virt/kvm/arm/vgic/vgic.c

   1 /*
   2  * Copyright (C) 2015, 2016 ARM Ltd.
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License version 2 as
   6  * published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  11  * GNU General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public License
  14  * along with this program.  If not, see <http://www.gnu.org/licenses/>.
  15  */
  16
  17 #include <linux/interrupt.h>
  18 #include <linux/irq.h>
  19 #include <linux/kvm.h>
  20 #include <linux/kvm_host.h>
  21 #include <linux/list_sort.h>
  22 #include <linux/nospec.h>
  23
  24 #include <asm/kvm_hyp.h>
  25
  26 #include "vgic.h"
  27
  28 #define CREATE_TRACE_POINTS
  29 #include "trace.h"
  30
  31 #ifdef CONFIG_DEBUG_SPINLOCK
  32 #define DEBUG_SPINLOCK_BUG_ON(p) BUG_ON(p)
  33 #else
  34 #define DEBUG_SPINLOCK_BUG_ON(p)
  35 #endif
  36
  37 struct vgic_global kvm_vgic_global_state __ro_after_init = {
  38         .gicv3_cpuif = STATIC_KEY_FALSE_INIT,
  39 };
  40
  41 /*
  42  * Locking order is always:
  43  * kvm->lock (mutex)
  44  *   its->cmd_lock (mutex)
  45  *     its->its_lock (mutex)
  46  *       vgic_cpu->ap_list_lock         must be taken with IRQs disabled
  47  *         kvm->lpi_list_lock           must be taken with IRQs disabled
  48  *           vgic_irq->irq_lock         must be taken with IRQs disabled
  49  *
  50  * As the ap_list_lock might be taken from the timer interrupt handler,
  51  * we have to disable IRQs before taking this lock and everything lower
  52  * than it.
  53  *
  54  * If you need to take multiple locks, always take the upper lock first,
  55  * then the lower ones, e.g. first take the its_lock, then the irq_lock.
  56  * If you are already holding a lock and need to take a higher one, you
  57  * have to drop the lower ranking lock first and re-aquire it after having
  58  * taken the upper one.
  59  *
  60  * When taking more than one ap_list_lock at the same time, always take the
  61  * lowest numbered VCPU's ap_list_lock first, so:
  62  *   vcpuX->vcpu_id < vcpuY->vcpu_id:
  63  *     spin_lock(vcpuX->arch.vgic_cpu.ap_list_lock);
  64  *     spin_lock(vcpuY->arch.vgic_cpu.ap_list_lock);
  65  *
  66  * Since the VGIC must support injecting virtual interrupts from ISRs, we have
  67  * to use the spin_lock_irqsave/spin_unlock_irqrestore versions of outer
  68  * spinlocks for any lock that may be taken while injecting an interrupt.
  69  */
  70
  71 /*
  72  * Iterate over the VM's list of mapped LPIs to find the one with a
  73  * matching interrupt ID and return a reference to the IRQ structure.
  74  */
  75 static struct vgic_irq *vgic_get_lpi(struct kvm *kvm, u32 intid)
  76 {
  77         struct vgic_dist *dist = &kvm->arch.vgic;
  78         struct vgic_irq *irq = NULL;
  79         unsigned long flags;
  80
  81         spin_lock_irqsave(&dist->lpi_list_lock, flags);
  82
  83         list_for_each_entry(irq, &dist->lpi_list_head, lpi_list) {
  84                 if (irq->intid != intid)
  85                         continue;
  86
  87                 /*
  88                  * This increases the refcount, the caller is expected to
  89                  * call vgic_put_irq() later once it's finished with the IRQ.
  90                  */
  91                 vgic_get_irq_kref(irq);
  92                 goto out_unlock;
  93         }
  94         irq = NULL;
  95
  96 out_unlock:
  97         spin_unlock_irqrestore(&dist->lpi_list_lock, flags);
  98
  99         return irq;
 100 }
 101
 102 /*
 103  * This looks up the virtual interrupt ID to get the corresponding
 104  * struct vgic_irq. It also increases the refcount, so any caller is expected
 105  * to call vgic_put_irq() once it's finished with this IRQ.
 106  */
 107 struct vgic_irq *vgic_get_irq(struct kvm *kvm, struct kvm_vcpu *vcpu,
 108                               u32 intid)
 109 {
 110         /* SGIs and PPIs */
 111         if (intid <= VGIC_MAX_PRIVATE) {
 112                 intid = array_index_nospec(intid, VGIC_MAX_PRIVATE);
 113                 return &vcpu->arch.vgic_cpu.private_irqs[intid];
 114         }
 115
 116         /* SPIs */
 117         if (intid <= VGIC_MAX_SPI) {
 118                 intid = array_index_nospec(intid, VGIC_MAX_SPI);
 119                 return &kvm->arch.vgic.spis[intid - VGIC_NR_PRIVATE_IRQS];
 120         }
 121
 122         /* LPIs */
 123         if (intid >= VGIC_MIN_LPI)
 124                 return vgic_get_lpi(kvm, intid);
 125
 126         WARN(1, "Looking up struct vgic_irq for reserved INTID");
 127         return NULL;
 128 }
 129
 130 /*
 131  * We can't do anything in here, because we lack the kvm pointer to
 132  * lock and remove the item from the lpi_list. So we keep this function
 133  * empty and use the return value of kref_put() to trigger the freeing.
 134  */
 135 static void vgic_irq_release(struct kref *ref)
 136 {
 137 }
 138
 139 void vgic_put_irq(struct kvm *kvm, struct vgic_irq *irq)
 140 {
 141         struct vgic_dist *dist = &kvm->arch.vgic;
 142         unsigned long flags;
 143
 144         if (irq->intid < VGIC_MIN_LPI)
 145                 return;
 146
 147         spin_lock_irqsave(&dist->lpi_list_lock, flags);
 148         if (!kref_put(&irq->refcount, vgic_irq_release)) {
 149                 spin_unlock_irqrestore(&dist->lpi_list_lock, flags);
 150                 return;
 151         };
 152
 153         list_del(&irq->lpi_list);
 154         dist->lpi_list_count--;
 155         spin_unlock_irqrestore(&dist->lpi_list_lock, flags);
 156
 157         kfree(irq);
 158 }
 159
 160 void vgic_irq_set_phys_pending(struct vgic_irq *irq, bool pending)
 161 {
 162         WARN_ON(irq_set_irqchip_state(irq->host_irq,
 163                                       IRQCHIP_STATE_PENDING,
 164                                       pending));
 165 }
 166
 167 bool vgic_get_phys_line_level(struct vgic_irq *irq)
 168 {
 169         bool line_level;
 170
 171         BUG_ON(!irq->hw);
 172
 173         if (irq->get_input_level)
 174                 return irq->get_input_level(irq->intid);
 175
 176         WARN_ON(irq_get_irqchip_state(irq->host_irq,
 177                                       IRQCHIP_STATE_PENDING,
 178                                       &line_level));
 179         return line_level;
 180 }
 181
 182 /* Set/Clear the physical active state */
 183 void vgic_irq_set_phys_active(struct vgic_irq *irq, bool active)
 184 {
 185
 186         BUG_ON(!irq->hw);
 187         WARN_ON(irq_set_irqchip_state(irq->host_irq,
 188                                       IRQCHIP_STATE_ACTIVE,
 189                                       active));
 190 }
 191
 192 /**
 193  * kvm_vgic_target_oracle - compute the target vcpu for an irq
 194  *
 195  * @irq:        The irq to route. Must be already locked.
 196  *
 197  * Based on the current state of the interrupt (enabled, pending,
 198  * active, vcpu and target_vcpu), compute the next vcpu this should be
 199  * given to. Return NULL if this shouldn't be injected at all.
 200  *
 201  * Requires the IRQ lock to be held.
 202  */
 203 static struct kvm_vcpu *vgic_target_oracle(struct vgic_irq *irq)
 204 {
 205         DEBUG_SPINLOCK_BUG_ON(!spin_is_locked(&irq->irq_lock));
 206
 207         /* If the interrupt is active, it must stay on the current vcpu */
 208         if (irq->active)
 209                 return irq->vcpu ? : irq->target_vcpu;
 210
 211         /*
 212          * If the IRQ is not active but enabled and pending, we should direct
 213          * it to its configured target VCPU.
 214          * If the distributor is disabled, pending interrupts shouldn't be
 215          * forwarded.
 216          */
 217         if (irq->enabled && irq_is_pending(irq)) {
 218                 if (unlikely(irq->target_vcpu &&
 219                              !irq->target_vcpu->kvm->arch.vgic.enabled))
 220                         return NULL;
 221
 222                 return irq->target_vcpu;
 223         }
 224
 225         /* If neither active nor pending and enabled, then this IRQ should not
 226          * be queued to any VCPU.
 227          */
 228         return NULL;
 229 }
 230
 231 /*
 232  * The order of items in the ap_lists defines how we'll pack things in LRs as
 233  * well, the first items in the list being the first things populated in the
 234  * LRs.
 235  *
 236  * A hard rule is that active interrupts can never be pushed out of the LRs
 237  * (and therefore take priority) since we cannot reliably trap on deactivation
 238  * of IRQs and therefore they have to be present in the LRs.
 239  *
 240  * Otherwise things should be sorted by the priority field and the GIC
 241  * hardware support will take care of preemption of priority groups etc.
 242  *
 243  * Return negative if "a" sorts before "b", 0 to preserve order, and positive
 244  * to sort "b" before "a".
 245  */
 246 static int vgic_irq_cmp(void *priv, struct list_head *a, struct list_head *b)
 247 {
 248         struct vgic_irq *irqa = container_of(a, struct vgic_irq, ap_list);
 249         struct vgic_irq *irqb = container_of(b, struct vgic_irq, ap_list);
 250         bool penda, pendb;
 251         int ret;
 252
 253         spin_lock(&irqa->irq_lock);
 254         spin_lock_nested(&irqb->irq_lock, SINGLE_DEPTH_NESTING);
 255
 256         if (irqa->active || irqb->active) {
 257                 ret = (int)irqb->active - (int)irqa->active;
 258                 goto out;
 259         }
 260
 261         penda = irqa->enabled && irq_is_pending(irqa);
 262         pendb = irqb->enabled && irq_is_pending(irqb);
 263
 264         if (!penda || !pendb) {
 265                 ret = (int)pendb - (int)penda;
 266                 goto out;
 267         }
 268
 269         /* Both pending and enabled, sort by priority */
 270         ret = irqa->priority - irqb->priority;
 271 out:
 272         spin_unlock(&irqb->irq_lock);
 273         spin_unlock(&irqa->irq_lock);
 274         return ret;
 275 }
 276
 277 /* Must be called with the ap_list_lock held */
 278 static void vgic_sort_ap_list(struct kvm_vcpu *vcpu)
 279 {
 280         struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
 281
 282         DEBUG_SPINLOCK_BUG_ON(!spin_is_locked(&vgic_cpu->ap_list_lock));
 283
 284         list_sort(NULL, &vgic_cpu->ap_list_head, vgic_irq_cmp);
 285 }
 286
 287 /*
 288  * Only valid injection if changing level for level-triggered IRQs or for a
 289  * rising edge, and in-kernel connected IRQ lines can only be controlled by
 290  * their owner.
 291  */
 292 static bool vgic_validate_injection(struct vgic_irq *irq, bool level, void *owner)
 293 {
 294         if (irq->owner != owner)
 295                 return false;
 296
 297         switch (irq->config) {
 298         case VGIC_CONFIG_LEVEL:
 299                 return irq->line_level != level;
 300         case VGIC_CONFIG_EDGE:
 301                 return level;
 302         }
 303
 304         return false;
 305 }
 306
 307 /*
 308  * Check whether an IRQ needs to (and can) be queued to a VCPU's ap list.
 309  * Do the queuing if necessary, taking the right locks in the right order.
 310  * Returns true when the IRQ was queued, false otherwise.
 311  *
 312  * Needs to be entered with the IRQ lock already held, but will return
 313  * with all locks dropped.
 314  */
 315 bool vgic_queue_irq_unlock(struct kvm *kvm, struct vgic_irq *irq,
 316                            unsigned long flags)
 317 {
 318         struct kvm_vcpu *vcpu;
 319
 320         DEBUG_SPINLOCK_BUG_ON(!spin_is_locked(&irq->irq_lock));
 321
 322 retry:
 323         vcpu = vgic_target_oracle(irq);
 324         if (irq->vcpu || !vcpu) {
 325                 /*
 326                  * If this IRQ is already on a VCPU's ap_list, then it
 327                  * cannot be moved or modified and there is no more work for
 328                  * us to do.
 329                  *
 330                  * Otherwise, if the irq is not pending and enabled, it does
 331                  * not need to be inserted into an ap_list and there is also
 332                  * no more work for us to do.
 333                  */
 334                 spin_unlock_irqrestore(&irq->irq_lock, flags);
 335
 336                 /*
 337                  * We have to kick the VCPU here, because we could be
 338                  * queueing an edge-triggered interrupt for which we
 339                  * get no EOI maintenance interrupt. In that case,
 340                  * while the IRQ is already on the VCPU's AP list, the
 341                  * VCPU could have EOI'ed the original interrupt and
 342                  * won't see this one until it exits for some other
 343                  * reason.
 344                  */
 345                 if (vcpu) {
 346                         kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
 347                         kvm_vcpu_kick(vcpu);
 348                 }
 349                 return false;
 350         }
 351
 352         /*
 353          * We must unlock the irq lock to take the ap_list_lock where
 354          * we are going to insert this new pending interrupt.
 355          */
 356         spin_unlock_irqrestore(&irq->irq_lock, flags);
 357
 358         /* someone can do stuff here, which we re-check below */
 359
 360         spin_lock_irqsave(&vcpu->arch.vgic_cpu.ap_list_lock, flags);
 361         spin_lock(&irq->irq_lock);
 362
 363         /*
 364          * Did something change behind our backs?
 365          *
 366          * There are two cases:
 367          * 1) The irq lost its pending state or was disabled behind our
 368          *    backs and/or it was queued to another VCPU's ap_list.
 369          * 2) Someone changed the affinity on this irq behind our
 370          *    backs and we are now holding the wrong ap_list_lock.
 371          *
 372          * In both cases, drop the locks and retry.
 373          */
 374
 375         if (unlikely(irq->vcpu || vcpu != vgic_target_oracle(irq))) {
 376                 spin_unlock(&irq->irq_lock);
 377                 spin_unlock_irqrestore(&vcpu->arch.vgic_cpu.ap_list_lock, flags);
 378
 379                 spin_lock_irqsave(&irq->irq_lock, flags);
 380                 goto retry;
 381         }
 382
 383         /*
 384          * Grab a reference to the irq to reflect the fact that it is
 385          * now in the ap_list.
 386          */
 387         vgic_get_irq_kref(irq);
 388         list_add_tail(&irq->ap_list, &vcpu->arch.vgic_cpu.ap_list_head);
 389         irq->vcpu = vcpu;
 390
 391         spin_unlock(&irq->irq_lock);
 392         spin_unlock_irqrestore(&vcpu->arch.vgic_cpu.ap_list_lock, flags);
 393
 394         kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
 395         kvm_vcpu_kick(vcpu);
 396
 397         return true;
 398 }
 399
 400 /**
 401  * kvm_vgic_inject_irq - Inject an IRQ from a device to the vgic
 402  * @kvm:     The VM structure pointer
 403  * @cpuid:   The CPU for PPIs
 404  * @intid:   The INTID to inject a new state to.
 405  * @level:   Edge-triggered:  true:  to trigger the interrupt
 406  *                            false: to ignore the call
 407  *           Level-sensitive  true:  raise the input signal
 408  *                            false: lower the input signal
 409  * @owner:   The opaque pointer to the owner of the IRQ being raised to verify
 410  *           that the caller is allowed to inject this IRQ.  Userspace
 411  *           injections will have owner == NULL.
 412  *
 413  * The VGIC is not concerned with devices being active-LOW or active-HIGH for
 414  * level-sensitive interrupts.  You can think of the level parameter as 1
 415  * being HIGH and 0 being LOW and all devices being active-HIGH.
 416  */
 417 int kvm_vgic_inject_irq(struct kvm *kvm, int cpuid, unsigned int intid,
 418                         bool level, void *owner)
 419 {
 420         struct kvm_vcpu *vcpu;
 421         struct vgic_irq *irq;
 422         unsigned long flags;
 423         int ret;
 424
 425         trace_vgic_update_irq_pending(cpuid, intid, level);
 426
 427         ret = vgic_lazy_init(kvm);
 428         if (ret)
 429                 return ret;
 430
 431         vcpu = kvm_get_vcpu(kvm, cpuid);
 432         if (!vcpu && intid < VGIC_NR_PRIVATE_IRQS)
 433                 return -EINVAL;
 434
 435         irq = vgic_get_irq(kvm, vcpu, intid);
 436         if (!irq)
 437                 return -EINVAL;
 438
 439         spin_lock_irqsave(&irq->irq_lock, flags);
 440
 441         if (!vgic_validate_injection(irq, level, owner)) {
 442                 /* Nothing to see here, move along... */
 443                 spin_unlock_irqrestore(&irq->irq_lock, flags);
 444                 vgic_put_irq(kvm, irq);
 445                 return 0;
 446         }
 447
 448         if (irq->config == VGIC_CONFIG_LEVEL)
 449                 irq->line_level = level;
 450         else
 451                 irq->pending_latch = true;
 452
 453         vgic_queue_irq_unlock(kvm, irq, flags);
 454         vgic_put_irq(kvm, irq);
 455
 456         return 0;
 457 }
 458
 459 /* @irq->irq_lock must be held */
 460 static int kvm_vgic_map_irq(struct kvm_vcpu *vcpu, struct vgic_irq *irq,
 461                             unsigned int host_irq,
 462                             bool (*get_input_level)(int vindid))
 463 {
 464         struct irq_desc *desc;
 465         struct irq_data *data;
 466
 467         /*
 468          * Find the physical IRQ number corresponding to @host_irq
 469          */
 470         desc = irq_to_desc(host_irq);
 471         if (!desc) {
 472                 kvm_err("%s: no interrupt descriptor\n", __func__);
 473                 return -EINVAL;
 474         }
 475         data = irq_desc_get_irq_data(desc);
 476         while (data->parent_data)
 477                 data = data->parent_data;
 478
 479         irq->hw = true;
 480         irq->host_irq = host_irq;
 481         irq->hwintid = data->hwirq;
 482         irq->get_input_level = get_input_level;
 483         return 0;
 484 }
 485
 486 /* @irq->irq_lock must be held */
 487 static inline void kvm_vgic_unmap_irq(struct vgic_irq *irq)
 488 {
 489         irq->hw = false;
 490         irq->hwintid = 0;
 491         irq->get_input_level = NULL;
 492 }
 493
 494 int kvm_vgic_map_phys_irq(struct kvm_vcpu *vcpu, unsigned int host_irq,
 495                           u32 vintid, bool (*get_input_level)(int vindid))
 496 {
 497         struct vgic_irq *irq = vgic_get_irq(vcpu->kvm, vcpu, vintid);
 498         unsigned long flags;
 499         int ret;
 500
 501         BUG_ON(!irq);
 502
 503         spin_lock_irqsave(&irq->irq_lock, flags);
 504         ret = kvm_vgic_map_irq(vcpu, irq, host_irq, get_input_level);
 505         spin_unlock_irqrestore(&irq->irq_lock, flags);
 506         vgic_put_irq(vcpu->kvm, irq);
 507
 508         return ret;
 509 }
 510
 511 /**
 512  * kvm_vgic_reset_mapped_irq - Reset a mapped IRQ
 513  * @vcpu: The VCPU pointer
 514  * @vintid: The INTID of the interrupt
 515  *
 516  * Reset the active and pending states of a mapped interrupt.  Kernel
 517  * subsystems injecting mapped interrupts should reset their interrupt lines
 518  * when we are doing a reset of the VM.
 519  */
 520 void kvm_vgic_reset_mapped_irq(struct kvm_vcpu *vcpu, u32 vintid)
 521 {
 522         struct vgic_irq *irq = vgic_get_irq(vcpu->kvm, vcpu, vintid);
 523         unsigned long flags;
 524
 525         if (!irq->hw)
 526                 goto out;
 527
 528         spin_lock_irqsave(&irq->irq_lock, flags);
 529         irq->active = false;
 530         irq->pending_latch = false;
 531         irq->line_level = false;
 532         spin_unlock_irqrestore(&irq->irq_lock, flags);
 533 out:
 534         vgic_put_irq(vcpu->kvm, irq);
 535 }
 536
 537 int kvm_vgic_unmap_phys_irq(struct kvm_vcpu *vcpu, unsigned int vintid)
 538 {
 539         struct vgic_irq *irq;
 540         unsigned long flags;
 541
 542         if (!vgic_initialized(vcpu->kvm))
 543                 return -EAGAIN;
 544
 545         irq = vgic_get_irq(vcpu->kvm, vcpu, vintid);
 546         BUG_ON(!irq);
 547
 548         spin_lock_irqsave(&irq->irq_lock, flags);
 549         kvm_vgic_unmap_irq(irq);
 550         spin_unlock_irqrestore(&irq->irq_lock, flags);
 551         vgic_put_irq(vcpu->kvm, irq);
 552
 553         return 0;
 554 }
 555
 556 /**
 557  * kvm_vgic_set_owner - Set the owner of an interrupt for a VM
 558  *
 559  * @vcpu:   Pointer to the VCPU (used for PPIs)
 560  * @intid:  The virtual INTID identifying the interrupt (PPI or SPI)
 561  * @owner:  Opaque pointer to the owner
 562  *
 563  * Returns 0 if intid is not already used by another in-kernel device and the
 564  * owner is set, otherwise returns an error code.
 565  */
 566 int kvm_vgic_set_owner(struct kvm_vcpu *vcpu, unsigned int intid, void *owner)
 567 {
 568         struct vgic_irq *irq;
 569         unsigned long flags;
 570         int ret = 0;
 571
 572         if (!vgic_initialized(vcpu->kvm))
 573                 return -EAGAIN;
 574
 575         /* SGIs and LPIs cannot be wired up to any device */
 576         if (!irq_is_ppi(intid) && !vgic_valid_spi(vcpu->kvm, intid))
 577                 return -EINVAL;
 578
 579         irq = vgic_get_irq(vcpu->kvm, vcpu, intid);
 580         spin_lock_irqsave(&irq->irq_lock, flags);
 581         if (irq->owner && irq->owner != owner)
 582                 ret = -EEXIST;
 583         else
 584                 irq->owner = owner;
 585         spin_unlock_irqrestore(&irq->irq_lock, flags);
 586
 587         return ret;
 588 }
 589
 590 /**
 591  * vgic_prune_ap_list - Remove non-relevant interrupts from the list
 592  *
 593  * @vcpu: The VCPU pointer
 594  *
 595  * Go over the list of "interesting" interrupts, and prune those that we
 596  * won't have to consider in the near future.
 597  */
 598 static void vgic_prune_ap_list(struct kvm_vcpu *vcpu)
 599 {
 600         struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
 601         struct vgic_irq *irq, *tmp;
 602         unsigned long flags;
 603
 604 retry:
 605         spin_lock_irqsave(&vgic_cpu->ap_list_lock, flags);
 606
 607         list_for_each_entry_safe(irq, tmp, &vgic_cpu->ap_list_head, ap_list) {
 608                 struct kvm_vcpu *target_vcpu, *vcpuA, *vcpuB;
 609                 bool target_vcpu_needs_kick = false;
 610
 611                 spin_lock(&irq->irq_lock);
 612
 613                 BUG_ON(vcpu != irq->vcpu);
 614
 615                 target_vcpu = vgic_target_oracle(irq);
 616
 617                 if (!target_vcpu) {
 618                         /*
 619                          * We don't need to process this interrupt any
 620                          * further, move it off the list.
 621                          */
 622                         list_del(&irq->ap_list);
 623                         irq->vcpu = NULL;
 624                         spin_unlock(&irq->irq_lock);
 625
 626                         /*
 627                          * This vgic_put_irq call matches the
 628                          * vgic_get_irq_kref in vgic_queue_irq_unlock,
 629                          * where we added the LPI to the ap_list. As
 630                          * we remove the irq from the list, we drop
 631                          * also drop the refcount.
 632                          */
 633                         vgic_put_irq(vcpu->kvm, irq);
 634                         continue;
 635                 }
 636
 637                 if (target_vcpu == vcpu) {
 638                         /* We're on the right CPU */
 639                         spin_unlock(&irq->irq_lock);
 640                         continue;
 641                 }
 642
 643                 /* This interrupt looks like it has to be migrated. */
 644
 645                 spin_unlock(&irq->irq_lock);
 646                 spin_unlock_irqrestore(&vgic_cpu->ap_list_lock, flags);
 647
 648                 /*
 649                  * Ensure locking order by always locking the smallest
 650                  * ID first.
 651                  */
 652                 if (vcpu->vcpu_id < target_vcpu->vcpu_id) {
 653                         vcpuA = vcpu;
 654                         vcpuB = target_vcpu;
 655                 } else {
 656                         vcpuA = target_vcpu;
 657                         vcpuB = vcpu;
 658                 }
 659
 660                 spin_lock_irqsave(&vcpuA->arch.vgic_cpu.ap_list_lock, flags);
 661                 spin_lock_nested(&vcpuB->arch.vgic_cpu.ap_list_lock,
 662                                  SINGLE_DEPTH_NESTING);
 663                 spin_lock(&irq->irq_lock);
 664
 665                 /*
 666                  * If the affinity has been preserved, move the
 667                  * interrupt around. Otherwise, it means things have
 668                  * changed while the interrupt was unlocked, and we
 669                  * need to replay this.
 670                  *
 671                  * In all cases, we cannot trust the list not to have
 672                  * changed, so we restart from the beginning.
 673                  */
 674                 if (target_vcpu == vgic_target_oracle(irq)) {
 675                         struct vgic_cpu *new_cpu = &target_vcpu->arch.vgic_cpu;
 676
 677                         list_del(&irq->ap_list);
 678                         irq->vcpu = target_vcpu;
 679                         list_add_tail(&irq->ap_list, &new_cpu->ap_list_head);
 680                         target_vcpu_needs_kick = true;
 681                 }
 682
 683                 spin_unlock(&irq->irq_lock);
 684                 spin_unlock(&vcpuB->arch.vgic_cpu.ap_list_lock);
 685                 spin_unlock_irqrestore(&vcpuA->arch.vgic_cpu.ap_list_lock, flags);
 686
 687                 if (target_vcpu_needs_kick) {
 688                         kvm_make_request(KVM_REQ_IRQ_PENDING, target_vcpu);
 689                         kvm_vcpu_kick(target_vcpu);
 690                 }
 691
 692                 goto retry;
 693         }
 694
 695         spin_unlock_irqrestore(&vgic_cpu->ap_list_lock, flags);
 696 }
 697
 698 static inline void vgic_fold_lr_state(struct kvm_vcpu *vcpu)
 699 {
 700         if (kvm_vgic_global_state.type == VGIC_V2)
 701                 vgic_v2_fold_lr_state(vcpu);
 702         else
 703                 vgic_v3_fold_lr_state(vcpu);
 704 }
 705
 706 /* Requires the irq_lock to be held. */
 707 static inline void vgic_populate_lr(struct kvm_vcpu *vcpu,
 708                                     struct vgic_irq *irq, int lr)
 709 {
 710         DEBUG_SPINLOCK_BUG_ON(!spin_is_locked(&irq->irq_lock));
 711
 712         if (kvm_vgic_global_state.type == VGIC_V2)
 713                 vgic_v2_populate_lr(vcpu, irq, lr);
 714         else
 715                 vgic_v3_populate_lr(vcpu, irq, lr);
 716 }
 717
 718 static inline void vgic_clear_lr(struct kvm_vcpu *vcpu, int lr)
 719 {
 720         if (kvm_vgic_global_state.type == VGIC_V2)
 721                 vgic_v2_clear_lr(vcpu, lr);
 722         else
 723                 vgic_v3_clear_lr(vcpu, lr);
 724 }
 725
 726 static inline void vgic_set_underflow(struct kvm_vcpu *vcpu)
 727 {
 728         if (kvm_vgic_global_state.type == VGIC_V2)
 729                 vgic_v2_set_underflow(vcpu);
 730         else
 731                 vgic_v3_set_underflow(vcpu);
 732 }
 733
 734 /* Requires the ap_list_lock to be held. */
 735 static int compute_ap_list_depth(struct kvm_vcpu *vcpu,
 736                                  bool *multi_sgi)
 737 {
 738         struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
 739         struct vgic_irq *irq;
 740         int count = 0;
 741
 742         *multi_sgi = false;
 743
 744         DEBUG_SPINLOCK_BUG_ON(!spin_is_locked(&vgic_cpu->ap_list_lock));
 745
 746         list_for_each_entry(irq, &vgic_cpu->ap_list_head, ap_list) {
 747                 int w;
 748
 749                 spin_lock(&irq->irq_lock);
 750                 /* GICv2 SGIs can count for more than one... */
 751                 w = vgic_irq_get_lr_count(irq);
 752                 spin_unlock(&irq->irq_lock);
 753
 754                 count += w;
 755                 *multi_sgi |= (w > 1);
 756         }
 757         return count;
 758 }
 759
 760 /* Requires the VCPU's ap_list_lock to be held. */
 761 static void vgic_flush_lr_state(struct kvm_vcpu *vcpu)
 762 {
 763         struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
 764         struct vgic_irq *irq;
 765         int count;
 766         bool multi_sgi;
 767         u8 prio = 0xff;
 768
 769         DEBUG_SPINLOCK_BUG_ON(!spin_is_locked(&vgic_cpu->ap_list_lock));
 770
 771         count = compute_ap_list_depth(vcpu, &multi_sgi);
 772         if (count > kvm_vgic_global_state.nr_lr || multi_sgi)
 773                 vgic_sort_ap_list(vcpu);
 774
 775         count = 0;
 776
 777         list_for_each_entry(irq, &vgic_cpu->ap_list_head, ap_list) {
 778                 spin_lock(&irq->irq_lock);
 779
 780                 /*
 781                  * If we have multi-SGIs in the pipeline, we need to
 782                  * guarantee that they are all seen before any IRQ of
 783                  * lower priority. In that case, we need to filter out
 784                  * these interrupts by exiting early. This is easy as
 785                  * the AP list has been sorted already.
 786                  */
 787                 if (multi_sgi && irq->priority > prio) {
 788                         spin_unlock(&irq->irq_lock);
 789                         break;
 790                 }
 791
 792                 if (likely(vgic_target_oracle(irq) == vcpu)) {
 793                         vgic_populate_lr(vcpu, irq, count++);
 794
 795                         if (irq->source)
 796                                 prio = irq->priority;
 797                 }
 798
 799                 spin_unlock(&irq->irq_lock);
 800
 801                 if (count == kvm_vgic_global_state.nr_lr) {
 802                         if (!list_is_last(&irq->ap_list,
 803                                           &vgic_cpu->ap_list_head))
 804                                 vgic_set_underflow(vcpu);
 805                         break;
 806                 }
 807         }
 808
 809         vcpu->arch.vgic_cpu.used_lrs = count;
 810
 811         /* Nuke remaining LRs */
 812         for ( ; count < kvm_vgic_global_state.nr_lr; count++)
 813                 vgic_clear_lr(vcpu, count);
 814 }
 815
 816 static inline bool can_access_vgic_from_kernel(void)
 817 {
 818         /*
 819          * GICv2 can always be accessed from the kernel because it is
 820          * memory-mapped, and VHE systems can access GICv3 EL2 system
 821          * registers.
 822          */
 823         return !static_branch_unlikely(&kvm_vgic_global_state.gicv3_cpuif) || has_vhe();
 824 }
 825
 826 static inline void vgic_save_state(struct kvm_vcpu *vcpu)
 827 {
 828         if (!static_branch_unlikely(&kvm_vgic_global_state.gicv3_cpuif))
 829                 vgic_v2_save_state(vcpu);
 830         else
 831                 __vgic_v3_save_state(vcpu);
 832 }
 833
 834 /* Sync back the hardware VGIC state into our emulation after a guest's run. */
 835 void kvm_vgic_sync_hwstate(struct kvm_vcpu *vcpu)
 836 {
 837         struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
 838
 839         WARN_ON(vgic_v4_sync_hwstate(vcpu));
 840
 841         /* An empty ap_list_head implies used_lrs == 0 */
 842         if (list_empty(&vcpu->arch.vgic_cpu.ap_list_head))
 843                 return;
 844
 845         if (can_access_vgic_from_kernel())
 846                 vgic_save_state(vcpu);
 847
 848         if (vgic_cpu->used_lrs)
 849                 vgic_fold_lr_state(vcpu);
 850         vgic_prune_ap_list(vcpu);
 851 }
 852
 853 static inline void vgic_restore_state(struct kvm_vcpu *vcpu)
 854 {
 855         if (!static_branch_unlikely(&kvm_vgic_global_state.gicv3_cpuif))
 856                 vgic_v2_restore_state(vcpu);
 857         else
 858                 __vgic_v3_restore_state(vcpu);
 859 }
 860
 861 /* Flush our emulation state into the GIC hardware before entering the guest. */
 862 void kvm_vgic_flush_hwstate(struct kvm_vcpu *vcpu)
 863 {
 864         WARN_ON(vgic_v4_flush_hwstate(vcpu));
 865
 866         /*
 867          * If there are no virtual interrupts active or pending for this
 868          * VCPU, then there is no work to do and we can bail out without
 869          * taking any lock.  There is a potential race with someone injecting
 870          * interrupts to the VCPU, but it is a benign race as the VCPU will
 871          * either observe the new interrupt before or after doing this check,
 872          * and introducing additional synchronization mechanism doesn't change
 873          * this.
 874          */
 875         if (list_empty(&vcpu->arch.vgic_cpu.ap_list_head))
 876                 return;
 877
 878         DEBUG_SPINLOCK_BUG_ON(!irqs_disabled());
 879
 880         spin_lock(&vcpu->arch.vgic_cpu.ap_list_lock);
 881         vgic_flush_lr_state(vcpu);
 882         spin_unlock(&vcpu->arch.vgic_cpu.ap_list_lock);
 883
 884         if (can_access_vgic_from_kernel())
 885                 vgic_restore_state(vcpu);
 886 }
 887
 888 void kvm_vgic_load(struct kvm_vcpu *vcpu)
 889 {
 890         if (unlikely(!vgic_initialized(vcpu->kvm)))
 891                 return;
 892
 893         if (kvm_vgic_global_state.type == VGIC_V2)
 894                 vgic_v2_load(vcpu);
 895         else
 896                 vgic_v3_load(vcpu);
 897 }
 898
 899 void kvm_vgic_put(struct kvm_vcpu *vcpu)
 900 {
 901         if (unlikely(!vgic_initialized(vcpu->kvm)))
 902                 return;
 903
 904         if (kvm_vgic_global_state.type == VGIC_V2)
 905                 vgic_v2_put(vcpu);
 906         else
 907                 vgic_v3_put(vcpu);
 908 }
 909
 910 int kvm_vgic_vcpu_pending_irq(struct kvm_vcpu *vcpu)
 911 {
 912         struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
 913         struct vgic_irq *irq;
 914         bool pending = false;
 915         unsigned long flags;
 916
 917         if (!vcpu->kvm->arch.vgic.enabled)
 918                 return false;
 919
 920         if (vcpu->arch.vgic_cpu.vgic_v3.its_vpe.pending_last)
 921                 return true;
 922
 923         spin_lock_irqsave(&vgic_cpu->ap_list_lock, flags);
 924
 925         list_for_each_entry(irq, &vgic_cpu->ap_list_head, ap_list) {
 926                 spin_lock(&irq->irq_lock);
 927                 pending = irq_is_pending(irq) && irq->enabled;
 928                 spin_unlock(&irq->irq_lock);
 929
 930                 if (pending)
 931                         break;
 932         }
 933
 934         spin_unlock_irqrestore(&vgic_cpu->ap_list_lock, flags);
 935
 936         return pending;
 937 }
 938
 939 void vgic_kick_vcpus(struct kvm *kvm)
 940 {
 941         struct kvm_vcpu *vcpu;
 942         int c;
 943
 944         /*
 945          * We've injected an interrupt, time to find out who deserves
 946          * a good kick...
 947          */
 948         kvm_for_each_vcpu(c, vcpu, kvm) {
 949                 if (kvm_vgic_vcpu_pending_irq(vcpu)) {
 950                         kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
 951                         kvm_vcpu_kick(vcpu);
 952                 }
 953         }
 954 }
 955
 956 bool kvm_vgic_map_is_active(struct kvm_vcpu *vcpu, unsigned int vintid)
 957 {
 958         struct vgic_irq *irq;
 959         bool map_is_active;
 960         unsigned long flags;
 961
 962         if (!vgic_initialized(vcpu->kvm))
 963                 return false;
 964
 965         irq = vgic_get_irq(vcpu->kvm, vcpu, vintid);
 966         spin_lock_irqsave(&irq->irq_lock, flags);
 967         map_is_active = irq->hw && irq->active;
 968         spin_unlock_irqrestore(&irq->irq_lock, flags);
 969         vgic_put_irq(vcpu->kvm, irq);
 970
 971         return map_is_active;
 972 }
 973