entry, root);
1774
1775out:
1776	spin_unlock(&dist->irq_phys_map_lock);
1777	/* If we've found a hit in the existing list, free the useless
1778	 * entry */
1779	if (IS_ERR(map) || map != &entry->map)
1780		kfree(entry);
1781	return map;
1782}
1783
1784static struct irq_phys_map *vgic_irq_map_search(struct kvm_vcpu *vcpu,
1785						int virt_irq)
1786{
1787	struct list_head *root = vgic_get_irq_phys_map_list(vcpu, virt_irq);
1788	struct irq_phys_map_entry *entry;
1789	struct irq_phys_map *map;
1790
1791	rcu_read_lock();
1792
1793	list_for_each_entry_rcu(entry, root, entry) {
1794		map = &entry->map;
1795		if (map->virt_irq == virt_irq) {
1796			rcu_read_unlock();
1797			return map;
1798		}
1799	}
1800
1801	rcu_read_unlock();
1802
1803	return NULL;
1804}
1805
1806static void vgic_free_phys_irq_map_rcu(struct rcu_head *rcu)
1807{
1808	struct irq_phys_map_entry *entry;
1809
1810	entry = container_of(rcu, struct irq_phys_map_entry, rcu);
1811	kfree(entry);
1812}
1813
1814/**
1815 * kvm_vgic_unmap_phys_irq - Remove a virtual to physical IRQ mapping
1816 * @vcpu: The VCPU pointer
1817 * @map: The pointer to a mapping obtained through kvm_vgic_map_phys_irq
1818 *
1819 * Remove an existing mapping between virtual and physical interrupts.
1820 */
1821int kvm_vgic_unmap_phys_irq(struct kvm_vcpu *vcpu, struct irq_phys_map *map)
1822{
1823	struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1824	struct irq_phys_map_entry *entry;
1825	struct list_head *root;
1826
1827	if (!map)
1828		return -EINVAL;
1829
1830	root = vgic_get_irq_phys_map_list(vcpu, map->virt_irq);
1831
1832	spin_lock(&dist->irq_phys_map_lock);
1833
1834	list_for_each_entry(entry, root, entry) {
1835		if (&entry->map == map) {
1836			list_del_rcu(&entry->entry);
1837			call_rcu(&entry->rcu, vgic_free_phys_irq_map_rcu);
1838			break;
1839		}
1840	}
1841
1842	spin_unlock(&dist->irq_phys_map_lock);
1843
1844	return 0;
1845}
1846
1847static void vgic_destroy_irq_phys_map(struct kvm *kvm, struct list_head *root)
1848{
1849	struct vgic_dist *dist = &kvm->arch.vgic;
1850	struct irq_phys_map_entry *entry;
1851
1852	spin_lock(&dist->irq_phys_map_lock);
1853
1854	list_for_each_entry(entry, root, entry) {
1855		list_del_rcu(&entry->entry);
1856		call_rcu(&entry->rcu, vgic_free_phys_irq_map_rcu);
1857	}
1858
1859	spin_unlock(&dist->irq_phys_map_lock);
1860}
1861
1862void kvm_vgic_vcpu_destroy(struct kvm_vcpu *vcpu)
1863{
1864	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
1865
1866	kfree(vgic_cpu->pending_shared);
1867	kfree(vgic_cpu->active_shared);
1868	kfree(vgic_cpu->pend_act_shared);
1869	vgic_destroy_irq_phys_map(vcpu->kvm, &vgic_cpu->irq_phys_map_list);
1870	vgic_cpu->pending_shared = NULL;
1871	vgic_cpu->active_shared = NULL;
1872	vgic_cpu->pend_act_shared = NULL;
1873}
1874
1875static int vgic_vcpu_init_maps(struct kvm_vcpu *vcpu, int nr_irqs)
1876{
1877	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
1878	int nr_longs = BITS_TO_LONGS(nr_irqs - VGIC_NR_PRIVATE_IRQS);
1879	int sz = nr_longs * sizeof(unsigned long);
1880	vgic_cpu->pending_shared = kzalloc(sz, GFP_KERNEL);
1881	vgic_cpu->active_shared = kzalloc(sz, GFP_KERNEL);
1882	vgic_cpu->pend_act_shared = kzalloc(sz, GFP_KERNEL);
1883
1884	if (!vgic_cpu->pending_shared
1885		|| !vgic_cpu->active_shared
1886		|| !vgic_cpu->pend_act_shared) {
1887		kvm_vgic_vcpu_destroy(vcpu);
1888		return -ENOMEM;
1889	}
1890
1891	/*
1892	 * Store the number of LRs per vcpu, so we don't have to go
1893	 * all the way to the distributor structure to find out. Only
1894	 * assembly code should use this one.
1895	 */
1896	vgic_cpu->nr_lr = vgic->nr_lr;
1897
1898	return 0;
1899}
1900
1901/**
1902 * kvm_vgic_vcpu_early_init - Earliest possible per-vcpu vgic init stage
1903 *
1904 * No memory allocation should be performed here, only static init.
1905 */
1906void kvm_vgic_vcpu_early_init(struct kvm_vcpu *vcpu)
1907{
1908	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
1909	INIT_LIST_HEAD(&vgic_cpu->irq_phys_map_list);
1910}
1911
1912/**
1913 * kvm_vgic_get_max_vcpus - Get the maximum number of VCPUs allowed by HW
1914 *
1915 * The host's GIC naturally limits the maximum amount of VCPUs a guest
1916 * can use.
1917 */
1918int kvm_vgic_get_max_vcpus(void)
1919{
1920	return vgic->max_gic_vcpus;
1921}
1922
1923void kvm_vgic_destroy(struct kvm *kvm)
1924{
1925	struct vgic_dist *dist = &kvm->arch.vgic;
1926	struct kvm_vcpu *vcpu;
1927	int i;
1928
1929	kvm_for_each_vcpu(i, vcpu, kvm)
1930		kvm_vgic_vcpu_destroy(vcpu);
1931
1932	vgic_free_bitmap(&dist->irq_enabled);
1933	vgic_free_bitmap(&dist->irq_level);
1934	vgic_free_bitmap(&dist->irq_pending);
1935	vgic_free_bitmap(&dist->irq_soft_pend);
1936	vgic_free_bitmap(&dist->irq_queued);
1937	vgic_free_bitmap(&dist->irq_cfg);
1938	vgic_free_bytemap(&dist->irq_priority);
1939	if (dist->irq_spi_target) {
1940		for (i = 0; i < dist->nr_cpus; i++)
1941			vgic_free_bitmap(&dist->irq_spi_target[i]);
1942	}
1943	kfree(dist->irq_sgi_sources);
1944	kfree(dist->irq_spi_cpu);
1945	kfree(dist->irq_spi_mpidr);
1946	kfree(dist->irq_spi_target);
1947	kfree(dist->irq_pending_on_cpu);
1948	kfree(dist->irq_active_on_cpu);
1949	vgic_destroy_irq_phys_map(kvm, &dist->irq_phys_map_list);
1950	dist->irq_sgi_sources = NULL;
1951	dist->irq_spi_cpu = NULL;
1952	dist->irq_spi_target = NULL;
1953	dist->irq_pending_on_cpu = NULL;
1954	dist->irq_active_on_cpu = NULL;
1955	dist->nr_cpus = 0;
1956}
1957
1958/*
1959 * Allocate and initialize the various data structures. Must be called
1960 * with kvm->lock held!
1961 */
1962int vgic_init(struct kvm *kvm)
1963{
1964	struct vgic_dist *dist = &kvm->arch.vgic;
1965	struct kvm_vcpu *vcpu;
1966	int nr_cpus, nr_irqs;
1967	int ret, i, vcpu_id;
1968
1969	if (vgic_initialized(kvm))
1970		return 0;
1971
1972	nr_cpus = dist->nr_cpus = atomic_read(&kvm->online_vcpus);
1973	if (!nr_cpus)		/* No vcpus? Can't be good... */
1974		return -ENODEV;
1975
1976	/*
1977	 * If nobody configured the number of interrupts, use the
1978	 * legacy one.
1979	 */
1980	if (!dist->nr_irqs)
1981		dist->nr_irqs = VGIC_NR_IRQS_LEGACY;
1982
1983	nr_irqs = dist->nr_irqs;
1984
1985	ret  = vgic_init_bitmap(&dist->irq_enabled, nr_cpus, nr_irqs);
1986	ret |= vgic_init_bitmap(&dist->irq_level, nr_cpus, nr_irqs);
1987	ret |= vgic_init_bitmap(&dist->irq_pending, nr_cpus, nr_irqs);
1988	ret |= vgic_init_bitmap(&dist->irq_soft_pend, nr_cpus, nr_irqs);
1989	ret |= vgic_init_bitmap(&dist->irq_queued, nr_cpus, nr_irqs);
1990	ret |= vgic_init_bitmap(&dist->irq_active, nr_cpus, nr_irqs);
1991	ret |= vgic_init_bitmap(&dist->irq_cfg, nr_cpus, nr_irqs);
1992	ret |= vgic_init_bytemap(&dist->irq_priority, nr_cpus, nr_irqs);
1993
1994	if (ret)
1995		goto out;
1996
1997	dist->irq_sgi_sources = kzalloc(nr_cpus * VGIC_NR_SGIS, GFP_KERNEL);
1998	dist->irq_spi_cpu = kzalloc(nr_irqs - VGIC_NR_PRIVATE_IRQS, GFP_KERNEL);
1999	dist->irq_spi_target = kzalloc(sizeof(*dist->irq_spi_target) * nr_cpus,
2000				       GFP_KERNEL);
2001	dist->irq_pending_on_cpu = kzalloc(BITS_TO_LONGS(nr_cpus) * sizeof(long),
2002					   GFP_KERNEL);
2003	dist->irq_active_on_cpu = kzalloc(BITS_TO_LONGS(nr_cpus) * sizeof(long),
2004					   GFP_KERNEL);
2005	if (!dist->irq_sgi_sources ||
2006	    !dist->irq_spi_cpu ||
2007	    !dist->irq_spi_target ||
2008	    !dist->irq_pending_on_cpu ||
2009	    !dist->irq_active_on_cpu) {
2010		ret = -ENOMEM;
2011		goto out;
2012	}
2013
2014	for (i = 0; i < nr_cpus; i++)
2015		ret |= vgic_init_bitmap(&dist->irq_spi_target[i],
2016					nr_cpus, nr_irqs);
2017
2018	if (ret)
2019		goto out;
2020
2021	ret = kvm->arch.vgic.vm_ops.init_model(kvm);
2022	if (ret)
2023		goto out;
2024
2025	kvm_for_each_vcpu(vcpu_id, vcpu, kvm) {
2026		ret = vgic_vcpu_init_maps(vcpu, nr_irqs);
2027		if (ret) {
2028			kvm_err("VGIC: Failed to allocate vcpu memory\n");
2029			break;
2030		}
2031
2032		/*
2033		 * Enable and configure all SGIs to be edge-triggere and
2034		 * configure all PPIs as level-triggered.
2035		 */
2036		for (i = 0; i < VGIC_NR_PRIVATE_IRQS; i++) {
2037			if (i < VGIC_NR_SGIS) {
2038				/* SGIs */
2039				vgic_bitmap_set_irq_val(&dist->irq_enabled,
2040							vcpu->vcpu_id, i, 1);
2041				vgic_bitmap_set_irq_val(&dist->irq_cfg,
2042							vcpu->vcpu_id, i,
2043							VGIC_CFG_EDGE);
2044			} else if (i < VGIC_NR_PRIVATE_IRQS) {
2045				/* PPIs */
2046				vgic_bitmap_set_irq_val(&dist->irq_cfg,
2047							vcpu->vcpu_id, i,
2048							VGIC_CFG_LEVEL);
2049			}
2050		}
2051
2052		vgic_enable(vcpu);
2053	}
2054
2055out:
2056	if (ret)
2057		kvm_vgic_destroy(kvm);
2058
2059	return ret;
2060}
2061
2062static int init_vgic_model(struct kvm *kvm, int type)
2063{
2064	switch (type) {
2065	case KVM_DEV_TYPE_ARM_VGIC_V2:
2066		vgic_v2_init_emulation(kvm);
2067		break;
2068#ifdef CONFIG_KVM_ARM_VGIC_V3
2069	case KVM_DEV_TYPE_ARM_VGIC_V3:
2070		vgic_v3_init_emulation(kvm);
2071		break;
2072#endif
2073	default:
2074		return -ENODEV;
2075	}
2076
2077	if (atomic_read(&kvm->online_vcpus) > kvm->arch.max_vcpus)
2078		return -E2BIG;
2079
2080	return 0;
2081}
2082
2083/**
2084 * kvm_vgic_early_init - Earliest possible vgic initialization stage
2085 *
2086 * No memory allocation should be performed here, only static init.
2087 */
2088void kvm_vgic_early_init(struct kvm *kvm)
2089{
2090	spin_lock_init(&kvm->arch.vgic.lock);
2091	spin_lock_init(&kvm->arch.vgic.irq_phys_map_lock);
2092	INIT_LIST_HEAD(&kvm->arch.vgic.irq_phys_map_list);
2093}
2094
2095int kvm_vgic_create(struct kvm *kvm, u32 type)
2096{
2097	int i, vcpu_lock_idx = -1, ret;
2098	struct kvm_vcpu *vcpu;
2099
2100	mutex_lock(&kvm->lock);
2101
2102	if (irqchip_in_kernel(kvm)) {
2103		ret = -EEXIST;
2104		goto out;
2105	}
2106
2107	/*
2108	 * This function is also called by the KVM_CREATE_IRQCHIP handler,
2109	 * which had no chance yet to check the availability of the GICv2
2110	 * emulation. So check this here again. KVM_CREATE_DEVICE does
2111	 * the proper checks already.
2112	 */
2113	if (type == KVM_DEV_TYPE_ARM_VGIC_V2 && !vgic->can_emulate_gicv2) {
2114		ret = -ENODEV;
2115		goto out;
2116	}
2117
2118	/*
2119	 * Any time a vcpu is run, vcpu_load is called which tries to grab the
2120	 * vcpu->mutex.  By grabbing the vcpu->mutex of all VCPUs we ensure
2121	 * that no other VCPUs are run while we create the vgic.
2122	 */
2123	ret = -EBUSY;
2124	kvm_for_each_vcpu(i, vcpu, kvm) {
2125		if (!mutex_trylock(&vcpu->mutex))
2126			goto out_unlock;
2127		vcpu_lock_idx = i;
2128	}
2129
2130	kvm_for_each_vcpu(i, vcpu, kvm) {
2131		if (vcpu->arch.has_run_once)
2132			goto out_unlock;
2133	}
2134	ret = 0;
2135
2136	ret = init_vgic_model(kvm, type);
2137	if (ret)
2138		goto out_unlock;
2139
2140	kvm->arch.vgic.in_kernel = true;
2141	kvm->arch.vgic.vgic_model = type;
2142	kvm->arch.vgic.vctrl_base = vgic->vctrl_base;
2143	kvm->arch.vgic.vgic_dist_base = VGIC_ADDR_UNDEF;
2144	kvm->arch.vgic.vgic_cpu_base = VGIC_ADDR_UNDEF;
2145	kvm->arch.vgic.vgic_redist_base = VGIC_ADDR_UNDEF;
2146
2147out_unlock:
2148	for (; vcpu_lock_idx >= 0; vcpu_lock_idx--) {
2149		vcpu = kvm_get_vcpu(kvm, vcpu_lock_idx);
2150		mutex_unlock(&vcpu->mutex);
2151	}
2152
2153out:
2154	mutex_unlock(&kvm->lock);
2155	return ret;
2156}
2157
2158static int vgic_ioaddr_overlap(struct kvm *kvm)
2159{
2160	phys_addr_t dist = kvm->arch.vgic.vgic_dist_base;
2161	phys_addr_t cpu = kvm->arch.vgic.vgic_cpu_base;
2162
2163	if (IS_VGIC_ADDR_UNDEF(dist) || IS_VGIC_ADDR_UNDEF(cpu))
2164		return 0;
2165	if ((dist <= cpu && dist + KVM_VGIC_V2_DIST_SIZE > cpu) ||
2166	    (cpu <= dist && cpu + KVM_VGIC_V2_CPU_SIZE > dist))
2167		return -EBUSY;
2168	return 0;
2169}
2170
2171static int vgic_ioaddr_assign(struct kvm *kvm, phys_addr_t *ioaddr,
2172			      phys_addr_t addr, phys_addr_t size)
2173{
2174	int ret;
2175
2176	if (addr & ~KVM_PHYS_MASK)
2177		return -E2BIG;
2178
2179	if (addr & (SZ_4K - 1))
2180		return -EINVAL;
2181
2182	if (!IS_VGIC_ADDR_UNDEF(*ioaddr))
2183		return -EEXIST;
2184	if (addr + size < addr)
2185		return -EINVAL;
2186
2187	*ioaddr = addr;
2188	ret = vgic_ioaddr_overlap(kvm);
2189	if (ret)
2190		*ioaddr = VGIC_ADDR_UNDEF;
2191
2192	return ret;
2193}
2194
2195/**
2196 * kvm_vgic_addr - set or get vgic VM base addresses
2197 * @kvm:   pointer to the vm struct
2198 * @type:  the VGIC addr type, one of KVM_VGIC_V[23]_ADDR_TYPE_XXX
2199 * @addr:  pointer to address value
2200 * @write: if true set the address in the VM address space, if false read the
2201 *          address
2202 *
2203 * Set or get the vgic base addresses for the distributor and the virtual CPU
2204 * interface in the VM physical address space.  These addresses are properties
2205 * of the emulated core/SoC and therefore user space initially knows this
2206 * information.
2207 */
2208int kvm_vgic_addr(struct kvm *kvm, unsigned long type, u64 *addr, bool write)
2209{
2210	int r = 0;
2211	struct vgic_dist *vgic = &kvm->arch.vgic;
2212	int type_needed;
2213	phys_addr_t *addr_ptr, block_size;
2214	phys_addr_t alignment;
2215
2216	mutex_lock(&kvm->lock);
2217	switch (type) {
2218	case KVM_VGIC_V2_ADDR_TYPE_DIST:
2219		type_needed = KVM_DEV_TYPE_ARM_VGIC_V2;
2220		addr_ptr = &vgic->vgic_dist_base;
2221		block_size = KVM_VGIC_V2_DIST_SIZE;
2222		alignment = SZ_4K;
2223		break;
2224	case KVM_VGIC_V2_ADDR_TYPE_CPU:
2225		type_needed = KVM_DEV_TYPE_ARM_VGIC_V2;
2226		addr_ptr = &vgic->vgic_cpu_base;
2227		block_size = KVM_VGIC_V2_CPU_SIZE;
2228		alignment = SZ_4K;
2229		break;
2230#ifdef CONFIG_KVM_ARM_VGIC_V3
2231	case KVM_VGIC_V3_ADDR_TYPE_DIST:
2232		type_needed = KVM_DEV_TYPE_ARM_VGIC_V3;
2233		addr_ptr = &vgic->vgic_dist_base;
2234		block_size = KVM_VGIC_V3_DIST_SIZE;
2235		alignment = SZ_64K;
2236		break;
2237	case KVM_VGIC_V3_ADDR_TYPE_REDIST:
2238		type_needed = KVM_DEV_TYPE_ARM_VGIC_V3;
2239		addr_ptr = &vgic->vgic_redist_base;
2240		block_size = KVM_VGIC_V3_REDIST_SIZE;
2241		alignment = SZ_64K;
2242		break;
2243#endif
2244	default:
2245		r = -ENODEV;
2246		goto out;
2247	}
2248
2249	if (vgic->vgic_model != type_needed) {
2250		r = -ENODEV;
2251		goto out;
2252	}
2253
2254	if (write) {
2255		if (!IS_ALIGNED(*addr, alignment))
2256			r = -EINVAL;
2257		else
2258			r = vgic_ioaddr_assign(kvm, addr_ptr, *addr,
2259					       block_size);
2260	} else {
2261		*addr = *addr_ptr;
2262	}
2263
2264out:
2265	mutex_unlock(&kvm->lock);
2266	return r;
2267}
2268
2269int vgic_set_common_attr(struct kvm_device *dev, struct kvm_device_attr *attr)
2270{
2271	int r;
2272
2273	switch (attr->group) {
2274	case KVM_DEV_ARM_VGIC_GRP_ADDR: {
2275		u64 __user *uaddr = (u64 __user *)(long)attr->addr;
2276		u64 addr;
2277		unsigned long type = (unsigned long)attr->attr;
2278
2279		if (copy_from_user(&addr, uaddr, sizeof(addr)))
2280			return -EFAULT;
2281
2282		r = kvm_vgic_addr(dev->kvm, type, &addr, true);
2283		return (r == -ENODEV) ? -ENXIO : r;
2284	}
2285	case KVM_DEV_ARM_VGIC_GRP_NR_IRQS: {
2286		u32 __user *uaddr = (u32 __user *)(long)attr->addr;
2287		u32 val;
2288		int ret = 0;
2289
2290		if (get_user(val, uaddr))
2291			return -EFAULT;
2292
2293		/*
2294		 * We require:
2295		 * - at least 32 SPIs on top of the 16 SGIs and 16 PPIs
2296		 * - at most 1024 interrupts
2297		 * - a multiple of 32 interrupts
2298		 */
2299		if (val < (VGIC_NR_PRIVATE_IRQS + 32) ||
2300		    val > VGIC_MAX_IRQS ||
2301		    (val & 31))
2302			return -EINVAL;
2303
2304		mutex_lock(&dev->kvm->lock);
2305
2306		if (vgic_ready(dev->kvm) || dev->kvm->arch.vgic.nr_irqs)
2307			ret = -EBUSY;
2308		else
2309			dev->kvm->arch.vgic.nr_irqs = val;
2310
2311		mutex_unlock(&dev->kvm->lock);
2312
2313		return ret;
2314	}
2315	case KVM_DEV_ARM_VGIC_GRP_CTRL: {
2316		switch (attr->attr) {
2317		case KVM_DEV_ARM_VGIC_CTRL_INIT:
2318			r = vgic_init(dev->kvm);
2319			return r;
2320		}
2321		break;
2322	}
2323	}
2324
2325	return -ENXIO;
2326}
2327
2328int vgic_get_common_attr(struct kvm_device *dev, struct kvm_device_attr *attr)
2329{
2330	int r = -ENXIO;
2331
2332	switch (attr->group) {
2333	case KVM_DEV_ARM_VGIC_GRP_ADDR: {
2334		u64 __user *uaddr = (u64 __user *)(long)attr->addr;
2335		u64 addr;
2336		unsigned long type = (unsigned long)attr->attr;
2337
2338		r = kvm_vgic_addr(dev->kvm, type, &addr, false);
2339		if (r)
2340			return (r == -ENODEV) ? -ENXIO : r;
2341
2342		if (copy_to_user(uaddr, &addr, sizeof(addr)))
2343			return -EFAULT;
2344		break;
2345	}
2346	case KVM_DEV_ARM_VGIC_GRP_NR_IRQS: {
2347		u32 __user *uaddr = (u32 __user *)(long)attr->addr;
2348
2349		r = put_user(dev->kvm->arch.vgic.nr_irqs, uaddr);
2350		break;
2351	}
2352
2353	}
2354
2355	return r;
2356}
2357
2358int vgic_has_attr_regs(const struct vgic_io_range *ranges, phys_addr_t offset)
2359{
2360	if (vgic_find_range(ranges, 4, offset))
2361		return 0;
2362	else
2363		return -ENXIO;
2364}
2365
2366static void vgic_init_maintenance_interrupt(void *info)
2367{
2368	enable_percpu_irq(vgic->maint_irq, 0);
2369}
2370
2371static int vgic_cpu_notify(struct notifier_block *self,
2372			   unsigned long action, void *cpu)
2373{
2374	switch (action) {
2375	case CPU_STARTING:
2376	case CPU_STARTING_FROZEN:
2377		vgic_init_maintenance_interrupt(NULL);
2378		break;
2379	case CPU_DYING:
2380	case CPU_DYING_FROZEN:
2381		disable_percpu_irq(vgic->maint_irq);
2382		break;
2383	}
2384
2385	return NOTIFY_OK;
2386}
2387
2388static struct notifier_block vgic_cpu_nb = {
2389	.notifier_call = vgic_cpu_notify,
2390};
2391
2392static const struct of_device_id vgic_ids[] = {
2393	{ .compatible = "arm,cortex-a15-gic",	.data = vgic_v2_probe, },
2394	{ .compatible = "arm,cortex-a7-gic",	.data = vgic_v2_probe, },
2395	{ .compatible = "arm,gic-400",		.data = vgic_v2_probe, },
2396	{ .compatible = "arm,gic-v3",		.data = vgic_v3_probe, },
2397	{},
2398};
2399
2400int kvm_vgic_hyp_init(void)
2401{
2402	const struct of_device_id *matched_id;
2403	const int (*vgic_probe)(struct device_node *,const struct vgic_ops **,
2404				const struct vgic_params **);
2405	struct device_node *vgic_node;
2406	int ret;
2407
2408	vgic_node = of_find_matching_node_and_match(NULL,
2409						    vgic_ids, &matched_id);
2410	if (!vgic_node) {
2411		kvm_err("error: no compatible GIC node found\n");
2412		return -ENODEV;
2413	}
2414
2415	vgic_probe = matched_id->data;
2416	ret = vgic_probe(vgic_node, &vgic_ops, &vgic);
2417	if (ret)
2418		return ret;
2419
2420	ret = request_percpu_irq(vgic->maint_irq, vgic_maintenance_handler,
2421				 "vgic", kvm_get_running_vcpus());
2422	if (ret) {
2423		kvm_err("Cannot register interrupt %d\n", vgic->maint_irq);
2424		return ret;
2425	}
2426
2427	ret = __register_cpu_notifier(&vgic_cpu_nb);
2428	if (ret) {
2429		kvm_err("Cannot register vgic CPU notifier\n");
2430		goto out_free_irq;
2431	}
2432
2433	on_each_cpu(vgic_init_maintenance_interrupt, NULL, 1);
2434
2435	return 0;
2436
2437out_free_irq:
2438	free_percpu_irq(vgic->maint_irq, kvm_get_running_vcpus());
2439	return ret;
2440}
2441
2442int kvm_irq_map_gsi(struct kvm *kvm,
2443		    struct kvm_kernel_irq_routing_entry *entries,
2444		    int gsi)
2445{
2446	return 0;
2447}
2448
2449int kvm_irq_map_chip_pin(struct kvm *kvm, unsigned irqchip, unsigned pin)
2450{
2451	return pin;
2452}
2453
2454int kvm_set_irq(struct kvm *kvm, int irq_source_id,
2455		u32 irq, int level, bool line_status)
2456{
2457	unsigned int spi = irq + VGIC_NR_PRIVATE_IRQS;
2458
2459	trace_kvm_set_irq(irq, level, irq_source_id);
2460
2461	BUG_ON(!vgic_initialized(kvm));
2462
2463	return kvm_vgic_inject_irq(kvm, 0, spi, level);
2464}
2465
2466/* MSI not implemented yet */
2467int kvm_set_msi(struct kvm_kernel_irq_routing_entry *e,
2468		struct kvm *kvm, int irq_source_id,
2469		int level, bool line_status)
2470{
2471	return 0;
2472}
2473