arch/arm64/kvm/vgic/vgic-v4.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2017 ARM Ltd.
   4  * Author: Marc Zyngier <marc.zyngier@arm.com>
   5  */
   6
   7 #include <linux/interrupt.h>
   8 #include <linux/irq.h>
   9 #include <linux/irqdomain.h>
  10 #include <linux/kvm_host.h>
  11 #include <linux/irqchip/arm-gic-v3.h>
  12
  13 #include "vgic.h"
  14
  15 /*
  16  * How KVM uses GICv4 (insert rude comments here):
  17  *
  18  * The vgic-v4 layer acts as a bridge between several entities:
  19  * - The GICv4 ITS representation offered by the ITS driver
  20  * - VFIO, which is in charge of the PCI endpoint
  21  * - The virtual ITS, which is the only thing the guest sees
  22  *
  23  * The configuration of VLPIs is triggered by a callback from VFIO,
  24  * instructing KVM that a PCI device has been configured to deliver
  25  * MSIs to a vITS.
  26  *
  27  * kvm_vgic_v4_set_forwarding() is thus called with the routing entry,
  28  * and this is used to find the corresponding vITS data structures
  29  * (ITS instance, device, event and irq) using a process that is
  30  * extremely similar to the injection of an MSI.
  31  *
  32  * At this stage, we can link the guest's view of an LPI (uniquely
  33  * identified by the routing entry) and the host irq, using the GICv4
  34  * driver mapping operation. Should the mapping succeed, we've then
  35  * successfully upgraded the guest's LPI to a VLPI. We can then start
  36  * with updating GICv4's view of the property table and generating an
  37  * INValidation in order to kickstart the delivery of this VLPI to the
  38  * guest directly, without software intervention. Well, almost.
  39  *
  40  * When the PCI endpoint is deconfigured, this operation is reversed
  41  * with VFIO calling kvm_vgic_v4_unset_forwarding().
  42  *
  43  * Once the VLPI has been mapped, it needs to follow any change the
  44  * guest performs on its LPI through the vITS. For that, a number of
  45  * command handlers have hooks to communicate these changes to the HW:
  46  * - Any invalidation triggers a call to its_prop_update_vlpi()
  47  * - The INT command results in a irq_set_irqchip_state(), which
  48  *   generates an INT on the corresponding VLPI.
  49  * - The CLEAR command results in a irq_set_irqchip_state(), which
  50  *   generates an CLEAR on the corresponding VLPI.
  51  * - DISCARD translates into an unmap, similar to a call to
  52  *   kvm_vgic_v4_unset_forwarding().
  53  * - MOVI is translated by an update of the existing mapping, changing
  54  *   the target vcpu, resulting in a VMOVI being generated.
  55  * - MOVALL is translated by a string of mapping updates (similar to
  56  *   the handling of MOVI). MOVALL is horrible.
  57  *
  58  * Note that a DISCARD/MAPTI sequence emitted from the guest without
  59  * reprogramming the PCI endpoint after MAPTI does not result in a
  60  * VLPI being mapped, as there is no callback from VFIO (the guest
  61  * will get the interrupt via the normal SW injection). Fixing this is
  62  * not trivial, and requires some horrible messing with the VFIO
  63  * internals. Not fun. Don't do that.
  64  *
  65  * Then there is the scheduling. Each time a vcpu is about to run on a
  66  * physical CPU, KVM must tell the corresponding redistributor about
  67  * it. And if we've migrated our vcpu from one CPU to another, we must
  68  * tell the ITS (so that the messages reach the right redistributor).
  69  * This is done in two steps: first issue a irq_set_affinity() on the
  70  * irq corresponding to the vcpu, then call its_make_vpe_resident().
  71  * You must be in a non-preemptible context. On exit, a call to
  72  * its_make_vpe_non_resident() tells the redistributor that we're done
  73  * with the vcpu.
  74  *
  75  * Finally, the doorbell handling: Each vcpu is allocated an interrupt
  76  * which will fire each time a VLPI is made pending whilst the vcpu is
  77  * not running. Each time the vcpu gets blocked, the doorbell
  78  * interrupt gets enabled. When the vcpu is unblocked (for whatever
  79  * reason), the doorbell interrupt is disabled.
  80  */
  81
  82 #define DB_IRQ_FLAGS    (IRQ_NOAUTOEN | IRQ_DISABLE_UNLAZY | IRQ_NO_BALANCING)
  83
  84 static irqreturn_t vgic_v4_doorbell_handler(int irq, void *info)
  85 {
  86         struct kvm_vcpu *vcpu = info;
  87
  88         /* We got the message, no need to fire again */
  89         if (!kvm_vgic_global_state.has_gicv4_1 &&
  90             !irqd_irq_disabled(&irq_to_desc(irq)->irq_data))
  91                 disable_irq_nosync(irq);
  92
  93         /*
  94          * The v4.1 doorbell can fire concurrently with the vPE being
  95          * made non-resident. Ensure we only update pending_last
  96          * *after* the non-residency sequence has completed.
  97          */
  98         raw_spin_lock(&vcpu->arch.vgic_cpu.vgic_v3.its_vpe.vpe_lock);
  99         vcpu->arch.vgic_cpu.vgic_v3.its_vpe.pending_last = true;
 100         raw_spin_unlock(&vcpu->arch.vgic_cpu.vgic_v3.its_vpe.vpe_lock);
 101
 102         kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
 103         kvm_vcpu_kick(vcpu);
 104
 105         return IRQ_HANDLED;
 106 }
 107
 108 static void vgic_v4_sync_sgi_config(struct its_vpe *vpe, struct vgic_irq *irq)
 109 {
 110         vpe->sgi_config[irq->intid].enabled     = irq->enabled;
 111         vpe->sgi_config[irq->intid].group       = irq->group;
 112         vpe->sgi_config[irq->intid].priority    = irq->priority;
 113 }
 114
 115 static void vgic_v4_enable_vsgis(struct kvm_vcpu *vcpu)
 116 {
 117         struct its_vpe *vpe = &vcpu->arch.vgic_cpu.vgic_v3.its_vpe;
 118         int i;
 119
 120         /*
 121          * With GICv4.1, every virtual SGI can be directly injected. So
 122          * let's pretend that they are HW interrupts, tied to a host
 123          * IRQ. The SGI code will do its magic.
 124          */
 125         for (i = 0; i < VGIC_NR_SGIS; i++) {
 126                 struct vgic_irq *irq = vgic_get_irq(vcpu->kvm, vcpu, i);
 127                 struct irq_desc *desc;
 128                 unsigned long flags;
 129                 int ret;
 130
 131                 raw_spin_lock_irqsave(&irq->irq_lock, flags);
 132
 133                 if (irq->hw)
 134                         goto unlock;
 135
 136                 irq->hw = true;
 137                 irq->host_irq = irq_find_mapping(vpe->sgi_domain, i);
 138
 139                 /* Transfer the full irq state to the vPE */
 140                 vgic_v4_sync_sgi_config(vpe, irq);
 141                 desc = irq_to_desc(irq->host_irq);
 142                 ret = irq_domain_activate_irq(irq_desc_get_irq_data(desc),
 143                                               false);
 144                 if (!WARN_ON(ret)) {
 145                         /* Transfer pending state */
 146                         ret = irq_set_irqchip_state(irq->host_irq,
 147                                                     IRQCHIP_STATE_PENDING,
 148                                                     irq->pending_latch);
 149                         WARN_ON(ret);
 150                         irq->pending_latch = false;
 151                 }
 152         unlock:
 153                 raw_spin_unlock_irqrestore(&irq->irq_lock, flags);
 154                 vgic_put_irq(vcpu->kvm, irq);
 155         }
 156 }
 157
 158 static void vgic_v4_disable_vsgis(struct kvm_vcpu *vcpu)
 159 {
 160         int i;
 161
 162         for (i = 0; i < VGIC_NR_SGIS; i++) {
 163                 struct vgic_irq *irq = vgic_get_irq(vcpu->kvm, vcpu, i);
 164                 struct irq_desc *desc;
 165                 unsigned long flags;
 166                 int ret;
 167
 168                 raw_spin_lock_irqsave(&irq->irq_lock, flags);
 169
 170                 if (!irq->hw)
 171                         goto unlock;
 172
 173                 irq->hw = false;
 174                 ret = irq_get_irqchip_state(irq->host_irq,
 175                                             IRQCHIP_STATE_PENDING,
 176                                             &irq->pending_latch);
 177                 WARN_ON(ret);
 178
 179                 desc = irq_to_desc(irq->host_irq);
 180                 irq_domain_deactivate_irq(irq_desc_get_irq_data(desc));
 181         unlock:
 182                 raw_spin_unlock_irqrestore(&irq->irq_lock, flags);
 183                 vgic_put_irq(vcpu->kvm, irq);
 184         }
 185 }
 186
 187 /* Must be called with the kvm lock held */
 188 void vgic_v4_configure_vsgis(struct kvm *kvm)
 189 {
 190         struct vgic_dist *dist = &kvm->arch.vgic;
 191         struct kvm_vcpu *vcpu;
 192         unsigned long i;
 193
 194         kvm_arm_halt_guest(kvm);
 195
 196         kvm_for_each_vcpu(i, vcpu, kvm) {
 197                 if (dist->nassgireq)
 198                         vgic_v4_enable_vsgis(vcpu);
 199                 else
 200                         vgic_v4_disable_vsgis(vcpu);
 201         }
 202
 203         kvm_arm_resume_guest(kvm);
 204 }
 205
 206 /*
 207  * Must be called with GICv4.1 and the vPE unmapped, which
 208  * indicates the invalidation of any VPT caches associated
 209  * with the vPE, thus we can get the VLPI state by peeking
 210  * at the VPT.
 211  */
 212 void vgic_v4_get_vlpi_state(struct vgic_irq *irq, bool *val)
 213 {
 214         struct its_vpe *vpe = &irq->target_vcpu->arch.vgic_cpu.vgic_v3.its_vpe;
 215         int mask = BIT(irq->intid % BITS_PER_BYTE);
 216         void *va;
 217         u8 *ptr;
 218
 219         va = page_address(vpe->vpt_page);
 220         ptr = va + irq->intid / BITS_PER_BYTE;
 221
 222         *val = !!(*ptr & mask);
 223 }
 224
 225 int vgic_v4_request_vpe_irq(struct kvm_vcpu *vcpu, int irq)
 226 {
 227         return request_irq(irq, vgic_v4_doorbell_handler, 0, "vcpu", vcpu);
 228 }
 229
 230 /**
 231  * vgic_v4_init - Initialize the GICv4 data structures
 232  * @kvm:        Pointer to the VM being initialized
 233  *
 234  * We may be called each time a vITS is created, or when the
 235  * vgic is initialized. In both cases, the number of vcpus
 236  * should now be fixed.
 237  */
 238 int vgic_v4_init(struct kvm *kvm)
 239 {
 240         struct vgic_dist *dist = &kvm->arch.vgic;
 241         struct kvm_vcpu *vcpu;
 242         int nr_vcpus, ret;
 243         unsigned long i;
 244
 245         lockdep_assert_held(&kvm->arch.config_lock);
 246
 247         if (!kvm_vgic_global_state.has_gicv4)
 248                 return 0; /* Nothing to see here... move along. */
 249
 250         if (dist->its_vm.vpes)
 251                 return 0;
 252
 253         nr_vcpus = atomic_read(&kvm->online_vcpus);
 254
 255         dist->its_vm.vpes = kcalloc(nr_vcpus, sizeof(*dist->its_vm.vpes),
 256                                     GFP_KERNEL_ACCOUNT);
 257         if (!dist->its_vm.vpes)
 258                 return -ENOMEM;
 259
 260         dist->its_vm.nr_vpes = nr_vcpus;
 261
 262         kvm_for_each_vcpu(i, vcpu, kvm)
 263                 dist->its_vm.vpes[i] = &vcpu->arch.vgic_cpu.vgic_v3.its_vpe;
 264
 265         ret = its_alloc_vcpu_irqs(&dist->its_vm);
 266         if (ret < 0) {
 267                 kvm_err("VPE IRQ allocation failure\n");
 268                 kfree(dist->its_vm.vpes);
 269                 dist->its_vm.nr_vpes = 0;
 270                 dist->its_vm.vpes = NULL;
 271                 return ret;
 272         }
 273
 274         kvm_for_each_vcpu(i, vcpu, kvm) {
 275                 int irq = dist->its_vm.vpes[i]->irq;
 276                 unsigned long irq_flags = DB_IRQ_FLAGS;
 277
 278                 /*
 279                  * Don't automatically enable the doorbell, as we're
 280                  * flipping it back and forth when the vcpu gets
 281                  * blocked. Also disable the lazy disabling, as the
 282                  * doorbell could kick us out of the guest too
 283                  * early...
 284                  *
 285                  * On GICv4.1, the doorbell is managed in HW and must
 286                  * be left enabled.
 287                  */
 288                 if (kvm_vgic_global_state.has_gicv4_1)
 289                         irq_flags &= ~IRQ_NOAUTOEN;
 290                 irq_set_status_flags(irq, irq_flags);
 291
 292                 ret = vgic_v4_request_vpe_irq(vcpu, irq);
 293                 if (ret) {
 294                         kvm_err("failed to allocate vcpu IRQ%d\n", irq);
 295                         /*
 296                          * Trick: adjust the number of vpes so we know
 297                          * how many to nuke on teardown...
 298                          */
 299                         dist->its_vm.nr_vpes = i;
 300                         break;
 301                 }
 302         }
 303
 304         if (ret)
 305                 vgic_v4_teardown(kvm);
 306
 307         return ret;
 308 }
 309
 310 /**
 311  * vgic_v4_teardown - Free the GICv4 data structures
 312  * @kvm:        Pointer to the VM being destroyed
 313  */
 314 void vgic_v4_teardown(struct kvm *kvm)
 315 {
 316         struct its_vm *its_vm = &kvm->arch.vgic.its_vm;
 317         int i;
 318
 319         lockdep_assert_held(&kvm->arch.config_lock);
 320
 321         if (!its_vm->vpes)
 322                 return;
 323
 324         for (i = 0; i < its_vm->nr_vpes; i++) {
 325                 struct kvm_vcpu *vcpu = kvm_get_vcpu(kvm, i);
 326                 int irq = its_vm->vpes[i]->irq;
 327
 328                 irq_clear_status_flags(irq, DB_IRQ_FLAGS);
 329                 free_irq(irq, vcpu);
 330         }
 331
 332         its_free_vcpu_irqs(its_vm);
 333         kfree(its_vm->vpes);
 334         its_vm->nr_vpes = 0;
 335         its_vm->vpes = NULL;
 336 }
 337
 338 int vgic_v4_put(struct kvm_vcpu *vcpu, bool need_db)
 339 {
 340         struct its_vpe *vpe = &vcpu->arch.vgic_cpu.vgic_v3.its_vpe;
 341
 342         if (!vgic_supports_direct_msis(vcpu->kvm) || !vpe->resident)
 343                 return 0;
 344
 345         return its_make_vpe_non_resident(vpe, need_db);
 346 }
 347
 348 int vgic_v4_load(struct kvm_vcpu *vcpu)
 349 {
 350         struct its_vpe *vpe = &vcpu->arch.vgic_cpu.vgic_v3.its_vpe;
 351         int err;
 352
 353         if (!vgic_supports_direct_msis(vcpu->kvm) || vpe->resident)
 354                 return 0;
 355
 356         /*
 357          * Before making the VPE resident, make sure the redistributor
 358          * corresponding to our current CPU expects us here. See the
 359          * doc in drivers/irqchip/irq-gic-v4.c to understand how this
 360          * turns into a VMOVP command at the ITS level.
 361          */
 362         err = irq_set_affinity(vpe->irq, cpumask_of(smp_processor_id()));
 363         if (err)
 364                 return err;
 365
 366         err = its_make_vpe_resident(vpe, false, vcpu->kvm->arch.vgic.enabled);
 367         if (err)
 368                 return err;
 369
 370         /*
 371          * Now that the VPE is resident, let's get rid of a potential
 372          * doorbell interrupt that would still be pending. This is a
 373          * GICv4.0 only "feature"...
 374          */
 375         if (!kvm_vgic_global_state.has_gicv4_1)
 376                 err = irq_set_irqchip_state(vpe->irq, IRQCHIP_STATE_PENDING, false);
 377
 378         return err;
 379 }
 380
 381 void vgic_v4_commit(struct kvm_vcpu *vcpu)
 382 {
 383         struct its_vpe *vpe = &vcpu->arch.vgic_cpu.vgic_v3.its_vpe;
 384
 385         /*
 386          * No need to wait for the vPE to be ready across a shallow guest
 387          * exit, as only a vcpu_put will invalidate it.
 388          */
 389         if (!vpe->ready)
 390                 its_commit_vpe(vpe);
 391 }
 392
 393 static struct vgic_its *vgic_get_its(struct kvm *kvm,
 394                                      struct kvm_kernel_irq_routing_entry *irq_entry)
 395 {
 396         struct kvm_msi msi  = (struct kvm_msi) {
 397                 .address_lo     = irq_entry->msi.address_lo,
 398                 .address_hi     = irq_entry->msi.address_hi,
 399                 .data           = irq_entry->msi.data,
 400                 .flags          = irq_entry->msi.flags,
 401                 .devid          = irq_entry->msi.devid,
 402         };
 403
 404         return vgic_msi_to_its(kvm, &msi);
 405 }
 406
 407 int kvm_vgic_v4_set_forwarding(struct kvm *kvm, int virq,
 408                                struct kvm_kernel_irq_routing_entry *irq_entry)
 409 {
 410         struct vgic_its *its;
 411         struct vgic_irq *irq;
 412         struct its_vlpi_map map;
 413         unsigned long flags;
 414         int ret;
 415
 416         if (!vgic_supports_direct_msis(kvm))
 417                 return 0;
 418
 419         /*
 420          * Get the ITS, and escape early on error (not a valid
 421          * doorbell for any of our vITSs).
 422          */
 423         its = vgic_get_its(kvm, irq_entry);
 424         if (IS_ERR(its))
 425                 return 0;
 426
 427         mutex_lock(&its->its_lock);
 428
 429         /* Perform the actual DevID/EventID -> LPI translation. */
 430         ret = vgic_its_resolve_lpi(kvm, its, irq_entry->msi.devid,
 431                                    irq_entry->msi.data, &irq);
 432         if (ret)
 433                 goto out;
 434
 435         /*
 436          * Emit the mapping request. If it fails, the ITS probably
 437          * isn't v4 compatible, so let's silently bail out. Holding
 438          * the ITS lock should ensure that nothing can modify the
 439          * target vcpu.
 440          */
 441         map = (struct its_vlpi_map) {
 442                 .vm             = &kvm->arch.vgic.its_vm,
 443                 .vpe            = &irq->target_vcpu->arch.vgic_cpu.vgic_v3.its_vpe,
 444                 .vintid         = irq->intid,
 445                 .properties     = ((irq->priority & 0xfc) |
 446                                    (irq->enabled ? LPI_PROP_ENABLED : 0) |
 447                                    LPI_PROP_GROUP1),
 448                 .db_enabled     = true,
 449         };
 450
 451         ret = its_map_vlpi(virq, &map);
 452         if (ret)
 453                 goto out;
 454
 455         irq->hw         = true;
 456         irq->host_irq   = virq;
 457         atomic_inc(&map.vpe->vlpi_count);
 458
 459         /* Transfer pending state */
 460         raw_spin_lock_irqsave(&irq->irq_lock, flags);
 461         if (irq->pending_latch) {
 462                 ret = irq_set_irqchip_state(irq->host_irq,
 463                                             IRQCHIP_STATE_PENDING,
 464                                             irq->pending_latch);
 465                 WARN_RATELIMIT(ret, "IRQ %d", irq->host_irq);
 466
 467                 /*
 468                  * Clear pending_latch and communicate this state
 469                  * change via vgic_queue_irq_unlock.
 470                  */
 471                 irq->pending_latch = false;
 472                 vgic_queue_irq_unlock(kvm, irq, flags);
 473         } else {
 474                 raw_spin_unlock_irqrestore(&irq->irq_lock, flags);
 475         }
 476
 477 out:
 478         mutex_unlock(&its->its_lock);
 479         return ret;
 480 }
 481
 482 int kvm_vgic_v4_unset_forwarding(struct kvm *kvm, int virq,
 483                                  struct kvm_kernel_irq_routing_entry *irq_entry)
 484 {
 485         struct vgic_its *its;
 486         struct vgic_irq *irq;
 487         int ret;
 488
 489         if (!vgic_supports_direct_msis(kvm))
 490                 return 0;
 491
 492         /*
 493          * Get the ITS, and escape early on error (not a valid
 494          * doorbell for any of our vITSs).
 495          */
 496         its = vgic_get_its(kvm, irq_entry);
 497         if (IS_ERR(its))
 498                 return 0;
 499
 500         mutex_lock(&its->its_lock);
 501
 502         ret = vgic_its_resolve_lpi(kvm, its, irq_entry->msi.devid,
 503                                    irq_entry->msi.data, &irq);
 504         if (ret)
 505                 goto out;
 506
 507         WARN_ON(!(irq->hw && irq->host_irq == virq));
 508         if (irq->hw) {
 509                 atomic_dec(&irq->target_vcpu->arch.vgic_cpu.vgic_v3.its_vpe.vlpi_count);
 510                 irq->hw = false;
 511                 ret = its_unmap_vlpi(virq);
 512         }
 513
 514 out:
 515         mutex_unlock(&its->its_lock);
 516         return ret;
 517 }