4 * Copyright (C) 2015,2016 ARM Ltd.
5 * Author: Andre Przywara <andre.przywara@arm.com>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program. If not, see <http://www.gnu.org/licenses/>.
20 #include <linux/cpu.h>
21 #include <linux/kvm.h>
22 #include <linux/kvm_host.h>
23 #include <linux/interrupt.h>
24 #include <linux/list.h>
25 #include <linux/uaccess.h>
27 #include <linux/irqchip/arm-gic-v3.h>
29 #include <asm/kvm_emulate.h>
30 #include <asm/kvm_arm.h>
31 #include <asm/kvm_mmu.h>
34 #include "vgic-mmio.h"
37 * Creates a new (reference to a) struct vgic_irq for a given LPI.
38 * If this LPI is already mapped on another ITS, we increase its refcount
39 * and return a pointer to the existing structure.
40 * If this is a "new" LPI, we allocate and initialize a new struct vgic_irq.
41 * This function returns a pointer to the _unlocked_ structure.
43 static struct vgic_irq *vgic_add_lpi(struct kvm *kvm, u32 intid)
45 struct vgic_dist *dist = &kvm->arch.vgic;
46 struct vgic_irq *irq = vgic_get_irq(kvm, NULL, intid), *oldirq;
48 /* In this case there is no put, since we keep the reference. */
52 irq = kzalloc(sizeof(struct vgic_irq), GFP_KERNEL);
54 return ERR_PTR(-ENOMEM);
56 INIT_LIST_HEAD(&irq->lpi_list);
57 INIT_LIST_HEAD(&irq->ap_list);
58 spin_lock_init(&irq->irq_lock);
60 irq->config = VGIC_CONFIG_EDGE;
61 kref_init(&irq->refcount);
64 spin_lock(&dist->lpi_list_lock);
67 * There could be a race with another vgic_add_lpi(), so we need to
68 * check that we don't add a second list entry with the same LPI.
70 list_for_each_entry(oldirq, &dist->lpi_list_head, lpi_list) {
71 if (oldirq->intid != intid)
74 /* Someone was faster with adding this LPI, lets use that. */
79 * This increases the refcount, the caller is expected to
80 * call vgic_put_irq() on the returned pointer once it's
81 * finished with the IRQ.
83 vgic_get_irq_kref(irq);
88 list_add_tail(&irq->lpi_list, &dist->lpi_list_head);
89 dist->lpi_list_count++;
92 spin_unlock(&dist->lpi_list_lock);
98 struct list_head dev_list;
100 /* the head for the list of ITTEs */
101 struct list_head itt_head;
105 #define COLLECTION_NOT_MAPPED ((u32)~0)
107 struct its_collection {
108 struct list_head coll_list;
114 #define its_is_collection_mapped(coll) ((coll) && \
115 ((coll)->target_addr != COLLECTION_NOT_MAPPED))
118 struct list_head itte_list;
120 struct vgic_irq *irq;
121 struct its_collection *collection;
127 * Find and returns a device in the device table for an ITS.
128 * Must be called with the its_lock mutex held.
130 static struct its_device *find_its_device(struct vgic_its *its, u32 device_id)
132 struct its_device *device;
134 list_for_each_entry(device, &its->device_list, dev_list)
135 if (device_id == device->device_id)
142 * Find and returns an interrupt translation table entry (ITTE) for a given
143 * Device ID/Event ID pair on an ITS.
144 * Must be called with the its_lock mutex held.
146 static struct its_itte *find_itte(struct vgic_its *its, u32 device_id,
149 struct its_device *device;
150 struct its_itte *itte;
152 device = find_its_device(its, device_id);
156 list_for_each_entry(itte, &device->itt_head, itte_list)
157 if (itte->event_id == event_id)
163 /* To be used as an iterator this macro misses the enclosing parentheses */
164 #define for_each_lpi_its(dev, itte, its) \
165 list_for_each_entry(dev, &(its)->device_list, dev_list) \
166 list_for_each_entry(itte, &(dev)->itt_head, itte_list)
169 * We only implement 48 bits of PA at the moment, although the ITS
170 * supports more. Let's be restrictive here.
172 #define BASER_ADDRESS(x) ((x) & GENMASK_ULL(47, 16))
173 #define CBASER_ADDRESS(x) ((x) & GENMASK_ULL(47, 12))
174 #define PENDBASER_ADDRESS(x) ((x) & GENMASK_ULL(47, 16))
175 #define PROPBASER_ADDRESS(x) ((x) & GENMASK_ULL(47, 12))
177 #define GIC_LPI_OFFSET 8192
180 * Finds and returns a collection in the ITS collection table.
181 * Must be called with the its_lock mutex held.
183 static struct its_collection *find_collection(struct vgic_its *its, int coll_id)
185 struct its_collection *collection;
187 list_for_each_entry(collection, &its->collection_list, coll_list) {
188 if (coll_id == collection->collection_id)
195 #define LPI_PROP_ENABLE_BIT(p) ((p) & LPI_PROP_ENABLED)
196 #define LPI_PROP_PRIORITY(p) ((p) & 0xfc)
199 * Reads the configuration data for a given LPI from guest memory and
200 * updates the fields in struct vgic_irq.
201 * If filter_vcpu is not NULL, applies only if the IRQ is targeting this
202 * VCPU. Unconditionally applies if filter_vcpu is NULL.
204 static int update_lpi_config(struct kvm *kvm, struct vgic_irq *irq,
205 struct kvm_vcpu *filter_vcpu)
207 u64 propbase = PROPBASER_ADDRESS(kvm->arch.vgic.propbaser);
211 ret = kvm_read_guest_lock(kvm, propbase + irq->intid - GIC_LPI_OFFSET,
217 spin_lock(&irq->irq_lock);
219 if (!filter_vcpu || filter_vcpu == irq->target_vcpu) {
220 irq->priority = LPI_PROP_PRIORITY(prop);
221 irq->enabled = LPI_PROP_ENABLE_BIT(prop);
223 vgic_queue_irq_unlock(kvm, irq);
225 spin_unlock(&irq->irq_lock);
232 * Create a snapshot of the current LPI list, so that we can enumerate all
233 * LPIs without holding any lock.
234 * Returns the array length and puts the kmalloc'ed array into intid_ptr.
236 static int vgic_copy_lpi_list(struct kvm *kvm, u32 **intid_ptr)
238 struct vgic_dist *dist = &kvm->arch.vgic;
239 struct vgic_irq *irq;
241 int irq_count = dist->lpi_list_count, i = 0;
244 * We use the current value of the list length, which may change
245 * after the kmalloc. We don't care, because the guest shouldn't
246 * change anything while the command handling is still running,
247 * and in the worst case we would miss a new IRQ, which one wouldn't
248 * expect to be covered by this command anyway.
250 intids = kmalloc_array(irq_count, sizeof(intids[0]), GFP_KERNEL);
254 spin_lock(&dist->lpi_list_lock);
255 list_for_each_entry(irq, &dist->lpi_list_head, lpi_list) {
256 /* We don't need to "get" the IRQ, as we hold the list lock. */
257 intids[i] = irq->intid;
258 if (++i == irq_count)
261 spin_unlock(&dist->lpi_list_lock);
268 * Promotes the ITS view of affinity of an ITTE (which redistributor this LPI
269 * is targeting) to the VGIC's view, which deals with target VCPUs.
270 * Needs to be called whenever either the collection for a LPIs has
271 * changed or the collection itself got retargeted.
273 static void update_affinity_itte(struct kvm *kvm, struct its_itte *itte)
275 struct kvm_vcpu *vcpu;
277 if (!its_is_collection_mapped(itte->collection))
280 vcpu = kvm_get_vcpu(kvm, itte->collection->target_addr);
282 spin_lock(&itte->irq->irq_lock);
283 itte->irq->target_vcpu = vcpu;
284 spin_unlock(&itte->irq->irq_lock);
288 * Updates the target VCPU for every LPI targeting this collection.
289 * Must be called with the its_lock mutex held.
291 static void update_affinity_collection(struct kvm *kvm, struct vgic_its *its,
292 struct its_collection *coll)
294 struct its_device *device;
295 struct its_itte *itte;
297 for_each_lpi_its(device, itte, its) {
298 if (!itte->collection || coll != itte->collection)
301 update_affinity_itte(kvm, itte);
305 static u32 max_lpis_propbaser(u64 propbaser)
307 int nr_idbits = (propbaser & 0x1f) + 1;
309 return 1U << min(nr_idbits, INTERRUPT_ID_BITS_ITS);
313 * Scan the whole LPI pending table and sync the pending bit in there
314 * with our own data structures. This relies on the LPI being
317 static int its_sync_lpi_pending_table(struct kvm_vcpu *vcpu)
319 gpa_t pendbase = PENDBASER_ADDRESS(vcpu->arch.vgic_cpu.pendbaser);
320 struct vgic_irq *irq;
321 int last_byte_offset = -1;
327 nr_irqs = vgic_copy_lpi_list(vcpu->kvm, &intids);
331 for (i = 0; i < nr_irqs; i++) {
332 int byte_offset, bit_nr;
334 byte_offset = intids[i] / BITS_PER_BYTE;
335 bit_nr = intids[i] % BITS_PER_BYTE;
338 * For contiguously allocated LPIs chances are we just read
339 * this very same byte in the last iteration. Reuse that.
341 if (byte_offset != last_byte_offset) {
342 ret = kvm_read_guest_lock(vcpu->kvm,
343 pendbase + byte_offset,
349 last_byte_offset = byte_offset;
352 irq = vgic_get_irq(vcpu->kvm, NULL, intids[i]);
353 spin_lock(&irq->irq_lock);
354 irq->pending = pendmask & (1U << bit_nr);
355 vgic_queue_irq_unlock(vcpu->kvm, irq);
356 vgic_put_irq(vcpu->kvm, irq);
364 static unsigned long vgic_mmio_read_its_typer(struct kvm *kvm,
365 struct vgic_its *its,
366 gpa_t addr, unsigned int len)
368 u64 reg = GITS_TYPER_PLPIS;
371 * We use linear CPU numbers for redistributor addressing,
372 * so GITS_TYPER.PTA is 0.
373 * Also we force all PROPBASER registers to be the same, so
374 * CommonLPIAff is 0 as well.
375 * To avoid memory waste in the guest, we keep the number of IDBits and
376 * DevBits low - as least for the time being.
378 reg |= 0x0f << GITS_TYPER_DEVBITS_SHIFT;
379 reg |= 0x0f << GITS_TYPER_IDBITS_SHIFT;
381 return extract_bytes(reg, addr & 7, len);
384 static unsigned long vgic_mmio_read_its_iidr(struct kvm *kvm,
385 struct vgic_its *its,
386 gpa_t addr, unsigned int len)
388 return (PRODUCT_ID_KVM << 24) | (IMPLEMENTER_ARM << 0);
391 static unsigned long vgic_mmio_read_its_idregs(struct kvm *kvm,
392 struct vgic_its *its,
393 gpa_t addr, unsigned int len)
395 switch (addr & 0xffff) {
397 return 0x92; /* part number, bits[7:0] */
399 return 0xb4; /* part number, bits[11:8] */
401 return GIC_PIDR2_ARCH_GICv3 | 0x0b;
403 return 0x40; /* This is a 64K software visible page */
404 /* The following are the ID registers for (any) GIC. */
419 * Find the target VCPU and the LPI number for a given devid/eventid pair
420 * and make this IRQ pending, possibly injecting it.
421 * Must be called with the its_lock mutex held.
422 * Returns 0 on success, a positive error value for any ITS mapping
423 * related errors and negative error values for generic errors.
425 static int vgic_its_trigger_msi(struct kvm *kvm, struct vgic_its *its,
426 u32 devid, u32 eventid)
428 struct kvm_vcpu *vcpu;
429 struct its_itte *itte;
434 itte = find_itte(its, devid, eventid);
435 if (!itte || !its_is_collection_mapped(itte->collection))
436 return E_ITS_INT_UNMAPPED_INTERRUPT;
438 vcpu = kvm_get_vcpu(kvm, itte->collection->target_addr);
440 return E_ITS_INT_UNMAPPED_INTERRUPT;
442 if (!vcpu->arch.vgic_cpu.lpis_enabled)
445 spin_lock(&itte->irq->irq_lock);
446 itte->irq->pending = true;
447 vgic_queue_irq_unlock(kvm, itte->irq);
452 static struct vgic_io_device *vgic_get_its_iodev(struct kvm_io_device *dev)
454 struct vgic_io_device *iodev;
456 if (dev->ops != &kvm_io_gic_ops)
459 iodev = container_of(dev, struct vgic_io_device, dev);
461 if (iodev->iodev_type != IODEV_ITS)
468 * Queries the KVM IO bus framework to get the ITS pointer from the given
470 * We then call vgic_its_trigger_msi() with the decoded data.
471 * According to the KVM_SIGNAL_MSI API description returns 1 on success.
473 int vgic_its_inject_msi(struct kvm *kvm, struct kvm_msi *msi)
476 struct kvm_io_device *kvm_io_dev;
477 struct vgic_io_device *iodev;
480 if (!vgic_has_its(kvm))
483 if (!(msi->flags & KVM_MSI_VALID_DEVID))
486 address = (u64)msi->address_hi << 32 | msi->address_lo;
488 kvm_io_dev = kvm_io_bus_get_dev(kvm, KVM_MMIO_BUS, address);
492 iodev = vgic_get_its_iodev(kvm_io_dev);
496 mutex_lock(&iodev->its->its_lock);
497 ret = vgic_its_trigger_msi(kvm, iodev->its, msi->devid, msi->data);
498 mutex_unlock(&iodev->its->its_lock);
504 * KVM_SIGNAL_MSI demands a return value > 0 for success and 0
505 * if the guest has blocked the MSI. So we map any LPI mapping
506 * related error to that.
514 /* Requires the its_lock to be held. */
515 static void its_free_itte(struct kvm *kvm, struct its_itte *itte)
517 list_del(&itte->itte_list);
519 /* This put matches the get in vgic_add_lpi. */
521 vgic_put_irq(kvm, itte->irq);
526 static u64 its_cmd_mask_field(u64 *its_cmd, int word, int shift, int size)
528 return (le64_to_cpu(its_cmd[word]) >> shift) & (BIT_ULL(size) - 1);
531 #define its_cmd_get_command(cmd) its_cmd_mask_field(cmd, 0, 0, 8)
532 #define its_cmd_get_deviceid(cmd) its_cmd_mask_field(cmd, 0, 32, 32)
533 #define its_cmd_get_id(cmd) its_cmd_mask_field(cmd, 1, 0, 32)
534 #define its_cmd_get_physical_id(cmd) its_cmd_mask_field(cmd, 1, 32, 32)
535 #define its_cmd_get_collection(cmd) its_cmd_mask_field(cmd, 2, 0, 16)
536 #define its_cmd_get_target_addr(cmd) its_cmd_mask_field(cmd, 2, 16, 32)
537 #define its_cmd_get_validbit(cmd) its_cmd_mask_field(cmd, 2, 63, 1)
540 * The DISCARD command frees an Interrupt Translation Table Entry (ITTE).
541 * Must be called with the its_lock mutex held.
543 static int vgic_its_cmd_handle_discard(struct kvm *kvm, struct vgic_its *its,
546 u32 device_id = its_cmd_get_deviceid(its_cmd);
547 u32 event_id = its_cmd_get_id(its_cmd);
548 struct its_itte *itte;
551 itte = find_itte(its, device_id, event_id);
552 if (itte && itte->collection) {
554 * Though the spec talks about removing the pending state, we
555 * don't bother here since we clear the ITTE anyway and the
556 * pending state is a property of the ITTE struct.
558 its_free_itte(kvm, itte);
562 return E_ITS_DISCARD_UNMAPPED_INTERRUPT;
566 * The MOVI command moves an ITTE to a different collection.
567 * Must be called with the its_lock mutex held.
569 static int vgic_its_cmd_handle_movi(struct kvm *kvm, struct vgic_its *its,
572 u32 device_id = its_cmd_get_deviceid(its_cmd);
573 u32 event_id = its_cmd_get_id(its_cmd);
574 u32 coll_id = its_cmd_get_collection(its_cmd);
575 struct kvm_vcpu *vcpu;
576 struct its_itte *itte;
577 struct its_collection *collection;
579 itte = find_itte(its, device_id, event_id);
581 return E_ITS_MOVI_UNMAPPED_INTERRUPT;
583 if (!its_is_collection_mapped(itte->collection))
584 return E_ITS_MOVI_UNMAPPED_COLLECTION;
586 collection = find_collection(its, coll_id);
587 if (!its_is_collection_mapped(collection))
588 return E_ITS_MOVI_UNMAPPED_COLLECTION;
590 itte->collection = collection;
591 vcpu = kvm_get_vcpu(kvm, collection->target_addr);
593 spin_lock(&itte->irq->irq_lock);
594 itte->irq->target_vcpu = vcpu;
595 spin_unlock(&itte->irq->irq_lock);
601 * Check whether an ID can be stored into the corresponding guest table.
602 * For a direct table this is pretty easy, but gets a bit nasty for
603 * indirect tables. We check whether the resulting guest physical address
604 * is actually valid (covered by a memslot and guest accessbible).
605 * For this we have to read the respective first level entry.
607 static bool vgic_its_check_id(struct vgic_its *its, u64 baser, int id)
609 int l1_tbl_size = GITS_BASER_NR_PAGES(baser) * SZ_64K;
614 if (!(baser & GITS_BASER_INDIRECT)) {
617 if (id >= (l1_tbl_size / GITS_BASER_ENTRY_SIZE(baser)))
620 addr = BASER_ADDRESS(baser) + id * GITS_BASER_ENTRY_SIZE(baser);
621 gfn = addr >> PAGE_SHIFT;
623 return kvm_is_visible_gfn(its->dev->kvm, gfn);
626 /* calculate and check the index into the 1st level */
627 index = id / (SZ_64K / GITS_BASER_ENTRY_SIZE(baser));
628 if (index >= (l1_tbl_size / sizeof(u64)))
631 /* Each 1st level entry is represented by a 64-bit value. */
632 if (kvm_read_guest_lock(its->dev->kvm,
633 BASER_ADDRESS(baser) + index * sizeof(indirect_ptr),
634 &indirect_ptr, sizeof(indirect_ptr)))
637 indirect_ptr = le64_to_cpu(indirect_ptr);
639 /* check the valid bit of the first level entry */
640 if (!(indirect_ptr & BIT_ULL(63)))
644 * Mask the guest physical address and calculate the frame number.
645 * Any address beyond our supported 48 bits of PA will be caught
646 * by the actual check in the final step.
648 indirect_ptr &= GENMASK_ULL(51, 16);
650 /* Find the address of the actual entry */
651 index = id % (SZ_64K / GITS_BASER_ENTRY_SIZE(baser));
652 indirect_ptr += index * GITS_BASER_ENTRY_SIZE(baser);
653 gfn = indirect_ptr >> PAGE_SHIFT;
655 return kvm_is_visible_gfn(its->dev->kvm, gfn);
658 static int vgic_its_alloc_collection(struct vgic_its *its,
659 struct its_collection **colp,
662 struct its_collection *collection;
664 if (!vgic_its_check_id(its, its->baser_coll_table, coll_id))
665 return E_ITS_MAPC_COLLECTION_OOR;
667 collection = kzalloc(sizeof(*collection), GFP_KERNEL);
671 collection->collection_id = coll_id;
672 collection->target_addr = COLLECTION_NOT_MAPPED;
674 list_add_tail(&collection->coll_list, &its->collection_list);
680 static void vgic_its_free_collection(struct vgic_its *its, u32 coll_id)
682 struct its_collection *collection;
683 struct its_device *device;
684 struct its_itte *itte;
687 * Clearing the mapping for that collection ID removes the
688 * entry from the list. If there wasn't any before, we can
691 collection = find_collection(its, coll_id);
695 for_each_lpi_its(device, itte, its)
696 if (itte->collection &&
697 itte->collection->collection_id == coll_id)
698 itte->collection = NULL;
700 list_del(&collection->coll_list);
705 * The MAPTI and MAPI commands map LPIs to ITTEs.
706 * Must be called with its_lock mutex held.
708 static int vgic_its_cmd_handle_mapi(struct kvm *kvm, struct vgic_its *its,
711 u32 device_id = its_cmd_get_deviceid(its_cmd);
712 u32 event_id = its_cmd_get_id(its_cmd);
713 u32 coll_id = its_cmd_get_collection(its_cmd);
714 struct its_itte *itte;
715 struct its_device *device;
716 struct its_collection *collection, *new_coll = NULL;
718 struct vgic_irq *irq;
720 device = find_its_device(its, device_id);
722 return E_ITS_MAPTI_UNMAPPED_DEVICE;
724 if (its_cmd_get_command(its_cmd) == GITS_CMD_MAPTI)
725 lpi_nr = its_cmd_get_physical_id(its_cmd);
728 if (lpi_nr < GIC_LPI_OFFSET ||
729 lpi_nr >= max_lpis_propbaser(kvm->arch.vgic.propbaser))
730 return E_ITS_MAPTI_PHYSICALID_OOR;
732 /* If there is an existing mapping, behavior is UNPREDICTABLE. */
733 if (find_itte(its, device_id, event_id))
736 collection = find_collection(its, coll_id);
738 int ret = vgic_its_alloc_collection(its, &collection, coll_id);
741 new_coll = collection;
744 itte = kzalloc(sizeof(struct its_itte), GFP_KERNEL);
747 vgic_its_free_collection(its, coll_id);
751 itte->event_id = event_id;
752 list_add_tail(&itte->itte_list, &device->itt_head);
754 itte->collection = collection;
757 irq = vgic_add_lpi(kvm, lpi_nr);
760 vgic_its_free_collection(its, coll_id);
761 its_free_itte(kvm, itte);
766 update_affinity_itte(kvm, itte);
769 * We "cache" the configuration table entries in out struct vgic_irq's.
770 * However we only have those structs for mapped IRQs, so we read in
771 * the respective config data from memory here upon mapping the LPI.
773 update_lpi_config(kvm, itte->irq, NULL);
778 /* Requires the its_lock to be held. */
779 static void vgic_its_unmap_device(struct kvm *kvm, struct its_device *device)
781 struct its_itte *itte, *temp;
784 * The spec says that unmapping a device with still valid
785 * ITTEs associated is UNPREDICTABLE. We remove all ITTEs,
786 * since we cannot leave the memory unreferenced.
788 list_for_each_entry_safe(itte, temp, &device->itt_head, itte_list)
789 its_free_itte(kvm, itte);
791 list_del(&device->dev_list);
796 * MAPD maps or unmaps a device ID to Interrupt Translation Tables (ITTs).
797 * Must be called with the its_lock mutex held.
799 static int vgic_its_cmd_handle_mapd(struct kvm *kvm, struct vgic_its *its,
802 u32 device_id = its_cmd_get_deviceid(its_cmd);
803 bool valid = its_cmd_get_validbit(its_cmd);
804 struct its_device *device;
806 if (!vgic_its_check_id(its, its->baser_device_table, device_id))
807 return E_ITS_MAPD_DEVICE_OOR;
809 device = find_its_device(its, device_id);
812 * The spec says that calling MAPD on an already mapped device
813 * invalidates all cached data for this device. We implement this
814 * by removing the mapping and re-establishing it.
817 vgic_its_unmap_device(kvm, device);
820 * The spec does not say whether unmapping a not-mapped device
821 * is an error, so we are done in any case.
826 device = kzalloc(sizeof(struct its_device), GFP_KERNEL);
830 device->device_id = device_id;
831 INIT_LIST_HEAD(&device->itt_head);
833 list_add_tail(&device->dev_list, &its->device_list);
839 * The MAPC command maps collection IDs to redistributors.
840 * Must be called with the its_lock mutex held.
842 static int vgic_its_cmd_handle_mapc(struct kvm *kvm, struct vgic_its *its,
847 struct its_collection *collection;
850 valid = its_cmd_get_validbit(its_cmd);
851 coll_id = its_cmd_get_collection(its_cmd);
852 target_addr = its_cmd_get_target_addr(its_cmd);
854 if (target_addr >= atomic_read(&kvm->online_vcpus))
855 return E_ITS_MAPC_PROCNUM_OOR;
858 vgic_its_free_collection(its, coll_id);
860 collection = find_collection(its, coll_id);
865 ret = vgic_its_alloc_collection(its, &collection,
869 collection->target_addr = target_addr;
871 collection->target_addr = target_addr;
872 update_affinity_collection(kvm, its, collection);
880 * The CLEAR command removes the pending state for a particular LPI.
881 * Must be called with the its_lock mutex held.
883 static int vgic_its_cmd_handle_clear(struct kvm *kvm, struct vgic_its *its,
886 u32 device_id = its_cmd_get_deviceid(its_cmd);
887 u32 event_id = its_cmd_get_id(its_cmd);
888 struct its_itte *itte;
891 itte = find_itte(its, device_id, event_id);
893 return E_ITS_CLEAR_UNMAPPED_INTERRUPT;
895 itte->irq->pending = false;
901 * The INV command syncs the configuration bits from the memory table.
902 * Must be called with the its_lock mutex held.
904 static int vgic_its_cmd_handle_inv(struct kvm *kvm, struct vgic_its *its,
907 u32 device_id = its_cmd_get_deviceid(its_cmd);
908 u32 event_id = its_cmd_get_id(its_cmd);
909 struct its_itte *itte;
912 itte = find_itte(its, device_id, event_id);
914 return E_ITS_INV_UNMAPPED_INTERRUPT;
916 return update_lpi_config(kvm, itte->irq, NULL);
920 * The INVALL command requests flushing of all IRQ data in this collection.
921 * Find the VCPU mapped to that collection, then iterate over the VM's list
922 * of mapped LPIs and update the configuration for each IRQ which targets
923 * the specified vcpu. The configuration will be read from the in-memory
924 * configuration table.
925 * Must be called with the its_lock mutex held.
927 static int vgic_its_cmd_handle_invall(struct kvm *kvm, struct vgic_its *its,
930 u32 coll_id = its_cmd_get_collection(its_cmd);
931 struct its_collection *collection;
932 struct kvm_vcpu *vcpu;
933 struct vgic_irq *irq;
937 collection = find_collection(its, coll_id);
938 if (!its_is_collection_mapped(collection))
939 return E_ITS_INVALL_UNMAPPED_COLLECTION;
941 vcpu = kvm_get_vcpu(kvm, collection->target_addr);
943 irq_count = vgic_copy_lpi_list(kvm, &intids);
947 for (i = 0; i < irq_count; i++) {
948 irq = vgic_get_irq(kvm, NULL, intids[i]);
951 update_lpi_config(kvm, irq, vcpu);
952 vgic_put_irq(kvm, irq);
961 * The MOVALL command moves the pending state of all IRQs targeting one
962 * redistributor to another. We don't hold the pending state in the VCPUs,
963 * but in the IRQs instead, so there is really not much to do for us here.
964 * However the spec says that no IRQ must target the old redistributor
965 * afterwards, so we make sure that no LPI is using the associated target_vcpu.
966 * This command affects all LPIs in the system that target that redistributor.
968 static int vgic_its_cmd_handle_movall(struct kvm *kvm, struct vgic_its *its,
971 struct vgic_dist *dist = &kvm->arch.vgic;
972 u32 target1_addr = its_cmd_get_target_addr(its_cmd);
973 u32 target2_addr = its_cmd_mask_field(its_cmd, 3, 16, 32);
974 struct kvm_vcpu *vcpu1, *vcpu2;
975 struct vgic_irq *irq;
977 if (target1_addr >= atomic_read(&kvm->online_vcpus) ||
978 target2_addr >= atomic_read(&kvm->online_vcpus))
979 return E_ITS_MOVALL_PROCNUM_OOR;
981 if (target1_addr == target2_addr)
984 vcpu1 = kvm_get_vcpu(kvm, target1_addr);
985 vcpu2 = kvm_get_vcpu(kvm, target2_addr);
987 spin_lock(&dist->lpi_list_lock);
989 list_for_each_entry(irq, &dist->lpi_list_head, lpi_list) {
990 spin_lock(&irq->irq_lock);
992 if (irq->target_vcpu == vcpu1)
993 irq->target_vcpu = vcpu2;
995 spin_unlock(&irq->irq_lock);
998 spin_unlock(&dist->lpi_list_lock);
1004 * The INT command injects the LPI associated with that DevID/EvID pair.
1005 * Must be called with the its_lock mutex held.
1007 static int vgic_its_cmd_handle_int(struct kvm *kvm, struct vgic_its *its,
1010 u32 msi_data = its_cmd_get_id(its_cmd);
1011 u64 msi_devid = its_cmd_get_deviceid(its_cmd);
1013 return vgic_its_trigger_msi(kvm, its, msi_devid, msi_data);
1017 * This function is called with the its_cmd lock held, but the ITS data
1018 * structure lock dropped.
1020 static int vgic_its_handle_command(struct kvm *kvm, struct vgic_its *its,
1025 mutex_lock(&its->its_lock);
1026 switch (its_cmd_get_command(its_cmd)) {
1028 ret = vgic_its_cmd_handle_mapd(kvm, its, its_cmd);
1031 ret = vgic_its_cmd_handle_mapc(kvm, its, its_cmd);
1034 ret = vgic_its_cmd_handle_mapi(kvm, its, its_cmd);
1036 case GITS_CMD_MAPTI:
1037 ret = vgic_its_cmd_handle_mapi(kvm, its, its_cmd);
1040 ret = vgic_its_cmd_handle_movi(kvm, its, its_cmd);
1042 case GITS_CMD_DISCARD:
1043 ret = vgic_its_cmd_handle_discard(kvm, its, its_cmd);
1045 case GITS_CMD_CLEAR:
1046 ret = vgic_its_cmd_handle_clear(kvm, its, its_cmd);
1048 case GITS_CMD_MOVALL:
1049 ret = vgic_its_cmd_handle_movall(kvm, its, its_cmd);
1052 ret = vgic_its_cmd_handle_int(kvm, its, its_cmd);
1055 ret = vgic_its_cmd_handle_inv(kvm, its, its_cmd);
1057 case GITS_CMD_INVALL:
1058 ret = vgic_its_cmd_handle_invall(kvm, its, its_cmd);
1061 /* we ignore this command: we are in sync all of the time */
1065 mutex_unlock(&its->its_lock);
1070 static u64 vgic_sanitise_its_baser(u64 reg)
1072 reg = vgic_sanitise_field(reg, GITS_BASER_SHAREABILITY_MASK,
1073 GITS_BASER_SHAREABILITY_SHIFT,
1074 vgic_sanitise_shareability);
1075 reg = vgic_sanitise_field(reg, GITS_BASER_INNER_CACHEABILITY_MASK,
1076 GITS_BASER_INNER_CACHEABILITY_SHIFT,
1077 vgic_sanitise_inner_cacheability);
1078 reg = vgic_sanitise_field(reg, GITS_BASER_OUTER_CACHEABILITY_MASK,
1079 GITS_BASER_OUTER_CACHEABILITY_SHIFT,
1080 vgic_sanitise_outer_cacheability);
1082 /* Bits 15:12 contain bits 51:48 of the PA, which we don't support. */
1083 reg &= ~GENMASK_ULL(15, 12);
1085 /* We support only one (ITS) page size: 64K */
1086 reg = (reg & ~GITS_BASER_PAGE_SIZE_MASK) | GITS_BASER_PAGE_SIZE_64K;
1091 static u64 vgic_sanitise_its_cbaser(u64 reg)
1093 reg = vgic_sanitise_field(reg, GITS_CBASER_SHAREABILITY_MASK,
1094 GITS_CBASER_SHAREABILITY_SHIFT,
1095 vgic_sanitise_shareability);
1096 reg = vgic_sanitise_field(reg, GITS_CBASER_INNER_CACHEABILITY_MASK,
1097 GITS_CBASER_INNER_CACHEABILITY_SHIFT,
1098 vgic_sanitise_inner_cacheability);
1099 reg = vgic_sanitise_field(reg, GITS_CBASER_OUTER_CACHEABILITY_MASK,
1100 GITS_CBASER_OUTER_CACHEABILITY_SHIFT,
1101 vgic_sanitise_outer_cacheability);
1104 * Sanitise the physical address to be 64k aligned.
1105 * Also limit the physical addresses to 48 bits.
1107 reg &= ~(GENMASK_ULL(51, 48) | GENMASK_ULL(15, 12));
1112 static unsigned long vgic_mmio_read_its_cbaser(struct kvm *kvm,
1113 struct vgic_its *its,
1114 gpa_t addr, unsigned int len)
1116 return extract_bytes(its->cbaser, addr & 7, len);
1119 static void vgic_mmio_write_its_cbaser(struct kvm *kvm, struct vgic_its *its,
1120 gpa_t addr, unsigned int len,
1123 /* When GITS_CTLR.Enable is 1, this register is RO. */
1127 mutex_lock(&its->cmd_lock);
1128 its->cbaser = update_64bit_reg(its->cbaser, addr & 7, len, val);
1129 its->cbaser = vgic_sanitise_its_cbaser(its->cbaser);
1132 * CWRITER is architecturally UNKNOWN on reset, but we need to reset
1133 * it to CREADR to make sure we start with an empty command buffer.
1135 its->cwriter = its->creadr;
1136 mutex_unlock(&its->cmd_lock);
1139 #define ITS_CMD_BUFFER_SIZE(baser) ((((baser) & 0xff) + 1) << 12)
1140 #define ITS_CMD_SIZE 32
1141 #define ITS_CMD_OFFSET(reg) ((reg) & GENMASK(19, 5))
1143 /* Must be called with the cmd_lock held. */
1144 static void vgic_its_process_commands(struct kvm *kvm, struct vgic_its *its)
1149 /* Commands are only processed when the ITS is enabled. */
1153 cbaser = CBASER_ADDRESS(its->cbaser);
1155 while (its->cwriter != its->creadr) {
1156 int ret = kvm_read_guest_lock(kvm, cbaser + its->creadr,
1157 cmd_buf, ITS_CMD_SIZE);
1159 * If kvm_read_guest() fails, this could be due to the guest
1160 * programming a bogus value in CBASER or something else going
1161 * wrong from which we cannot easily recover.
1162 * According to section 6.3.2 in the GICv3 spec we can just
1163 * ignore that command then.
1166 vgic_its_handle_command(kvm, its, cmd_buf);
1168 its->creadr += ITS_CMD_SIZE;
1169 if (its->creadr == ITS_CMD_BUFFER_SIZE(its->cbaser))
1175 * By writing to CWRITER the guest announces new commands to be processed.
1176 * To avoid any races in the first place, we take the its_cmd lock, which
1177 * protects our ring buffer variables, so that there is only one user
1178 * per ITS handling commands at a given time.
1180 static void vgic_mmio_write_its_cwriter(struct kvm *kvm, struct vgic_its *its,
1181 gpa_t addr, unsigned int len,
1189 mutex_lock(&its->cmd_lock);
1191 reg = update_64bit_reg(its->cwriter, addr & 7, len, val);
1192 reg = ITS_CMD_OFFSET(reg);
1193 if (reg >= ITS_CMD_BUFFER_SIZE(its->cbaser)) {
1194 mutex_unlock(&its->cmd_lock);
1199 vgic_its_process_commands(kvm, its);
1201 mutex_unlock(&its->cmd_lock);
1204 static unsigned long vgic_mmio_read_its_cwriter(struct kvm *kvm,
1205 struct vgic_its *its,
1206 gpa_t addr, unsigned int len)
1208 return extract_bytes(its->cwriter, addr & 0x7, len);
1211 static unsigned long vgic_mmio_read_its_creadr(struct kvm *kvm,
1212 struct vgic_its *its,
1213 gpa_t addr, unsigned int len)
1215 return extract_bytes(its->creadr, addr & 0x7, len);
1218 #define BASER_INDEX(addr) (((addr) / sizeof(u64)) & 0x7)
1219 static unsigned long vgic_mmio_read_its_baser(struct kvm *kvm,
1220 struct vgic_its *its,
1221 gpa_t addr, unsigned int len)
1225 switch (BASER_INDEX(addr)) {
1227 reg = its->baser_device_table;
1230 reg = its->baser_coll_table;
1237 return extract_bytes(reg, addr & 7, len);
1240 #define GITS_BASER_RO_MASK (GENMASK_ULL(52, 48) | GENMASK_ULL(58, 56))
1241 static void vgic_mmio_write_its_baser(struct kvm *kvm,
1242 struct vgic_its *its,
1243 gpa_t addr, unsigned int len,
1246 u64 entry_size, device_type;
1247 u64 reg, *regptr, clearbits = 0;
1249 /* When GITS_CTLR.Enable is 1, we ignore write accesses. */
1253 switch (BASER_INDEX(addr)) {
1255 regptr = &its->baser_device_table;
1257 device_type = GITS_BASER_TYPE_DEVICE;
1260 regptr = &its->baser_coll_table;
1262 device_type = GITS_BASER_TYPE_COLLECTION;
1263 clearbits = GITS_BASER_INDIRECT;
1269 reg = update_64bit_reg(*regptr, addr & 7, len, val);
1270 reg &= ~GITS_BASER_RO_MASK;
1273 reg |= (entry_size - 1) << GITS_BASER_ENTRY_SIZE_SHIFT;
1274 reg |= device_type << GITS_BASER_TYPE_SHIFT;
1275 reg = vgic_sanitise_its_baser(reg);
1280 static unsigned long vgic_mmio_read_its_ctlr(struct kvm *vcpu,
1281 struct vgic_its *its,
1282 gpa_t addr, unsigned int len)
1286 mutex_lock(&its->cmd_lock);
1287 if (its->creadr == its->cwriter)
1288 reg |= GITS_CTLR_QUIESCENT;
1290 reg |= GITS_CTLR_ENABLE;
1291 mutex_unlock(&its->cmd_lock);
1296 static void vgic_mmio_write_its_ctlr(struct kvm *kvm, struct vgic_its *its,
1297 gpa_t addr, unsigned int len,
1300 mutex_lock(&its->cmd_lock);
1302 its->enabled = !!(val & GITS_CTLR_ENABLE);
1305 * Try to process any pending commands. This function bails out early
1306 * if the ITS is disabled or no commands have been queued.
1308 vgic_its_process_commands(kvm, its);
1310 mutex_unlock(&its->cmd_lock);
1313 #define REGISTER_ITS_DESC(off, rd, wr, length, acc) \
1315 .reg_offset = off, \
1317 .access_flags = acc, \
1322 static void its_mmio_write_wi(struct kvm *kvm, struct vgic_its *its,
1323 gpa_t addr, unsigned int len, unsigned long val)
1328 static struct vgic_register_region its_registers[] = {
1329 REGISTER_ITS_DESC(GITS_CTLR,
1330 vgic_mmio_read_its_ctlr, vgic_mmio_write_its_ctlr, 4,
1332 REGISTER_ITS_DESC(GITS_IIDR,
1333 vgic_mmio_read_its_iidr, its_mmio_write_wi, 4,
1335 REGISTER_ITS_DESC(GITS_TYPER,
1336 vgic_mmio_read_its_typer, its_mmio_write_wi, 8,
1337 VGIC_ACCESS_64bit | VGIC_ACCESS_32bit),
1338 REGISTER_ITS_DESC(GITS_CBASER,
1339 vgic_mmio_read_its_cbaser, vgic_mmio_write_its_cbaser, 8,
1340 VGIC_ACCESS_64bit | VGIC_ACCESS_32bit),
1341 REGISTER_ITS_DESC(GITS_CWRITER,
1342 vgic_mmio_read_its_cwriter, vgic_mmio_write_its_cwriter, 8,
1343 VGIC_ACCESS_64bit | VGIC_ACCESS_32bit),
1344 REGISTER_ITS_DESC(GITS_CREADR,
1345 vgic_mmio_read_its_creadr, its_mmio_write_wi, 8,
1346 VGIC_ACCESS_64bit | VGIC_ACCESS_32bit),
1347 REGISTER_ITS_DESC(GITS_BASER,
1348 vgic_mmio_read_its_baser, vgic_mmio_write_its_baser, 0x40,
1349 VGIC_ACCESS_64bit | VGIC_ACCESS_32bit),
1350 REGISTER_ITS_DESC(GITS_IDREGS_BASE,
1351 vgic_mmio_read_its_idregs, its_mmio_write_wi, 0x30,
1355 /* This is called on setting the LPI enable bit in the redistributor. */
1356 void vgic_enable_lpis(struct kvm_vcpu *vcpu)
1358 if (!(vcpu->arch.vgic_cpu.pendbaser & GICR_PENDBASER_PTZ))
1359 its_sync_lpi_pending_table(vcpu);
1362 static int vgic_register_its_iodev(struct kvm *kvm, struct vgic_its *its)
1364 struct vgic_io_device *iodev = &its->iodev;
1367 if (!its->initialized)
1370 if (IS_VGIC_ADDR_UNDEF(its->vgic_its_base))
1373 iodev->regions = its_registers;
1374 iodev->nr_regions = ARRAY_SIZE(its_registers);
1375 kvm_iodevice_init(&iodev->dev, &kvm_io_gic_ops);
1377 iodev->base_addr = its->vgic_its_base;
1378 iodev->iodev_type = IODEV_ITS;
1380 mutex_lock(&kvm->slots_lock);
1381 ret = kvm_io_bus_register_dev(kvm, KVM_MMIO_BUS, iodev->base_addr,
1382 KVM_VGIC_V3_ITS_SIZE, &iodev->dev);
1383 mutex_unlock(&kvm->slots_lock);
1388 #define INITIAL_BASER_VALUE \
1389 (GIC_BASER_CACHEABILITY(GITS_BASER, INNER, RaWb) | \
1390 GIC_BASER_CACHEABILITY(GITS_BASER, OUTER, SameAsInner) | \
1391 GIC_BASER_SHAREABILITY(GITS_BASER, InnerShareable) | \
1392 ((8ULL - 1) << GITS_BASER_ENTRY_SIZE_SHIFT) | \
1393 GITS_BASER_PAGE_SIZE_64K)
1395 #define INITIAL_PROPBASER_VALUE \
1396 (GIC_BASER_CACHEABILITY(GICR_PROPBASER, INNER, RaWb) | \
1397 GIC_BASER_CACHEABILITY(GICR_PROPBASER, OUTER, SameAsInner) | \
1398 GIC_BASER_SHAREABILITY(GICR_PROPBASER, InnerShareable))
1400 static int vgic_its_create(struct kvm_device *dev, u32 type)
1402 struct vgic_its *its;
1404 if (type != KVM_DEV_TYPE_ARM_VGIC_ITS)
1407 its = kzalloc(sizeof(struct vgic_its), GFP_KERNEL);
1411 mutex_init(&its->its_lock);
1412 mutex_init(&its->cmd_lock);
1414 its->vgic_its_base = VGIC_ADDR_UNDEF;
1416 INIT_LIST_HEAD(&its->device_list);
1417 INIT_LIST_HEAD(&its->collection_list);
1419 dev->kvm->arch.vgic.has_its = true;
1420 its->initialized = false;
1421 its->enabled = false;
1424 its->baser_device_table = INITIAL_BASER_VALUE |
1425 ((u64)GITS_BASER_TYPE_DEVICE << GITS_BASER_TYPE_SHIFT);
1426 its->baser_coll_table = INITIAL_BASER_VALUE |
1427 ((u64)GITS_BASER_TYPE_COLLECTION << GITS_BASER_TYPE_SHIFT);
1428 dev->kvm->arch.vgic.propbaser = INITIAL_PROPBASER_VALUE;
1435 static void vgic_its_destroy(struct kvm_device *kvm_dev)
1437 struct kvm *kvm = kvm_dev->kvm;
1438 struct vgic_its *its = kvm_dev->private;
1439 struct its_device *dev;
1440 struct its_itte *itte;
1441 struct list_head *dev_cur, *dev_temp;
1442 struct list_head *cur, *temp;
1445 * We may end up here without the lists ever having been initialized.
1446 * Check this and bail out early to avoid dereferencing a NULL pointer.
1448 if (!its->device_list.next)
1451 mutex_lock(&its->its_lock);
1452 list_for_each_safe(dev_cur, dev_temp, &its->device_list) {
1453 dev = container_of(dev_cur, struct its_device, dev_list);
1454 list_for_each_safe(cur, temp, &dev->itt_head) {
1455 itte = (container_of(cur, struct its_itte, itte_list));
1456 its_free_itte(kvm, itte);
1462 list_for_each_safe(cur, temp, &its->collection_list) {
1464 kfree(container_of(cur, struct its_collection, coll_list));
1466 mutex_unlock(&its->its_lock);
1469 kfree(kvm_dev);/* alloc by kvm_ioctl_create_device, free by .destroy */
1472 static int vgic_its_has_attr(struct kvm_device *dev,
1473 struct kvm_device_attr *attr)
1475 switch (attr->group) {
1476 case KVM_DEV_ARM_VGIC_GRP_ADDR:
1477 switch (attr->attr) {
1478 case KVM_VGIC_ITS_ADDR_TYPE:
1482 case KVM_DEV_ARM_VGIC_GRP_CTRL:
1483 switch (attr->attr) {
1484 case KVM_DEV_ARM_VGIC_CTRL_INIT:
1492 static int vgic_its_set_attr(struct kvm_device *dev,
1493 struct kvm_device_attr *attr)
1495 struct vgic_its *its = dev->private;
1498 switch (attr->group) {
1499 case KVM_DEV_ARM_VGIC_GRP_ADDR: {
1500 u64 __user *uaddr = (u64 __user *)(long)attr->addr;
1501 unsigned long type = (unsigned long)attr->attr;
1504 if (type != KVM_VGIC_ITS_ADDR_TYPE)
1507 if (copy_from_user(&addr, uaddr, sizeof(addr)))
1510 ret = vgic_check_ioaddr(dev->kvm, &its->vgic_its_base,
1515 its->vgic_its_base = addr;
1519 case KVM_DEV_ARM_VGIC_GRP_CTRL:
1520 switch (attr->attr) {
1521 case KVM_DEV_ARM_VGIC_CTRL_INIT:
1522 its->initialized = true;
1531 static int vgic_its_get_attr(struct kvm_device *dev,
1532 struct kvm_device_attr *attr)
1534 switch (attr->group) {
1535 case KVM_DEV_ARM_VGIC_GRP_ADDR: {
1536 struct vgic_its *its = dev->private;
1537 u64 addr = its->vgic_its_base;
1538 u64 __user *uaddr = (u64 __user *)(long)attr->addr;
1539 unsigned long type = (unsigned long)attr->attr;
1541 if (type != KVM_VGIC_ITS_ADDR_TYPE)
1544 if (copy_to_user(uaddr, &addr, sizeof(addr)))
1555 static struct kvm_device_ops kvm_arm_vgic_its_ops = {
1556 .name = "kvm-arm-vgic-its",
1557 .create = vgic_its_create,
1558 .destroy = vgic_its_destroy,
1559 .set_attr = vgic_its_set_attr,
1560 .get_attr = vgic_its_get_attr,
1561 .has_attr = vgic_its_has_attr,
1564 int kvm_vgic_register_its_device(void)
1566 return kvm_register_device_ops(&kvm_arm_vgic_its_ops,
1567 KVM_DEV_TYPE_ARM_VGIC_ITS);
1571 * Registers all ITSes with the kvm_io_bus framework.
1572 * To follow the existing VGIC initialization sequence, this has to be
1573 * done as late as possible, just before the first VCPU runs.
1575 int vgic_register_its_iodevs(struct kvm *kvm)
1577 struct kvm_device *dev;
1580 list_for_each_entry(dev, &kvm->devices, vm_node) {
1581 if (dev->ops != &kvm_arm_vgic_its_ops)
1584 ret = vgic_register_its_iodev(kvm, dev->private);
1588 * We don't need to care about tearing down previously
1589 * registered ITSes, as the kvm_io_bus framework removes
1590 * them for us if the VM gets destroyed.