| 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | /* |
| 3 | * Copyright (C) 2015, 2016 ARM Ltd. |
| 4 | */ |
| 5 | |
| 6 | #include <linux/uaccess.h> |
| 7 | #include <linux/interrupt.h> |
| 8 | #include <linux/cpu.h> |
| 9 | #include <linux/kvm_host.h> |
| 10 | #include <kvm/arm_vgic.h> |
| 11 | #include <asm/kvm_emulate.h> |
| 12 | #include <asm/kvm_mmu.h> |
| 13 | #include "vgic.h" |
| 14 | |
| 15 | /* |
| 16 | * Initialization rules: there are multiple stages to the vgic |
| 17 | * initialization, both for the distributor and the CPU interfaces. The basic |
| 18 | * idea is that even though the VGIC is not functional or not requested from |
| 19 | * user space, the critical path of the run loop can still call VGIC functions |
| 20 | * that just won't do anything, without them having to check additional |
| 21 | * initialization flags to ensure they don't look at uninitialized data |
| 22 | * structures. |
| 23 | * |
| 24 | * Distributor: |
| 25 | * |
| 26 | * - kvm_vgic_early_init(): initialization of static data that doesn't |
| 27 | * depend on any sizing information or emulation type. No allocation |
| 28 | * is allowed there. |
| 29 | * |
| 30 | * - vgic_init(): allocation and initialization of the generic data |
| 31 | * structures that depend on sizing information (number of CPUs, |
| 32 | * number of interrupts). Also initializes the vcpu specific data |
| 33 | * structures. Can be executed lazily for GICv2. |
| 34 | * |
| 35 | * CPU Interface: |
| 36 | * |
| 37 | * - kvm_vgic_vcpu_init(): initialization of static data that doesn't depend |
| 38 | * on any sizing information. Private interrupts are allocated if not |
| 39 | * already allocated at vgic-creation time. |
| 40 | */ |
| 41 | |
| 42 | /* EARLY INIT */ |
| 43 | |
| 44 | /** |
| 45 | * kvm_vgic_early_init() - Initialize static VGIC VCPU data structures |
| 46 | * @kvm: The VM whose VGIC districutor should be initialized |
| 47 | * |
| 48 | * Only do initialization of static structures that don't require any |
| 49 | * allocation or sizing information from userspace. vgic_init() called |
| 50 | * kvm_vgic_dist_init() which takes care of the rest. |
| 51 | */ |
| 52 | void kvm_vgic_early_init(struct kvm *kvm) |
| 53 | { |
| 54 | struct vgic_dist *dist = &kvm->arch.vgic; |
| 55 | |
| 56 | xa_init_flags(xa: &dist->lpi_xa, XA_FLAGS_LOCK_IRQ); |
| 57 | } |
| 58 | |
| 59 | /* CREATION */ |
| 60 | |
| 61 | static int vgic_allocate_private_irqs_locked(struct kvm_vcpu *vcpu, u32 type); |
| 62 | |
| 63 | /** |
| 64 | * kvm_vgic_create: triggered by the instantiation of the VGIC device by |
| 65 | * user space, either through the legacy KVM_CREATE_IRQCHIP ioctl (v2 only) |
| 66 | * or through the generic KVM_CREATE_DEVICE API ioctl. |
| 67 | * irqchip_in_kernel() tells you if this function succeeded or not. |
| 68 | * @kvm: kvm struct pointer |
| 69 | * @type: KVM_DEV_TYPE_ARM_VGIC_V[23] |
| 70 | */ |
| 71 | int kvm_vgic_create(struct kvm *kvm, u32 type) |
| 72 | { |
| 73 | struct kvm_vcpu *vcpu; |
| 74 | unsigned long i; |
| 75 | int ret; |
| 76 | |
| 77 | /* |
| 78 | * This function is also called by the KVM_CREATE_IRQCHIP handler, |
| 79 | * which had no chance yet to check the availability of the GICv2 |
| 80 | * emulation. So check this here again. KVM_CREATE_DEVICE does |
| 81 | * the proper checks already. |
| 82 | */ |
| 83 | if (type == KVM_DEV_TYPE_ARM_VGIC_V2 && |
| 84 | !kvm_vgic_global_state.can_emulate_gicv2) |
| 85 | return -ENODEV; |
| 86 | |
| 87 | /* |
| 88 | * Ensure mutual exclusion with vCPU creation and any vCPU ioctls by: |
| 89 | * |
| 90 | * - Holding kvm->lock to prevent KVM_CREATE_VCPU from reaching |
| 91 | * kvm_arch_vcpu_precreate() and ensuring created_vcpus is stable. |
| 92 | * This alone is insufficient, as kvm_vm_ioctl_create_vcpu() drops |
| 93 | * the kvm->lock before completing the vCPU creation. |
| 94 | */ |
| 95 | lockdep_assert_held(&kvm->lock); |
| 96 | |
| 97 | /* |
| 98 | * - Acquiring the vCPU mutex for every *online* vCPU to prevent |
| 99 | * concurrent vCPU ioctls for vCPUs already visible to userspace. |
| 100 | */ |
| 101 | ret = -EBUSY; |
| 102 | if (kvm_trylock_all_vcpus(kvm)) |
| 103 | return ret; |
| 104 | |
| 105 | /* |
| 106 | * - Taking the config_lock which protects VGIC data structures such |
| 107 | * as the per-vCPU arrays of private IRQs (SGIs, PPIs). |
| 108 | */ |
| 109 | mutex_lock(&kvm->arch.config_lock); |
| 110 | |
| 111 | /* |
| 112 | * - Bailing on the entire thing if a vCPU is in the middle of creation, |
| 113 | * dropped the kvm->lock, but hasn't reached kvm_arch_vcpu_create(). |
| 114 | * |
| 115 | * The whole combination of this guarantees that no vCPU can get into |
| 116 | * KVM with a VGIC configuration inconsistent with the VM's VGIC. |
| 117 | */ |
| 118 | if (kvm->created_vcpus != atomic_read(v: &kvm->online_vcpus)) |
| 119 | goto out_unlock; |
| 120 | |
| 121 | if (irqchip_in_kernel(kvm)) { |
| 122 | ret = -EEXIST; |
| 123 | goto out_unlock; |
| 124 | } |
| 125 | |
| 126 | kvm_for_each_vcpu(i, vcpu, kvm) { |
| 127 | if (vcpu_has_run_once(vcpu)) |
| 128 | goto out_unlock; |
| 129 | } |
| 130 | ret = 0; |
| 131 | |
| 132 | if (type == KVM_DEV_TYPE_ARM_VGIC_V2) |
| 133 | kvm->max_vcpus = VGIC_V2_MAX_CPUS; |
| 134 | else |
| 135 | kvm->max_vcpus = VGIC_V3_MAX_CPUS; |
| 136 | |
| 137 | if (atomic_read(v: &kvm->online_vcpus) > kvm->max_vcpus) { |
| 138 | ret = -E2BIG; |
| 139 | goto out_unlock; |
| 140 | } |
| 141 | |
| 142 | kvm_for_each_vcpu(i, vcpu, kvm) { |
| 143 | ret = vgic_allocate_private_irqs_locked(vcpu, type); |
| 144 | if (ret) |
| 145 | break; |
| 146 | } |
| 147 | |
| 148 | if (ret) { |
| 149 | kvm_for_each_vcpu(i, vcpu, kvm) { |
| 150 | struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu; |
| 151 | kfree(objp: vgic_cpu->private_irqs); |
| 152 | vgic_cpu->private_irqs = NULL; |
| 153 | } |
| 154 | |
| 155 | goto out_unlock; |
| 156 | } |
| 157 | |
| 158 | kvm->arch.vgic.in_kernel = true; |
| 159 | kvm->arch.vgic.vgic_model = type; |
| 160 | |
| 161 | kvm->arch.vgic.vgic_dist_base = VGIC_ADDR_UNDEF; |
| 162 | |
| 163 | if (type == KVM_DEV_TYPE_ARM_VGIC_V2) |
| 164 | kvm->arch.vgic.vgic_cpu_base = VGIC_ADDR_UNDEF; |
| 165 | else |
| 166 | INIT_LIST_HEAD(list: &kvm->arch.vgic.rd_regions); |
| 167 | |
| 168 | out_unlock: |
| 169 | mutex_unlock(lock: &kvm->arch.config_lock); |
| 170 | kvm_unlock_all_vcpus(kvm); |
| 171 | return ret; |
| 172 | } |
| 173 | |
| 174 | /* INIT/DESTROY */ |
| 175 | |
| 176 | /** |
| 177 | * kvm_vgic_dist_init: initialize the dist data structures |
| 178 | * @kvm: kvm struct pointer |
| 179 | * @nr_spis: number of spis, frozen by caller |
| 180 | */ |
| 181 | static int kvm_vgic_dist_init(struct kvm *kvm, unsigned int nr_spis) |
| 182 | { |
| 183 | struct vgic_dist *dist = &kvm->arch.vgic; |
| 184 | struct kvm_vcpu *vcpu0 = kvm_get_vcpu(kvm, i: 0); |
| 185 | int i; |
| 186 | |
| 187 | dist->spis = kcalloc(nr_spis, sizeof(struct vgic_irq), GFP_KERNEL_ACCOUNT); |
| 188 | if (!dist->spis) |
| 189 | return -ENOMEM; |
| 190 | |
| 191 | /* |
| 192 | * In the following code we do not take the irq struct lock since |
| 193 | * no other action on irq structs can happen while the VGIC is |
| 194 | * not initialized yet: |
| 195 | * If someone wants to inject an interrupt or does a MMIO access, we |
| 196 | * require prior initialization in case of a virtual GICv3 or trigger |
| 197 | * initialization when using a virtual GICv2. |
| 198 | */ |
| 199 | for (i = 0; i < nr_spis; i++) { |
| 200 | struct vgic_irq *irq = &dist->spis[i]; |
| 201 | |
| 202 | irq->intid = i + VGIC_NR_PRIVATE_IRQS; |
| 203 | INIT_LIST_HEAD(list: &irq->ap_list); |
| 204 | raw_spin_lock_init(&irq->irq_lock); |
| 205 | irq->vcpu = NULL; |
| 206 | irq->target_vcpu = vcpu0; |
| 207 | kref_init(kref: &irq->refcount); |
| 208 | switch (dist->vgic_model) { |
| 209 | case KVM_DEV_TYPE_ARM_VGIC_V2: |
| 210 | irq->targets = 0; |
| 211 | irq->group = 0; |
| 212 | break; |
| 213 | case KVM_DEV_TYPE_ARM_VGIC_V3: |
| 214 | irq->mpidr = 0; |
| 215 | irq->group = 1; |
| 216 | break; |
| 217 | default: |
| 218 | kfree(objp: dist->spis); |
| 219 | dist->spis = NULL; |
| 220 | return -EINVAL; |
| 221 | } |
| 222 | } |
| 223 | return 0; |
| 224 | } |
| 225 | |
| 226 | /* Default GICv3 Maintenance Interrupt INTID, as per SBSA */ |
| 227 | #define DEFAULT_MI_INTID 25 |
| 228 | |
| 229 | int kvm_vgic_vcpu_nv_init(struct kvm_vcpu *vcpu) |
| 230 | { |
| 231 | int ret; |
| 232 | |
| 233 | guard(mutex)(T: &vcpu->kvm->arch.config_lock); |
| 234 | |
| 235 | /* |
| 236 | * Matching the tradition established with the timers, provide |
| 237 | * a default PPI for the maintenance interrupt. It makes |
| 238 | * things easier to reason about. |
| 239 | */ |
| 240 | if (vcpu->kvm->arch.vgic.mi_intid == 0) |
| 241 | vcpu->kvm->arch.vgic.mi_intid = DEFAULT_MI_INTID; |
| 242 | ret = kvm_vgic_set_owner(vcpu, vcpu->kvm->arch.intid: vgic.mi_intid, owner: vcpu); |
| 243 | |
| 244 | return ret; |
| 245 | } |
| 246 | |
| 247 | static int vgic_allocate_private_irqs_locked(struct kvm_vcpu *vcpu, u32 type) |
| 248 | { |
| 249 | struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu; |
| 250 | int i; |
| 251 | |
| 252 | lockdep_assert_held(&vcpu->kvm->arch.config_lock); |
| 253 | |
| 254 | if (vgic_cpu->private_irqs) |
| 255 | return 0; |
| 256 | |
| 257 | vgic_cpu->private_irqs = kcalloc(VGIC_NR_PRIVATE_IRQS, |
| 258 | sizeof(struct vgic_irq), |
| 259 | GFP_KERNEL_ACCOUNT); |
| 260 | |
| 261 | if (!vgic_cpu->private_irqs) |
| 262 | return -ENOMEM; |
| 263 | |
| 264 | /* |
| 265 | * Enable and configure all SGIs to be edge-triggered and |
| 266 | * configure all PPIs as level-triggered. |
| 267 | */ |
| 268 | for (i = 0; i < VGIC_NR_PRIVATE_IRQS; i++) { |
| 269 | struct vgic_irq *irq = &vgic_cpu->private_irqs[i]; |
| 270 | |
| 271 | INIT_LIST_HEAD(list: &irq->ap_list); |
| 272 | raw_spin_lock_init(&irq->irq_lock); |
| 273 | irq->intid = i; |
| 274 | irq->vcpu = NULL; |
| 275 | irq->target_vcpu = vcpu; |
| 276 | kref_init(kref: &irq->refcount); |
| 277 | if (vgic_irq_is_sgi(i)) { |
| 278 | /* SGIs */ |
| 279 | irq->enabled = 1; |
| 280 | irq->config = VGIC_CONFIG_EDGE; |
| 281 | } else { |
| 282 | /* PPIs */ |
| 283 | irq->config = VGIC_CONFIG_LEVEL; |
| 284 | } |
| 285 | |
| 286 | switch (type) { |
| 287 | case KVM_DEV_TYPE_ARM_VGIC_V3: |
| 288 | irq->group = 1; |
| 289 | irq->mpidr = kvm_vcpu_get_mpidr_aff(vcpu); |
| 290 | break; |
| 291 | case KVM_DEV_TYPE_ARM_VGIC_V2: |
| 292 | irq->group = 0; |
| 293 | irq->targets = BIT(vcpu->vcpu_id); |
| 294 | break; |
| 295 | } |
| 296 | } |
| 297 | |
| 298 | return 0; |
| 299 | } |
| 300 | |
| 301 | static int vgic_allocate_private_irqs(struct kvm_vcpu *vcpu, u32 type) |
| 302 | { |
| 303 | int ret; |
| 304 | |
| 305 | mutex_lock(&vcpu->kvm->arch.config_lock); |
| 306 | ret = vgic_allocate_private_irqs_locked(vcpu, type); |
| 307 | mutex_unlock(lock: &vcpu->kvm->arch.config_lock); |
| 308 | |
| 309 | return ret; |
| 310 | } |
| 311 | |
| 312 | /** |
| 313 | * kvm_vgic_vcpu_init() - Initialize static VGIC VCPU data |
| 314 | * structures and register VCPU-specific KVM iodevs |
| 315 | * |
| 316 | * @vcpu: pointer to the VCPU being created and initialized |
| 317 | * |
| 318 | * Only do initialization, but do not actually enable the |
| 319 | * VGIC CPU interface |
| 320 | */ |
| 321 | int kvm_vgic_vcpu_init(struct kvm_vcpu *vcpu) |
| 322 | { |
| 323 | struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu; |
| 324 | struct vgic_dist *dist = &vcpu->kvm->arch.vgic; |
| 325 | int ret = 0; |
| 326 | |
| 327 | vgic_cpu->rd_iodev.base_addr = VGIC_ADDR_UNDEF; |
| 328 | |
| 329 | INIT_LIST_HEAD(list: &vgic_cpu->ap_list_head); |
| 330 | raw_spin_lock_init(&vgic_cpu->ap_list_lock); |
| 331 | atomic_set(v: &vgic_cpu->vgic_v3.its_vpe.vlpi_count, i: 0); |
| 332 | |
| 333 | if (!irqchip_in_kernel(vcpu->kvm)) |
| 334 | return 0; |
| 335 | |
| 336 | ret = vgic_allocate_private_irqs(vcpu, type: dist->vgic_model); |
| 337 | if (ret) |
| 338 | return ret; |
| 339 | |
| 340 | /* |
| 341 | * If we are creating a VCPU with a GICv3 we must also register the |
| 342 | * KVM io device for the redistributor that belongs to this VCPU. |
| 343 | */ |
| 344 | if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) { |
| 345 | mutex_lock(&vcpu->kvm->slots_lock); |
| 346 | ret = vgic_register_redist_iodev(vcpu); |
| 347 | mutex_unlock(lock: &vcpu->kvm->slots_lock); |
| 348 | } |
| 349 | return ret; |
| 350 | } |
| 351 | |
| 352 | static void kvm_vgic_vcpu_enable(struct kvm_vcpu *vcpu) |
| 353 | { |
| 354 | if (kvm_vgic_global_state.type == VGIC_V2) |
| 355 | vgic_v2_enable(vcpu); |
| 356 | else |
| 357 | vgic_v3_enable(vcpu); |
| 358 | } |
| 359 | |
| 360 | /* |
| 361 | * vgic_init: allocates and initializes dist and vcpu data structures |
| 362 | * depending on two dimensioning parameters: |
| 363 | * - the number of spis |
| 364 | * - the number of vcpus |
| 365 | * The function is generally called when nr_spis has been explicitly set |
| 366 | * by the guest through the KVM DEVICE API. If not nr_spis is set to 256. |
| 367 | * vgic_initialized() returns true when this function has succeeded. |
| 368 | */ |
| 369 | int vgic_init(struct kvm *kvm) |
| 370 | { |
| 371 | struct vgic_dist *dist = &kvm->arch.vgic; |
| 372 | struct kvm_vcpu *vcpu; |
| 373 | int ret = 0; |
| 374 | unsigned long idx; |
| 375 | |
| 376 | lockdep_assert_held(&kvm->arch.config_lock); |
| 377 | |
| 378 | if (vgic_initialized(kvm)) |
| 379 | return 0; |
| 380 | |
| 381 | /* Are we also in the middle of creating a VCPU? */ |
| 382 | if (kvm->created_vcpus != atomic_read(v: &kvm->online_vcpus)) |
| 383 | return -EBUSY; |
| 384 | |
| 385 | /* freeze the number of spis */ |
| 386 | if (!dist->nr_spis) |
| 387 | dist->nr_spis = VGIC_NR_IRQS_LEGACY - VGIC_NR_PRIVATE_IRQS; |
| 388 | |
| 389 | ret = kvm_vgic_dist_init(kvm, nr_spis: dist->nr_spis); |
| 390 | if (ret) |
| 391 | goto out; |
| 392 | |
| 393 | /* |
| 394 | * If we have GICv4.1 enabled, unconditionally request enable the |
| 395 | * v4 support so that we get HW-accelerated vSGIs. Otherwise, only |
| 396 | * enable it if we present a virtual ITS to the guest. |
| 397 | */ |
| 398 | if (vgic_supports_direct_msis(kvm)) { |
| 399 | ret = vgic_v4_init(kvm); |
| 400 | if (ret) |
| 401 | goto out; |
| 402 | } |
| 403 | |
| 404 | kvm_for_each_vcpu(idx, vcpu, kvm) |
| 405 | kvm_vgic_vcpu_enable(vcpu); |
| 406 | |
| 407 | ret = kvm_vgic_setup_default_irq_routing(kvm); |
| 408 | if (ret) |
| 409 | goto out; |
| 410 | |
| 411 | vgic_debug_init(kvm); |
| 412 | |
| 413 | /* |
| 414 | * If userspace didn't set the GIC implementation revision, |
| 415 | * default to the latest and greatest. You know want it. |
| 416 | */ |
| 417 | if (!dist->implementation_rev) |
| 418 | dist->implementation_rev = KVM_VGIC_IMP_REV_LATEST; |
| 419 | dist->initialized = true; |
| 420 | |
| 421 | out: |
| 422 | return ret; |
| 423 | } |
| 424 | |
| 425 | static void kvm_vgic_dist_destroy(struct kvm *kvm) |
| 426 | { |
| 427 | struct vgic_dist *dist = &kvm->arch.vgic; |
| 428 | struct vgic_redist_region *rdreg, *next; |
| 429 | |
| 430 | dist->ready = false; |
| 431 | dist->initialized = false; |
| 432 | |
| 433 | kfree(objp: dist->spis); |
| 434 | dist->spis = NULL; |
| 435 | dist->nr_spis = 0; |
| 436 | dist->vgic_dist_base = VGIC_ADDR_UNDEF; |
| 437 | |
| 438 | if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) { |
| 439 | list_for_each_entry_safe(rdreg, next, &dist->rd_regions, list) |
| 440 | vgic_v3_free_redist_region(kvm, rdreg); |
| 441 | INIT_LIST_HEAD(list: &dist->rd_regions); |
| 442 | } else { |
| 443 | dist->vgic_cpu_base = VGIC_ADDR_UNDEF; |
| 444 | } |
| 445 | |
| 446 | if (vgic_supports_direct_msis(kvm)) |
| 447 | vgic_v4_teardown(kvm); |
| 448 | |
| 449 | xa_destroy(&dist->lpi_xa); |
| 450 | } |
| 451 | |
| 452 | static void __kvm_vgic_vcpu_destroy(struct kvm_vcpu *vcpu) |
| 453 | { |
| 454 | struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu; |
| 455 | |
| 456 | /* |
| 457 | * Retire all pending LPIs on this vcpu anyway as we're |
| 458 | * going to destroy it. |
| 459 | */ |
| 460 | vgic_flush_pending_lpis(vcpu); |
| 461 | |
| 462 | INIT_LIST_HEAD(list: &vgic_cpu->ap_list_head); |
| 463 | kfree(objp: vgic_cpu->private_irqs); |
| 464 | vgic_cpu->private_irqs = NULL; |
| 465 | |
| 466 | if (vcpu->kvm->arch.vgic.vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) { |
| 467 | /* |
| 468 | * If this vCPU is being destroyed because of a failed creation |
| 469 | * then unregister the redistributor to avoid leaving behind a |
| 470 | * dangling pointer to the vCPU struct. |
| 471 | * |
| 472 | * vCPUs that have been successfully created (i.e. added to |
| 473 | * kvm->vcpu_array) get unregistered in kvm_vgic_destroy(), as |
| 474 | * this function gets called while holding kvm->arch.config_lock |
| 475 | * in the VM teardown path and would otherwise introduce a lock |
| 476 | * inversion w.r.t. kvm->srcu. |
| 477 | * |
| 478 | * vCPUs that failed creation are torn down outside of the |
| 479 | * kvm->arch.config_lock and do not get unregistered in |
| 480 | * kvm_vgic_destroy(), meaning it is both safe and necessary to |
| 481 | * do so here. |
| 482 | */ |
| 483 | if (kvm_get_vcpu_by_id(kvm: vcpu->kvm, id: vcpu->vcpu_id) != vcpu) |
| 484 | vgic_unregister_redist_iodev(vcpu); |
| 485 | |
| 486 | vgic_cpu->rd_iodev.base_addr = VGIC_ADDR_UNDEF; |
| 487 | } |
| 488 | } |
| 489 | |
| 490 | void kvm_vgic_vcpu_destroy(struct kvm_vcpu *vcpu) |
| 491 | { |
| 492 | struct kvm *kvm = vcpu->kvm; |
| 493 | |
| 494 | mutex_lock(&kvm->slots_lock); |
| 495 | __kvm_vgic_vcpu_destroy(vcpu); |
| 496 | mutex_unlock(lock: &kvm->slots_lock); |
| 497 | } |
| 498 | |
| 499 | void kvm_vgic_destroy(struct kvm *kvm) |
| 500 | { |
| 501 | struct kvm_vcpu *vcpu; |
| 502 | unsigned long i; |
| 503 | |
| 504 | mutex_lock(&kvm->slots_lock); |
| 505 | mutex_lock(&kvm->arch.config_lock); |
| 506 | |
| 507 | vgic_debug_destroy(kvm); |
| 508 | |
| 509 | kvm_for_each_vcpu(i, vcpu, kvm) |
| 510 | __kvm_vgic_vcpu_destroy(vcpu); |
| 511 | |
| 512 | kvm_vgic_dist_destroy(kvm); |
| 513 | |
| 514 | mutex_unlock(lock: &kvm->arch.config_lock); |
| 515 | |
| 516 | if (kvm->arch.vgic.vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3) |
| 517 | kvm_for_each_vcpu(i, vcpu, kvm) |
| 518 | vgic_unregister_redist_iodev(vcpu); |
| 519 | |
| 520 | mutex_unlock(lock: &kvm->slots_lock); |
| 521 | } |
| 522 | |
| 523 | /** |
| 524 | * vgic_lazy_init: Lazy init is only allowed if the GIC exposed to the guest |
| 525 | * is a GICv2. A GICv3 must be explicitly initialized by userspace using the |
| 526 | * KVM_DEV_ARM_VGIC_GRP_CTRL KVM_DEVICE group. |
| 527 | * @kvm: kvm struct pointer |
| 528 | */ |
| 529 | int vgic_lazy_init(struct kvm *kvm) |
| 530 | { |
| 531 | int ret = 0; |
| 532 | |
| 533 | if (unlikely(!vgic_initialized(kvm))) { |
| 534 | /* |
| 535 | * We only provide the automatic initialization of the VGIC |
| 536 | * for the legacy case of a GICv2. Any other type must |
| 537 | * be explicitly initialized once setup with the respective |
| 538 | * KVM device call. |
| 539 | */ |
| 540 | if (kvm->arch.vgic.vgic_model != KVM_DEV_TYPE_ARM_VGIC_V2) |
| 541 | return -EBUSY; |
| 542 | |
| 543 | mutex_lock(&kvm->arch.config_lock); |
| 544 | ret = vgic_init(kvm); |
| 545 | mutex_unlock(lock: &kvm->arch.config_lock); |
| 546 | } |
| 547 | |
| 548 | return ret; |
| 549 | } |
| 550 | |
| 551 | /* RESOURCE MAPPING */ |
| 552 | |
| 553 | /** |
| 554 | * kvm_vgic_map_resources - map the MMIO regions |
| 555 | * @kvm: kvm struct pointer |
| 556 | * |
| 557 | * Map the MMIO regions depending on the VGIC model exposed to the guest |
| 558 | * called on the first VCPU run. |
| 559 | * Also map the virtual CPU interface into the VM. |
| 560 | * v2 calls vgic_init() if not already done. |
| 561 | * v3 and derivatives return an error if the VGIC is not initialized. |
| 562 | * vgic_ready() returns true if this function has succeeded. |
| 563 | */ |
| 564 | int kvm_vgic_map_resources(struct kvm *kvm) |
| 565 | { |
| 566 | struct vgic_dist *dist = &kvm->arch.vgic; |
| 567 | enum vgic_type type; |
| 568 | gpa_t dist_base; |
| 569 | int ret = 0; |
| 570 | |
| 571 | if (likely(vgic_ready(kvm))) |
| 572 | return 0; |
| 573 | |
| 574 | mutex_lock(&kvm->slots_lock); |
| 575 | mutex_lock(&kvm->arch.config_lock); |
| 576 | if (vgic_ready(kvm)) |
| 577 | goto out; |
| 578 | |
| 579 | if (!irqchip_in_kernel(kvm)) |
| 580 | goto out; |
| 581 | |
| 582 | if (dist->vgic_model == KVM_DEV_TYPE_ARM_VGIC_V2) { |
| 583 | ret = vgic_v2_map_resources(kvm); |
| 584 | type = VGIC_V2; |
| 585 | } else { |
| 586 | ret = vgic_v3_map_resources(kvm); |
| 587 | type = VGIC_V3; |
| 588 | } |
| 589 | |
| 590 | if (ret) |
| 591 | goto out; |
| 592 | |
| 593 | dist_base = dist->vgic_dist_base; |
| 594 | mutex_unlock(lock: &kvm->arch.config_lock); |
| 595 | |
| 596 | ret = vgic_register_dist_iodev(kvm, dist_base_address: dist_base, type); |
| 597 | if (ret) { |
| 598 | kvm_err("Unable to register VGIC dist MMIO regions\n" ); |
| 599 | goto out_slots; |
| 600 | } |
| 601 | |
| 602 | /* |
| 603 | * kvm_io_bus_register_dev() guarantees all readers see the new MMIO |
| 604 | * registration before returning through synchronize_srcu(), which also |
| 605 | * implies a full memory barrier. As such, marking the distributor as |
| 606 | * 'ready' here is guaranteed to be ordered after all vCPUs having seen |
| 607 | * a completely configured distributor. |
| 608 | */ |
| 609 | dist->ready = true; |
| 610 | goto out_slots; |
| 611 | out: |
| 612 | mutex_unlock(lock: &kvm->arch.config_lock); |
| 613 | out_slots: |
| 614 | if (ret) |
| 615 | kvm_vm_dead(kvm); |
| 616 | |
| 617 | mutex_unlock(lock: &kvm->slots_lock); |
| 618 | |
| 619 | return ret; |
| 620 | } |
| 621 | |
| 622 | /* GENERIC PROBE */ |
| 623 | |
| 624 | void kvm_vgic_cpu_up(void) |
| 625 | { |
| 626 | enable_percpu_irq(irq: kvm_vgic_global_state.maint_irq, type: 0); |
| 627 | } |
| 628 | |
| 629 | |
| 630 | void kvm_vgic_cpu_down(void) |
| 631 | { |
| 632 | disable_percpu_irq(irq: kvm_vgic_global_state.maint_irq); |
| 633 | } |
| 634 | |
| 635 | static irqreturn_t vgic_maintenance_handler(int irq, void *data) |
| 636 | { |
| 637 | struct kvm_vcpu *vcpu = *(struct kvm_vcpu **)data; |
| 638 | |
| 639 | /* |
| 640 | * We cannot rely on the vgic maintenance interrupt to be |
| 641 | * delivered synchronously. This means we can only use it to |
| 642 | * exit the VM, and we perform the handling of EOIed |
| 643 | * interrupts on the exit path (see vgic_fold_lr_state). |
| 644 | * |
| 645 | * Of course, NV throws a wrench in this plan, and needs |
| 646 | * something special. |
| 647 | */ |
| 648 | if (vcpu && vgic_state_is_nested(vcpu)) |
| 649 | vgic_v3_handle_nested_maint_irq(vcpu); |
| 650 | |
| 651 | return IRQ_HANDLED; |
| 652 | } |
| 653 | |
| 654 | static struct gic_kvm_info *gic_kvm_info; |
| 655 | |
| 656 | void __init vgic_set_kvm_info(const struct gic_kvm_info *info) |
| 657 | { |
| 658 | BUG_ON(gic_kvm_info != NULL); |
| 659 | gic_kvm_info = kmalloc(sizeof(*info), GFP_KERNEL); |
| 660 | if (gic_kvm_info) |
| 661 | *gic_kvm_info = *info; |
| 662 | } |
| 663 | |
| 664 | /** |
| 665 | * kvm_vgic_init_cpu_hardware - initialize the GIC VE hardware |
| 666 | * |
| 667 | * For a specific CPU, initialize the GIC VE hardware. |
| 668 | */ |
| 669 | void kvm_vgic_init_cpu_hardware(void) |
| 670 | { |
| 671 | BUG_ON(preemptible()); |
| 672 | |
| 673 | /* |
| 674 | * We want to make sure the list registers start out clear so that we |
| 675 | * only have the program the used registers. |
| 676 | */ |
| 677 | if (kvm_vgic_global_state.type == VGIC_V2) |
| 678 | vgic_v2_init_lrs(); |
| 679 | else |
| 680 | kvm_call_hyp(__vgic_v3_init_lrs); |
| 681 | } |
| 682 | |
| 683 | /** |
| 684 | * kvm_vgic_hyp_init: populates the kvm_vgic_global_state variable |
| 685 | * according to the host GIC model. Accordingly calls either |
| 686 | * vgic_v2/v3_probe which registers the KVM_DEVICE that can be |
| 687 | * instantiated by a guest later on . |
| 688 | */ |
| 689 | int kvm_vgic_hyp_init(void) |
| 690 | { |
| 691 | bool has_mask; |
| 692 | int ret; |
| 693 | |
| 694 | if (!gic_kvm_info) |
| 695 | return -ENODEV; |
| 696 | |
| 697 | has_mask = !gic_kvm_info->no_maint_irq_mask; |
| 698 | |
| 699 | if (has_mask && !gic_kvm_info->maint_irq) { |
| 700 | kvm_err("No vgic maintenance irq\n" ); |
| 701 | return -ENXIO; |
| 702 | } |
| 703 | |
| 704 | /* |
| 705 | * If we get one of these oddball non-GICs, taint the kernel, |
| 706 | * as we have no idea of how they *really* behave. |
| 707 | */ |
| 708 | if (gic_kvm_info->no_hw_deactivation) { |
| 709 | kvm_info("Non-architectural vgic, tainting kernel\n" ); |
| 710 | add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK); |
| 711 | kvm_vgic_global_state.no_hw_deactivation = true; |
| 712 | } |
| 713 | |
| 714 | switch (gic_kvm_info->type) { |
| 715 | case GIC_V2: |
| 716 | ret = vgic_v2_probe(info: gic_kvm_info); |
| 717 | break; |
| 718 | case GIC_V3: |
| 719 | ret = vgic_v3_probe(info: gic_kvm_info); |
| 720 | if (!ret) { |
| 721 | static_branch_enable(&kvm_vgic_global_state.gicv3_cpuif); |
| 722 | kvm_info("GIC system register CPU interface enabled\n" ); |
| 723 | } |
| 724 | break; |
| 725 | default: |
| 726 | ret = -ENODEV; |
| 727 | } |
| 728 | |
| 729 | kvm_vgic_global_state.maint_irq = gic_kvm_info->maint_irq; |
| 730 | |
| 731 | kfree(objp: gic_kvm_info); |
| 732 | gic_kvm_info = NULL; |
| 733 | |
| 734 | if (ret) |
| 735 | return ret; |
| 736 | |
| 737 | if (!has_mask && !kvm_vgic_global_state.maint_irq) |
| 738 | return 0; |
| 739 | |
| 740 | ret = request_percpu_irq(irq: kvm_vgic_global_state.maint_irq, |
| 741 | handler: vgic_maintenance_handler, |
| 742 | devname: "vgic" , percpu_dev_id: kvm_get_running_vcpus()); |
| 743 | if (ret) { |
| 744 | kvm_err("Cannot register interrupt %d\n" , |
| 745 | kvm_vgic_global_state.maint_irq); |
| 746 | return ret; |
| 747 | } |
| 748 | |
| 749 | kvm_info("vgic interrupt IRQ%d\n" , kvm_vgic_global_state.maint_irq); |
| 750 | return 0; |
| 751 | } |
| 752 | |