2 * Copyright (C) 2012 ARM Ltd.
3 * Author: Marc Zyngier <marc.zyngier@arm.com>
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.
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.
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
19 #include <linux/cpu.h>
20 #include <linux/kvm.h>
21 #include <linux/kvm_host.h>
22 #include <linux/interrupt.h>
25 #include <linux/of_address.h>
26 #include <linux/of_irq.h>
27 #include <linux/uaccess.h>
29 #include <linux/irqchip/arm-gic.h>
31 #include <asm/kvm_emulate.h>
32 #include <asm/kvm_arm.h>
33 #include <asm/kvm_mmu.h>
34 #include <trace/events/kvm.h>
36 #include <kvm/iodev.h>
39 * How the whole thing works (courtesy of Christoffer Dall):
41 * - At any time, the dist->irq_pending_on_cpu is the oracle that knows if
42 * something is pending on the CPU interface.
43 * - Interrupts that are pending on the distributor are stored on the
44 * vgic.irq_pending vgic bitmap (this bitmap is updated by both user land
45 * ioctls and guest mmio ops, and other in-kernel peripherals such as the
47 * - Every time the bitmap changes, the irq_pending_on_cpu oracle is
49 * - To calculate the oracle, we need info for each cpu from
50 * compute_pending_for_cpu, which considers:
51 * - PPI: dist->irq_pending & dist->irq_enable
52 * - SPI: dist->irq_pending & dist->irq_enable & dist->irq_spi_target
53 * - irq_spi_target is a 'formatted' version of the GICD_ITARGETSRn
54 * registers, stored on each vcpu. We only keep one bit of
55 * information per interrupt, making sure that only one vcpu can
56 * accept the interrupt.
57 * - If any of the above state changes, we must recalculate the oracle.
58 * - The same is true when injecting an interrupt, except that we only
59 * consider a single interrupt at a time. The irq_spi_cpu array
60 * contains the target CPU for each SPI.
62 * The handling of level interrupts adds some extra complexity. We
63 * need to track when the interrupt has been EOIed, so we can sample
64 * the 'line' again. This is achieved as such:
66 * - When a level interrupt is moved onto a vcpu, the corresponding
67 * bit in irq_queued is set. As long as this bit is set, the line
68 * will be ignored for further interrupts. The interrupt is injected
69 * into the vcpu with the GICH_LR_EOI bit set (generate a
70 * maintenance interrupt on EOI).
71 * - When the interrupt is EOIed, the maintenance interrupt fires,
72 * and clears the corresponding bit in irq_queued. This allows the
73 * interrupt line to be sampled again.
74 * - Note that level-triggered interrupts can also be set to pending from
75 * writes to GICD_ISPENDRn and lowering the external input line does not
76 * cause the interrupt to become inactive in such a situation.
77 * Conversely, writes to GICD_ICPENDRn do not cause the interrupt to become
78 * inactive as long as the external input line is held high.
83 static void vgic_retire_disabled_irqs(struct kvm_vcpu *vcpu);
84 static void vgic_retire_lr(int lr_nr, int irq, struct kvm_vcpu *vcpu);
85 static struct vgic_lr vgic_get_lr(const struct kvm_vcpu *vcpu, int lr);
86 static void vgic_set_lr(struct kvm_vcpu *vcpu, int lr, struct vgic_lr lr_desc);
88 static const struct vgic_ops *vgic_ops;
89 static const struct vgic_params *vgic;
91 static void add_sgi_source(struct kvm_vcpu *vcpu, int irq, int source)
93 vcpu->kvm->arch.vgic.vm_ops.add_sgi_source(vcpu, irq, source);
96 static bool queue_sgi(struct kvm_vcpu *vcpu, int irq)
98 return vcpu->kvm->arch.vgic.vm_ops.queue_sgi(vcpu, irq);
101 int kvm_vgic_map_resources(struct kvm *kvm)
103 return kvm->arch.vgic.vm_ops.map_resources(kvm, vgic);
107 * struct vgic_bitmap contains a bitmap made of unsigned longs, but
108 * extracts u32s out of them.
110 * This does not work on 64-bit BE systems, because the bitmap access
111 * will store two consecutive 32-bit words with the higher-addressed
112 * register's bits at the lower index and the lower-addressed register's
113 * bits at the higher index.
115 * Therefore, swizzle the register index when accessing the 32-bit word
116 * registers to access the right register's value.
118 #if defined(CONFIG_CPU_BIG_ENDIAN) && BITS_PER_LONG == 64
119 #define REG_OFFSET_SWIZZLE 1
121 #define REG_OFFSET_SWIZZLE 0
124 static int vgic_init_bitmap(struct vgic_bitmap *b, int nr_cpus, int nr_irqs)
128 nr_longs = nr_cpus + BITS_TO_LONGS(nr_irqs - VGIC_NR_PRIVATE_IRQS);
130 b->private = kzalloc(sizeof(unsigned long) * nr_longs, GFP_KERNEL);
134 b->shared = b->private + nr_cpus;
139 static void vgic_free_bitmap(struct vgic_bitmap *b)
147 * Call this function to convert a u64 value to an unsigned long * bitmask
148 * in a way that works on both 32-bit and 64-bit LE and BE platforms.
150 * Warning: Calling this function may modify *val.
152 static unsigned long *u64_to_bitmask(u64 *val)
154 #if defined(CONFIG_CPU_BIG_ENDIAN) && BITS_PER_LONG == 32
155 *val = (*val >> 32) | (*val << 32);
157 return (unsigned long *)val;
160 u32 *vgic_bitmap_get_reg(struct vgic_bitmap *x, int cpuid, u32 offset)
164 return (u32 *)(x->private + cpuid) + REG_OFFSET_SWIZZLE;
166 return (u32 *)(x->shared) + ((offset - 1) ^ REG_OFFSET_SWIZZLE);
169 static int vgic_bitmap_get_irq_val(struct vgic_bitmap *x,
172 if (irq < VGIC_NR_PRIVATE_IRQS)
173 return test_bit(irq, x->private + cpuid);
175 return test_bit(irq - VGIC_NR_PRIVATE_IRQS, x->shared);
178 void vgic_bitmap_set_irq_val(struct vgic_bitmap *x, int cpuid,
183 if (irq < VGIC_NR_PRIVATE_IRQS) {
184 reg = x->private + cpuid;
187 irq -= VGIC_NR_PRIVATE_IRQS;
196 static unsigned long *vgic_bitmap_get_cpu_map(struct vgic_bitmap *x, int cpuid)
198 return x->private + cpuid;
201 unsigned long *vgic_bitmap_get_shared_map(struct vgic_bitmap *x)
206 static int vgic_init_bytemap(struct vgic_bytemap *x, int nr_cpus, int nr_irqs)
210 size = nr_cpus * VGIC_NR_PRIVATE_IRQS;
211 size += nr_irqs - VGIC_NR_PRIVATE_IRQS;
213 x->private = kzalloc(size, GFP_KERNEL);
217 x->shared = x->private + nr_cpus * VGIC_NR_PRIVATE_IRQS / sizeof(u32);
221 static void vgic_free_bytemap(struct vgic_bytemap *b)
228 u32 *vgic_bytemap_get_reg(struct vgic_bytemap *x, int cpuid, u32 offset)
232 if (offset < VGIC_NR_PRIVATE_IRQS) {
234 offset += cpuid * VGIC_NR_PRIVATE_IRQS;
237 offset -= VGIC_NR_PRIVATE_IRQS;
240 return reg + (offset / sizeof(u32));
243 #define VGIC_CFG_LEVEL 0
244 #define VGIC_CFG_EDGE 1
246 static bool vgic_irq_is_edge(struct kvm_vcpu *vcpu, int irq)
248 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
251 irq_val = vgic_bitmap_get_irq_val(&dist->irq_cfg, vcpu->vcpu_id, irq);
252 return irq_val == VGIC_CFG_EDGE;
255 static int vgic_irq_is_enabled(struct kvm_vcpu *vcpu, int irq)
257 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
259 return vgic_bitmap_get_irq_val(&dist->irq_enabled, vcpu->vcpu_id, irq);
262 static int vgic_irq_is_queued(struct kvm_vcpu *vcpu, int irq)
264 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
266 return vgic_bitmap_get_irq_val(&dist->irq_queued, vcpu->vcpu_id, irq);
269 static int vgic_irq_is_active(struct kvm_vcpu *vcpu, int irq)
271 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
273 return vgic_bitmap_get_irq_val(&dist->irq_active, vcpu->vcpu_id, irq);
276 static void vgic_irq_set_queued(struct kvm_vcpu *vcpu, int irq)
278 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
280 vgic_bitmap_set_irq_val(&dist->irq_queued, vcpu->vcpu_id, irq, 1);
283 static void vgic_irq_clear_queued(struct kvm_vcpu *vcpu, int irq)
285 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
287 vgic_bitmap_set_irq_val(&dist->irq_queued, vcpu->vcpu_id, irq, 0);
290 static void vgic_irq_set_active(struct kvm_vcpu *vcpu, int irq)
292 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
294 vgic_bitmap_set_irq_val(&dist->irq_active, vcpu->vcpu_id, irq, 1);
297 static void vgic_irq_clear_active(struct kvm_vcpu *vcpu, int irq)
299 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
301 vgic_bitmap_set_irq_val(&dist->irq_active, vcpu->vcpu_id, irq, 0);
304 static int vgic_dist_irq_get_level(struct kvm_vcpu *vcpu, int irq)
306 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
308 return vgic_bitmap_get_irq_val(&dist->irq_level, vcpu->vcpu_id, irq);
311 static void vgic_dist_irq_set_level(struct kvm_vcpu *vcpu, int irq)
313 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
315 vgic_bitmap_set_irq_val(&dist->irq_level, vcpu->vcpu_id, irq, 1);
318 static void vgic_dist_irq_clear_level(struct kvm_vcpu *vcpu, int irq)
320 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
322 vgic_bitmap_set_irq_val(&dist->irq_level, vcpu->vcpu_id, irq, 0);
325 static int vgic_dist_irq_soft_pend(struct kvm_vcpu *vcpu, int irq)
327 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
329 return vgic_bitmap_get_irq_val(&dist->irq_soft_pend, vcpu->vcpu_id, irq);
332 static void vgic_dist_irq_clear_soft_pend(struct kvm_vcpu *vcpu, int irq)
334 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
336 vgic_bitmap_set_irq_val(&dist->irq_soft_pend, vcpu->vcpu_id, irq, 0);
339 static int vgic_dist_irq_is_pending(struct kvm_vcpu *vcpu, int irq)
341 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
343 return vgic_bitmap_get_irq_val(&dist->irq_pending, vcpu->vcpu_id, irq);
346 void vgic_dist_irq_set_pending(struct kvm_vcpu *vcpu, int irq)
348 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
350 vgic_bitmap_set_irq_val(&dist->irq_pending, vcpu->vcpu_id, irq, 1);
353 void vgic_dist_irq_clear_pending(struct kvm_vcpu *vcpu, int irq)
355 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
357 vgic_bitmap_set_irq_val(&dist->irq_pending, vcpu->vcpu_id, irq, 0);
360 static void vgic_cpu_irq_set(struct kvm_vcpu *vcpu, int irq)
362 if (irq < VGIC_NR_PRIVATE_IRQS)
363 set_bit(irq, vcpu->arch.vgic_cpu.pending_percpu);
365 set_bit(irq - VGIC_NR_PRIVATE_IRQS,
366 vcpu->arch.vgic_cpu.pending_shared);
369 void vgic_cpu_irq_clear(struct kvm_vcpu *vcpu, int irq)
371 if (irq < VGIC_NR_PRIVATE_IRQS)
372 clear_bit(irq, vcpu->arch.vgic_cpu.pending_percpu);
374 clear_bit(irq - VGIC_NR_PRIVATE_IRQS,
375 vcpu->arch.vgic_cpu.pending_shared);
378 static bool vgic_can_sample_irq(struct kvm_vcpu *vcpu, int irq)
380 return vgic_irq_is_edge(vcpu, irq) || !vgic_irq_is_queued(vcpu, irq);
384 * vgic_reg_access - access vgic register
385 * @mmio: pointer to the data describing the mmio access
386 * @reg: pointer to the virtual backing of vgic distributor data
387 * @offset: least significant 2 bits used for word offset
388 * @mode: ACCESS_ mode (see defines above)
390 * Helper to make vgic register access easier using one of the access
391 * modes defined for vgic register access
392 * (read,raz,write-ignored,setbit,clearbit,write)
394 void vgic_reg_access(struct kvm_exit_mmio *mmio, u32 *reg,
395 phys_addr_t offset, int mode)
397 int word_offset = (offset & 3) * 8;
398 u32 mask = (1UL << (mmio->len * 8)) - 1;
402 * Any alignment fault should have been delivered to the guest
403 * directly (ARM ARM B3.12.7 "Prioritization of aborts").
409 BUG_ON(mode != (ACCESS_READ_RAZ | ACCESS_WRITE_IGNORED));
413 if (mmio->is_write) {
414 u32 data = mmio_data_read(mmio, mask) << word_offset;
415 switch (ACCESS_WRITE_MASK(mode)) {
416 case ACCESS_WRITE_IGNORED:
419 case ACCESS_WRITE_SETBIT:
423 case ACCESS_WRITE_CLEARBIT:
427 case ACCESS_WRITE_VALUE:
428 regval = (regval & ~(mask << word_offset)) | data;
433 switch (ACCESS_READ_MASK(mode)) {
434 case ACCESS_READ_RAZ:
438 case ACCESS_READ_VALUE:
439 mmio_data_write(mmio, mask, regval >> word_offset);
444 bool handle_mmio_raz_wi(struct kvm_vcpu *vcpu, struct kvm_exit_mmio *mmio,
447 vgic_reg_access(mmio, NULL, offset,
448 ACCESS_READ_RAZ | ACCESS_WRITE_IGNORED);
452 bool vgic_handle_enable_reg(struct kvm *kvm, struct kvm_exit_mmio *mmio,
453 phys_addr_t offset, int vcpu_id, int access)
456 int mode = ACCESS_READ_VALUE | access;
457 struct kvm_vcpu *target_vcpu = kvm_get_vcpu(kvm, vcpu_id);
459 reg = vgic_bitmap_get_reg(&kvm->arch.vgic.irq_enabled, vcpu_id, offset);
460 vgic_reg_access(mmio, reg, offset, mode);
461 if (mmio->is_write) {
462 if (access & ACCESS_WRITE_CLEARBIT) {
463 if (offset < 4) /* Force SGI enabled */
465 vgic_retire_disabled_irqs(target_vcpu);
467 vgic_update_state(kvm);
474 bool vgic_handle_set_pending_reg(struct kvm *kvm,
475 struct kvm_exit_mmio *mmio,
476 phys_addr_t offset, int vcpu_id)
480 int mode = ACCESS_READ_VALUE | ACCESS_WRITE_SETBIT;
481 struct vgic_dist *dist = &kvm->arch.vgic;
483 reg = vgic_bitmap_get_reg(&dist->irq_cfg, vcpu_id, offset);
484 level_mask = (~(*reg));
486 /* Mark both level and edge triggered irqs as pending */
487 reg = vgic_bitmap_get_reg(&dist->irq_pending, vcpu_id, offset);
489 vgic_reg_access(mmio, reg, offset, mode);
491 if (mmio->is_write) {
492 /* Set the soft-pending flag only for level-triggered irqs */
493 reg = vgic_bitmap_get_reg(&dist->irq_soft_pend,
495 vgic_reg_access(mmio, reg, offset, mode);
498 /* Ignore writes to SGIs */
501 *reg |= orig & 0xffff;
504 vgic_update_state(kvm);
511 bool vgic_handle_clear_pending_reg(struct kvm *kvm,
512 struct kvm_exit_mmio *mmio,
513 phys_addr_t offset, int vcpu_id)
517 int mode = ACCESS_READ_VALUE | ACCESS_WRITE_CLEARBIT;
518 struct vgic_dist *dist = &kvm->arch.vgic;
520 reg = vgic_bitmap_get_reg(&dist->irq_pending, vcpu_id, offset);
522 vgic_reg_access(mmio, reg, offset, mode);
523 if (mmio->is_write) {
524 /* Re-set level triggered level-active interrupts */
525 level_active = vgic_bitmap_get_reg(&dist->irq_level,
527 reg = vgic_bitmap_get_reg(&dist->irq_pending, vcpu_id, offset);
528 *reg |= *level_active;
530 /* Ignore writes to SGIs */
533 *reg |= orig & 0xffff;
536 /* Clear soft-pending flags */
537 reg = vgic_bitmap_get_reg(&dist->irq_soft_pend,
539 vgic_reg_access(mmio, reg, offset, mode);
541 vgic_update_state(kvm);
547 bool vgic_handle_set_active_reg(struct kvm *kvm,
548 struct kvm_exit_mmio *mmio,
549 phys_addr_t offset, int vcpu_id)
552 struct vgic_dist *dist = &kvm->arch.vgic;
554 reg = vgic_bitmap_get_reg(&dist->irq_active, vcpu_id, offset);
555 vgic_reg_access(mmio, reg, offset,
556 ACCESS_READ_VALUE | ACCESS_WRITE_SETBIT);
558 if (mmio->is_write) {
559 vgic_update_state(kvm);
566 bool vgic_handle_clear_active_reg(struct kvm *kvm,
567 struct kvm_exit_mmio *mmio,
568 phys_addr_t offset, int vcpu_id)
571 struct vgic_dist *dist = &kvm->arch.vgic;
573 reg = vgic_bitmap_get_reg(&dist->irq_active, vcpu_id, offset);
574 vgic_reg_access(mmio, reg, offset,
575 ACCESS_READ_VALUE | ACCESS_WRITE_CLEARBIT);
577 if (mmio->is_write) {
578 vgic_update_state(kvm);
585 static u32 vgic_cfg_expand(u16 val)
591 * Turn a 16bit value like abcd...mnop into a 32bit word
592 * a0b0c0d0...m0n0o0p0, which is what the HW cfg register is.
594 for (i = 0; i < 16; i++)
595 res |= ((val >> i) & VGIC_CFG_EDGE) << (2 * i + 1);
600 static u16 vgic_cfg_compress(u32 val)
606 * Turn a 32bit word a0b0c0d0...m0n0o0p0 into 16bit value like
607 * abcd...mnop which is what we really care about.
609 for (i = 0; i < 16; i++)
610 res |= ((val >> (i * 2 + 1)) & VGIC_CFG_EDGE) << i;
616 * The distributor uses 2 bits per IRQ for the CFG register, but the
617 * LSB is always 0. As such, we only keep the upper bit, and use the
618 * two above functions to compress/expand the bits
620 bool vgic_handle_cfg_reg(u32 *reg, struct kvm_exit_mmio *mmio,
630 val = vgic_cfg_expand(val);
631 vgic_reg_access(mmio, &val, offset,
632 ACCESS_READ_VALUE | ACCESS_WRITE_VALUE);
633 if (mmio->is_write) {
635 *reg = ~0U; /* Force PPIs/SGIs to 1 */
639 val = vgic_cfg_compress(val);
644 *reg &= 0xffff << 16;
653 * vgic_unqueue_irqs - move pending/active IRQs from LRs to the distributor
654 * @vgic_cpu: Pointer to the vgic_cpu struct holding the LRs
656 * Move any IRQs that have already been assigned to LRs back to the
657 * emulated distributor state so that the complete emulated state can be read
658 * from the main emulation structures without investigating the LRs.
660 void vgic_unqueue_irqs(struct kvm_vcpu *vcpu)
662 struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
665 for_each_set_bit(i, vgic_cpu->lr_used, vgic_cpu->nr_lr) {
666 struct vgic_lr lr = vgic_get_lr(vcpu, i);
669 * There are three options for the state bits:
673 * 11: pending and active
675 BUG_ON(!(lr.state & LR_STATE_MASK));
677 /* Reestablish SGI source for pending and active IRQs */
678 if (lr.irq < VGIC_NR_SGIS)
679 add_sgi_source(vcpu, lr.irq, lr.source);
682 * If the LR holds an active (10) or a pending and active (11)
683 * interrupt then move the active state to the
684 * distributor tracking bit.
686 if (lr.state & LR_STATE_ACTIVE) {
687 vgic_irq_set_active(vcpu, lr.irq);
688 lr.state &= ~LR_STATE_ACTIVE;
692 * Reestablish the pending state on the distributor and the
693 * CPU interface. It may have already been pending, but that
694 * is fine, then we are only setting a few bits that were
697 if (lr.state & LR_STATE_PENDING) {
698 vgic_dist_irq_set_pending(vcpu, lr.irq);
699 lr.state &= ~LR_STATE_PENDING;
702 vgic_set_lr(vcpu, i, lr);
705 * Mark the LR as free for other use.
707 BUG_ON(lr.state & LR_STATE_MASK);
708 vgic_retire_lr(i, lr.irq, vcpu);
709 vgic_irq_clear_queued(vcpu, lr.irq);
711 /* Finally update the VGIC state. */
712 vgic_update_state(vcpu->kvm);
717 struct vgic_io_range *vgic_find_range(const struct vgic_io_range *ranges,
718 int len, gpa_t offset)
720 while (ranges->len) {
721 if (offset >= ranges->base &&
722 (offset + len) <= (ranges->base + ranges->len))
730 static bool vgic_validate_access(const struct vgic_dist *dist,
731 const struct vgic_io_range *range,
732 unsigned long offset)
736 if (!range->bits_per_irq)
737 return true; /* Not an irq-based access */
739 irq = offset * 8 / range->bits_per_irq;
740 if (irq >= dist->nr_irqs)
747 * Call the respective handler function for the given range.
748 * We split up any 64 bit accesses into two consecutive 32 bit
749 * handler calls and merge the result afterwards.
750 * We do this in a little endian fashion regardless of the host's
751 * or guest's endianness, because the GIC is always LE and the rest of
752 * the code (vgic_reg_access) also puts it in a LE fashion already.
753 * At this point we have already identified the handle function, so
754 * range points to that one entry and offset is relative to this.
756 static bool call_range_handler(struct kvm_vcpu *vcpu,
757 struct kvm_exit_mmio *mmio,
758 unsigned long offset,
759 const struct vgic_io_range *range)
761 struct kvm_exit_mmio mmio32;
764 if (likely(mmio->len <= 4))
765 return range->handle_mmio(vcpu, mmio, offset);
768 * Any access bigger than 4 bytes (that we currently handle in KVM)
769 * is actually 8 bytes long, caused by a 64-bit access
773 mmio32.is_write = mmio->is_write;
774 mmio32.private = mmio->private;
776 mmio32.phys_addr = mmio->phys_addr + 4;
777 mmio32.data = &((u32 *)mmio->data)[1];
778 ret = range->handle_mmio(vcpu, &mmio32, offset + 4);
780 mmio32.phys_addr = mmio->phys_addr;
781 mmio32.data = &((u32 *)mmio->data)[0];
782 ret |= range->handle_mmio(vcpu, &mmio32, offset);
788 * vgic_handle_mmio_access - handle an in-kernel MMIO access
789 * This is called by the read/write KVM IO device wrappers below.
790 * @vcpu: pointer to the vcpu performing the access
791 * @this: pointer to the KVM IO device in charge
792 * @addr: guest physical address of the access
793 * @len: size of the access
794 * @val: pointer to the data region
795 * @is_write: read or write access
797 * returns true if the MMIO access could be performed
799 static int vgic_handle_mmio_access(struct kvm_vcpu *vcpu,
800 struct kvm_io_device *this, gpa_t addr,
801 int len, void *val, bool is_write)
803 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
804 struct vgic_io_device *iodev = container_of(this,
805 struct vgic_io_device, dev);
806 struct kvm_run *run = vcpu->run;
807 const struct vgic_io_range *range;
808 struct kvm_exit_mmio mmio;
812 offset = addr - iodev->addr;
813 range = vgic_find_range(iodev->reg_ranges, len, offset);
814 if (unlikely(!range || !range->handle_mmio)) {
815 pr_warn("Unhandled access %d %08llx %d\n", is_write, addr, len);
819 mmio.phys_addr = addr;
821 mmio.is_write = is_write;
823 mmio.private = iodev->redist_vcpu;
825 spin_lock(&dist->lock);
826 offset -= range->base;
827 if (vgic_validate_access(dist, range, offset)) {
828 updated_state = call_range_handler(vcpu, &mmio, offset, range);
832 updated_state = false;
834 spin_unlock(&dist->lock);
835 run->mmio.is_write = is_write;
837 run->mmio.phys_addr = addr;
838 memcpy(run->mmio.data, val, len);
840 kvm_handle_mmio_return(vcpu, run);
843 vgic_kick_vcpus(vcpu->kvm);
848 static int vgic_handle_mmio_read(struct kvm_vcpu *vcpu,
849 struct kvm_io_device *this,
850 gpa_t addr, int len, void *val)
852 return vgic_handle_mmio_access(vcpu, this, addr, len, val, false);
855 static int vgic_handle_mmio_write(struct kvm_vcpu *vcpu,
856 struct kvm_io_device *this,
857 gpa_t addr, int len, const void *val)
859 return vgic_handle_mmio_access(vcpu, this, addr, len, (void *)val,
863 struct kvm_io_device_ops vgic_io_ops = {
864 .read = vgic_handle_mmio_read,
865 .write = vgic_handle_mmio_write,
869 * vgic_register_kvm_io_dev - register VGIC register frame on the KVM I/O bus
870 * @kvm: The VM structure pointer
871 * @base: The (guest) base address for the register frame
872 * @len: Length of the register frame window
873 * @ranges: Describing the handler functions for each register
874 * @redist_vcpu_id: The VCPU ID to pass on to the handlers on call
875 * @iodev: Points to memory to be passed on to the handler
877 * @iodev stores the parameters of this function to be usable by the handler
878 * respectively the dispatcher function (since the KVM I/O bus framework lacks
879 * an opaque parameter). Initialization is done in this function, but the
880 * reference should be valid and unique for the whole VGIC lifetime.
881 * If the register frame is not mapped for a specific VCPU, pass -1 to
884 int vgic_register_kvm_io_dev(struct kvm *kvm, gpa_t base, int len,
885 const struct vgic_io_range *ranges,
887 struct vgic_io_device *iodev)
889 struct kvm_vcpu *vcpu = NULL;
892 if (redist_vcpu_id >= 0)
893 vcpu = kvm_get_vcpu(kvm, redist_vcpu_id);
897 iodev->reg_ranges = ranges;
898 iodev->redist_vcpu = vcpu;
900 kvm_iodevice_init(&iodev->dev, &vgic_io_ops);
902 mutex_lock(&kvm->slots_lock);
904 ret = kvm_io_bus_register_dev(kvm, KVM_MMIO_BUS, base, len,
906 mutex_unlock(&kvm->slots_lock);
908 /* Mark the iodev as invalid if registration fails. */
910 iodev->dev.ops = NULL;
915 static int vgic_nr_shared_irqs(struct vgic_dist *dist)
917 return dist->nr_irqs - VGIC_NR_PRIVATE_IRQS;
920 static int compute_active_for_cpu(struct kvm_vcpu *vcpu)
922 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
923 unsigned long *active, *enabled, *act_percpu, *act_shared;
924 unsigned long active_private, active_shared;
925 int nr_shared = vgic_nr_shared_irqs(dist);
928 vcpu_id = vcpu->vcpu_id;
929 act_percpu = vcpu->arch.vgic_cpu.active_percpu;
930 act_shared = vcpu->arch.vgic_cpu.active_shared;
932 active = vgic_bitmap_get_cpu_map(&dist->irq_active, vcpu_id);
933 enabled = vgic_bitmap_get_cpu_map(&dist->irq_enabled, vcpu_id);
934 bitmap_and(act_percpu, active, enabled, VGIC_NR_PRIVATE_IRQS);
936 active = vgic_bitmap_get_shared_map(&dist->irq_active);
937 enabled = vgic_bitmap_get_shared_map(&dist->irq_enabled);
938 bitmap_and(act_shared, active, enabled, nr_shared);
939 bitmap_and(act_shared, act_shared,
940 vgic_bitmap_get_shared_map(&dist->irq_spi_target[vcpu_id]),
943 active_private = find_first_bit(act_percpu, VGIC_NR_PRIVATE_IRQS);
944 active_shared = find_first_bit(act_shared, nr_shared);
946 return (active_private < VGIC_NR_PRIVATE_IRQS ||
947 active_shared < nr_shared);
950 static int compute_pending_for_cpu(struct kvm_vcpu *vcpu)
952 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
953 unsigned long *pending, *enabled, *pend_percpu, *pend_shared;
954 unsigned long pending_private, pending_shared;
955 int nr_shared = vgic_nr_shared_irqs(dist);
958 vcpu_id = vcpu->vcpu_id;
959 pend_percpu = vcpu->arch.vgic_cpu.pending_percpu;
960 pend_shared = vcpu->arch.vgic_cpu.pending_shared;
962 pending = vgic_bitmap_get_cpu_map(&dist->irq_pending, vcpu_id);
963 enabled = vgic_bitmap_get_cpu_map(&dist->irq_enabled, vcpu_id);
964 bitmap_and(pend_percpu, pending, enabled, VGIC_NR_PRIVATE_IRQS);
966 pending = vgic_bitmap_get_shared_map(&dist->irq_pending);
967 enabled = vgic_bitmap_get_shared_map(&dist->irq_enabled);
968 bitmap_and(pend_shared, pending, enabled, nr_shared);
969 bitmap_and(pend_shared, pend_shared,
970 vgic_bitmap_get_shared_map(&dist->irq_spi_target[vcpu_id]),
973 pending_private = find_first_bit(pend_percpu, VGIC_NR_PRIVATE_IRQS);
974 pending_shared = find_first_bit(pend_shared, nr_shared);
975 return (pending_private < VGIC_NR_PRIVATE_IRQS ||
976 pending_shared < vgic_nr_shared_irqs(dist));
980 * Update the interrupt state and determine which CPUs have pending
981 * or active interrupts. Must be called with distributor lock held.
983 void vgic_update_state(struct kvm *kvm)
985 struct vgic_dist *dist = &kvm->arch.vgic;
986 struct kvm_vcpu *vcpu;
989 if (!dist->enabled) {
990 set_bit(0, dist->irq_pending_on_cpu);
994 kvm_for_each_vcpu(c, vcpu, kvm) {
995 if (compute_pending_for_cpu(vcpu))
996 set_bit(c, dist->irq_pending_on_cpu);
998 if (compute_active_for_cpu(vcpu))
999 set_bit(c, dist->irq_active_on_cpu);
1001 clear_bit(c, dist->irq_active_on_cpu);
1005 static struct vgic_lr vgic_get_lr(const struct kvm_vcpu *vcpu, int lr)
1007 return vgic_ops->get_lr(vcpu, lr);
1010 static void vgic_set_lr(struct kvm_vcpu *vcpu, int lr,
1013 vgic_ops->set_lr(vcpu, lr, vlr);
1016 static void vgic_sync_lr_elrsr(struct kvm_vcpu *vcpu, int lr,
1019 vgic_ops->sync_lr_elrsr(vcpu, lr, vlr);
1022 static inline u64 vgic_get_elrsr(struct kvm_vcpu *vcpu)
1024 return vgic_ops->get_elrsr(vcpu);
1027 static inline u64 vgic_get_eisr(struct kvm_vcpu *vcpu)
1029 return vgic_ops->get_eisr(vcpu);
1032 static inline u32 vgic_get_interrupt_status(struct kvm_vcpu *vcpu)
1034 return vgic_ops->get_interrupt_status(vcpu);
1037 static inline void vgic_enable_underflow(struct kvm_vcpu *vcpu)
1039 vgic_ops->enable_underflow(vcpu);
1042 static inline void vgic_disable_underflow(struct kvm_vcpu *vcpu)
1044 vgic_ops->disable_underflow(vcpu);
1047 void vgic_get_vmcr(struct kvm_vcpu *vcpu, struct vgic_vmcr *vmcr)
1049 vgic_ops->get_vmcr(vcpu, vmcr);
1052 void vgic_set_vmcr(struct kvm_vcpu *vcpu, struct vgic_vmcr *vmcr)
1054 vgic_ops->set_vmcr(vcpu, vmcr);
1057 static inline void vgic_enable(struct kvm_vcpu *vcpu)
1059 vgic_ops->enable(vcpu);
1062 static void vgic_retire_lr(int lr_nr, int irq, struct kvm_vcpu *vcpu)
1064 struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
1065 struct vgic_lr vlr = vgic_get_lr(vcpu, lr_nr);
1068 vgic_set_lr(vcpu, lr_nr, vlr);
1069 clear_bit(lr_nr, vgic_cpu->lr_used);
1070 vgic_cpu->vgic_irq_lr_map[irq] = LR_EMPTY;
1074 * An interrupt may have been disabled after being made pending on the
1075 * CPU interface (the classic case is a timer running while we're
1076 * rebooting the guest - the interrupt would kick as soon as the CPU
1077 * interface gets enabled, with deadly consequences).
1079 * The solution is to examine already active LRs, and check the
1080 * interrupt is still enabled. If not, just retire it.
1082 static void vgic_retire_disabled_irqs(struct kvm_vcpu *vcpu)
1084 struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
1087 for_each_set_bit(lr, vgic_cpu->lr_used, vgic->nr_lr) {
1088 struct vgic_lr vlr = vgic_get_lr(vcpu, lr);
1090 if (!vgic_irq_is_enabled(vcpu, vlr.irq)) {
1091 vgic_retire_lr(lr, vlr.irq, vcpu);
1092 if (vgic_irq_is_queued(vcpu, vlr.irq))
1093 vgic_irq_clear_queued(vcpu, vlr.irq);
1098 static void vgic_queue_irq_to_lr(struct kvm_vcpu *vcpu, int irq,
1099 int lr_nr, struct vgic_lr vlr)
1101 if (vgic_irq_is_active(vcpu, irq)) {
1102 vlr.state |= LR_STATE_ACTIVE;
1103 kvm_debug("Set active, clear distributor: 0x%x\n", vlr.state);
1104 vgic_irq_clear_active(vcpu, irq);
1105 vgic_update_state(vcpu->kvm);
1106 } else if (vgic_dist_irq_is_pending(vcpu, irq)) {
1107 vlr.state |= LR_STATE_PENDING;
1108 kvm_debug("Set pending: 0x%x\n", vlr.state);
1111 if (!vgic_irq_is_edge(vcpu, irq))
1112 vlr.state |= LR_EOI_INT;
1114 vgic_set_lr(vcpu, lr_nr, vlr);
1118 * Queue an interrupt to a CPU virtual interface. Return true on success,
1119 * or false if it wasn't possible to queue it.
1120 * sgi_source must be zero for any non-SGI interrupts.
1122 bool vgic_queue_irq(struct kvm_vcpu *vcpu, u8 sgi_source_id, int irq)
1124 struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
1125 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1129 /* Sanitize the input... */
1130 BUG_ON(sgi_source_id & ~7);
1131 BUG_ON(sgi_source_id && irq >= VGIC_NR_SGIS);
1132 BUG_ON(irq >= dist->nr_irqs);
1134 kvm_debug("Queue IRQ%d\n", irq);
1136 lr = vgic_cpu->vgic_irq_lr_map[irq];
1138 /* Do we have an active interrupt for the same CPUID? */
1139 if (lr != LR_EMPTY) {
1140 vlr = vgic_get_lr(vcpu, lr);
1141 if (vlr.source == sgi_source_id) {
1142 kvm_debug("LR%d piggyback for IRQ%d\n", lr, vlr.irq);
1143 BUG_ON(!test_bit(lr, vgic_cpu->lr_used));
1144 vgic_queue_irq_to_lr(vcpu, irq, lr, vlr);
1149 /* Try to use another LR for this interrupt */
1150 lr = find_first_zero_bit((unsigned long *)vgic_cpu->lr_used,
1152 if (lr >= vgic->nr_lr)
1155 kvm_debug("LR%d allocated for IRQ%d %x\n", lr, irq, sgi_source_id);
1156 vgic_cpu->vgic_irq_lr_map[irq] = lr;
1157 set_bit(lr, vgic_cpu->lr_used);
1160 vlr.source = sgi_source_id;
1162 vgic_queue_irq_to_lr(vcpu, irq, lr, vlr);
1167 static bool vgic_queue_hwirq(struct kvm_vcpu *vcpu, int irq)
1169 if (!vgic_can_sample_irq(vcpu, irq))
1170 return true; /* level interrupt, already queued */
1172 if (vgic_queue_irq(vcpu, 0, irq)) {
1173 if (vgic_irq_is_edge(vcpu, irq)) {
1174 vgic_dist_irq_clear_pending(vcpu, irq);
1175 vgic_cpu_irq_clear(vcpu, irq);
1177 vgic_irq_set_queued(vcpu, irq);
1187 * Fill the list registers with pending interrupts before running the
1190 static void __kvm_vgic_flush_hwstate(struct kvm_vcpu *vcpu)
1192 struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
1193 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1194 unsigned long *pa_percpu, *pa_shared;
1197 int nr_shared = vgic_nr_shared_irqs(dist);
1199 vcpu_id = vcpu->vcpu_id;
1201 pa_percpu = vcpu->arch.vgic_cpu.pend_act_percpu;
1202 pa_shared = vcpu->arch.vgic_cpu.pend_act_shared;
1204 bitmap_or(pa_percpu, vgic_cpu->pending_percpu, vgic_cpu->active_percpu,
1205 VGIC_NR_PRIVATE_IRQS);
1206 bitmap_or(pa_shared, vgic_cpu->pending_shared, vgic_cpu->active_shared,
1209 * We may not have any pending interrupt, or the interrupts
1210 * may have been serviced from another vcpu. In all cases,
1213 if (!kvm_vgic_vcpu_pending_irq(vcpu) && !kvm_vgic_vcpu_active_irq(vcpu))
1217 for_each_set_bit(i, pa_percpu, VGIC_NR_SGIS) {
1218 if (!queue_sgi(vcpu, i))
1223 for_each_set_bit_from(i, pa_percpu, VGIC_NR_PRIVATE_IRQS) {
1224 if (!vgic_queue_hwirq(vcpu, i))
1229 for_each_set_bit(i, pa_shared, nr_shared) {
1230 if (!vgic_queue_hwirq(vcpu, i + VGIC_NR_PRIVATE_IRQS))
1239 vgic_enable_underflow(vcpu);
1241 vgic_disable_underflow(vcpu);
1243 * We're about to run this VCPU, and we've consumed
1244 * everything the distributor had in store for
1245 * us. Claim we don't have anything pending. We'll
1246 * adjust that if needed while exiting.
1248 clear_bit(vcpu_id, dist->irq_pending_on_cpu);
1252 static bool vgic_process_maintenance(struct kvm_vcpu *vcpu)
1254 u32 status = vgic_get_interrupt_status(vcpu);
1255 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1256 bool level_pending = false;
1257 struct kvm *kvm = vcpu->kvm;
1259 kvm_debug("STATUS = %08x\n", status);
1261 if (status & INT_STATUS_EOI) {
1263 * Some level interrupts have been EOIed. Clear their
1266 u64 eisr = vgic_get_eisr(vcpu);
1267 unsigned long *eisr_ptr = u64_to_bitmask(&eisr);
1270 for_each_set_bit(lr, eisr_ptr, vgic->nr_lr) {
1271 struct vgic_lr vlr = vgic_get_lr(vcpu, lr);
1272 WARN_ON(vgic_irq_is_edge(vcpu, vlr.irq));
1274 spin_lock(&dist->lock);
1275 vgic_irq_clear_queued(vcpu, vlr.irq);
1276 WARN_ON(vlr.state & LR_STATE_MASK);
1278 vgic_set_lr(vcpu, lr, vlr);
1281 * If the IRQ was EOIed it was also ACKed and we we
1282 * therefore assume we can clear the soft pending
1283 * state (should it had been set) for this interrupt.
1285 * Note: if the IRQ soft pending state was set after
1286 * the IRQ was acked, it actually shouldn't be
1287 * cleared, but we have no way of knowing that unless
1288 * we start trapping ACKs when the soft-pending state
1291 vgic_dist_irq_clear_soft_pend(vcpu, vlr.irq);
1294 * kvm_notify_acked_irq calls kvm_set_irq()
1295 * to reset the IRQ level. Need to release the
1296 * lock for kvm_set_irq to grab it.
1298 spin_unlock(&dist->lock);
1300 kvm_notify_acked_irq(kvm, 0,
1301 vlr.irq - VGIC_NR_PRIVATE_IRQS);
1302 spin_lock(&dist->lock);
1304 /* Any additional pending interrupt? */
1305 if (vgic_dist_irq_get_level(vcpu, vlr.irq)) {
1306 vgic_cpu_irq_set(vcpu, vlr.irq);
1307 level_pending = true;
1309 vgic_dist_irq_clear_pending(vcpu, vlr.irq);
1310 vgic_cpu_irq_clear(vcpu, vlr.irq);
1313 spin_unlock(&dist->lock);
1316 * Despite being EOIed, the LR may not have
1317 * been marked as empty.
1319 vgic_sync_lr_elrsr(vcpu, lr, vlr);
1323 if (status & INT_STATUS_UNDERFLOW)
1324 vgic_disable_underflow(vcpu);
1326 return level_pending;
1329 /* Sync back the VGIC state after a guest run */
1330 static void __kvm_vgic_sync_hwstate(struct kvm_vcpu *vcpu)
1332 struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
1333 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1335 unsigned long *elrsr_ptr;
1339 level_pending = vgic_process_maintenance(vcpu);
1340 elrsr = vgic_get_elrsr(vcpu);
1341 elrsr_ptr = u64_to_bitmask(&elrsr);
1343 /* Clear mappings for empty LRs */
1344 for_each_set_bit(lr, elrsr_ptr, vgic->nr_lr) {
1347 if (!test_and_clear_bit(lr, vgic_cpu->lr_used))
1350 vlr = vgic_get_lr(vcpu, lr);
1352 BUG_ON(vlr.irq >= dist->nr_irqs);
1353 vgic_cpu->vgic_irq_lr_map[vlr.irq] = LR_EMPTY;
1356 /* Check if we still have something up our sleeve... */
1357 pending = find_first_zero_bit(elrsr_ptr, vgic->nr_lr);
1358 if (level_pending || pending < vgic->nr_lr)
1359 set_bit(vcpu->vcpu_id, dist->irq_pending_on_cpu);
1362 void kvm_vgic_flush_hwstate(struct kvm_vcpu *vcpu)
1364 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1366 if (!irqchip_in_kernel(vcpu->kvm))
1369 spin_lock(&dist->lock);
1370 __kvm_vgic_flush_hwstate(vcpu);
1371 spin_unlock(&dist->lock);
1374 void kvm_vgic_sync_hwstate(struct kvm_vcpu *vcpu)
1376 if (!irqchip_in_kernel(vcpu->kvm))
1379 __kvm_vgic_sync_hwstate(vcpu);
1382 int kvm_vgic_vcpu_pending_irq(struct kvm_vcpu *vcpu)
1384 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1386 if (!irqchip_in_kernel(vcpu->kvm))
1389 return test_bit(vcpu->vcpu_id, dist->irq_pending_on_cpu);
1392 int kvm_vgic_vcpu_active_irq(struct kvm_vcpu *vcpu)
1394 struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
1396 if (!irqchip_in_kernel(vcpu->kvm))
1399 return test_bit(vcpu->vcpu_id, dist->irq_active_on_cpu);
1403 void vgic_kick_vcpus(struct kvm *kvm)
1405 struct kvm_vcpu *vcpu;
1409 * We've injected an interrupt, time to find out who deserves
1412 kvm_for_each_vcpu(c, vcpu, kvm) {
1413 if (kvm_vgic_vcpu_pending_irq(vcpu))
1414 kvm_vcpu_kick(vcpu);
1418 static int vgic_validate_injection(struct kvm_vcpu *vcpu, int irq, int level)
1420 int edge_triggered = vgic_irq_is_edge(vcpu, irq);
1423 * Only inject an interrupt if:
1424 * - edge triggered and we have a rising edge
1425 * - level triggered and we change level
1427 if (edge_triggered) {
1428 int state = vgic_dist_irq_is_pending(vcpu, irq);
1429 return level > state;
1431 int state = vgic_dist_irq_get_level(vcpu, irq);
1432 return level != state;
1436 static int vgic_update_irq_pending(struct kvm *kvm, int cpuid,
1437 unsigned int irq_num, bool level)
1439 struct vgic_dist *dist = &kvm->arch.vgic;
1440 struct kvm_vcpu *vcpu;
1441 int edge_triggered, level_triggered;
1443 bool ret = true, can_inject = true;
1445 spin_lock(&dist->lock);
1447 vcpu = kvm_get_vcpu(kvm, cpuid);
1448 edge_triggered = vgic_irq_is_edge(vcpu, irq_num);
1449 level_triggered = !edge_triggered;
1451 if (!vgic_validate_injection(vcpu, irq_num, level)) {
1456 if (irq_num >= VGIC_NR_PRIVATE_IRQS) {
1457 cpuid = dist->irq_spi_cpu[irq_num - VGIC_NR_PRIVATE_IRQS];
1458 if (cpuid == VCPU_NOT_ALLOCATED) {
1459 /* Pretend we use CPU0, and prevent injection */
1463 vcpu = kvm_get_vcpu(kvm, cpuid);
1466 kvm_debug("Inject IRQ%d level %d CPU%d\n", irq_num, level, cpuid);
1469 if (level_triggered)
1470 vgic_dist_irq_set_level(vcpu, irq_num);
1471 vgic_dist_irq_set_pending(vcpu, irq_num);
1473 if (level_triggered) {
1474 vgic_dist_irq_clear_level(vcpu, irq_num);
1475 if (!vgic_dist_irq_soft_pend(vcpu, irq_num))
1476 vgic_dist_irq_clear_pending(vcpu, irq_num);
1483 enabled = vgic_irq_is_enabled(vcpu, irq_num);
1485 if (!enabled || !can_inject) {
1490 if (!vgic_can_sample_irq(vcpu, irq_num)) {
1492 * Level interrupt in progress, will be picked up
1500 vgic_cpu_irq_set(vcpu, irq_num);
1501 set_bit(cpuid, dist->irq_pending_on_cpu);
1505 spin_unlock(&dist->lock);
1507 return ret ? cpuid : -EINVAL;
1511 * kvm_vgic_inject_irq - Inject an IRQ from a device to the vgic
1512 * @kvm: The VM structure pointer
1513 * @cpuid: The CPU for PPIs
1514 * @irq_num: The IRQ number that is assigned to the device
1515 * @level: Edge-triggered: true: to trigger the interrupt
1516 * false: to ignore the call
1517 * Level-sensitive true: activates an interrupt
1518 * false: deactivates an interrupt
1520 * The GIC is not concerned with devices being active-LOW or active-HIGH for
1521 * level-sensitive interrupts. You can think of the level parameter as 1
1522 * being HIGH and 0 being LOW and all devices being active-HIGH.
1524 int kvm_vgic_inject_irq(struct kvm *kvm, int cpuid, unsigned int irq_num,
1530 if (unlikely(!vgic_initialized(kvm))) {
1532 * We only provide the automatic initialization of the VGIC
1533 * for the legacy case of a GICv2. Any other type must
1534 * be explicitly initialized once setup with the respective
1537 if (kvm->arch.vgic.vgic_model != KVM_DEV_TYPE_ARM_VGIC_V2) {
1541 mutex_lock(&kvm->lock);
1542 ret = vgic_init(kvm);
1543 mutex_unlock(&kvm->lock);
1549 vcpu_id = vgic_update_irq_pending(kvm, cpuid, irq_num, level);
1551 /* kick the specified vcpu */
1552 kvm_vcpu_kick(kvm_get_vcpu(kvm, vcpu_id));
1559 static irqreturn_t vgic_maintenance_handler(int irq, void *data)
1562 * We cannot rely on the vgic maintenance interrupt to be
1563 * delivered synchronously. This means we can only use it to
1564 * exit the VM, and we perform the handling of EOIed
1565 * interrupts on the exit path (see vgic_process_maintenance).
1570 void kvm_vgic_vcpu_destroy(struct kvm_vcpu *vcpu)
1572 struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
1574 kfree(vgic_cpu->pending_shared);
1575 kfree(vgic_cpu->active_shared);
1576 kfree(vgic_cpu->pend_act_shared);
1577 kfree(vgic_cpu->vgic_irq_lr_map);
1578 vgic_cpu->pending_shared = NULL;
1579 vgic_cpu->active_shared = NULL;
1580 vgic_cpu->pend_act_shared = NULL;
1581 vgic_cpu->vgic_irq_lr_map = NULL;
1584 static int vgic_vcpu_init_maps(struct kvm_vcpu *vcpu, int nr_irqs)
1586 struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
1588 int sz = (nr_irqs - VGIC_NR_PRIVATE_IRQS) / 8;
1589 vgic_cpu->pending_shared = kzalloc(sz, GFP_KERNEL);
1590 vgic_cpu->active_shared = kzalloc(sz, GFP_KERNEL);
1591 vgic_cpu->pend_act_shared = kzalloc(sz, GFP_KERNEL);
1592 vgic_cpu->vgic_irq_lr_map = kmalloc(nr_irqs, GFP_KERNEL);
1594 if (!vgic_cpu->pending_shared
1595 || !vgic_cpu->active_shared
1596 || !vgic_cpu->pend_act_shared
1597 || !vgic_cpu->vgic_irq_lr_map) {
1598 kvm_vgic_vcpu_destroy(vcpu);
1602 memset(vgic_cpu->vgic_irq_lr_map, LR_EMPTY, nr_irqs);
1605 * Store the number of LRs per vcpu, so we don't have to go
1606 * all the way to the distributor structure to find out. Only
1607 * assembly code should use this one.
1609 vgic_cpu->nr_lr = vgic->nr_lr;
1615 * kvm_vgic_get_max_vcpus - Get the maximum number of VCPUs allowed by HW
1617 * The host's GIC naturally limits the maximum amount of VCPUs a guest
1620 int kvm_vgic_get_max_vcpus(void)
1622 return vgic->max_gic_vcpus;
1625 void kvm_vgic_destroy(struct kvm *kvm)
1627 struct vgic_dist *dist = &kvm->arch.vgic;
1628 struct kvm_vcpu *vcpu;
1631 kvm_for_each_vcpu(i, vcpu, kvm)
1632 kvm_vgic_vcpu_destroy(vcpu);
1634 vgic_free_bitmap(&dist->irq_enabled);
1635 vgic_free_bitmap(&dist->irq_level);
1636 vgic_free_bitmap(&dist->irq_pending);
1637 vgic_free_bitmap(&dist->irq_soft_pend);
1638 vgic_free_bitmap(&dist->irq_queued);
1639 vgic_free_bitmap(&dist->irq_cfg);
1640 vgic_free_bytemap(&dist->irq_priority);
1641 if (dist->irq_spi_target) {
1642 for (i = 0; i < dist->nr_cpus; i++)
1643 vgic_free_bitmap(&dist->irq_spi_target[i]);
1645 kfree(dist->irq_sgi_sources);
1646 kfree(dist->irq_spi_cpu);
1647 kfree(dist->irq_spi_mpidr);
1648 kfree(dist->irq_spi_target);
1649 kfree(dist->irq_pending_on_cpu);
1650 kfree(dist->irq_active_on_cpu);
1651 dist->irq_sgi_sources = NULL;
1652 dist->irq_spi_cpu = NULL;
1653 dist->irq_spi_target = NULL;
1654 dist->irq_pending_on_cpu = NULL;
1655 dist->irq_active_on_cpu = NULL;
1660 * Allocate and initialize the various data structures. Must be called
1661 * with kvm->lock held!
1663 int vgic_init(struct kvm *kvm)
1665 struct vgic_dist *dist = &kvm->arch.vgic;
1666 struct kvm_vcpu *vcpu;
1667 int nr_cpus, nr_irqs;
1668 int ret, i, vcpu_id;
1670 if (vgic_initialized(kvm))
1673 nr_cpus = dist->nr_cpus = atomic_read(&kvm->online_vcpus);
1674 if (!nr_cpus) /* No vcpus? Can't be good... */
1678 * If nobody configured the number of interrupts, use the
1682 dist->nr_irqs = VGIC_NR_IRQS_LEGACY;
1684 nr_irqs = dist->nr_irqs;
1686 ret = vgic_init_bitmap(&dist->irq_enabled, nr_cpus, nr_irqs);
1687 ret |= vgic_init_bitmap(&dist->irq_level, nr_cpus, nr_irqs);
1688 ret |= vgic_init_bitmap(&dist->irq_pending, nr_cpus, nr_irqs);
1689 ret |= vgic_init_bitmap(&dist->irq_soft_pend, nr_cpus, nr_irqs);
1690 ret |= vgic_init_bitmap(&dist->irq_queued, nr_cpus, nr_irqs);
1691 ret |= vgic_init_bitmap(&dist->irq_active, nr_cpus, nr_irqs);
1692 ret |= vgic_init_bitmap(&dist->irq_cfg, nr_cpus, nr_irqs);
1693 ret |= vgic_init_bytemap(&dist->irq_priority, nr_cpus, nr_irqs);
1698 dist->irq_sgi_sources = kzalloc(nr_cpus * VGIC_NR_SGIS, GFP_KERNEL);
1699 dist->irq_spi_cpu = kzalloc(nr_irqs - VGIC_NR_PRIVATE_IRQS, GFP_KERNEL);
1700 dist->irq_spi_target = kzalloc(sizeof(*dist->irq_spi_target) * nr_cpus,
1702 dist->irq_pending_on_cpu = kzalloc(BITS_TO_LONGS(nr_cpus) * sizeof(long),
1704 dist->irq_active_on_cpu = kzalloc(BITS_TO_LONGS(nr_cpus) * sizeof(long),
1706 if (!dist->irq_sgi_sources ||
1707 !dist->irq_spi_cpu ||
1708 !dist->irq_spi_target ||
1709 !dist->irq_pending_on_cpu ||
1710 !dist->irq_active_on_cpu) {
1715 for (i = 0; i < nr_cpus; i++)
1716 ret |= vgic_init_bitmap(&dist->irq_spi_target[i],
1722 ret = kvm->arch.vgic.vm_ops.init_model(kvm);
1726 kvm_for_each_vcpu(vcpu_id, vcpu, kvm) {
1727 ret = vgic_vcpu_init_maps(vcpu, nr_irqs);
1729 kvm_err("VGIC: Failed to allocate vcpu memory\n");
1733 for (i = 0; i < dist->nr_irqs; i++) {
1734 if (i < VGIC_NR_PPIS)
1735 vgic_bitmap_set_irq_val(&dist->irq_enabled,
1736 vcpu->vcpu_id, i, 1);
1737 if (i < VGIC_NR_PRIVATE_IRQS)
1738 vgic_bitmap_set_irq_val(&dist->irq_cfg,
1748 kvm_vgic_destroy(kvm);
1753 static int init_vgic_model(struct kvm *kvm, int type)
1756 case KVM_DEV_TYPE_ARM_VGIC_V2:
1757 vgic_v2_init_emulation(kvm);
1759 #ifdef CONFIG_ARM_GIC_V3
1760 case KVM_DEV_TYPE_ARM_VGIC_V3:
1761 vgic_v3_init_emulation(kvm);
1768 if (atomic_read(&kvm->online_vcpus) > kvm->arch.max_vcpus)
1774 int kvm_vgic_create(struct kvm *kvm, u32 type)
1776 int i, vcpu_lock_idx = -1, ret;
1777 struct kvm_vcpu *vcpu;
1779 mutex_lock(&kvm->lock);
1781 if (irqchip_in_kernel(kvm)) {
1787 * This function is also called by the KVM_CREATE_IRQCHIP handler,
1788 * which had no chance yet to check the availability of the GICv2
1789 * emulation. So check this here again. KVM_CREATE_DEVICE does
1790 * the proper checks already.
1792 if (type == KVM_DEV_TYPE_ARM_VGIC_V2 && !vgic->can_emulate_gicv2)
1796 * Any time a vcpu is run, vcpu_load is called which tries to grab the
1797 * vcpu->mutex. By grabbing the vcpu->mutex of all VCPUs we ensure
1798 * that no other VCPUs are run while we create the vgic.
1801 kvm_for_each_vcpu(i, vcpu, kvm) {
1802 if (!mutex_trylock(&vcpu->mutex))
1807 kvm_for_each_vcpu(i, vcpu, kvm) {
1808 if (vcpu->arch.has_run_once)
1813 ret = init_vgic_model(kvm, type);
1817 spin_lock_init(&kvm->arch.vgic.lock);
1818 kvm->arch.vgic.in_kernel = true;
1819 kvm->arch.vgic.vgic_model = type;
1820 kvm->arch.vgic.vctrl_base = vgic->vctrl_base;
1821 kvm->arch.vgic.vgic_dist_base = VGIC_ADDR_UNDEF;
1822 kvm->arch.vgic.vgic_cpu_base = VGIC_ADDR_UNDEF;
1823 kvm->arch.vgic.vgic_redist_base = VGIC_ADDR_UNDEF;
1826 for (; vcpu_lock_idx >= 0; vcpu_lock_idx--) {
1827 vcpu = kvm_get_vcpu(kvm, vcpu_lock_idx);
1828 mutex_unlock(&vcpu->mutex);
1832 mutex_unlock(&kvm->lock);
1836 static int vgic_ioaddr_overlap(struct kvm *kvm)
1838 phys_addr_t dist = kvm->arch.vgic.vgic_dist_base;
1839 phys_addr_t cpu = kvm->arch.vgic.vgic_cpu_base;
1841 if (IS_VGIC_ADDR_UNDEF(dist) || IS_VGIC_ADDR_UNDEF(cpu))
1843 if ((dist <= cpu && dist + KVM_VGIC_V2_DIST_SIZE > cpu) ||
1844 (cpu <= dist && cpu + KVM_VGIC_V2_CPU_SIZE > dist))
1849 static int vgic_ioaddr_assign(struct kvm *kvm, phys_addr_t *ioaddr,
1850 phys_addr_t addr, phys_addr_t size)
1854 if (addr & ~KVM_PHYS_MASK)
1857 if (addr & (SZ_4K - 1))
1860 if (!IS_VGIC_ADDR_UNDEF(*ioaddr))
1862 if (addr + size < addr)
1866 ret = vgic_ioaddr_overlap(kvm);
1868 *ioaddr = VGIC_ADDR_UNDEF;
1874 * kvm_vgic_addr - set or get vgic VM base addresses
1875 * @kvm: pointer to the vm struct
1876 * @type: the VGIC addr type, one of KVM_VGIC_V[23]_ADDR_TYPE_XXX
1877 * @addr: pointer to address value
1878 * @write: if true set the address in the VM address space, if false read the
1881 * Set or get the vgic base addresses for the distributor and the virtual CPU
1882 * interface in the VM physical address space. These addresses are properties
1883 * of the emulated core/SoC and therefore user space initially knows this
1886 int kvm_vgic_addr(struct kvm *kvm, unsigned long type, u64 *addr, bool write)
1889 struct vgic_dist *vgic = &kvm->arch.vgic;
1891 phys_addr_t *addr_ptr, block_size;
1892 phys_addr_t alignment;
1894 mutex_lock(&kvm->lock);
1896 case KVM_VGIC_V2_ADDR_TYPE_DIST:
1897 type_needed = KVM_DEV_TYPE_ARM_VGIC_V2;
1898 addr_ptr = &vgic->vgic_dist_base;
1899 block_size = KVM_VGIC_V2_DIST_SIZE;
1902 case KVM_VGIC_V2_ADDR_TYPE_CPU:
1903 type_needed = KVM_DEV_TYPE_ARM_VGIC_V2;
1904 addr_ptr = &vgic->vgic_cpu_base;
1905 block_size = KVM_VGIC_V2_CPU_SIZE;
1908 #ifdef CONFIG_ARM_GIC_V3
1909 case KVM_VGIC_V3_ADDR_TYPE_DIST:
1910 type_needed = KVM_DEV_TYPE_ARM_VGIC_V3;
1911 addr_ptr = &vgic->vgic_dist_base;
1912 block_size = KVM_VGIC_V3_DIST_SIZE;
1915 case KVM_VGIC_V3_ADDR_TYPE_REDIST:
1916 type_needed = KVM_DEV_TYPE_ARM_VGIC_V3;
1917 addr_ptr = &vgic->vgic_redist_base;
1918 block_size = KVM_VGIC_V3_REDIST_SIZE;
1927 if (vgic->vgic_model != type_needed) {
1933 if (!IS_ALIGNED(*addr, alignment))
1936 r = vgic_ioaddr_assign(kvm, addr_ptr, *addr,
1943 mutex_unlock(&kvm->lock);
1947 int vgic_set_common_attr(struct kvm_device *dev, struct kvm_device_attr *attr)
1951 switch (attr->group) {
1952 case KVM_DEV_ARM_VGIC_GRP_ADDR: {
1953 u64 __user *uaddr = (u64 __user *)(long)attr->addr;
1955 unsigned long type = (unsigned long)attr->attr;
1957 if (copy_from_user(&addr, uaddr, sizeof(addr)))
1960 r = kvm_vgic_addr(dev->kvm, type, &addr, true);
1961 return (r == -ENODEV) ? -ENXIO : r;
1963 case KVM_DEV_ARM_VGIC_GRP_NR_IRQS: {
1964 u32 __user *uaddr = (u32 __user *)(long)attr->addr;
1968 if (get_user(val, uaddr))
1973 * - at least 32 SPIs on top of the 16 SGIs and 16 PPIs
1974 * - at most 1024 interrupts
1975 * - a multiple of 32 interrupts
1977 if (val < (VGIC_NR_PRIVATE_IRQS + 32) ||
1978 val > VGIC_MAX_IRQS ||
1982 mutex_lock(&dev->kvm->lock);
1984 if (vgic_ready(dev->kvm) || dev->kvm->arch.vgic.nr_irqs)
1987 dev->kvm->arch.vgic.nr_irqs = val;
1989 mutex_unlock(&dev->kvm->lock);
1993 case KVM_DEV_ARM_VGIC_GRP_CTRL: {
1994 switch (attr->attr) {
1995 case KVM_DEV_ARM_VGIC_CTRL_INIT:
1996 r = vgic_init(dev->kvm);
2006 int vgic_get_common_attr(struct kvm_device *dev, struct kvm_device_attr *attr)
2010 switch (attr->group) {
2011 case KVM_DEV_ARM_VGIC_GRP_ADDR: {
2012 u64 __user *uaddr = (u64 __user *)(long)attr->addr;
2014 unsigned long type = (unsigned long)attr->attr;
2016 r = kvm_vgic_addr(dev->kvm, type, &addr, false);
2018 return (r == -ENODEV) ? -ENXIO : r;
2020 if (copy_to_user(uaddr, &addr, sizeof(addr)))
2024 case KVM_DEV_ARM_VGIC_GRP_NR_IRQS: {
2025 u32 __user *uaddr = (u32 __user *)(long)attr->addr;
2027 r = put_user(dev->kvm->arch.vgic.nr_irqs, uaddr);
2036 int vgic_has_attr_regs(const struct vgic_io_range *ranges, phys_addr_t offset)
2038 if (vgic_find_range(ranges, 4, offset))
2044 static void vgic_init_maintenance_interrupt(void *info)
2046 enable_percpu_irq(vgic->maint_irq, 0);
2049 static int vgic_cpu_notify(struct notifier_block *self,
2050 unsigned long action, void *cpu)
2054 case CPU_STARTING_FROZEN:
2055 vgic_init_maintenance_interrupt(NULL);
2058 case CPU_DYING_FROZEN:
2059 disable_percpu_irq(vgic->maint_irq);
2066 static struct notifier_block vgic_cpu_nb = {
2067 .notifier_call = vgic_cpu_notify,
2070 static const struct of_device_id vgic_ids[] = {
2071 { .compatible = "arm,cortex-a15-gic", .data = vgic_v2_probe, },
2072 { .compatible = "arm,cortex-a7-gic", .data = vgic_v2_probe, },
2073 { .compatible = "arm,gic-400", .data = vgic_v2_probe, },
2074 { .compatible = "arm,gic-v3", .data = vgic_v3_probe, },
2078 int kvm_vgic_hyp_init(void)
2080 const struct of_device_id *matched_id;
2081 const int (*vgic_probe)(struct device_node *,const struct vgic_ops **,
2082 const struct vgic_params **);
2083 struct device_node *vgic_node;
2086 vgic_node = of_find_matching_node_and_match(NULL,
2087 vgic_ids, &matched_id);
2089 kvm_err("error: no compatible GIC node found\n");
2093 vgic_probe = matched_id->data;
2094 ret = vgic_probe(vgic_node, &vgic_ops, &vgic);
2098 ret = request_percpu_irq(vgic->maint_irq, vgic_maintenance_handler,
2099 "vgic", kvm_get_running_vcpus());
2101 kvm_err("Cannot register interrupt %d\n", vgic->maint_irq);
2105 ret = __register_cpu_notifier(&vgic_cpu_nb);
2107 kvm_err("Cannot register vgic CPU notifier\n");
2111 /* Callback into for arch code for setup */
2112 vgic_arch_setup(vgic);
2114 on_each_cpu(vgic_init_maintenance_interrupt, NULL, 1);
2119 free_percpu_irq(vgic->maint_irq, kvm_get_running_vcpus());
2123 int kvm_irq_map_gsi(struct kvm *kvm,
2124 struct kvm_kernel_irq_routing_entry *entries,
2130 int kvm_irq_map_chip_pin(struct kvm *kvm, unsigned irqchip, unsigned pin)
2135 int kvm_set_irq(struct kvm *kvm, int irq_source_id,
2136 u32 irq, int level, bool line_status)
2138 unsigned int spi = irq + VGIC_NR_PRIVATE_IRQS;
2140 trace_kvm_set_irq(irq, level, irq_source_id);
2142 BUG_ON(!vgic_initialized(kvm));
2144 if (spi > kvm->arch.vgic.nr_irqs)
2146 return kvm_vgic_inject_irq(kvm, 0, spi, level);
2150 /* MSI not implemented yet */
2151 int kvm_set_msi(struct kvm_kernel_irq_routing_entry *e,
2152 struct kvm *kvm, int irq_source_id,
2153 int level, bool line_status)