linux/drivers/iommu/hyperv-iommu.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0
   2
   3/*
   4 * Hyper-V stub IOMMU driver.
   5 *
   6 * Copyright (C) 2019, Microsoft, Inc.
   7 *
   8 * Author : Lan Tianyu <Tianyu.Lan@microsoft.com>
   9 */
  10
  11#include <linux/types.h>
  12#include <linux/interrupt.h>
  13#include <linux/irq.h>
  14#include <linux/iommu.h>
  15#include <linux/module.h>
  16
  17#include <asm/apic.h>
  18#include <asm/cpu.h>
  19#include <asm/hw_irq.h>
  20#include <asm/io_apic.h>
  21#include <asm/irq_remapping.h>
  22#include <asm/hypervisor.h>
  23#include <asm/mshyperv.h>
  24
  25#include "irq_remapping.h"
  26
  27#ifdef CONFIG_IRQ_REMAP
  28
  29/*
  30 * According 82093AA IO-APIC spec , IO APIC has a 24-entry Interrupt
  31 * Redirection Table. Hyper-V exposes one single IO-APIC and so define
  32 * 24 IO APIC remmapping entries.
  33 */
  34#define IOAPIC_REMAPPING_ENTRY 24
  35
  36static cpumask_t ioapic_max_cpumask = { CPU_BITS_NONE };
  37static struct irq_domain *ioapic_ir_domain;
  38
  39static int hyperv_ir_set_affinity(struct irq_data *data,
  40                const struct cpumask *mask, bool force)
  41{
  42        struct irq_data *parent = data->parent_data;
  43        struct irq_cfg *cfg = irqd_cfg(data);
  44        int ret;
  45
  46        /* Return error If new irq affinity is out of ioapic_max_cpumask. */
  47        if (!cpumask_subset(mask, &ioapic_max_cpumask))
  48                return -EINVAL;
  49
  50        ret = parent->chip->irq_set_affinity(parent, mask, force);
  51        if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
  52                return ret;
  53
  54        send_cleanup_vector(cfg);
  55
  56        return 0;
  57}
  58
  59static struct irq_chip hyperv_ir_chip = {
  60        .name                   = "HYPERV-IR",
  61        .irq_ack                = apic_ack_irq,
  62        .irq_set_affinity       = hyperv_ir_set_affinity,
  63};
  64
  65static int hyperv_irq_remapping_alloc(struct irq_domain *domain,
  66                                     unsigned int virq, unsigned int nr_irqs,
  67                                     void *arg)
  68{
  69        struct irq_alloc_info *info = arg;
  70        struct irq_data *irq_data;
  71        struct irq_desc *desc;
  72        int ret = 0;
  73
  74        if (!info || info->type != X86_IRQ_ALLOC_TYPE_IOAPIC || nr_irqs > 1)
  75                return -EINVAL;
  76
  77        ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
  78        if (ret < 0)
  79                return ret;
  80
  81        irq_data = irq_domain_get_irq_data(domain, virq);
  82        if (!irq_data) {
  83                irq_domain_free_irqs_common(domain, virq, nr_irqs);
  84                return -EINVAL;
  85        }
  86
  87        irq_data->chip = &hyperv_ir_chip;
  88
  89        /*
  90         * Hypver-V IO APIC irq affinity should be in the scope of
  91         * ioapic_max_cpumask because no irq remapping support.
  92         */
  93        desc = irq_data_to_desc(irq_data);
  94        cpumask_copy(desc->irq_common_data.affinity, &ioapic_max_cpumask);
  95
  96        return 0;
  97}
  98
  99static void hyperv_irq_remapping_free(struct irq_domain *domain,
 100                                 unsigned int virq, unsigned int nr_irqs)
 101{
 102        irq_domain_free_irqs_common(domain, virq, nr_irqs);
 103}
 104
 105static int hyperv_irq_remapping_select(struct irq_domain *d,
 106                                       struct irq_fwspec *fwspec,
 107                                       enum irq_domain_bus_token bus_token)
 108{
 109        /* Claim the only I/O APIC emulated by Hyper-V */
 110        return x86_fwspec_is_ioapic(fwspec);
 111}
 112
 113static const struct irq_domain_ops hyperv_ir_domain_ops = {
 114        .select = hyperv_irq_remapping_select,
 115        .alloc = hyperv_irq_remapping_alloc,
 116        .free = hyperv_irq_remapping_free,
 117};
 118
 119static const struct irq_domain_ops hyperv_root_ir_domain_ops;
 120static int __init hyperv_prepare_irq_remapping(void)
 121{
 122        struct fwnode_handle *fn;
 123        int i;
 124        const char *name;
 125        const struct irq_domain_ops *ops;
 126
 127        if (!hypervisor_is_type(X86_HYPER_MS_HYPERV) ||
 128            x86_init.hyper.msi_ext_dest_id() ||
 129            !x2apic_supported())
 130                return -ENODEV;
 131
 132        if (hv_root_partition) {
 133                name = "HYPERV-ROOT-IR";
 134                ops = &hyperv_root_ir_domain_ops;
 135        } else {
 136                name = "HYPERV-IR";
 137                ops = &hyperv_ir_domain_ops;
 138        }
 139
 140        fn = irq_domain_alloc_named_id_fwnode(name, 0);
 141        if (!fn)
 142                return -ENOMEM;
 143
 144        ioapic_ir_domain =
 145                irq_domain_create_hierarchy(arch_get_ir_parent_domain(),
 146                                0, IOAPIC_REMAPPING_ENTRY, fn, ops, NULL);
 147
 148        if (!ioapic_ir_domain) {
 149                irq_domain_free_fwnode(fn);
 150                return -ENOMEM;
 151        }
 152
 153        if (hv_root_partition)
 154                return 0; /* The rest is only relevant to guests */
 155
 156        /*
 157         * Hyper-V doesn't provide irq remapping function for
 158         * IO-APIC and so IO-APIC only accepts 8-bit APIC ID.
 159         * Cpu's APIC ID is read from ACPI MADT table and APIC IDs
 160         * in the MADT table on Hyper-v are sorted monotonic increasingly.
 161         * APIC ID reflects cpu topology. There maybe some APIC ID
 162         * gaps when cpu number in a socket is not power of two. Prepare
 163         * max cpu affinity for IOAPIC irqs. Scan cpu 0-255 and set cpu
 164         * into ioapic_max_cpumask if its APIC ID is less than 256.
 165         */
 166        for (i = min_t(unsigned int, num_possible_cpus() - 1, 255); i >= 0; i--)
 167                if (cpu_physical_id(i) < 256)
 168                        cpumask_set_cpu(i, &ioapic_max_cpumask);
 169
 170        return 0;
 171}
 172
 173static int __init hyperv_enable_irq_remapping(void)
 174{
 175        return IRQ_REMAP_X2APIC_MODE;
 176}
 177
 178struct irq_remap_ops hyperv_irq_remap_ops = {
 179        .prepare                = hyperv_prepare_irq_remapping,
 180        .enable                 = hyperv_enable_irq_remapping,
 181};
 182
 183/* IRQ remapping domain when Linux runs as the root partition */
 184struct hyperv_root_ir_data {
 185        u8 ioapic_id;
 186        bool is_level;
 187        struct hv_interrupt_entry entry;
 188};
 189
 190static void
 191hyperv_root_ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
 192{
 193        u64 status;
 194        u32 vector;
 195        struct irq_cfg *cfg;
 196        int ioapic_id;
 197        struct cpumask *affinity;
 198        int cpu;
 199        struct hv_interrupt_entry entry;
 200        struct hyperv_root_ir_data *data = irq_data->chip_data;
 201        struct IO_APIC_route_entry e;
 202
 203        cfg = irqd_cfg(irq_data);
 204        affinity = irq_data_get_effective_affinity_mask(irq_data);
 205        cpu = cpumask_first_and(affinity, cpu_online_mask);
 206
 207        vector = cfg->vector;
 208        ioapic_id = data->ioapic_id;
 209
 210        if (data->entry.source == HV_DEVICE_TYPE_IOAPIC
 211            && data->entry.ioapic_rte.as_uint64) {
 212                entry = data->entry;
 213
 214                status = hv_unmap_ioapic_interrupt(ioapic_id, &entry);
 215
 216                if (status != HV_STATUS_SUCCESS)
 217                        pr_debug("%s: unexpected unmap status %lld\n", __func__, status);
 218
 219                data->entry.ioapic_rte.as_uint64 = 0;
 220                data->entry.source = 0; /* Invalid source */
 221        }
 222
 223
 224        status = hv_map_ioapic_interrupt(ioapic_id, data->is_level, cpu,
 225                                        vector, &entry);
 226
 227        if (status != HV_STATUS_SUCCESS) {
 228                pr_err("%s: map hypercall failed, status %lld\n", __func__, status);
 229                return;
 230        }
 231
 232        data->entry = entry;
 233
 234        /* Turn it into an IO_APIC_route_entry, and generate MSI MSG. */
 235        e.w1 = entry.ioapic_rte.low_uint32;
 236        e.w2 = entry.ioapic_rte.high_uint32;
 237
 238        memset(msg, 0, sizeof(*msg));
 239        msg->arch_data.vector = e.vector;
 240        msg->arch_data.delivery_mode = e.delivery_mode;
 241        msg->arch_addr_lo.dest_mode_logical = e.dest_mode_logical;
 242        msg->arch_addr_lo.dmar_format = e.ir_format;
 243        msg->arch_addr_lo.dmar_index_0_14 = e.ir_index_0_14;
 244}
 245
 246static int hyperv_root_ir_set_affinity(struct irq_data *data,
 247                const struct cpumask *mask, bool force)
 248{
 249        struct irq_data *parent = data->parent_data;
 250        struct irq_cfg *cfg = irqd_cfg(data);
 251        int ret;
 252
 253        ret = parent->chip->irq_set_affinity(parent, mask, force);
 254        if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
 255                return ret;
 256
 257        send_cleanup_vector(cfg);
 258
 259        return 0;
 260}
 261
 262static struct irq_chip hyperv_root_ir_chip = {
 263        .name                   = "HYPERV-ROOT-IR",
 264        .irq_ack                = apic_ack_irq,
 265        .irq_set_affinity       = hyperv_root_ir_set_affinity,
 266        .irq_compose_msi_msg    = hyperv_root_ir_compose_msi_msg,
 267};
 268
 269static int hyperv_root_irq_remapping_alloc(struct irq_domain *domain,
 270                                     unsigned int virq, unsigned int nr_irqs,
 271                                     void *arg)
 272{
 273        struct irq_alloc_info *info = arg;
 274        struct irq_data *irq_data;
 275        struct hyperv_root_ir_data *data;
 276        int ret = 0;
 277
 278        if (!info || info->type != X86_IRQ_ALLOC_TYPE_IOAPIC || nr_irqs > 1)
 279                return -EINVAL;
 280
 281        ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
 282        if (ret < 0)
 283                return ret;
 284
 285        data = kzalloc(sizeof(*data), GFP_KERNEL);
 286        if (!data) {
 287                irq_domain_free_irqs_common(domain, virq, nr_irqs);
 288                return -ENOMEM;
 289        }
 290
 291        irq_data = irq_domain_get_irq_data(domain, virq);
 292        if (!irq_data) {
 293                kfree(data);
 294                irq_domain_free_irqs_common(domain, virq, nr_irqs);
 295                return -EINVAL;
 296        }
 297
 298        data->ioapic_id = info->devid;
 299        data->is_level = info->ioapic.is_level;
 300
 301        irq_data->chip = &hyperv_root_ir_chip;
 302        irq_data->chip_data = data;
 303
 304        return 0;
 305}
 306
 307static void hyperv_root_irq_remapping_free(struct irq_domain *domain,
 308                                 unsigned int virq, unsigned int nr_irqs)
 309{
 310        struct irq_data *irq_data;
 311        struct hyperv_root_ir_data *data;
 312        struct hv_interrupt_entry *e;
 313        int i;
 314
 315        for (i = 0; i < nr_irqs; i++) {
 316                irq_data = irq_domain_get_irq_data(domain, virq + i);
 317
 318                if (irq_data && irq_data->chip_data) {
 319                        data = irq_data->chip_data;
 320                        e = &data->entry;
 321
 322                        if (e->source == HV_DEVICE_TYPE_IOAPIC
 323                              && e->ioapic_rte.as_uint64)
 324                                hv_unmap_ioapic_interrupt(data->ioapic_id,
 325                                                        &data->entry);
 326
 327                        kfree(data);
 328                }
 329        }
 330
 331        irq_domain_free_irqs_common(domain, virq, nr_irqs);
 332}
 333
 334static const struct irq_domain_ops hyperv_root_ir_domain_ops = {
 335        .select = hyperv_irq_remapping_select,
 336        .alloc = hyperv_root_irq_remapping_alloc,
 337        .free = hyperv_root_irq_remapping_free,
 338};
 339
 340#endif
 341