]> git.kernelconcepts.de Git - karo-tx-linux.git/blobdiff - virt/kvm/arm/vgic/vgic-its.c
Merge tag 'mac80211-for-davem-2017-05-23' of git://git.kernel.org/pub/scm/linux/kerne...
[karo-tx-linux.git] / virt / kvm / arm / vgic / vgic-its.c
index 8d1da1af4b09e47c174cf7151b37b98666f03953..2dff288b3a668e401d924e52cee10b9e85ecd8df 100644 (file)
@@ -23,6 +23,7 @@
 #include <linux/interrupt.h>
 #include <linux/list.h>
 #include <linux/uaccess.h>
+#include <linux/list_sort.h>
 
 #include <linux/irqchip/arm-gic-v3.h>
 
 #include "vgic.h"
 #include "vgic-mmio.h"
 
+static int vgic_its_save_tables_v0(struct vgic_its *its);
+static int vgic_its_restore_tables_v0(struct vgic_its *its);
+static int vgic_its_commit_v0(struct vgic_its *its);
+static int update_lpi_config(struct kvm *kvm, struct vgic_irq *irq,
+                            struct kvm_vcpu *filter_vcpu);
+
 /*
  * Creates a new (reference to a) struct vgic_irq for a given LPI.
  * If this LPI is already mapped on another ITS, we increase its refcount
  * If this is a "new" LPI, we allocate and initialize a new struct vgic_irq.
  * This function returns a pointer to the _unlocked_ structure.
  */
-static struct vgic_irq *vgic_add_lpi(struct kvm *kvm, u32 intid)
+static struct vgic_irq *vgic_add_lpi(struct kvm *kvm, u32 intid,
+                                    struct kvm_vcpu *vcpu)
 {
        struct vgic_dist *dist = &kvm->arch.vgic;
        struct vgic_irq *irq = vgic_get_irq(kvm, NULL, intid), *oldirq;
+       int ret;
 
        /* In this case there is no put, since we keep the reference. */
        if (irq)
@@ -60,6 +69,7 @@ static struct vgic_irq *vgic_add_lpi(struct kvm *kvm, u32 intid)
        irq->config = VGIC_CONFIG_EDGE;
        kref_init(&irq->refcount);
        irq->intid = intid;
+       irq->target_vcpu = vcpu;
 
        spin_lock(&dist->lpi_list_lock);
 
@@ -91,6 +101,19 @@ static struct vgic_irq *vgic_add_lpi(struct kvm *kvm, u32 intid)
 out_unlock:
        spin_unlock(&dist->lpi_list_lock);
 
+       /*
+        * We "cache" the configuration table entries in our struct vgic_irq's.
+        * However we only have those structs for mapped IRQs, so we read in
+        * the respective config data from memory here upon mapping the LPI.
+        */
+       ret = update_lpi_config(kvm, irq, NULL);
+       if (ret)
+               return ERR_PTR(ret);
+
+       ret = vgic_v3_lpi_sync_pending_status(kvm, irq);
+       if (ret)
+               return ERR_PTR(ret);
+
        return irq;
 }
 
@@ -99,6 +122,8 @@ struct its_device {
 
        /* the head for the list of ITTEs */
        struct list_head itt_head;
+       u32 num_eventid_bits;
+       gpa_t itt_addr;
        u32 device_id;
 };
 
@@ -114,8 +139,8 @@ struct its_collection {
 #define its_is_collection_mapped(coll) ((coll) && \
                                ((coll)->target_addr != COLLECTION_NOT_MAPPED))
 
-struct its_itte {
-       struct list_head itte_list;
+struct its_ite {
+       struct list_head ite_list;
 
        struct vgic_irq *irq;
        struct its_collection *collection;
@@ -123,6 +148,50 @@ struct its_itte {
        u32 event_id;
 };
 
+/**
+ * struct vgic_its_abi - ITS abi ops and settings
+ * @cte_esz: collection table entry size
+ * @dte_esz: device table entry size
+ * @ite_esz: interrupt translation table entry size
+ * @save tables: save the ITS tables into guest RAM
+ * @restore_tables: restore the ITS internal structs from tables
+ *  stored in guest RAM
+ * @commit: initialize the registers which expose the ABI settings,
+ *  especially the entry sizes
+ */
+struct vgic_its_abi {
+       int cte_esz;
+       int dte_esz;
+       int ite_esz;
+       int (*save_tables)(struct vgic_its *its);
+       int (*restore_tables)(struct vgic_its *its);
+       int (*commit)(struct vgic_its *its);
+};
+
+static const struct vgic_its_abi its_table_abi_versions[] = {
+       [0] = {.cte_esz = 8, .dte_esz = 8, .ite_esz = 8,
+        .save_tables = vgic_its_save_tables_v0,
+        .restore_tables = vgic_its_restore_tables_v0,
+        .commit = vgic_its_commit_v0,
+       },
+};
+
+#define NR_ITS_ABIS    ARRAY_SIZE(its_table_abi_versions)
+
+inline const struct vgic_its_abi *vgic_its_get_abi(struct vgic_its *its)
+{
+       return &its_table_abi_versions[its->abi_rev];
+}
+
+int vgic_its_set_abi(struct vgic_its *its, int rev)
+{
+       const struct vgic_its_abi *abi;
+
+       its->abi_rev = rev;
+       abi = vgic_its_get_abi(its);
+       return abi->commit(its);
+}
+
 /*
  * Find and returns a device in the device table for an ITS.
  * Must be called with the its_lock mutex held.
@@ -143,27 +212,27 @@ static struct its_device *find_its_device(struct vgic_its *its, u32 device_id)
  * Device ID/Event ID pair on an ITS.
  * Must be called with the its_lock mutex held.
  */
-static struct its_itte *find_itte(struct vgic_its *its, u32 device_id,
+static struct its_ite *find_ite(struct vgic_its *its, u32 device_id,
                                  u32 event_id)
 {
        struct its_device *device;
-       struct its_itte *itte;
+       struct its_ite *ite;
 
        device = find_its_device(its, device_id);
        if (device == NULL)
                return NULL;
 
-       list_for_each_entry(itte, &device->itt_head, itte_list)
-               if (itte->event_id == event_id)
-                       return itte;
+       list_for_each_entry(ite, &device->itt_head, ite_list)
+               if (ite->event_id == event_id)
+                       return ite;
 
        return NULL;
 }
 
 /* To be used as an iterator this macro misses the enclosing parentheses */
-#define for_each_lpi_its(dev, itte, its) \
+#define for_each_lpi_its(dev, ite, its) \
        list_for_each_entry(dev, &(its)->device_list, dev_list) \
-               list_for_each_entry(itte, &(dev)->itt_head, itte_list)
+               list_for_each_entry(ite, &(dev)->itt_head, ite_list)
 
 /*
  * We only implement 48 bits of PA at the moment, although the ITS
@@ -171,11 +240,14 @@ static struct its_itte *find_itte(struct vgic_its *its, u32 device_id,
  */
 #define BASER_ADDRESS(x)       ((x) & GENMASK_ULL(47, 16))
 #define CBASER_ADDRESS(x)      ((x) & GENMASK_ULL(47, 12))
-#define PENDBASER_ADDRESS(x)   ((x) & GENMASK_ULL(47, 16))
-#define PROPBASER_ADDRESS(x)   ((x) & GENMASK_ULL(47, 12))
 
 #define GIC_LPI_OFFSET 8192
 
+#define VITS_TYPER_IDBITS 16
+#define VITS_TYPER_DEVBITS 16
+#define VITS_DTE_MAX_DEVID_OFFSET      (BIT(14) - 1)
+#define VITS_ITE_MAX_EVENTID_OFFSET    (BIT(16) - 1)
+
 /*
  * Finds and returns a collection in the ITS collection table.
  * Must be called with the its_lock mutex held.
@@ -204,7 +276,7 @@ static struct its_collection *find_collection(struct vgic_its *its, int coll_id)
 static int update_lpi_config(struct kvm *kvm, struct vgic_irq *irq,
                             struct kvm_vcpu *filter_vcpu)
 {
-       u64 propbase = PROPBASER_ADDRESS(kvm->arch.vgic.propbaser);
+       u64 propbase = GICR_PROPBASER_ADDRESS(kvm->arch.vgic.propbaser);
        u8 prop;
        int ret;
 
@@ -229,13 +301,13 @@ static int update_lpi_config(struct kvm *kvm, struct vgic_irq *irq,
 }
 
 /*
- * Create a snapshot of the current LPI list, so that we can enumerate all
- * LPIs without holding any lock.
- * Returns the array length and puts the kmalloc'ed array into intid_ptr.
+ * Create a snapshot of the current LPIs targeting @vcpu, so that we can
+ * enumerate those LPIs without holding any lock.
+ * Returns their number and puts the kmalloc'ed array into intid_ptr.
  */
-static int vgic_copy_lpi_list(struct kvm *kvm, u32 **intid_ptr)
+static int vgic_copy_lpi_list(struct kvm_vcpu *vcpu, u32 **intid_ptr)
 {
-       struct vgic_dist *dist = &kvm->arch.vgic;
+       struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
        struct vgic_irq *irq;
        u32 *intids;
        int irq_count = dist->lpi_list_count, i = 0;
@@ -254,14 +326,14 @@ static int vgic_copy_lpi_list(struct kvm *kvm, u32 **intid_ptr)
        spin_lock(&dist->lpi_list_lock);
        list_for_each_entry(irq, &dist->lpi_list_head, lpi_list) {
                /* We don't need to "get" the IRQ, as we hold the list lock. */
-               intids[i] = irq->intid;
-               if (++i == irq_count)
-                       break;
+               if (irq->target_vcpu != vcpu)
+                       continue;
+               intids[i++] = irq->intid;
        }
        spin_unlock(&dist->lpi_list_lock);
 
        *intid_ptr = intids;
-       return irq_count;
+       return i;
 }
 
 /*
@@ -270,18 +342,18 @@ static int vgic_copy_lpi_list(struct kvm *kvm, u32 **intid_ptr)
  * Needs to be called whenever either the collection for a LPIs has
  * changed or the collection itself got retargeted.
  */
-static void update_affinity_itte(struct kvm *kvm, struct its_itte *itte)
+static void update_affinity_ite(struct kvm *kvm, struct its_ite *ite)
 {
        struct kvm_vcpu *vcpu;
 
-       if (!its_is_collection_mapped(itte->collection))
+       if (!its_is_collection_mapped(ite->collection))
                return;
 
-       vcpu = kvm_get_vcpu(kvm, itte->collection->target_addr);
+       vcpu = kvm_get_vcpu(kvm, ite->collection->target_addr);
 
-       spin_lock(&itte->irq->irq_lock);
-       itte->irq->target_vcpu = vcpu;
-       spin_unlock(&itte->irq->irq_lock);
+       spin_lock(&ite->irq->irq_lock);
+       ite->irq->target_vcpu = vcpu;
+       spin_unlock(&ite->irq->irq_lock);
 }
 
 /*
@@ -292,13 +364,13 @@ static void update_affinity_collection(struct kvm *kvm, struct vgic_its *its,
                                       struct its_collection *coll)
 {
        struct its_device *device;
-       struct its_itte *itte;
+       struct its_ite *ite;
 
-       for_each_lpi_its(device, itte, its) {
-               if (!itte->collection || coll != itte->collection)
+       for_each_lpi_its(device, ite, its) {
+               if (!ite->collection || coll != ite->collection)
                        continue;
 
-               update_affinity_itte(kvm, itte);
+               update_affinity_ite(kvm, ite);
        }
 }
 
@@ -310,20 +382,20 @@ static u32 max_lpis_propbaser(u64 propbaser)
 }
 
 /*
- * Scan the whole LPI pending table and sync the pending bit in there
+ * Sync the pending table pending bit of LPIs targeting @vcpu
  * with our own data structures. This relies on the LPI being
  * mapped before.
  */
 static int its_sync_lpi_pending_table(struct kvm_vcpu *vcpu)
 {
-       gpa_t pendbase = PENDBASER_ADDRESS(vcpu->arch.vgic_cpu.pendbaser);
+       gpa_t pendbase = GICR_PENDBASER_ADDRESS(vcpu->arch.vgic_cpu.pendbaser);
        struct vgic_irq *irq;
        int last_byte_offset = -1;
        int ret = 0;
        u32 *intids;
        int nr_irqs, i;
 
-       nr_irqs = vgic_copy_lpi_list(vcpu->kvm, &intids);
+       nr_irqs = vgic_copy_lpi_list(vcpu, &intids);
        if (nr_irqs < 0)
                return nr_irqs;
 
@@ -364,6 +436,7 @@ static unsigned long vgic_mmio_read_its_typer(struct kvm *kvm,
                                              struct vgic_its *its,
                                              gpa_t addr, unsigned int len)
 {
+       const struct vgic_its_abi *abi = vgic_its_get_abi(its);
        u64 reg = GITS_TYPER_PLPIS;
 
        /*
@@ -374,8 +447,9 @@ static unsigned long vgic_mmio_read_its_typer(struct kvm *kvm,
         * To avoid memory waste in the guest, we keep the number of IDBits and
         * DevBits low - as least for the time being.
         */
-       reg |= 0x0f << GITS_TYPER_DEVBITS_SHIFT;
-       reg |= 0x0f << GITS_TYPER_IDBITS_SHIFT;
+       reg |= GIC_ENCODE_SZ(VITS_TYPER_DEVBITS, 5) << GITS_TYPER_DEVBITS_SHIFT;
+       reg |= GIC_ENCODE_SZ(VITS_TYPER_IDBITS, 5) << GITS_TYPER_IDBITS_SHIFT;
+       reg |= GIC_ENCODE_SZ(abi->ite_esz, 4) << GITS_TYPER_ITT_ENTRY_SIZE_SHIFT;
 
        return extract_bytes(reg, addr & 7, len);
 }
@@ -384,7 +458,23 @@ static unsigned long vgic_mmio_read_its_iidr(struct kvm *kvm,
                                             struct vgic_its *its,
                                             gpa_t addr, unsigned int len)
 {
-       return (PRODUCT_ID_KVM << 24) | (IMPLEMENTER_ARM << 0);
+       u32 val;
+
+       val = (its->abi_rev << GITS_IIDR_REV_SHIFT) & GITS_IIDR_REV_MASK;
+       val |= (PRODUCT_ID_KVM << GITS_IIDR_PRODUCTID_SHIFT) | IMPLEMENTER_ARM;
+       return val;
+}
+
+static int vgic_mmio_uaccess_write_its_iidr(struct kvm *kvm,
+                                           struct vgic_its *its,
+                                           gpa_t addr, unsigned int len,
+                                           unsigned long val)
+{
+       u32 rev = GITS_IIDR_REV(val);
+
+       if (rev >= NR_ITS_ABIS)
+               return -EINVAL;
+       return vgic_its_set_abi(its, rev);
 }
 
 static unsigned long vgic_mmio_read_its_idregs(struct kvm *kvm,
@@ -425,25 +515,25 @@ static int vgic_its_trigger_msi(struct kvm *kvm, struct vgic_its *its,
                                u32 devid, u32 eventid)
 {
        struct kvm_vcpu *vcpu;
-       struct its_itte *itte;
+       struct its_ite *ite;
 
        if (!its->enabled)
                return -EBUSY;
 
-       itte = find_itte(its, devid, eventid);
-       if (!itte || !its_is_collection_mapped(itte->collection))
+       ite = find_ite(its, devid, eventid);
+       if (!ite || !its_is_collection_mapped(ite->collection))
                return E_ITS_INT_UNMAPPED_INTERRUPT;
 
-       vcpu = kvm_get_vcpu(kvm, itte->collection->target_addr);
+       vcpu = kvm_get_vcpu(kvm, ite->collection->target_addr);
        if (!vcpu)
                return E_ITS_INT_UNMAPPED_INTERRUPT;
 
        if (!vcpu->arch.vgic_cpu.lpis_enabled)
                return -EBUSY;
 
-       spin_lock(&itte->irq->irq_lock);
-       itte->irq->pending_latch = true;
-       vgic_queue_irq_unlock(kvm, itte->irq);
+       spin_lock(&ite->irq->irq_lock);
+       ite->irq->pending_latch = true;
+       vgic_queue_irq_unlock(kvm, ite->irq);
 
        return 0;
 }
@@ -511,15 +601,15 @@ int vgic_its_inject_msi(struct kvm *kvm, struct kvm_msi *msi)
 }
 
 /* Requires the its_lock to be held. */
-static void its_free_itte(struct kvm *kvm, struct its_itte *itte)
+static void its_free_ite(struct kvm *kvm, struct its_ite *ite)
 {
-       list_del(&itte->itte_list);
+       list_del(&ite->ite_list);
 
        /* This put matches the get in vgic_add_lpi. */
-       if (itte->irq)
-               vgic_put_irq(kvm, itte->irq);
+       if (ite->irq)
+               vgic_put_irq(kvm, ite->irq);
 
-       kfree(itte);
+       kfree(ite);
 }
 
 static u64 its_cmd_mask_field(u64 *its_cmd, int word, int shift, int size)
@@ -529,9 +619,11 @@ static u64 its_cmd_mask_field(u64 *its_cmd, int word, int shift, int size)
 
 #define its_cmd_get_command(cmd)       its_cmd_mask_field(cmd, 0,  0,  8)
 #define its_cmd_get_deviceid(cmd)      its_cmd_mask_field(cmd, 0, 32, 32)
+#define its_cmd_get_size(cmd)          (its_cmd_mask_field(cmd, 1,  0,  5) + 1)
 #define its_cmd_get_id(cmd)            its_cmd_mask_field(cmd, 1,  0, 32)
 #define its_cmd_get_physical_id(cmd)   its_cmd_mask_field(cmd, 1, 32, 32)
 #define its_cmd_get_collection(cmd)    its_cmd_mask_field(cmd, 2,  0, 16)
+#define its_cmd_get_ittaddr(cmd)       (its_cmd_mask_field(cmd, 2,  8, 44) << 8)
 #define its_cmd_get_target_addr(cmd)   its_cmd_mask_field(cmd, 2, 16, 32)
 #define its_cmd_get_validbit(cmd)      its_cmd_mask_field(cmd, 2, 63,  1)
 
@@ -544,17 +636,17 @@ static int vgic_its_cmd_handle_discard(struct kvm *kvm, struct vgic_its *its,
 {
        u32 device_id = its_cmd_get_deviceid(its_cmd);
        u32 event_id = its_cmd_get_id(its_cmd);
-       struct its_itte *itte;
+       struct its_ite *ite;
 
 
-       itte = find_itte(its, device_id, event_id);
-       if (itte && itte->collection) {
+       ite = find_ite(its, device_id, event_id);
+       if (ite && ite->collection) {
                /*
                 * Though the spec talks about removing the pending state, we
                 * don't bother here since we clear the ITTE anyway and the
                 * pending state is a property of the ITTE struct.
                 */
-               its_free_itte(kvm, itte);
+               its_free_ite(kvm, ite);
                return 0;
        }
 
@@ -572,26 +664,26 @@ static int vgic_its_cmd_handle_movi(struct kvm *kvm, struct vgic_its *its,
        u32 event_id = its_cmd_get_id(its_cmd);
        u32 coll_id = its_cmd_get_collection(its_cmd);
        struct kvm_vcpu *vcpu;
-       struct its_itte *itte;
+       struct its_ite *ite;
        struct its_collection *collection;
 
-       itte = find_itte(its, device_id, event_id);
-       if (!itte)
+       ite = find_ite(its, device_id, event_id);
+       if (!ite)
                return E_ITS_MOVI_UNMAPPED_INTERRUPT;
 
-       if (!its_is_collection_mapped(itte->collection))
+       if (!its_is_collection_mapped(ite->collection))
                return E_ITS_MOVI_UNMAPPED_COLLECTION;
 
        collection = find_collection(its, coll_id);
        if (!its_is_collection_mapped(collection))
                return E_ITS_MOVI_UNMAPPED_COLLECTION;
 
-       itte->collection = collection;
+       ite->collection = collection;
        vcpu = kvm_get_vcpu(kvm, collection->target_addr);
 
-       spin_lock(&itte->irq->irq_lock);
-       itte->irq->target_vcpu = vcpu;
-       spin_unlock(&itte->irq->irq_lock);
+       spin_lock(&ite->irq->irq_lock);
+       ite->irq->target_vcpu = vcpu;
+       spin_unlock(&ite->irq->irq_lock);
 
        return 0;
 }
@@ -600,16 +692,31 @@ static int vgic_its_cmd_handle_movi(struct kvm *kvm, struct vgic_its *its,
  * Check whether an ID can be stored into the corresponding guest table.
  * For a direct table this is pretty easy, but gets a bit nasty for
  * indirect tables. We check whether the resulting guest physical address
- * is actually valid (covered by a memslot and guest accessbible).
+ * is actually valid (covered by a memslot and guest accessible).
  * For this we have to read the respective first level entry.
  */
-static bool vgic_its_check_id(struct vgic_its *its, u64 baser, int id)
+static bool vgic_its_check_id(struct vgic_its *its, u64 baser, u32 id,
+                             gpa_t *eaddr)
 {
        int l1_tbl_size = GITS_BASER_NR_PAGES(baser) * SZ_64K;
+       u64 indirect_ptr, type = GITS_BASER_TYPE(baser);
+       int esz = GITS_BASER_ENTRY_SIZE(baser);
        int index;
-       u64 indirect_ptr;
        gfn_t gfn;
-       int esz = GITS_BASER_ENTRY_SIZE(baser);
+
+       switch (type) {
+       case GITS_BASER_TYPE_DEVICE:
+               if (id >= BIT_ULL(VITS_TYPER_DEVBITS))
+                       return false;
+               break;
+       case GITS_BASER_TYPE_COLLECTION:
+               /* as GITS_TYPER.CIL == 0, ITS supports 16-bit collection ID */
+               if (id >= BIT_ULL(16))
+                       return false;
+               break;
+       default:
+               return false;
+       }
 
        if (!(baser & GITS_BASER_INDIRECT)) {
                phys_addr_t addr;
@@ -620,6 +727,8 @@ static bool vgic_its_check_id(struct vgic_its *its, u64 baser, int id)
                addr = BASER_ADDRESS(baser) + id * esz;
                gfn = addr >> PAGE_SHIFT;
 
+               if (eaddr)
+                       *eaddr = addr;
                return kvm_is_visible_gfn(its->dev->kvm, gfn);
        }
 
@@ -652,6 +761,8 @@ static bool vgic_its_check_id(struct vgic_its *its, u64 baser, int id)
        indirect_ptr += index * esz;
        gfn = indirect_ptr >> PAGE_SHIFT;
 
+       if (eaddr)
+               *eaddr = indirect_ptr;
        return kvm_is_visible_gfn(its->dev->kvm, gfn);
 }
 
@@ -661,7 +772,7 @@ static int vgic_its_alloc_collection(struct vgic_its *its,
 {
        struct its_collection *collection;
 
-       if (!vgic_its_check_id(its, its->baser_coll_table, coll_id))
+       if (!vgic_its_check_id(its, its->baser_coll_table, coll_id, NULL))
                return E_ITS_MAPC_COLLECTION_OOR;
 
        collection = kzalloc(sizeof(*collection), GFP_KERNEL);
@@ -679,7 +790,7 @@ static void vgic_its_free_collection(struct vgic_its *its, u32 coll_id)
 {
        struct its_collection *collection;
        struct its_device *device;
-       struct its_itte *itte;
+       struct its_ite *ite;
 
        /*
         * Clearing the mapping for that collection ID removes the
@@ -690,15 +801,34 @@ static void vgic_its_free_collection(struct vgic_its *its, u32 coll_id)
        if (!collection)
                return;
 
-       for_each_lpi_its(device, itte, its)
-               if (itte->collection &&
-                   itte->collection->collection_id == coll_id)
-                       itte->collection = NULL;
+       for_each_lpi_its(device, ite, its)
+               if (ite->collection &&
+                   ite->collection->collection_id == coll_id)
+                       ite->collection = NULL;
 
        list_del(&collection->coll_list);
        kfree(collection);
 }
 
+/* Must be called with its_lock mutex held */
+static struct its_ite *vgic_its_alloc_ite(struct its_device *device,
+                                         struct its_collection *collection,
+                                         u32 lpi_id, u32 event_id)
+{
+       struct its_ite *ite;
+
+       ite = kzalloc(sizeof(*ite), GFP_KERNEL);
+       if (!ite)
+               return ERR_PTR(-ENOMEM);
+
+       ite->event_id   = event_id;
+       ite->collection = collection;
+       ite->lpi = lpi_id;
+
+       list_add_tail(&ite->ite_list, &device->itt_head);
+       return ite;
+}
+
 /*
  * The MAPTI and MAPI commands map LPIs to ITTEs.
  * Must be called with its_lock mutex held.
@@ -709,16 +839,20 @@ static int vgic_its_cmd_handle_mapi(struct kvm *kvm, struct vgic_its *its,
        u32 device_id = its_cmd_get_deviceid(its_cmd);
        u32 event_id = its_cmd_get_id(its_cmd);
        u32 coll_id = its_cmd_get_collection(its_cmd);
-       struct its_itte *itte;
+       struct its_ite *ite;
+       struct kvm_vcpu *vcpu = NULL;
        struct its_device *device;
        struct its_collection *collection, *new_coll = NULL;
-       int lpi_nr;
        struct vgic_irq *irq;
+       int lpi_nr;
 
        device = find_its_device(its, device_id);
        if (!device)
                return E_ITS_MAPTI_UNMAPPED_DEVICE;
 
+       if (event_id >= BIT_ULL(device->num_eventid_bits))
+               return E_ITS_MAPTI_ID_OOR;
+
        if (its_cmd_get_command(its_cmd) == GITS_CMD_MAPTI)
                lpi_nr = its_cmd_get_physical_id(its_cmd);
        else
@@ -728,7 +862,7 @@ static int vgic_its_cmd_handle_mapi(struct kvm *kvm, struct vgic_its *its,
                return E_ITS_MAPTI_PHYSICALID_OOR;
 
        /* If there is an existing mapping, behavior is UNPREDICTABLE. */
-       if (find_itte(its, device_id, event_id))
+       if (find_ite(its, device_id, event_id))
                return 0;
 
        collection = find_collection(its, coll_id);
@@ -739,36 +873,24 @@ static int vgic_its_cmd_handle_mapi(struct kvm *kvm, struct vgic_its *its,
                new_coll = collection;
        }
 
-       itte = kzalloc(sizeof(struct its_itte), GFP_KERNEL);
-       if (!itte) {
+       ite = vgic_its_alloc_ite(device, collection, lpi_nr, event_id);
+       if (IS_ERR(ite)) {
                if (new_coll)
                        vgic_its_free_collection(its, coll_id);
-               return -ENOMEM;
+               return PTR_ERR(ite);
        }
 
-       itte->event_id  = event_id;
-       list_add_tail(&itte->itte_list, &device->itt_head);
-
-       itte->collection = collection;
-       itte->lpi = lpi_nr;
+       if (its_is_collection_mapped(collection))
+               vcpu = kvm_get_vcpu(kvm, collection->target_addr);
 
-       irq = vgic_add_lpi(kvm, lpi_nr);
+       irq = vgic_add_lpi(kvm, lpi_nr, vcpu);
        if (IS_ERR(irq)) {
                if (new_coll)
                        vgic_its_free_collection(its, coll_id);
-               its_free_itte(kvm, itte);
+               its_free_ite(kvm, ite);
                return PTR_ERR(irq);
        }
-       itte->irq = irq;
-
-       update_affinity_itte(kvm, itte);
-
-       /*
-        * We "cache" the configuration table entries in out struct vgic_irq's.
-        * However we only have those structs for mapped IRQs, so we read in
-        * the respective config data from memory here upon mapping the LPI.
-        */
-       update_lpi_config(kvm, itte->irq, NULL);
+       ite->irq = irq;
 
        return 0;
 }
@@ -776,20 +898,40 @@ static int vgic_its_cmd_handle_mapi(struct kvm *kvm, struct vgic_its *its,
 /* Requires the its_lock to be held. */
 static void vgic_its_unmap_device(struct kvm *kvm, struct its_device *device)
 {
-       struct its_itte *itte, *temp;
+       struct its_ite *ite, *temp;
 
        /*
         * The spec says that unmapping a device with still valid
         * ITTEs associated is UNPREDICTABLE. We remove all ITTEs,
         * since we cannot leave the memory unreferenced.
         */
-       list_for_each_entry_safe(itte, temp, &device->itt_head, itte_list)
-               its_free_itte(kvm, itte);
+       list_for_each_entry_safe(ite, temp, &device->itt_head, ite_list)
+               its_free_ite(kvm, ite);
 
        list_del(&device->dev_list);
        kfree(device);
 }
 
+/* Must be called with its_lock mutex held */
+static struct its_device *vgic_its_alloc_device(struct vgic_its *its,
+                                               u32 device_id, gpa_t itt_addr,
+                                               u8 num_eventid_bits)
+{
+       struct its_device *device;
+
+       device = kzalloc(sizeof(*device), GFP_KERNEL);
+       if (!device)
+               return ERR_PTR(-ENOMEM);
+
+       device->device_id = device_id;
+       device->itt_addr = itt_addr;
+       device->num_eventid_bits = num_eventid_bits;
+       INIT_LIST_HEAD(&device->itt_head);
+
+       list_add_tail(&device->dev_list, &its->device_list);
+       return device;
+}
+
 /*
  * MAPD maps or unmaps a device ID to Interrupt Translation Tables (ITTs).
  * Must be called with the its_lock mutex held.
@@ -799,11 +941,16 @@ static int vgic_its_cmd_handle_mapd(struct kvm *kvm, struct vgic_its *its,
 {
        u32 device_id = its_cmd_get_deviceid(its_cmd);
        bool valid = its_cmd_get_validbit(its_cmd);
+       u8 num_eventid_bits = its_cmd_get_size(its_cmd);
+       gpa_t itt_addr = its_cmd_get_ittaddr(its_cmd);
        struct its_device *device;
 
-       if (!vgic_its_check_id(its, its->baser_device_table, device_id))
+       if (!vgic_its_check_id(its, its->baser_device_table, device_id, NULL))
                return E_ITS_MAPD_DEVICE_OOR;
 
+       if (valid && num_eventid_bits > VITS_TYPER_IDBITS)
+               return E_ITS_MAPD_ITTSIZE_OOR;
+
        device = find_its_device(its, device_id);
 
        /*
@@ -821,14 +968,10 @@ static int vgic_its_cmd_handle_mapd(struct kvm *kvm, struct vgic_its *its,
        if (!valid)
                return 0;
 
-       device = kzalloc(sizeof(struct its_device), GFP_KERNEL);
-       if (!device)
-               return -ENOMEM;
-
-       device->device_id = device_id;
-       INIT_LIST_HEAD(&device->itt_head);
-
-       list_add_tail(&device->dev_list, &its->device_list);
+       device = vgic_its_alloc_device(its, device_id, itt_addr,
+                                      num_eventid_bits);
+       if (IS_ERR(device))
+               return PTR_ERR(device);
 
        return 0;
 }
@@ -883,14 +1026,14 @@ static int vgic_its_cmd_handle_clear(struct kvm *kvm, struct vgic_its *its,
 {
        u32 device_id = its_cmd_get_deviceid(its_cmd);
        u32 event_id = its_cmd_get_id(its_cmd);
-       struct its_itte *itte;
+       struct its_ite *ite;
 
 
-       itte = find_itte(its, device_id, event_id);
-       if (!itte)
+       ite = find_ite(its, device_id, event_id);
+       if (!ite)
                return E_ITS_CLEAR_UNMAPPED_INTERRUPT;
 
-       itte->irq->pending_latch = false;
+       ite->irq->pending_latch = false;
 
        return 0;
 }
@@ -904,14 +1047,14 @@ static int vgic_its_cmd_handle_inv(struct kvm *kvm, struct vgic_its *its,
 {
        u32 device_id = its_cmd_get_deviceid(its_cmd);
        u32 event_id = its_cmd_get_id(its_cmd);
-       struct its_itte *itte;
+       struct its_ite *ite;
 
 
-       itte = find_itte(its, device_id, event_id);
-       if (!itte)
+       ite = find_ite(its, device_id, event_id);
+       if (!ite)
                return E_ITS_INV_UNMAPPED_INTERRUPT;
 
-       return update_lpi_config(kvm, itte->irq, NULL);
+       return update_lpi_config(kvm, ite->irq, NULL);
 }
 
 /*
@@ -938,7 +1081,7 @@ static int vgic_its_cmd_handle_invall(struct kvm *kvm, struct vgic_its *its,
 
        vcpu = kvm_get_vcpu(kvm, collection->target_addr);
 
-       irq_count = vgic_copy_lpi_list(kvm, &intids);
+       irq_count = vgic_copy_lpi_list(vcpu, &intids);
        if (irq_count < 0)
                return irq_count;
 
@@ -1213,6 +1356,33 @@ static unsigned long vgic_mmio_read_its_creadr(struct kvm *kvm,
        return extract_bytes(its->creadr, addr & 0x7, len);
 }
 
+static int vgic_mmio_uaccess_write_its_creadr(struct kvm *kvm,
+                                             struct vgic_its *its,
+                                             gpa_t addr, unsigned int len,
+                                             unsigned long val)
+{
+       u32 cmd_offset;
+       int ret = 0;
+
+       mutex_lock(&its->cmd_lock);
+
+       if (its->enabled) {
+               ret = -EBUSY;
+               goto out;
+       }
+
+       cmd_offset = ITS_CMD_OFFSET(val);
+       if (cmd_offset >= ITS_CMD_BUFFER_SIZE(its->cbaser)) {
+               ret = -EINVAL;
+               goto out;
+       }
+
+       its->creadr = cmd_offset;
+out:
+       mutex_unlock(&its->cmd_lock);
+       return ret;
+}
+
 #define BASER_INDEX(addr) (((addr) / sizeof(u64)) & 0x7)
 static unsigned long vgic_mmio_read_its_baser(struct kvm *kvm,
                                              struct vgic_its *its,
@@ -1241,6 +1411,7 @@ static void vgic_mmio_write_its_baser(struct kvm *kvm,
                                      gpa_t addr, unsigned int len,
                                      unsigned long val)
 {
+       const struct vgic_its_abi *abi = vgic_its_get_abi(its);
        u64 entry_size, device_type;
        u64 reg, *regptr, clearbits = 0;
 
@@ -1251,12 +1422,12 @@ static void vgic_mmio_write_its_baser(struct kvm *kvm,
        switch (BASER_INDEX(addr)) {
        case 0:
                regptr = &its->baser_device_table;
-               entry_size = 8;
+               entry_size = abi->dte_esz;
                device_type = GITS_BASER_TYPE_DEVICE;
                break;
        case 1:
                regptr = &its->baser_coll_table;
-               entry_size = 8;
+               entry_size = abi->cte_esz;
                device_type = GITS_BASER_TYPE_COLLECTION;
                clearbits = GITS_BASER_INDIRECT;
                break;
@@ -1317,6 +1488,16 @@ static void vgic_mmio_write_its_ctlr(struct kvm *kvm, struct vgic_its *its,
        .its_write = wr,                                        \
 }
 
+#define REGISTER_ITS_DESC_UACCESS(off, rd, wr, uwr, length, acc)\
+{                                                              \
+       .reg_offset = off,                                      \
+       .len = length,                                          \
+       .access_flags = acc,                                    \
+       .its_read = rd,                                         \
+       .its_write = wr,                                        \
+       .uaccess_its_write = uwr,                               \
+}
+
 static void its_mmio_write_wi(struct kvm *kvm, struct vgic_its *its,
                              gpa_t addr, unsigned int len, unsigned long val)
 {
@@ -1327,8 +1508,9 @@ static struct vgic_register_region its_registers[] = {
        REGISTER_ITS_DESC(GITS_CTLR,
                vgic_mmio_read_its_ctlr, vgic_mmio_write_its_ctlr, 4,
                VGIC_ACCESS_32bit),
-       REGISTER_ITS_DESC(GITS_IIDR,
-               vgic_mmio_read_its_iidr, its_mmio_write_wi, 4,
+       REGISTER_ITS_DESC_UACCESS(GITS_IIDR,
+               vgic_mmio_read_its_iidr, its_mmio_write_wi,
+               vgic_mmio_uaccess_write_its_iidr, 4,
                VGIC_ACCESS_32bit),
        REGISTER_ITS_DESC(GITS_TYPER,
                vgic_mmio_read_its_typer, its_mmio_write_wi, 8,
@@ -1339,8 +1521,9 @@ static struct vgic_register_region its_registers[] = {
        REGISTER_ITS_DESC(GITS_CWRITER,
                vgic_mmio_read_its_cwriter, vgic_mmio_write_its_cwriter, 8,
                VGIC_ACCESS_64bit | VGIC_ACCESS_32bit),
-       REGISTER_ITS_DESC(GITS_CREADR,
-               vgic_mmio_read_its_creadr, its_mmio_write_wi, 8,
+       REGISTER_ITS_DESC_UACCESS(GITS_CREADR,
+               vgic_mmio_read_its_creadr, its_mmio_write_wi,
+               vgic_mmio_uaccess_write_its_creadr, 8,
                VGIC_ACCESS_64bit | VGIC_ACCESS_32bit),
        REGISTER_ITS_DESC(GITS_BASER,
                vgic_mmio_read_its_baser, vgic_mmio_write_its_baser, 0x40,
@@ -1357,17 +1540,19 @@ void vgic_enable_lpis(struct kvm_vcpu *vcpu)
                its_sync_lpi_pending_table(vcpu);
 }
 
-static int vgic_register_its_iodev(struct kvm *kvm, struct vgic_its *its)
+static int vgic_register_its_iodev(struct kvm *kvm, struct vgic_its *its,
+                                  u64 addr)
 {
        struct vgic_io_device *iodev = &its->iodev;
        int ret;
 
-       if (!its->initialized)
-               return -EBUSY;
-
-       if (IS_VGIC_ADDR_UNDEF(its->vgic_its_base))
-               return -ENXIO;
+       mutex_lock(&kvm->slots_lock);
+       if (!IS_VGIC_ADDR_UNDEF(its->vgic_its_base)) {
+               ret = -EBUSY;
+               goto out;
+       }
 
+       its->vgic_its_base = addr;
        iodev->regions = its_registers;
        iodev->nr_regions = ARRAY_SIZE(its_registers);
        kvm_iodevice_init(&iodev->dev, &kvm_io_gic_ops);
@@ -1375,9 +1560,9 @@ static int vgic_register_its_iodev(struct kvm *kvm, struct vgic_its *its)
        iodev->base_addr = its->vgic_its_base;
        iodev->iodev_type = IODEV_ITS;
        iodev->its = its;
-       mutex_lock(&kvm->slots_lock);
        ret = kvm_io_bus_register_dev(kvm, KVM_MMIO_BUS, iodev->base_addr,
                                      KVM_VGIC_V3_ITS_SIZE, &iodev->dev);
+out:
        mutex_unlock(&kvm->slots_lock);
 
        return ret;
@@ -1387,7 +1572,6 @@ static int vgic_register_its_iodev(struct kvm *kvm, struct vgic_its *its)
        (GIC_BASER_CACHEABILITY(GITS_BASER, INNER, RaWb)                | \
         GIC_BASER_CACHEABILITY(GITS_BASER, OUTER, SameAsInner)         | \
         GIC_BASER_SHAREABILITY(GITS_BASER, InnerShareable)             | \
-        ((8ULL - 1) << GITS_BASER_ENTRY_SIZE_SHIFT)                    | \
         GITS_BASER_PAGE_SIZE_64K)
 
 #define INITIAL_PROPBASER_VALUE                                                  \
@@ -1415,7 +1599,6 @@ static int vgic_its_create(struct kvm_device *dev, u32 type)
        INIT_LIST_HEAD(&its->collection_list);
 
        dev->kvm->arch.vgic.has_its = true;
-       its->initialized = false;
        its->enabled = false;
        its->dev = dev;
 
@@ -1427,16 +1610,23 @@ static int vgic_its_create(struct kvm_device *dev, u32 type)
 
        dev->private = its;
 
-       return 0;
+       return vgic_its_set_abi(its, NR_ITS_ABIS - 1);
+}
+
+static void vgic_its_free_device(struct kvm *kvm, struct its_device *dev)
+{
+       struct its_ite *ite, *tmp;
+
+       list_for_each_entry_safe(ite, tmp, &dev->itt_head, ite_list)
+               its_free_ite(kvm, ite);
+       list_del(&dev->dev_list);
+       kfree(dev);
 }
 
 static void vgic_its_destroy(struct kvm_device *kvm_dev)
 {
        struct kvm *kvm = kvm_dev->kvm;
        struct vgic_its *its = kvm_dev->private;
-       struct its_device *dev;
-       struct its_itte *itte;
-       struct list_head *dev_cur, *dev_temp;
        struct list_head *cur, *temp;
 
        /*
@@ -1447,25 +1637,710 @@ static void vgic_its_destroy(struct kvm_device *kvm_dev)
                return;
 
        mutex_lock(&its->its_lock);
-       list_for_each_safe(dev_cur, dev_temp, &its->device_list) {
-               dev = container_of(dev_cur, struct its_device, dev_list);
-               list_for_each_safe(cur, temp, &dev->itt_head) {
-                       itte = (container_of(cur, struct its_itte, itte_list));
-                       its_free_itte(kvm, itte);
-               }
-               list_del(dev_cur);
-               kfree(dev);
+       list_for_each_safe(cur, temp, &its->device_list) {
+               struct its_device *dev;
+
+               dev = list_entry(cur, struct its_device, dev_list);
+               vgic_its_free_device(kvm, dev);
        }
 
        list_for_each_safe(cur, temp, &its->collection_list) {
+               struct its_collection *coll;
+
+               coll = list_entry(cur, struct its_collection, coll_list);
                list_del(cur);
-               kfree(container_of(cur, struct its_collection, coll_list));
+               kfree(coll);
        }
        mutex_unlock(&its->its_lock);
 
        kfree(its);
 }
 
+int vgic_its_has_attr_regs(struct kvm_device *dev,
+                          struct kvm_device_attr *attr)
+{
+       const struct vgic_register_region *region;
+       gpa_t offset = attr->attr;
+       int align;
+
+       align = (offset < GITS_TYPER) || (offset >= GITS_PIDR4) ? 0x3 : 0x7;
+
+       if (offset & align)
+               return -EINVAL;
+
+       region = vgic_find_mmio_region(its_registers,
+                                      ARRAY_SIZE(its_registers),
+                                      offset);
+       if (!region)
+               return -ENXIO;
+
+       return 0;
+}
+
+int vgic_its_attr_regs_access(struct kvm_device *dev,
+                             struct kvm_device_attr *attr,
+                             u64 *reg, bool is_write)
+{
+       const struct vgic_register_region *region;
+       struct vgic_its *its;
+       gpa_t addr, offset;
+       unsigned int len;
+       int align, ret = 0;
+
+       its = dev->private;
+       offset = attr->attr;
+
+       /*
+        * Although the spec supports upper/lower 32-bit accesses to
+        * 64-bit ITS registers, the userspace ABI requires 64-bit
+        * accesses to all 64-bit wide registers. We therefore only
+        * support 32-bit accesses to GITS_CTLR, GITS_IIDR and GITS ID
+        * registers
+        */
+       if ((offset < GITS_TYPER) || (offset >= GITS_PIDR4))
+               align = 0x3;
+       else
+               align = 0x7;
+
+       if (offset & align)
+               return -EINVAL;
+
+       mutex_lock(&dev->kvm->lock);
+
+       if (IS_VGIC_ADDR_UNDEF(its->vgic_its_base)) {
+               ret = -ENXIO;
+               goto out;
+       }
+
+       region = vgic_find_mmio_region(its_registers,
+                                      ARRAY_SIZE(its_registers),
+                                      offset);
+       if (!region) {
+               ret = -ENXIO;
+               goto out;
+       }
+
+       if (!lock_all_vcpus(dev->kvm)) {
+               ret = -EBUSY;
+               goto out;
+       }
+
+       addr = its->vgic_its_base + offset;
+
+       len = region->access_flags & VGIC_ACCESS_64bit ? 8 : 4;
+
+       if (is_write) {
+               if (region->uaccess_its_write)
+                       ret = region->uaccess_its_write(dev->kvm, its, addr,
+                                                       len, *reg);
+               else
+                       region->its_write(dev->kvm, its, addr, len, *reg);
+       } else {
+               *reg = region->its_read(dev->kvm, its, addr, len);
+       }
+       unlock_all_vcpus(dev->kvm);
+out:
+       mutex_unlock(&dev->kvm->lock);
+       return ret;
+}
+
+static u32 compute_next_devid_offset(struct list_head *h,
+                                    struct its_device *dev)
+{
+       struct its_device *next;
+       u32 next_offset;
+
+       if (list_is_last(&dev->dev_list, h))
+               return 0;
+       next = list_next_entry(dev, dev_list);
+       next_offset = next->device_id - dev->device_id;
+
+       return min_t(u32, next_offset, VITS_DTE_MAX_DEVID_OFFSET);
+}
+
+static u32 compute_next_eventid_offset(struct list_head *h, struct its_ite *ite)
+{
+       struct its_ite *next;
+       u32 next_offset;
+
+       if (list_is_last(&ite->ite_list, h))
+               return 0;
+       next = list_next_entry(ite, ite_list);
+       next_offset = next->event_id - ite->event_id;
+
+       return min_t(u32, next_offset, VITS_ITE_MAX_EVENTID_OFFSET);
+}
+
+/**
+ * entry_fn_t - Callback called on a table entry restore path
+ * @its: its handle
+ * @id: id of the entry
+ * @entry: pointer to the entry
+ * @opaque: pointer to an opaque data
+ *
+ * Return: < 0 on error, 0 if last element was identified, id offset to next
+ * element otherwise
+ */
+typedef int (*entry_fn_t)(struct vgic_its *its, u32 id, void *entry,
+                         void *opaque);
+
+/**
+ * scan_its_table - Scan a contiguous table in guest RAM and applies a function
+ * to each entry
+ *
+ * @its: its handle
+ * @base: base gpa of the table
+ * @size: size of the table in bytes
+ * @esz: entry size in bytes
+ * @start_id: the ID of the first entry in the table
+ * (non zero for 2d level tables)
+ * @fn: function to apply on each entry
+ *
+ * Return: < 0 on error, 0 if last element was identified, 1 otherwise
+ * (the last element may not be found on second level tables)
+ */
+static int scan_its_table(struct vgic_its *its, gpa_t base, int size, int esz,
+                         int start_id, entry_fn_t fn, void *opaque)
+{
+       void *entry = kzalloc(esz, GFP_KERNEL);
+       struct kvm *kvm = its->dev->kvm;
+       unsigned long len = size;
+       int id = start_id;
+       gpa_t gpa = base;
+       int ret;
+
+       while (len > 0) {
+               int next_offset;
+               size_t byte_offset;
+
+               ret = kvm_read_guest(kvm, gpa, entry, esz);
+               if (ret)
+                       goto out;
+
+               next_offset = fn(its, id, entry, opaque);
+               if (next_offset <= 0) {
+                       ret = next_offset;
+                       goto out;
+               }
+
+               byte_offset = next_offset * esz;
+               id += next_offset;
+               gpa += byte_offset;
+               len -= byte_offset;
+       }
+       ret =  1;
+
+out:
+       kfree(entry);
+       return ret;
+}
+
+/**
+ * vgic_its_save_ite - Save an interrupt translation entry at @gpa
+ */
+static int vgic_its_save_ite(struct vgic_its *its, struct its_device *dev,
+                             struct its_ite *ite, gpa_t gpa, int ite_esz)
+{
+       struct kvm *kvm = its->dev->kvm;
+       u32 next_offset;
+       u64 val;
+
+       next_offset = compute_next_eventid_offset(&dev->itt_head, ite);
+       val = ((u64)next_offset << KVM_ITS_ITE_NEXT_SHIFT) |
+              ((u64)ite->lpi << KVM_ITS_ITE_PINTID_SHIFT) |
+               ite->collection->collection_id;
+       val = cpu_to_le64(val);
+       return kvm_write_guest(kvm, gpa, &val, ite_esz);
+}
+
+/**
+ * vgic_its_restore_ite - restore an interrupt translation entry
+ * @event_id: id used for indexing
+ * @ptr: pointer to the ITE entry
+ * @opaque: pointer to the its_device
+ */
+static int vgic_its_restore_ite(struct vgic_its *its, u32 event_id,
+                               void *ptr, void *opaque)
+{
+       struct its_device *dev = (struct its_device *)opaque;
+       struct its_collection *collection;
+       struct kvm *kvm = its->dev->kvm;
+       struct kvm_vcpu *vcpu = NULL;
+       u64 val;
+       u64 *p = (u64 *)ptr;
+       struct vgic_irq *irq;
+       u32 coll_id, lpi_id;
+       struct its_ite *ite;
+       u32 offset;
+
+       val = *p;
+
+       val = le64_to_cpu(val);
+
+       coll_id = val & KVM_ITS_ITE_ICID_MASK;
+       lpi_id = (val & KVM_ITS_ITE_PINTID_MASK) >> KVM_ITS_ITE_PINTID_SHIFT;
+
+       if (!lpi_id)
+               return 1; /* invalid entry, no choice but to scan next entry */
+
+       if (lpi_id < VGIC_MIN_LPI)
+               return -EINVAL;
+
+       offset = val >> KVM_ITS_ITE_NEXT_SHIFT;
+       if (event_id + offset >= BIT_ULL(dev->num_eventid_bits))
+               return -EINVAL;
+
+       collection = find_collection(its, coll_id);
+       if (!collection)
+               return -EINVAL;
+
+       ite = vgic_its_alloc_ite(dev, collection, lpi_id, event_id);
+       if (IS_ERR(ite))
+               return PTR_ERR(ite);
+
+       if (its_is_collection_mapped(collection))
+               vcpu = kvm_get_vcpu(kvm, collection->target_addr);
+
+       irq = vgic_add_lpi(kvm, lpi_id, vcpu);
+       if (IS_ERR(irq))
+               return PTR_ERR(irq);
+       ite->irq = irq;
+
+       return offset;
+}
+
+static int vgic_its_ite_cmp(void *priv, struct list_head *a,
+                           struct list_head *b)
+{
+       struct its_ite *itea = container_of(a, struct its_ite, ite_list);
+       struct its_ite *iteb = container_of(b, struct its_ite, ite_list);
+
+       if (itea->event_id < iteb->event_id)
+               return -1;
+       else
+               return 1;
+}
+
+static int vgic_its_save_itt(struct vgic_its *its, struct its_device *device)
+{
+       const struct vgic_its_abi *abi = vgic_its_get_abi(its);
+       gpa_t base = device->itt_addr;
+       struct its_ite *ite;
+       int ret;
+       int ite_esz = abi->ite_esz;
+
+       list_sort(NULL, &device->itt_head, vgic_its_ite_cmp);
+
+       list_for_each_entry(ite, &device->itt_head, ite_list) {
+               gpa_t gpa = base + ite->event_id * ite_esz;
+
+               ret = vgic_its_save_ite(its, device, ite, gpa, ite_esz);
+               if (ret)
+                       return ret;
+       }
+       return 0;
+}
+
+static int vgic_its_restore_itt(struct vgic_its *its, struct its_device *dev)
+{
+       const struct vgic_its_abi *abi = vgic_its_get_abi(its);
+       gpa_t base = dev->itt_addr;
+       int ret;
+       int ite_esz = abi->ite_esz;
+       size_t max_size = BIT_ULL(dev->num_eventid_bits) * ite_esz;
+
+       ret = scan_its_table(its, base, max_size, ite_esz, 0,
+                            vgic_its_restore_ite, dev);
+
+       return ret;
+}
+
+/**
+ * vgic_its_save_dte - Save a device table entry at a given GPA
+ *
+ * @its: ITS handle
+ * @dev: ITS device
+ * @ptr: GPA
+ */
+static int vgic_its_save_dte(struct vgic_its *its, struct its_device *dev,
+                            gpa_t ptr, int dte_esz)
+{
+       struct kvm *kvm = its->dev->kvm;
+       u64 val, itt_addr_field;
+       u32 next_offset;
+
+       itt_addr_field = dev->itt_addr >> 8;
+       next_offset = compute_next_devid_offset(&its->device_list, dev);
+       val = (1ULL << KVM_ITS_DTE_VALID_SHIFT |
+              ((u64)next_offset << KVM_ITS_DTE_NEXT_SHIFT) |
+              (itt_addr_field << KVM_ITS_DTE_ITTADDR_SHIFT) |
+               (dev->num_eventid_bits - 1));
+       val = cpu_to_le64(val);
+       return kvm_write_guest(kvm, ptr, &val, dte_esz);
+}
+
+/**
+ * vgic_its_restore_dte - restore a device table entry
+ *
+ * @its: its handle
+ * @id: device id the DTE corresponds to
+ * @ptr: kernel VA where the 8 byte DTE is located
+ * @opaque: unused
+ *
+ * Return: < 0 on error, 0 if the dte is the last one, id offset to the
+ * next dte otherwise
+ */
+static int vgic_its_restore_dte(struct vgic_its *its, u32 id,
+                               void *ptr, void *opaque)
+{
+       struct its_device *dev;
+       gpa_t itt_addr;
+       u8 num_eventid_bits;
+       u64 entry = *(u64 *)ptr;
+       bool valid;
+       u32 offset;
+       int ret;
+
+       entry = le64_to_cpu(entry);
+
+       valid = entry >> KVM_ITS_DTE_VALID_SHIFT;
+       num_eventid_bits = (entry & KVM_ITS_DTE_SIZE_MASK) + 1;
+       itt_addr = ((entry & KVM_ITS_DTE_ITTADDR_MASK)
+                       >> KVM_ITS_DTE_ITTADDR_SHIFT) << 8;
+
+       if (!valid)
+               return 1;
+
+       /* dte entry is valid */
+       offset = (entry & KVM_ITS_DTE_NEXT_MASK) >> KVM_ITS_DTE_NEXT_SHIFT;
+
+       dev = vgic_its_alloc_device(its, id, itt_addr, num_eventid_bits);
+       if (IS_ERR(dev))
+               return PTR_ERR(dev);
+
+       ret = vgic_its_restore_itt(its, dev);
+       if (ret) {
+               vgic_its_free_device(its->dev->kvm, dev);
+               return ret;
+       }
+
+       return offset;
+}
+
+static int vgic_its_device_cmp(void *priv, struct list_head *a,
+                              struct list_head *b)
+{
+       struct its_device *deva = container_of(a, struct its_device, dev_list);
+       struct its_device *devb = container_of(b, struct its_device, dev_list);
+
+       if (deva->device_id < devb->device_id)
+               return -1;
+       else
+               return 1;
+}
+
+/**
+ * vgic_its_save_device_tables - Save the device table and all ITT
+ * into guest RAM
+ *
+ * L1/L2 handling is hidden by vgic_its_check_id() helper which directly
+ * returns the GPA of the device entry
+ */
+static int vgic_its_save_device_tables(struct vgic_its *its)
+{
+       const struct vgic_its_abi *abi = vgic_its_get_abi(its);
+       struct its_device *dev;
+       int dte_esz = abi->dte_esz;
+       u64 baser;
+
+       baser = its->baser_device_table;
+
+       list_sort(NULL, &its->device_list, vgic_its_device_cmp);
+
+       list_for_each_entry(dev, &its->device_list, dev_list) {
+               int ret;
+               gpa_t eaddr;
+
+               if (!vgic_its_check_id(its, baser,
+                                      dev->device_id, &eaddr))
+                       return -EINVAL;
+
+               ret = vgic_its_save_itt(its, dev);
+               if (ret)
+                       return ret;
+
+               ret = vgic_its_save_dte(its, dev, eaddr, dte_esz);
+               if (ret)
+                       return ret;
+       }
+       return 0;
+}
+
+/**
+ * handle_l1_dte - callback used for L1 device table entries (2 stage case)
+ *
+ * @its: its handle
+ * @id: index of the entry in the L1 table
+ * @addr: kernel VA
+ * @opaque: unused
+ *
+ * L1 table entries are scanned by steps of 1 entry
+ * Return < 0 if error, 0 if last dte was found when scanning the L2
+ * table, +1 otherwise (meaning next L1 entry must be scanned)
+ */
+static int handle_l1_dte(struct vgic_its *its, u32 id, void *addr,
+                        void *opaque)
+{
+       const struct vgic_its_abi *abi = vgic_its_get_abi(its);
+       int l2_start_id = id * (SZ_64K / abi->dte_esz);
+       u64 entry = *(u64 *)addr;
+       int dte_esz = abi->dte_esz;
+       gpa_t gpa;
+       int ret;
+
+       entry = le64_to_cpu(entry);
+
+       if (!(entry & KVM_ITS_L1E_VALID_MASK))
+               return 1;
+
+       gpa = entry & KVM_ITS_L1E_ADDR_MASK;
+
+       ret = scan_its_table(its, gpa, SZ_64K, dte_esz,
+                            l2_start_id, vgic_its_restore_dte, NULL);
+
+       if (ret <= 0)
+               return ret;
+
+       return 1;
+}
+
+/**
+ * vgic_its_restore_device_tables - Restore the device table and all ITT
+ * from guest RAM to internal data structs
+ */
+static int vgic_its_restore_device_tables(struct vgic_its *its)
+{
+       const struct vgic_its_abi *abi = vgic_its_get_abi(its);
+       u64 baser = its->baser_device_table;
+       int l1_esz, ret;
+       int l1_tbl_size = GITS_BASER_NR_PAGES(baser) * SZ_64K;
+       gpa_t l1_gpa;
+
+       if (!(baser & GITS_BASER_VALID))
+               return 0;
+
+       l1_gpa = BASER_ADDRESS(baser);
+
+       if (baser & GITS_BASER_INDIRECT) {
+               l1_esz = GITS_LVL1_ENTRY_SIZE;
+               ret = scan_its_table(its, l1_gpa, l1_tbl_size, l1_esz, 0,
+                                    handle_l1_dte, NULL);
+       } else {
+               l1_esz = abi->dte_esz;
+               ret = scan_its_table(its, l1_gpa, l1_tbl_size, l1_esz, 0,
+                                    vgic_its_restore_dte, NULL);
+       }
+
+       if (ret > 0)
+               ret = -EINVAL;
+
+       return ret;
+}
+
+static int vgic_its_save_cte(struct vgic_its *its,
+                            struct its_collection *collection,
+                            gpa_t gpa, int esz)
+{
+       u64 val;
+
+       val = (1ULL << KVM_ITS_CTE_VALID_SHIFT |
+              ((u64)collection->target_addr << KVM_ITS_CTE_RDBASE_SHIFT) |
+              collection->collection_id);
+       val = cpu_to_le64(val);
+       return kvm_write_guest(its->dev->kvm, gpa, &val, esz);
+}
+
+static int vgic_its_restore_cte(struct vgic_its *its, gpa_t gpa, int esz)
+{
+       struct its_collection *collection;
+       struct kvm *kvm = its->dev->kvm;
+       u32 target_addr, coll_id;
+       u64 val;
+       int ret;
+
+       BUG_ON(esz > sizeof(val));
+       ret = kvm_read_guest(kvm, gpa, &val, esz);
+       if (ret)
+               return ret;
+       val = le64_to_cpu(val);
+       if (!(val & KVM_ITS_CTE_VALID_MASK))
+               return 0;
+
+       target_addr = (u32)(val >> KVM_ITS_CTE_RDBASE_SHIFT);
+       coll_id = val & KVM_ITS_CTE_ICID_MASK;
+
+       if (target_addr >= atomic_read(&kvm->online_vcpus))
+               return -EINVAL;
+
+       collection = find_collection(its, coll_id);
+       if (collection)
+               return -EEXIST;
+       ret = vgic_its_alloc_collection(its, &collection, coll_id);
+       if (ret)
+               return ret;
+       collection->target_addr = target_addr;
+       return 1;
+}
+
+/**
+ * vgic_its_save_collection_table - Save the collection table into
+ * guest RAM
+ */
+static int vgic_its_save_collection_table(struct vgic_its *its)
+{
+       const struct vgic_its_abi *abi = vgic_its_get_abi(its);
+       struct its_collection *collection;
+       u64 val;
+       gpa_t gpa;
+       size_t max_size, filled = 0;
+       int ret, cte_esz = abi->cte_esz;
+
+       gpa = BASER_ADDRESS(its->baser_coll_table);
+       if (!gpa)
+               return 0;
+
+       max_size = GITS_BASER_NR_PAGES(its->baser_coll_table) * SZ_64K;
+
+       list_for_each_entry(collection, &its->collection_list, coll_list) {
+               ret = vgic_its_save_cte(its, collection, gpa, cte_esz);
+               if (ret)
+                       return ret;
+               gpa += cte_esz;
+               filled += cte_esz;
+       }
+
+       if (filled == max_size)
+               return 0;
+
+       /*
+        * table is not fully filled, add a last dummy element
+        * with valid bit unset
+        */
+       val = 0;
+       BUG_ON(cte_esz > sizeof(val));
+       ret = kvm_write_guest(its->dev->kvm, gpa, &val, cte_esz);
+       return ret;
+}
+
+/**
+ * vgic_its_restore_collection_table - reads the collection table
+ * in guest memory and restores the ITS internal state. Requires the
+ * BASER registers to be restored before.
+ */
+static int vgic_its_restore_collection_table(struct vgic_its *its)
+{
+       const struct vgic_its_abi *abi = vgic_its_get_abi(its);
+       int cte_esz = abi->cte_esz;
+       size_t max_size, read = 0;
+       gpa_t gpa;
+       int ret;
+
+       if (!(its->baser_coll_table & GITS_BASER_VALID))
+               return 0;
+
+       gpa = BASER_ADDRESS(its->baser_coll_table);
+
+       max_size = GITS_BASER_NR_PAGES(its->baser_coll_table) * SZ_64K;
+
+       while (read < max_size) {
+               ret = vgic_its_restore_cte(its, gpa, cte_esz);
+               if (ret <= 0)
+                       break;
+               gpa += cte_esz;
+               read += cte_esz;
+       }
+       return ret;
+}
+
+/**
+ * vgic_its_save_tables_v0 - Save the ITS tables into guest ARM
+ * according to v0 ABI
+ */
+static int vgic_its_save_tables_v0(struct vgic_its *its)
+{
+       struct kvm *kvm = its->dev->kvm;
+       int ret;
+
+       mutex_lock(&kvm->lock);
+       mutex_lock(&its->its_lock);
+
+       if (!lock_all_vcpus(kvm)) {
+               mutex_unlock(&its->its_lock);
+               mutex_unlock(&kvm->lock);
+               return -EBUSY;
+       }
+
+       ret = vgic_its_save_device_tables(its);
+       if (ret)
+               goto out;
+
+       ret = vgic_its_save_collection_table(its);
+
+out:
+       unlock_all_vcpus(kvm);
+       mutex_unlock(&its->its_lock);
+       mutex_unlock(&kvm->lock);
+       return ret;
+}
+
+/**
+ * vgic_its_restore_tables_v0 - Restore the ITS tables from guest RAM
+ * to internal data structs according to V0 ABI
+ *
+ */
+static int vgic_its_restore_tables_v0(struct vgic_its *its)
+{
+       struct kvm *kvm = its->dev->kvm;
+       int ret;
+
+       mutex_lock(&kvm->lock);
+       mutex_lock(&its->its_lock);
+
+       if (!lock_all_vcpus(kvm)) {
+               mutex_unlock(&its->its_lock);
+               mutex_unlock(&kvm->lock);
+               return -EBUSY;
+       }
+
+       ret = vgic_its_restore_collection_table(its);
+       if (ret)
+               goto out;
+
+       ret = vgic_its_restore_device_tables(its);
+out:
+       unlock_all_vcpus(kvm);
+       mutex_unlock(&its->its_lock);
+       mutex_unlock(&kvm->lock);
+
+       return ret;
+}
+
+static int vgic_its_commit_v0(struct vgic_its *its)
+{
+       const struct vgic_its_abi *abi;
+
+       abi = vgic_its_get_abi(its);
+       its->baser_coll_table &= ~GITS_BASER_ENTRY_SIZE_MASK;
+       its->baser_device_table &= ~GITS_BASER_ENTRY_SIZE_MASK;
+
+       its->baser_coll_table |= (GIC_ENCODE_SZ(abi->cte_esz, 5)
+                                       << GITS_BASER_ENTRY_SIZE_SHIFT);
+
+       its->baser_device_table |= (GIC_ENCODE_SZ(abi->dte_esz, 5)
+                                       << GITS_BASER_ENTRY_SIZE_SHIFT);
+       return 0;
+}
+
 static int vgic_its_has_attr(struct kvm_device *dev,
                             struct kvm_device_attr *attr)
 {
@@ -1480,8 +2355,14 @@ static int vgic_its_has_attr(struct kvm_device *dev,
                switch (attr->attr) {
                case KVM_DEV_ARM_VGIC_CTRL_INIT:
                        return 0;
+               case KVM_DEV_ARM_ITS_SAVE_TABLES:
+                       return 0;
+               case KVM_DEV_ARM_ITS_RESTORE_TABLES:
+                       return 0;
                }
                break;
+       case KVM_DEV_ARM_VGIC_GRP_ITS_REGS:
+               return vgic_its_has_attr_regs(dev, attr);
        }
        return -ENXIO;
 }
@@ -1509,18 +2390,30 @@ static int vgic_its_set_attr(struct kvm_device *dev,
                if (ret)
                        return ret;
 
-               its->vgic_its_base = addr;
-
-               return 0;
+               return vgic_register_its_iodev(dev->kvm, its, addr);
        }
-       case KVM_DEV_ARM_VGIC_GRP_CTRL:
+       case KVM_DEV_ARM_VGIC_GRP_CTRL: {
+               const struct vgic_its_abi *abi = vgic_its_get_abi(its);
+
                switch (attr->attr) {
                case KVM_DEV_ARM_VGIC_CTRL_INIT:
-                       its->initialized = true;
-
+                       /* Nothing to do */
                        return 0;
+               case KVM_DEV_ARM_ITS_SAVE_TABLES:
+                       return abi->save_tables(its);
+               case KVM_DEV_ARM_ITS_RESTORE_TABLES:
+                       return abi->restore_tables(its);
                }
-               break;
+       }
+       case KVM_DEV_ARM_VGIC_GRP_ITS_REGS: {
+               u64 __user *uaddr = (u64 __user *)(long)attr->addr;
+               u64 reg;
+
+               if (get_user(reg, uaddr))
+                       return -EFAULT;
+
+               return vgic_its_attr_regs_access(dev, attr, &reg, true);
+       }
        }
        return -ENXIO;
 }
@@ -1541,10 +2434,20 @@ static int vgic_its_get_attr(struct kvm_device *dev,
                if (copy_to_user(uaddr, &addr, sizeof(addr)))
                        return -EFAULT;
                break;
+       }
+       case KVM_DEV_ARM_VGIC_GRP_ITS_REGS: {
+               u64 __user *uaddr = (u64 __user *)(long)attr->addr;
+               u64 reg;
+               int ret;
+
+               ret = vgic_its_attr_regs_access(dev, attr, &reg, false);
+               if (ret)
+                       return ret;
+               return put_user(reg, uaddr);
+       }
        default:
                return -ENXIO;
        }
-       }
 
        return 0;
 }
@@ -1563,30 +2466,3 @@ int kvm_vgic_register_its_device(void)
        return kvm_register_device_ops(&kvm_arm_vgic_its_ops,
                                       KVM_DEV_TYPE_ARM_VGIC_ITS);
 }
-
-/*
- * Registers all ITSes with the kvm_io_bus framework.
- * To follow the existing VGIC initialization sequence, this has to be
- * done as late as possible, just before the first VCPU runs.
- */
-int vgic_register_its_iodevs(struct kvm *kvm)
-{
-       struct kvm_device *dev;
-       int ret = 0;
-
-       list_for_each_entry(dev, &kvm->devices, vm_node) {
-               if (dev->ops != &kvm_arm_vgic_its_ops)
-                       continue;
-
-               ret = vgic_register_its_iodev(kvm, dev->private);
-               if (ret)
-                       return ret;
-               /*
-                * We don't need to care about tearing down previously
-                * registered ITSes, as the kvm_io_bus framework removes
-                * them for us if the VM gets destroyed.
-                */
-       }
-
-       return ret;
-}