]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - virt/kvm/arm/vgic.c
Merge remote-tracking branch 'asoc/topic/tlv320aic26' into asoc-next
[karo-tx-linux.git] / virt / kvm / arm / vgic.c
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
28 #include <linux/irqchip/arm-gic.h>
29
30 #include <asm/kvm_emulate.h>
31 #include <asm/kvm_arm.h>
32 #include <asm/kvm_mmu.h>
33
34 /*
35  * How the whole thing works (courtesy of Christoffer Dall):
36  *
37  * - At any time, the dist->irq_pending_on_cpu is the oracle that knows if
38  *   something is pending
39  * - VGIC pending interrupts are stored on the vgic.irq_state vgic
40  *   bitmap (this bitmap is updated by both user land ioctls and guest
41  *   mmio ops, and other in-kernel peripherals such as the
42  *   arch. timers) and indicate the 'wire' state.
43  * - Every time the bitmap changes, the irq_pending_on_cpu oracle is
44  *   recalculated
45  * - To calculate the oracle, we need info for each cpu from
46  *   compute_pending_for_cpu, which considers:
47  *   - PPI: dist->irq_state & dist->irq_enable
48  *   - SPI: dist->irq_state & dist->irq_enable & dist->irq_spi_target
49  *   - irq_spi_target is a 'formatted' version of the GICD_ICFGR
50  *     registers, stored on each vcpu. We only keep one bit of
51  *     information per interrupt, making sure that only one vcpu can
52  *     accept the interrupt.
53  * - The same is true when injecting an interrupt, except that we only
54  *   consider a single interrupt at a time. The irq_spi_cpu array
55  *   contains the target CPU for each SPI.
56  *
57  * The handling of level interrupts adds some extra complexity. We
58  * need to track when the interrupt has been EOIed, so we can sample
59  * the 'line' again. This is achieved as such:
60  *
61  * - When a level interrupt is moved onto a vcpu, the corresponding
62  *   bit in irq_active is set. As long as this bit is set, the line
63  *   will be ignored for further interrupts. The interrupt is injected
64  *   into the vcpu with the GICH_LR_EOI bit set (generate a
65  *   maintenance interrupt on EOI).
66  * - When the interrupt is EOIed, the maintenance interrupt fires,
67  *   and clears the corresponding bit in irq_active. This allow the
68  *   interrupt line to be sampled again.
69  */
70
71 #define VGIC_ADDR_UNDEF         (-1)
72 #define IS_VGIC_ADDR_UNDEF(_x)  ((_x) == VGIC_ADDR_UNDEF)
73
74 /* Physical address of vgic virtual cpu interface */
75 static phys_addr_t vgic_vcpu_base;
76
77 /* Virtual control interface base address */
78 static void __iomem *vgic_vctrl_base;
79
80 static struct device_node *vgic_node;
81
82 #define ACCESS_READ_VALUE       (1 << 0)
83 #define ACCESS_READ_RAZ         (0 << 0)
84 #define ACCESS_READ_MASK(x)     ((x) & (1 << 0))
85 #define ACCESS_WRITE_IGNORED    (0 << 1)
86 #define ACCESS_WRITE_SETBIT     (1 << 1)
87 #define ACCESS_WRITE_CLEARBIT   (2 << 1)
88 #define ACCESS_WRITE_VALUE      (3 << 1)
89 #define ACCESS_WRITE_MASK(x)    ((x) & (3 << 1))
90
91 static void vgic_retire_disabled_irqs(struct kvm_vcpu *vcpu);
92 static void vgic_update_state(struct kvm *kvm);
93 static void vgic_kick_vcpus(struct kvm *kvm);
94 static void vgic_dispatch_sgi(struct kvm_vcpu *vcpu, u32 reg);
95 static u32 vgic_nr_lr;
96
97 static unsigned int vgic_maint_irq;
98
99 static u32 *vgic_bitmap_get_reg(struct vgic_bitmap *x,
100                                 int cpuid, u32 offset)
101 {
102         offset >>= 2;
103         if (!offset)
104                 return x->percpu[cpuid].reg;
105         else
106                 return x->shared.reg + offset - 1;
107 }
108
109 static int vgic_bitmap_get_irq_val(struct vgic_bitmap *x,
110                                    int cpuid, int irq)
111 {
112         if (irq < VGIC_NR_PRIVATE_IRQS)
113                 return test_bit(irq, x->percpu[cpuid].reg_ul);
114
115         return test_bit(irq - VGIC_NR_PRIVATE_IRQS, x->shared.reg_ul);
116 }
117
118 static void vgic_bitmap_set_irq_val(struct vgic_bitmap *x, int cpuid,
119                                     int irq, int val)
120 {
121         unsigned long *reg;
122
123         if (irq < VGIC_NR_PRIVATE_IRQS) {
124                 reg = x->percpu[cpuid].reg_ul;
125         } else {
126                 reg =  x->shared.reg_ul;
127                 irq -= VGIC_NR_PRIVATE_IRQS;
128         }
129
130         if (val)
131                 set_bit(irq, reg);
132         else
133                 clear_bit(irq, reg);
134 }
135
136 static unsigned long *vgic_bitmap_get_cpu_map(struct vgic_bitmap *x, int cpuid)
137 {
138         if (unlikely(cpuid >= VGIC_MAX_CPUS))
139                 return NULL;
140         return x->percpu[cpuid].reg_ul;
141 }
142
143 static unsigned long *vgic_bitmap_get_shared_map(struct vgic_bitmap *x)
144 {
145         return x->shared.reg_ul;
146 }
147
148 static u32 *vgic_bytemap_get_reg(struct vgic_bytemap *x, int cpuid, u32 offset)
149 {
150         offset >>= 2;
151         BUG_ON(offset > (VGIC_NR_IRQS / 4));
152         if (offset < 8)
153                 return x->percpu[cpuid] + offset;
154         else
155                 return x->shared + offset - 8;
156 }
157
158 #define VGIC_CFG_LEVEL  0
159 #define VGIC_CFG_EDGE   1
160
161 static bool vgic_irq_is_edge(struct kvm_vcpu *vcpu, int irq)
162 {
163         struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
164         int irq_val;
165
166         irq_val = vgic_bitmap_get_irq_val(&dist->irq_cfg, vcpu->vcpu_id, irq);
167         return irq_val == VGIC_CFG_EDGE;
168 }
169
170 static int vgic_irq_is_enabled(struct kvm_vcpu *vcpu, int irq)
171 {
172         struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
173
174         return vgic_bitmap_get_irq_val(&dist->irq_enabled, vcpu->vcpu_id, irq);
175 }
176
177 static int vgic_irq_is_active(struct kvm_vcpu *vcpu, int irq)
178 {
179         struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
180
181         return vgic_bitmap_get_irq_val(&dist->irq_active, vcpu->vcpu_id, irq);
182 }
183
184 static void vgic_irq_set_active(struct kvm_vcpu *vcpu, int irq)
185 {
186         struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
187
188         vgic_bitmap_set_irq_val(&dist->irq_active, vcpu->vcpu_id, irq, 1);
189 }
190
191 static void vgic_irq_clear_active(struct kvm_vcpu *vcpu, int irq)
192 {
193         struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
194
195         vgic_bitmap_set_irq_val(&dist->irq_active, vcpu->vcpu_id, irq, 0);
196 }
197
198 static int vgic_dist_irq_is_pending(struct kvm_vcpu *vcpu, int irq)
199 {
200         struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
201
202         return vgic_bitmap_get_irq_val(&dist->irq_state, vcpu->vcpu_id, irq);
203 }
204
205 static void vgic_dist_irq_set(struct kvm_vcpu *vcpu, int irq)
206 {
207         struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
208
209         vgic_bitmap_set_irq_val(&dist->irq_state, vcpu->vcpu_id, irq, 1);
210 }
211
212 static void vgic_dist_irq_clear(struct kvm_vcpu *vcpu, int irq)
213 {
214         struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
215
216         vgic_bitmap_set_irq_val(&dist->irq_state, vcpu->vcpu_id, irq, 0);
217 }
218
219 static void vgic_cpu_irq_set(struct kvm_vcpu *vcpu, int irq)
220 {
221         if (irq < VGIC_NR_PRIVATE_IRQS)
222                 set_bit(irq, vcpu->arch.vgic_cpu.pending_percpu);
223         else
224                 set_bit(irq - VGIC_NR_PRIVATE_IRQS,
225                         vcpu->arch.vgic_cpu.pending_shared);
226 }
227
228 static void vgic_cpu_irq_clear(struct kvm_vcpu *vcpu, int irq)
229 {
230         if (irq < VGIC_NR_PRIVATE_IRQS)
231                 clear_bit(irq, vcpu->arch.vgic_cpu.pending_percpu);
232         else
233                 clear_bit(irq - VGIC_NR_PRIVATE_IRQS,
234                           vcpu->arch.vgic_cpu.pending_shared);
235 }
236
237 static u32 mmio_data_read(struct kvm_exit_mmio *mmio, u32 mask)
238 {
239         return *((u32 *)mmio->data) & mask;
240 }
241
242 static void mmio_data_write(struct kvm_exit_mmio *mmio, u32 mask, u32 value)
243 {
244         *((u32 *)mmio->data) = value & mask;
245 }
246
247 /**
248  * vgic_reg_access - access vgic register
249  * @mmio:   pointer to the data describing the mmio access
250  * @reg:    pointer to the virtual backing of vgic distributor data
251  * @offset: least significant 2 bits used for word offset
252  * @mode:   ACCESS_ mode (see defines above)
253  *
254  * Helper to make vgic register access easier using one of the access
255  * modes defined for vgic register access
256  * (read,raz,write-ignored,setbit,clearbit,write)
257  */
258 static void vgic_reg_access(struct kvm_exit_mmio *mmio, u32 *reg,
259                             phys_addr_t offset, int mode)
260 {
261         int word_offset = (offset & 3) * 8;
262         u32 mask = (1UL << (mmio->len * 8)) - 1;
263         u32 regval;
264
265         /*
266          * Any alignment fault should have been delivered to the guest
267          * directly (ARM ARM B3.12.7 "Prioritization of aborts").
268          */
269
270         if (reg) {
271                 regval = *reg;
272         } else {
273                 BUG_ON(mode != (ACCESS_READ_RAZ | ACCESS_WRITE_IGNORED));
274                 regval = 0;
275         }
276
277         if (mmio->is_write) {
278                 u32 data = mmio_data_read(mmio, mask) << word_offset;
279                 switch (ACCESS_WRITE_MASK(mode)) {
280                 case ACCESS_WRITE_IGNORED:
281                         return;
282
283                 case ACCESS_WRITE_SETBIT:
284                         regval |= data;
285                         break;
286
287                 case ACCESS_WRITE_CLEARBIT:
288                         regval &= ~data;
289                         break;
290
291                 case ACCESS_WRITE_VALUE:
292                         regval = (regval & ~(mask << word_offset)) | data;
293                         break;
294                 }
295                 *reg = regval;
296         } else {
297                 switch (ACCESS_READ_MASK(mode)) {
298                 case ACCESS_READ_RAZ:
299                         regval = 0;
300                         /* fall through */
301
302                 case ACCESS_READ_VALUE:
303                         mmio_data_write(mmio, mask, regval >> word_offset);
304                 }
305         }
306 }
307
308 static bool handle_mmio_misc(struct kvm_vcpu *vcpu,
309                              struct kvm_exit_mmio *mmio, phys_addr_t offset)
310 {
311         u32 reg;
312         u32 word_offset = offset & 3;
313
314         switch (offset & ~3) {
315         case 0:                 /* CTLR */
316                 reg = vcpu->kvm->arch.vgic.enabled;
317                 vgic_reg_access(mmio, &reg, word_offset,
318                                 ACCESS_READ_VALUE | ACCESS_WRITE_VALUE);
319                 if (mmio->is_write) {
320                         vcpu->kvm->arch.vgic.enabled = reg & 1;
321                         vgic_update_state(vcpu->kvm);
322                         return true;
323                 }
324                 break;
325
326         case 4:                 /* TYPER */
327                 reg  = (atomic_read(&vcpu->kvm->online_vcpus) - 1) << 5;
328                 reg |= (VGIC_NR_IRQS >> 5) - 1;
329                 vgic_reg_access(mmio, &reg, word_offset,
330                                 ACCESS_READ_VALUE | ACCESS_WRITE_IGNORED);
331                 break;
332
333         case 8:                 /* IIDR */
334                 reg = 0x4B00043B;
335                 vgic_reg_access(mmio, &reg, word_offset,
336                                 ACCESS_READ_VALUE | ACCESS_WRITE_IGNORED);
337                 break;
338         }
339
340         return false;
341 }
342
343 static bool handle_mmio_raz_wi(struct kvm_vcpu *vcpu,
344                                struct kvm_exit_mmio *mmio, phys_addr_t offset)
345 {
346         vgic_reg_access(mmio, NULL, offset,
347                         ACCESS_READ_RAZ | ACCESS_WRITE_IGNORED);
348         return false;
349 }
350
351 static bool handle_mmio_set_enable_reg(struct kvm_vcpu *vcpu,
352                                        struct kvm_exit_mmio *mmio,
353                                        phys_addr_t offset)
354 {
355         u32 *reg = vgic_bitmap_get_reg(&vcpu->kvm->arch.vgic.irq_enabled,
356                                        vcpu->vcpu_id, offset);
357         vgic_reg_access(mmio, reg, offset,
358                         ACCESS_READ_VALUE | ACCESS_WRITE_SETBIT);
359         if (mmio->is_write) {
360                 vgic_update_state(vcpu->kvm);
361                 return true;
362         }
363
364         return false;
365 }
366
367 static bool handle_mmio_clear_enable_reg(struct kvm_vcpu *vcpu,
368                                          struct kvm_exit_mmio *mmio,
369                                          phys_addr_t offset)
370 {
371         u32 *reg = vgic_bitmap_get_reg(&vcpu->kvm->arch.vgic.irq_enabled,
372                                        vcpu->vcpu_id, offset);
373         vgic_reg_access(mmio, reg, offset,
374                         ACCESS_READ_VALUE | ACCESS_WRITE_CLEARBIT);
375         if (mmio->is_write) {
376                 if (offset < 4) /* Force SGI enabled */
377                         *reg |= 0xffff;
378                 vgic_retire_disabled_irqs(vcpu);
379                 vgic_update_state(vcpu->kvm);
380                 return true;
381         }
382
383         return false;
384 }
385
386 static bool handle_mmio_set_pending_reg(struct kvm_vcpu *vcpu,
387                                         struct kvm_exit_mmio *mmio,
388                                         phys_addr_t offset)
389 {
390         u32 *reg = vgic_bitmap_get_reg(&vcpu->kvm->arch.vgic.irq_state,
391                                        vcpu->vcpu_id, offset);
392         vgic_reg_access(mmio, reg, offset,
393                         ACCESS_READ_VALUE | ACCESS_WRITE_SETBIT);
394         if (mmio->is_write) {
395                 vgic_update_state(vcpu->kvm);
396                 return true;
397         }
398
399         return false;
400 }
401
402 static bool handle_mmio_clear_pending_reg(struct kvm_vcpu *vcpu,
403                                           struct kvm_exit_mmio *mmio,
404                                           phys_addr_t offset)
405 {
406         u32 *reg = vgic_bitmap_get_reg(&vcpu->kvm->arch.vgic.irq_state,
407                                        vcpu->vcpu_id, offset);
408         vgic_reg_access(mmio, reg, offset,
409                         ACCESS_READ_VALUE | ACCESS_WRITE_CLEARBIT);
410         if (mmio->is_write) {
411                 vgic_update_state(vcpu->kvm);
412                 return true;
413         }
414
415         return false;
416 }
417
418 static bool handle_mmio_priority_reg(struct kvm_vcpu *vcpu,
419                                      struct kvm_exit_mmio *mmio,
420                                      phys_addr_t offset)
421 {
422         u32 *reg = vgic_bytemap_get_reg(&vcpu->kvm->arch.vgic.irq_priority,
423                                         vcpu->vcpu_id, offset);
424         vgic_reg_access(mmio, reg, offset,
425                         ACCESS_READ_VALUE | ACCESS_WRITE_VALUE);
426         return false;
427 }
428
429 #define GICD_ITARGETSR_SIZE     32
430 #define GICD_CPUTARGETS_BITS    8
431 #define GICD_IRQS_PER_ITARGETSR (GICD_ITARGETSR_SIZE / GICD_CPUTARGETS_BITS)
432 static u32 vgic_get_target_reg(struct kvm *kvm, int irq)
433 {
434         struct vgic_dist *dist = &kvm->arch.vgic;
435         int i;
436         u32 val = 0;
437
438         irq -= VGIC_NR_PRIVATE_IRQS;
439
440         for (i = 0; i < GICD_IRQS_PER_ITARGETSR; i++)
441                 val |= 1 << (dist->irq_spi_cpu[irq + i] + i * 8);
442
443         return val;
444 }
445
446 static void vgic_set_target_reg(struct kvm *kvm, u32 val, int irq)
447 {
448         struct vgic_dist *dist = &kvm->arch.vgic;
449         struct kvm_vcpu *vcpu;
450         int i, c;
451         unsigned long *bmap;
452         u32 target;
453
454         irq -= VGIC_NR_PRIVATE_IRQS;
455
456         /*
457          * Pick the LSB in each byte. This ensures we target exactly
458          * one vcpu per IRQ. If the byte is null, assume we target
459          * CPU0.
460          */
461         for (i = 0; i < GICD_IRQS_PER_ITARGETSR; i++) {
462                 int shift = i * GICD_CPUTARGETS_BITS;
463                 target = ffs((val >> shift) & 0xffU);
464                 target = target ? (target - 1) : 0;
465                 dist->irq_spi_cpu[irq + i] = target;
466                 kvm_for_each_vcpu(c, vcpu, kvm) {
467                         bmap = vgic_bitmap_get_shared_map(&dist->irq_spi_target[c]);
468                         if (c == target)
469                                 set_bit(irq + i, bmap);
470                         else
471                                 clear_bit(irq + i, bmap);
472                 }
473         }
474 }
475
476 static bool handle_mmio_target_reg(struct kvm_vcpu *vcpu,
477                                    struct kvm_exit_mmio *mmio,
478                                    phys_addr_t offset)
479 {
480         u32 reg;
481
482         /* We treat the banked interrupts targets as read-only */
483         if (offset < 32) {
484                 u32 roreg = 1 << vcpu->vcpu_id;
485                 roreg |= roreg << 8;
486                 roreg |= roreg << 16;
487
488                 vgic_reg_access(mmio, &roreg, offset,
489                                 ACCESS_READ_VALUE | ACCESS_WRITE_IGNORED);
490                 return false;
491         }
492
493         reg = vgic_get_target_reg(vcpu->kvm, offset & ~3U);
494         vgic_reg_access(mmio, &reg, offset,
495                         ACCESS_READ_VALUE | ACCESS_WRITE_VALUE);
496         if (mmio->is_write) {
497                 vgic_set_target_reg(vcpu->kvm, reg, offset & ~3U);
498                 vgic_update_state(vcpu->kvm);
499                 return true;
500         }
501
502         return false;
503 }
504
505 static u32 vgic_cfg_expand(u16 val)
506 {
507         u32 res = 0;
508         int i;
509
510         /*
511          * Turn a 16bit value like abcd...mnop into a 32bit word
512          * a0b0c0d0...m0n0o0p0, which is what the HW cfg register is.
513          */
514         for (i = 0; i < 16; i++)
515                 res |= ((val >> i) & VGIC_CFG_EDGE) << (2 * i + 1);
516
517         return res;
518 }
519
520 static u16 vgic_cfg_compress(u32 val)
521 {
522         u16 res = 0;
523         int i;
524
525         /*
526          * Turn a 32bit word a0b0c0d0...m0n0o0p0 into 16bit value like
527          * abcd...mnop which is what we really care about.
528          */
529         for (i = 0; i < 16; i++)
530                 res |= ((val >> (i * 2 + 1)) & VGIC_CFG_EDGE) << i;
531
532         return res;
533 }
534
535 /*
536  * The distributor uses 2 bits per IRQ for the CFG register, but the
537  * LSB is always 0. As such, we only keep the upper bit, and use the
538  * two above functions to compress/expand the bits
539  */
540 static bool handle_mmio_cfg_reg(struct kvm_vcpu *vcpu,
541                                 struct kvm_exit_mmio *mmio, phys_addr_t offset)
542 {
543         u32 val;
544         u32 *reg;
545
546         offset >>= 1;
547         reg = vgic_bitmap_get_reg(&vcpu->kvm->arch.vgic.irq_cfg,
548                                   vcpu->vcpu_id, offset);
549
550         if (offset & 2)
551                 val = *reg >> 16;
552         else
553                 val = *reg & 0xffff;
554
555         val = vgic_cfg_expand(val);
556         vgic_reg_access(mmio, &val, offset,
557                         ACCESS_READ_VALUE | ACCESS_WRITE_VALUE);
558         if (mmio->is_write) {
559                 if (offset < 4) {
560                         *reg = ~0U; /* Force PPIs/SGIs to 1 */
561                         return false;
562                 }
563
564                 val = vgic_cfg_compress(val);
565                 if (offset & 2) {
566                         *reg &= 0xffff;
567                         *reg |= val << 16;
568                 } else {
569                         *reg &= 0xffff << 16;
570                         *reg |= val;
571                 }
572         }
573
574         return false;
575 }
576
577 static bool handle_mmio_sgi_reg(struct kvm_vcpu *vcpu,
578                                 struct kvm_exit_mmio *mmio, phys_addr_t offset)
579 {
580         u32 reg;
581         vgic_reg_access(mmio, &reg, offset,
582                         ACCESS_READ_RAZ | ACCESS_WRITE_VALUE);
583         if (mmio->is_write) {
584                 vgic_dispatch_sgi(vcpu, reg);
585                 vgic_update_state(vcpu->kvm);
586                 return true;
587         }
588
589         return false;
590 }
591
592 /*
593  * I would have liked to use the kvm_bus_io_*() API instead, but it
594  * cannot cope with banked registers (only the VM pointer is passed
595  * around, and we need the vcpu). One of these days, someone please
596  * fix it!
597  */
598 struct mmio_range {
599         phys_addr_t base;
600         unsigned long len;
601         bool (*handle_mmio)(struct kvm_vcpu *vcpu, struct kvm_exit_mmio *mmio,
602                             phys_addr_t offset);
603 };
604
605 static const struct mmio_range vgic_ranges[] = {
606         {
607                 .base           = GIC_DIST_CTRL,
608                 .len            = 12,
609                 .handle_mmio    = handle_mmio_misc,
610         },
611         {
612                 .base           = GIC_DIST_IGROUP,
613                 .len            = VGIC_NR_IRQS / 8,
614                 .handle_mmio    = handle_mmio_raz_wi,
615         },
616         {
617                 .base           = GIC_DIST_ENABLE_SET,
618                 .len            = VGIC_NR_IRQS / 8,
619                 .handle_mmio    = handle_mmio_set_enable_reg,
620         },
621         {
622                 .base           = GIC_DIST_ENABLE_CLEAR,
623                 .len            = VGIC_NR_IRQS / 8,
624                 .handle_mmio    = handle_mmio_clear_enable_reg,
625         },
626         {
627                 .base           = GIC_DIST_PENDING_SET,
628                 .len            = VGIC_NR_IRQS / 8,
629                 .handle_mmio    = handle_mmio_set_pending_reg,
630         },
631         {
632                 .base           = GIC_DIST_PENDING_CLEAR,
633                 .len            = VGIC_NR_IRQS / 8,
634                 .handle_mmio    = handle_mmio_clear_pending_reg,
635         },
636         {
637                 .base           = GIC_DIST_ACTIVE_SET,
638                 .len            = VGIC_NR_IRQS / 8,
639                 .handle_mmio    = handle_mmio_raz_wi,
640         },
641         {
642                 .base           = GIC_DIST_ACTIVE_CLEAR,
643                 .len            = VGIC_NR_IRQS / 8,
644                 .handle_mmio    = handle_mmio_raz_wi,
645         },
646         {
647                 .base           = GIC_DIST_PRI,
648                 .len            = VGIC_NR_IRQS,
649                 .handle_mmio    = handle_mmio_priority_reg,
650         },
651         {
652                 .base           = GIC_DIST_TARGET,
653                 .len            = VGIC_NR_IRQS,
654                 .handle_mmio    = handle_mmio_target_reg,
655         },
656         {
657                 .base           = GIC_DIST_CONFIG,
658                 .len            = VGIC_NR_IRQS / 4,
659                 .handle_mmio    = handle_mmio_cfg_reg,
660         },
661         {
662                 .base           = GIC_DIST_SOFTINT,
663                 .len            = 4,
664                 .handle_mmio    = handle_mmio_sgi_reg,
665         },
666         {}
667 };
668
669 static const
670 struct mmio_range *find_matching_range(const struct mmio_range *ranges,
671                                        struct kvm_exit_mmio *mmio,
672                                        phys_addr_t base)
673 {
674         const struct mmio_range *r = ranges;
675         phys_addr_t addr = mmio->phys_addr - base;
676
677         while (r->len) {
678                 if (addr >= r->base &&
679                     (addr + mmio->len) <= (r->base + r->len))
680                         return r;
681                 r++;
682         }
683
684         return NULL;
685 }
686
687 /**
688  * vgic_handle_mmio - handle an in-kernel MMIO access
689  * @vcpu:       pointer to the vcpu performing the access
690  * @run:        pointer to the kvm_run structure
691  * @mmio:       pointer to the data describing the access
692  *
693  * returns true if the MMIO access has been performed in kernel space,
694  * and false if it needs to be emulated in user space.
695  */
696 bool vgic_handle_mmio(struct kvm_vcpu *vcpu, struct kvm_run *run,
697                       struct kvm_exit_mmio *mmio)
698 {
699         const struct mmio_range *range;
700         struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
701         unsigned long base = dist->vgic_dist_base;
702         bool updated_state;
703         unsigned long offset;
704
705         if (!irqchip_in_kernel(vcpu->kvm) ||
706             mmio->phys_addr < base ||
707             (mmio->phys_addr + mmio->len) > (base + KVM_VGIC_V2_DIST_SIZE))
708                 return false;
709
710         /* We don't support ldrd / strd or ldm / stm to the emulated vgic */
711         if (mmio->len > 4) {
712                 kvm_inject_dabt(vcpu, mmio->phys_addr);
713                 return true;
714         }
715
716         range = find_matching_range(vgic_ranges, mmio, base);
717         if (unlikely(!range || !range->handle_mmio)) {
718                 pr_warn("Unhandled access %d %08llx %d\n",
719                         mmio->is_write, mmio->phys_addr, mmio->len);
720                 return false;
721         }
722
723         spin_lock(&vcpu->kvm->arch.vgic.lock);
724         offset = mmio->phys_addr - range->base - base;
725         updated_state = range->handle_mmio(vcpu, mmio, offset);
726         spin_unlock(&vcpu->kvm->arch.vgic.lock);
727         kvm_prepare_mmio(run, mmio);
728         kvm_handle_mmio_return(vcpu, run);
729
730         if (updated_state)
731                 vgic_kick_vcpus(vcpu->kvm);
732
733         return true;
734 }
735
736 static void vgic_dispatch_sgi(struct kvm_vcpu *vcpu, u32 reg)
737 {
738         struct kvm *kvm = vcpu->kvm;
739         struct vgic_dist *dist = &kvm->arch.vgic;
740         int nrcpus = atomic_read(&kvm->online_vcpus);
741         u8 target_cpus;
742         int sgi, mode, c, vcpu_id;
743
744         vcpu_id = vcpu->vcpu_id;
745
746         sgi = reg & 0xf;
747         target_cpus = (reg >> 16) & 0xff;
748         mode = (reg >> 24) & 3;
749
750         switch (mode) {
751         case 0:
752                 if (!target_cpus)
753                         return;
754
755         case 1:
756                 target_cpus = ((1 << nrcpus) - 1) & ~(1 << vcpu_id) & 0xff;
757                 break;
758
759         case 2:
760                 target_cpus = 1 << vcpu_id;
761                 break;
762         }
763
764         kvm_for_each_vcpu(c, vcpu, kvm) {
765                 if (target_cpus & 1) {
766                         /* Flag the SGI as pending */
767                         vgic_dist_irq_set(vcpu, sgi);
768                         dist->irq_sgi_sources[c][sgi] |= 1 << vcpu_id;
769                         kvm_debug("SGI%d from CPU%d to CPU%d\n", sgi, vcpu_id, c);
770                 }
771
772                 target_cpus >>= 1;
773         }
774 }
775
776 static int compute_pending_for_cpu(struct kvm_vcpu *vcpu)
777 {
778         struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
779         unsigned long *pending, *enabled, *pend_percpu, *pend_shared;
780         unsigned long pending_private, pending_shared;
781         int vcpu_id;
782
783         vcpu_id = vcpu->vcpu_id;
784         pend_percpu = vcpu->arch.vgic_cpu.pending_percpu;
785         pend_shared = vcpu->arch.vgic_cpu.pending_shared;
786
787         pending = vgic_bitmap_get_cpu_map(&dist->irq_state, vcpu_id);
788         enabled = vgic_bitmap_get_cpu_map(&dist->irq_enabled, vcpu_id);
789         bitmap_and(pend_percpu, pending, enabled, VGIC_NR_PRIVATE_IRQS);
790
791         pending = vgic_bitmap_get_shared_map(&dist->irq_state);
792         enabled = vgic_bitmap_get_shared_map(&dist->irq_enabled);
793         bitmap_and(pend_shared, pending, enabled, VGIC_NR_SHARED_IRQS);
794         bitmap_and(pend_shared, pend_shared,
795                    vgic_bitmap_get_shared_map(&dist->irq_spi_target[vcpu_id]),
796                    VGIC_NR_SHARED_IRQS);
797
798         pending_private = find_first_bit(pend_percpu, VGIC_NR_PRIVATE_IRQS);
799         pending_shared = find_first_bit(pend_shared, VGIC_NR_SHARED_IRQS);
800         return (pending_private < VGIC_NR_PRIVATE_IRQS ||
801                 pending_shared < VGIC_NR_SHARED_IRQS);
802 }
803
804 /*
805  * Update the interrupt state and determine which CPUs have pending
806  * interrupts. Must be called with distributor lock held.
807  */
808 static void vgic_update_state(struct kvm *kvm)
809 {
810         struct vgic_dist *dist = &kvm->arch.vgic;
811         struct kvm_vcpu *vcpu;
812         int c;
813
814         if (!dist->enabled) {
815                 set_bit(0, &dist->irq_pending_on_cpu);
816                 return;
817         }
818
819         kvm_for_each_vcpu(c, vcpu, kvm) {
820                 if (compute_pending_for_cpu(vcpu)) {
821                         pr_debug("CPU%d has pending interrupts\n", c);
822                         set_bit(c, &dist->irq_pending_on_cpu);
823                 }
824         }
825 }
826
827 #define LR_CPUID(lr)    \
828         (((lr) & GICH_LR_PHYSID_CPUID) >> GICH_LR_PHYSID_CPUID_SHIFT)
829 #define MK_LR_PEND(src, irq)    \
830         (GICH_LR_PENDING_BIT | ((src) << GICH_LR_PHYSID_CPUID_SHIFT) | (irq))
831
832 /*
833  * An interrupt may have been disabled after being made pending on the
834  * CPU interface (the classic case is a timer running while we're
835  * rebooting the guest - the interrupt would kick as soon as the CPU
836  * interface gets enabled, with deadly consequences).
837  *
838  * The solution is to examine already active LRs, and check the
839  * interrupt is still enabled. If not, just retire it.
840  */
841 static void vgic_retire_disabled_irqs(struct kvm_vcpu *vcpu)
842 {
843         struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
844         int lr;
845
846         for_each_set_bit(lr, vgic_cpu->lr_used, vgic_cpu->nr_lr) {
847                 int irq = vgic_cpu->vgic_lr[lr] & GICH_LR_VIRTUALID;
848
849                 if (!vgic_irq_is_enabled(vcpu, irq)) {
850                         vgic_cpu->vgic_irq_lr_map[irq] = LR_EMPTY;
851                         clear_bit(lr, vgic_cpu->lr_used);
852                         vgic_cpu->vgic_lr[lr] &= ~GICH_LR_STATE;
853                         if (vgic_irq_is_active(vcpu, irq))
854                                 vgic_irq_clear_active(vcpu, irq);
855                 }
856         }
857 }
858
859 /*
860  * Queue an interrupt to a CPU virtual interface. Return true on success,
861  * or false if it wasn't possible to queue it.
862  */
863 static bool vgic_queue_irq(struct kvm_vcpu *vcpu, u8 sgi_source_id, int irq)
864 {
865         struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
866         int lr;
867
868         /* Sanitize the input... */
869         BUG_ON(sgi_source_id & ~7);
870         BUG_ON(sgi_source_id && irq >= VGIC_NR_SGIS);
871         BUG_ON(irq >= VGIC_NR_IRQS);
872
873         kvm_debug("Queue IRQ%d\n", irq);
874
875         lr = vgic_cpu->vgic_irq_lr_map[irq];
876
877         /* Do we have an active interrupt for the same CPUID? */
878         if (lr != LR_EMPTY &&
879             (LR_CPUID(vgic_cpu->vgic_lr[lr]) == sgi_source_id)) {
880                 kvm_debug("LR%d piggyback for IRQ%d %x\n",
881                           lr, irq, vgic_cpu->vgic_lr[lr]);
882                 BUG_ON(!test_bit(lr, vgic_cpu->lr_used));
883                 vgic_cpu->vgic_lr[lr] |= GICH_LR_PENDING_BIT;
884                 return true;
885         }
886
887         /* Try to use another LR for this interrupt */
888         lr = find_first_zero_bit((unsigned long *)vgic_cpu->lr_used,
889                                vgic_cpu->nr_lr);
890         if (lr >= vgic_cpu->nr_lr)
891                 return false;
892
893         kvm_debug("LR%d allocated for IRQ%d %x\n", lr, irq, sgi_source_id);
894         vgic_cpu->vgic_lr[lr] = MK_LR_PEND(sgi_source_id, irq);
895         vgic_cpu->vgic_irq_lr_map[irq] = lr;
896         set_bit(lr, vgic_cpu->lr_used);
897
898         if (!vgic_irq_is_edge(vcpu, irq))
899                 vgic_cpu->vgic_lr[lr] |= GICH_LR_EOI;
900
901         return true;
902 }
903
904 static bool vgic_queue_sgi(struct kvm_vcpu *vcpu, int irq)
905 {
906         struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
907         unsigned long sources;
908         int vcpu_id = vcpu->vcpu_id;
909         int c;
910
911         sources = dist->irq_sgi_sources[vcpu_id][irq];
912
913         for_each_set_bit(c, &sources, VGIC_MAX_CPUS) {
914                 if (vgic_queue_irq(vcpu, c, irq))
915                         clear_bit(c, &sources);
916         }
917
918         dist->irq_sgi_sources[vcpu_id][irq] = sources;
919
920         /*
921          * If the sources bitmap has been cleared it means that we
922          * could queue all the SGIs onto link registers (see the
923          * clear_bit above), and therefore we are done with them in
924          * our emulated gic and can get rid of them.
925          */
926         if (!sources) {
927                 vgic_dist_irq_clear(vcpu, irq);
928                 vgic_cpu_irq_clear(vcpu, irq);
929                 return true;
930         }
931
932         return false;
933 }
934
935 static bool vgic_queue_hwirq(struct kvm_vcpu *vcpu, int irq)
936 {
937         if (vgic_irq_is_active(vcpu, irq))
938                 return true; /* level interrupt, already queued */
939
940         if (vgic_queue_irq(vcpu, 0, irq)) {
941                 if (vgic_irq_is_edge(vcpu, irq)) {
942                         vgic_dist_irq_clear(vcpu, irq);
943                         vgic_cpu_irq_clear(vcpu, irq);
944                 } else {
945                         vgic_irq_set_active(vcpu, irq);
946                 }
947
948                 return true;
949         }
950
951         return false;
952 }
953
954 /*
955  * Fill the list registers with pending interrupts before running the
956  * guest.
957  */
958 static void __kvm_vgic_flush_hwstate(struct kvm_vcpu *vcpu)
959 {
960         struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
961         struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
962         int i, vcpu_id;
963         int overflow = 0;
964
965         vcpu_id = vcpu->vcpu_id;
966
967         /*
968          * We may not have any pending interrupt, or the interrupts
969          * may have been serviced from another vcpu. In all cases,
970          * move along.
971          */
972         if (!kvm_vgic_vcpu_pending_irq(vcpu)) {
973                 pr_debug("CPU%d has no pending interrupt\n", vcpu_id);
974                 goto epilog;
975         }
976
977         /* SGIs */
978         for_each_set_bit(i, vgic_cpu->pending_percpu, VGIC_NR_SGIS) {
979                 if (!vgic_queue_sgi(vcpu, i))
980                         overflow = 1;
981         }
982
983         /* PPIs */
984         for_each_set_bit_from(i, vgic_cpu->pending_percpu, VGIC_NR_PRIVATE_IRQS) {
985                 if (!vgic_queue_hwirq(vcpu, i))
986                         overflow = 1;
987         }
988
989         /* SPIs */
990         for_each_set_bit(i, vgic_cpu->pending_shared, VGIC_NR_SHARED_IRQS) {
991                 if (!vgic_queue_hwirq(vcpu, i + VGIC_NR_PRIVATE_IRQS))
992                         overflow = 1;
993         }
994
995 epilog:
996         if (overflow) {
997                 vgic_cpu->vgic_hcr |= GICH_HCR_UIE;
998         } else {
999                 vgic_cpu->vgic_hcr &= ~GICH_HCR_UIE;
1000                 /*
1001                  * We're about to run this VCPU, and we've consumed
1002                  * everything the distributor had in store for
1003                  * us. Claim we don't have anything pending. We'll
1004                  * adjust that if needed while exiting.
1005                  */
1006                 clear_bit(vcpu_id, &dist->irq_pending_on_cpu);
1007         }
1008 }
1009
1010 static bool vgic_process_maintenance(struct kvm_vcpu *vcpu)
1011 {
1012         struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
1013         bool level_pending = false;
1014
1015         kvm_debug("MISR = %08x\n", vgic_cpu->vgic_misr);
1016
1017         if (vgic_cpu->vgic_misr & GICH_MISR_EOI) {
1018                 /*
1019                  * Some level interrupts have been EOIed. Clear their
1020                  * active bit.
1021                  */
1022                 int lr, irq;
1023
1024                 for_each_set_bit(lr, (unsigned long *)vgic_cpu->vgic_eisr,
1025                                  vgic_cpu->nr_lr) {
1026                         irq = vgic_cpu->vgic_lr[lr] & GICH_LR_VIRTUALID;
1027
1028                         vgic_irq_clear_active(vcpu, irq);
1029                         vgic_cpu->vgic_lr[lr] &= ~GICH_LR_EOI;
1030
1031                         /* Any additional pending interrupt? */
1032                         if (vgic_dist_irq_is_pending(vcpu, irq)) {
1033                                 vgic_cpu_irq_set(vcpu, irq);
1034                                 level_pending = true;
1035                         } else {
1036                                 vgic_cpu_irq_clear(vcpu, irq);
1037                         }
1038
1039                         /*
1040                          * Despite being EOIed, the LR may not have
1041                          * been marked as empty.
1042                          */
1043                         set_bit(lr, (unsigned long *)vgic_cpu->vgic_elrsr);
1044                         vgic_cpu->vgic_lr[lr] &= ~GICH_LR_ACTIVE_BIT;
1045                 }
1046         }
1047
1048         if (vgic_cpu->vgic_misr & GICH_MISR_U)
1049                 vgic_cpu->vgic_hcr &= ~GICH_HCR_UIE;
1050
1051         return level_pending;
1052 }
1053
1054 /*
1055  * Sync back the VGIC state after a guest run. The distributor lock is
1056  * needed so we don't get preempted in the middle of the state processing.
1057  */
1058 static void __kvm_vgic_sync_hwstate(struct kvm_vcpu *vcpu)
1059 {
1060         struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
1061         struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1062         int lr, pending;
1063         bool level_pending;
1064
1065         level_pending = vgic_process_maintenance(vcpu);
1066
1067         /* Clear mappings for empty LRs */
1068         for_each_set_bit(lr, (unsigned long *)vgic_cpu->vgic_elrsr,
1069                          vgic_cpu->nr_lr) {
1070                 int irq;
1071
1072                 if (!test_and_clear_bit(lr, vgic_cpu->lr_used))
1073                         continue;
1074
1075                 irq = vgic_cpu->vgic_lr[lr] & GICH_LR_VIRTUALID;
1076
1077                 BUG_ON(irq >= VGIC_NR_IRQS);
1078                 vgic_cpu->vgic_irq_lr_map[irq] = LR_EMPTY;
1079         }
1080
1081         /* Check if we still have something up our sleeve... */
1082         pending = find_first_zero_bit((unsigned long *)vgic_cpu->vgic_elrsr,
1083                                       vgic_cpu->nr_lr);
1084         if (level_pending || pending < vgic_cpu->nr_lr)
1085                 set_bit(vcpu->vcpu_id, &dist->irq_pending_on_cpu);
1086 }
1087
1088 void kvm_vgic_flush_hwstate(struct kvm_vcpu *vcpu)
1089 {
1090         struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1091
1092         if (!irqchip_in_kernel(vcpu->kvm))
1093                 return;
1094
1095         spin_lock(&dist->lock);
1096         __kvm_vgic_flush_hwstate(vcpu);
1097         spin_unlock(&dist->lock);
1098 }
1099
1100 void kvm_vgic_sync_hwstate(struct kvm_vcpu *vcpu)
1101 {
1102         struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1103
1104         if (!irqchip_in_kernel(vcpu->kvm))
1105                 return;
1106
1107         spin_lock(&dist->lock);
1108         __kvm_vgic_sync_hwstate(vcpu);
1109         spin_unlock(&dist->lock);
1110 }
1111
1112 int kvm_vgic_vcpu_pending_irq(struct kvm_vcpu *vcpu)
1113 {
1114         struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1115
1116         if (!irqchip_in_kernel(vcpu->kvm))
1117                 return 0;
1118
1119         return test_bit(vcpu->vcpu_id, &dist->irq_pending_on_cpu);
1120 }
1121
1122 static void vgic_kick_vcpus(struct kvm *kvm)
1123 {
1124         struct kvm_vcpu *vcpu;
1125         int c;
1126
1127         /*
1128          * We've injected an interrupt, time to find out who deserves
1129          * a good kick...
1130          */
1131         kvm_for_each_vcpu(c, vcpu, kvm) {
1132                 if (kvm_vgic_vcpu_pending_irq(vcpu))
1133                         kvm_vcpu_kick(vcpu);
1134         }
1135 }
1136
1137 static int vgic_validate_injection(struct kvm_vcpu *vcpu, int irq, int level)
1138 {
1139         int is_edge = vgic_irq_is_edge(vcpu, irq);
1140         int state = vgic_dist_irq_is_pending(vcpu, irq);
1141
1142         /*
1143          * Only inject an interrupt if:
1144          * - edge triggered and we have a rising edge
1145          * - level triggered and we change level
1146          */
1147         if (is_edge)
1148                 return level > state;
1149         else
1150                 return level != state;
1151 }
1152
1153 static bool vgic_update_irq_state(struct kvm *kvm, int cpuid,
1154                                   unsigned int irq_num, bool level)
1155 {
1156         struct vgic_dist *dist = &kvm->arch.vgic;
1157         struct kvm_vcpu *vcpu;
1158         int is_edge, is_level;
1159         int enabled;
1160         bool ret = true;
1161
1162         spin_lock(&dist->lock);
1163
1164         vcpu = kvm_get_vcpu(kvm, cpuid);
1165         is_edge = vgic_irq_is_edge(vcpu, irq_num);
1166         is_level = !is_edge;
1167
1168         if (!vgic_validate_injection(vcpu, irq_num, level)) {
1169                 ret = false;
1170                 goto out;
1171         }
1172
1173         if (irq_num >= VGIC_NR_PRIVATE_IRQS) {
1174                 cpuid = dist->irq_spi_cpu[irq_num - VGIC_NR_PRIVATE_IRQS];
1175                 vcpu = kvm_get_vcpu(kvm, cpuid);
1176         }
1177
1178         kvm_debug("Inject IRQ%d level %d CPU%d\n", irq_num, level, cpuid);
1179
1180         if (level)
1181                 vgic_dist_irq_set(vcpu, irq_num);
1182         else
1183                 vgic_dist_irq_clear(vcpu, irq_num);
1184
1185         enabled = vgic_irq_is_enabled(vcpu, irq_num);
1186
1187         if (!enabled) {
1188                 ret = false;
1189                 goto out;
1190         }
1191
1192         if (is_level && vgic_irq_is_active(vcpu, irq_num)) {
1193                 /*
1194                  * Level interrupt in progress, will be picked up
1195                  * when EOId.
1196                  */
1197                 ret = false;
1198                 goto out;
1199         }
1200
1201         if (level) {
1202                 vgic_cpu_irq_set(vcpu, irq_num);
1203                 set_bit(cpuid, &dist->irq_pending_on_cpu);
1204         }
1205
1206 out:
1207         spin_unlock(&dist->lock);
1208
1209         return ret;
1210 }
1211
1212 /**
1213  * kvm_vgic_inject_irq - Inject an IRQ from a device to the vgic
1214  * @kvm:     The VM structure pointer
1215  * @cpuid:   The CPU for PPIs
1216  * @irq_num: The IRQ number that is assigned to the device
1217  * @level:   Edge-triggered:  true:  to trigger the interrupt
1218  *                            false: to ignore the call
1219  *           Level-sensitive  true:  activates an interrupt
1220  *                            false: deactivates an interrupt
1221  *
1222  * The GIC is not concerned with devices being active-LOW or active-HIGH for
1223  * level-sensitive interrupts.  You can think of the level parameter as 1
1224  * being HIGH and 0 being LOW and all devices being active-HIGH.
1225  */
1226 int kvm_vgic_inject_irq(struct kvm *kvm, int cpuid, unsigned int irq_num,
1227                         bool level)
1228 {
1229         if (vgic_update_irq_state(kvm, cpuid, irq_num, level))
1230                 vgic_kick_vcpus(kvm);
1231
1232         return 0;
1233 }
1234
1235 static irqreturn_t vgic_maintenance_handler(int irq, void *data)
1236 {
1237         /*
1238          * We cannot rely on the vgic maintenance interrupt to be
1239          * delivered synchronously. This means we can only use it to
1240          * exit the VM, and we perform the handling of EOIed
1241          * interrupts on the exit path (see vgic_process_maintenance).
1242          */
1243         return IRQ_HANDLED;
1244 }
1245
1246 int kvm_vgic_vcpu_init(struct kvm_vcpu *vcpu)
1247 {
1248         struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
1249         struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1250         int i;
1251
1252         if (!irqchip_in_kernel(vcpu->kvm))
1253                 return 0;
1254
1255         if (vcpu->vcpu_id >= VGIC_MAX_CPUS)
1256                 return -EBUSY;
1257
1258         for (i = 0; i < VGIC_NR_IRQS; i++) {
1259                 if (i < VGIC_NR_PPIS)
1260                         vgic_bitmap_set_irq_val(&dist->irq_enabled,
1261                                                 vcpu->vcpu_id, i, 1);
1262                 if (i < VGIC_NR_PRIVATE_IRQS)
1263                         vgic_bitmap_set_irq_val(&dist->irq_cfg,
1264                                                 vcpu->vcpu_id, i, VGIC_CFG_EDGE);
1265
1266                 vgic_cpu->vgic_irq_lr_map[i] = LR_EMPTY;
1267         }
1268
1269         /*
1270          * By forcing VMCR to zero, the GIC will restore the binary
1271          * points to their reset values. Anything else resets to zero
1272          * anyway.
1273          */
1274         vgic_cpu->vgic_vmcr = 0;
1275
1276         vgic_cpu->nr_lr = vgic_nr_lr;
1277         vgic_cpu->vgic_hcr = GICH_HCR_EN; /* Get the show on the road... */
1278
1279         return 0;
1280 }
1281
1282 static void vgic_init_maintenance_interrupt(void *info)
1283 {
1284         enable_percpu_irq(vgic_maint_irq, 0);
1285 }
1286
1287 static int vgic_cpu_notify(struct notifier_block *self,
1288                            unsigned long action, void *cpu)
1289 {
1290         switch (action) {
1291         case CPU_STARTING:
1292         case CPU_STARTING_FROZEN:
1293                 vgic_init_maintenance_interrupt(NULL);
1294                 break;
1295         case CPU_DYING:
1296         case CPU_DYING_FROZEN:
1297                 disable_percpu_irq(vgic_maint_irq);
1298                 break;
1299         }
1300
1301         return NOTIFY_OK;
1302 }
1303
1304 static struct notifier_block vgic_cpu_nb = {
1305         .notifier_call = vgic_cpu_notify,
1306 };
1307
1308 int kvm_vgic_hyp_init(void)
1309 {
1310         int ret;
1311         struct resource vctrl_res;
1312         struct resource vcpu_res;
1313
1314         vgic_node = of_find_compatible_node(NULL, NULL, "arm,cortex-a15-gic");
1315         if (!vgic_node) {
1316                 kvm_err("error: no compatible vgic node in DT\n");
1317                 return -ENODEV;
1318         }
1319
1320         vgic_maint_irq = irq_of_parse_and_map(vgic_node, 0);
1321         if (!vgic_maint_irq) {
1322                 kvm_err("error getting vgic maintenance irq from DT\n");
1323                 ret = -ENXIO;
1324                 goto out;
1325         }
1326
1327         ret = request_percpu_irq(vgic_maint_irq, vgic_maintenance_handler,
1328                                  "vgic", kvm_get_running_vcpus());
1329         if (ret) {
1330                 kvm_err("Cannot register interrupt %d\n", vgic_maint_irq);
1331                 goto out;
1332         }
1333
1334         ret = register_cpu_notifier(&vgic_cpu_nb);
1335         if (ret) {
1336                 kvm_err("Cannot register vgic CPU notifier\n");
1337                 goto out_free_irq;
1338         }
1339
1340         ret = of_address_to_resource(vgic_node, 2, &vctrl_res);
1341         if (ret) {
1342                 kvm_err("Cannot obtain VCTRL resource\n");
1343                 goto out_free_irq;
1344         }
1345
1346         vgic_vctrl_base = of_iomap(vgic_node, 2);
1347         if (!vgic_vctrl_base) {
1348                 kvm_err("Cannot ioremap VCTRL\n");
1349                 ret = -ENOMEM;
1350                 goto out_free_irq;
1351         }
1352
1353         vgic_nr_lr = readl_relaxed(vgic_vctrl_base + GICH_VTR);
1354         vgic_nr_lr = (vgic_nr_lr & 0x3f) + 1;
1355
1356         ret = create_hyp_io_mappings(vgic_vctrl_base,
1357                                      vgic_vctrl_base + resource_size(&vctrl_res),
1358                                      vctrl_res.start);
1359         if (ret) {
1360                 kvm_err("Cannot map VCTRL into hyp\n");
1361                 goto out_unmap;
1362         }
1363
1364         kvm_info("%s@%llx IRQ%d\n", vgic_node->name,
1365                  vctrl_res.start, vgic_maint_irq);
1366         on_each_cpu(vgic_init_maintenance_interrupt, NULL, 1);
1367
1368         if (of_address_to_resource(vgic_node, 3, &vcpu_res)) {
1369                 kvm_err("Cannot obtain VCPU resource\n");
1370                 ret = -ENXIO;
1371                 goto out_unmap;
1372         }
1373         vgic_vcpu_base = vcpu_res.start;
1374
1375         goto out;
1376
1377 out_unmap:
1378         iounmap(vgic_vctrl_base);
1379 out_free_irq:
1380         free_percpu_irq(vgic_maint_irq, kvm_get_running_vcpus());
1381 out:
1382         of_node_put(vgic_node);
1383         return ret;
1384 }
1385
1386 int kvm_vgic_init(struct kvm *kvm)
1387 {
1388         int ret = 0, i;
1389
1390         mutex_lock(&kvm->lock);
1391
1392         if (vgic_initialized(kvm))
1393                 goto out;
1394
1395         if (IS_VGIC_ADDR_UNDEF(kvm->arch.vgic.vgic_dist_base) ||
1396             IS_VGIC_ADDR_UNDEF(kvm->arch.vgic.vgic_cpu_base)) {
1397                 kvm_err("Need to set vgic cpu and dist addresses first\n");
1398                 ret = -ENXIO;
1399                 goto out;
1400         }
1401
1402         ret = kvm_phys_addr_ioremap(kvm, kvm->arch.vgic.vgic_cpu_base,
1403                                     vgic_vcpu_base, KVM_VGIC_V2_CPU_SIZE);
1404         if (ret) {
1405                 kvm_err("Unable to remap VGIC CPU to VCPU\n");
1406                 goto out;
1407         }
1408
1409         for (i = VGIC_NR_PRIVATE_IRQS; i < VGIC_NR_IRQS; i += 4)
1410                 vgic_set_target_reg(kvm, 0, i);
1411
1412         kvm_timer_init(kvm);
1413         kvm->arch.vgic.ready = true;
1414 out:
1415         mutex_unlock(&kvm->lock);
1416         return ret;
1417 }
1418
1419 int kvm_vgic_create(struct kvm *kvm)
1420 {
1421         int ret = 0;
1422
1423         mutex_lock(&kvm->lock);
1424
1425         if (atomic_read(&kvm->online_vcpus) || kvm->arch.vgic.vctrl_base) {
1426                 ret = -EEXIST;
1427                 goto out;
1428         }
1429
1430         spin_lock_init(&kvm->arch.vgic.lock);
1431         kvm->arch.vgic.vctrl_base = vgic_vctrl_base;
1432         kvm->arch.vgic.vgic_dist_base = VGIC_ADDR_UNDEF;
1433         kvm->arch.vgic.vgic_cpu_base = VGIC_ADDR_UNDEF;
1434
1435 out:
1436         mutex_unlock(&kvm->lock);
1437         return ret;
1438 }
1439
1440 static bool vgic_ioaddr_overlap(struct kvm *kvm)
1441 {
1442         phys_addr_t dist = kvm->arch.vgic.vgic_dist_base;
1443         phys_addr_t cpu = kvm->arch.vgic.vgic_cpu_base;
1444
1445         if (IS_VGIC_ADDR_UNDEF(dist) || IS_VGIC_ADDR_UNDEF(cpu))
1446                 return 0;
1447         if ((dist <= cpu && dist + KVM_VGIC_V2_DIST_SIZE > cpu) ||
1448             (cpu <= dist && cpu + KVM_VGIC_V2_CPU_SIZE > dist))
1449                 return -EBUSY;
1450         return 0;
1451 }
1452
1453 static int vgic_ioaddr_assign(struct kvm *kvm, phys_addr_t *ioaddr,
1454                               phys_addr_t addr, phys_addr_t size)
1455 {
1456         int ret;
1457
1458         if (!IS_VGIC_ADDR_UNDEF(*ioaddr))
1459                 return -EEXIST;
1460         if (addr + size < addr)
1461                 return -EINVAL;
1462
1463         ret = vgic_ioaddr_overlap(kvm);
1464         if (ret)
1465                 return ret;
1466         *ioaddr = addr;
1467         return ret;
1468 }
1469
1470 int kvm_vgic_set_addr(struct kvm *kvm, unsigned long type, u64 addr)
1471 {
1472         int r = 0;
1473         struct vgic_dist *vgic = &kvm->arch.vgic;
1474
1475         if (addr & ~KVM_PHYS_MASK)
1476                 return -E2BIG;
1477
1478         if (addr & (SZ_4K - 1))
1479                 return -EINVAL;
1480
1481         mutex_lock(&kvm->lock);
1482         switch (type) {
1483         case KVM_VGIC_V2_ADDR_TYPE_DIST:
1484                 r = vgic_ioaddr_assign(kvm, &vgic->vgic_dist_base,
1485                                        addr, KVM_VGIC_V2_DIST_SIZE);
1486                 break;
1487         case KVM_VGIC_V2_ADDR_TYPE_CPU:
1488                 r = vgic_ioaddr_assign(kvm, &vgic->vgic_cpu_base,
1489                                        addr, KVM_VGIC_V2_CPU_SIZE);
1490                 break;
1491         default:
1492                 r = -ENODEV;
1493         }
1494
1495         mutex_unlock(&kvm->lock);
1496         return r;
1497 }