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/kvm.h>
  18 #include <linux/kvm_host.h>
  19 #include <linux/list_sort.h>
  20
  21 #include "vgic.h"
  22
  23 #define CREATE_TRACE_POINTS
  24 #include "../trace.h"
  25
  26 #ifdef CONFIG_DEBUG_SPINLOCK
  27 #define DEBUG_SPINLOCK_BUG_ON(p) BUG_ON(p)
  28 #else
  29 #define DEBUG_SPINLOCK_BUG_ON(p)
  30 #endif
  31
  32 struct vgic_global __section(.hyp.text) kvm_vgic_global_state;
  33
  34 /*
  35  * Locking order is always:
  36  *   vgic_cpu->ap_list_lock
  37  *     vgic_irq->irq_lock
  38  *
  39  * (that is, always take the ap_list_lock before the struct vgic_irq lock).
  40  *
  41  * When taking more than one ap_list_lock at the same time, always take the
  42  * lowest numbered VCPU's ap_list_lock first, so:
  43  *   vcpuX->vcpu_id < vcpuY->vcpu_id:
  44  *     spin_lock(vcpuX->arch.vgic_cpu.ap_list_lock);
  45  *     spin_lock(vcpuY->arch.vgic_cpu.ap_list_lock);
  46  */
  47
  48 struct vgic_irq *vgic_get_irq(struct kvm *kvm, struct kvm_vcpu *vcpu,
  49                               u32 intid)
  50 {
  51         /* SGIs and PPIs */
  52         if (intid <= VGIC_MAX_PRIVATE)
  53                 return &vcpu->arch.vgic_cpu.private_irqs[intid];
  54
  55         /* SPIs */
  56         if (intid <= VGIC_MAX_SPI)
  57                 return &kvm->arch.vgic.spis[intid - VGIC_NR_PRIVATE_IRQS];
  58
  59         /* LPIs are not yet covered */
  60         if (intid >= VGIC_MIN_LPI)
  61                 return NULL;
  62
  63         WARN(1, "Looking up struct vgic_irq for reserved INTID");
  64         return NULL;
  65 }
  66
  67 /**
  68  * kvm_vgic_target_oracle - compute the target vcpu for an irq
  69  *
  70  * @irq:        The irq to route. Must be already locked.
  71  *
  72  * Based on the current state of the interrupt (enabled, pending,
  73  * active, vcpu and target_vcpu), compute the next vcpu this should be
  74  * given to. Return NULL if this shouldn't be injected at all.
  75  *
  76  * Requires the IRQ lock to be held.
  77  */
  78 static struct kvm_vcpu *vgic_target_oracle(struct vgic_irq *irq)
  79 {
  80         DEBUG_SPINLOCK_BUG_ON(!spin_is_locked(&irq->irq_lock));
  81
  82         /* If the interrupt is active, it must stay on the current vcpu */
  83         if (irq->active)
  84                 return irq->vcpu ? : irq->target_vcpu;
  85
  86         /*
  87          * If the IRQ is not active but enabled and pending, we should direct
  88          * it to its configured target VCPU.
  89          * If the distributor is disabled, pending interrupts shouldn't be
  90          * forwarded.
  91          */
  92         if (irq->enabled && irq->pending) {
  93                 if (unlikely(irq->target_vcpu &&
  94                              !irq->target_vcpu->kvm->arch.vgic.enabled))
  95                         return NULL;
  96
  97                 return irq->target_vcpu;
  98         }
  99
 100         /* If neither active nor pending and enabled, then this IRQ should not
 101          * be queued to any VCPU.
 102          */
 103         return NULL;
 104 }
 105
 106 /*
 107  * The order of items in the ap_lists defines how we'll pack things in LRs as
 108  * well, the first items in the list being the first things populated in the
 109  * LRs.
 110  *
 111  * A hard rule is that active interrupts can never be pushed out of the LRs
 112  * (and therefore take priority) since we cannot reliably trap on deactivation
 113  * of IRQs and therefore they have to be present in the LRs.
 114  *
 115  * Otherwise things should be sorted by the priority field and the GIC
 116  * hardware support will take care of preemption of priority groups etc.
 117  *
 118  * Return negative if "a" sorts before "b", 0 to preserve order, and positive
 119  * to sort "b" before "a".
 120  */
 121 static int vgic_irq_cmp(void *priv, struct list_head *a, struct list_head *b)
 122 {
 123         struct vgic_irq *irqa = container_of(a, struct vgic_irq, ap_list);
 124         struct vgic_irq *irqb = container_of(b, struct vgic_irq, ap_list);
 125         bool penda, pendb;
 126         int ret;
 127
 128         spin_lock(&irqa->irq_lock);
 129         spin_lock_nested(&irqb->irq_lock, SINGLE_DEPTH_NESTING);
 130
 131         if (irqa->active || irqb->active) {
 132                 ret = (int)irqb->active - (int)irqa->active;
 133                 goto out;
 134         }
 135
 136         penda = irqa->enabled && irqa->pending;
 137         pendb = irqb->enabled && irqb->pending;
 138
 139         if (!penda || !pendb) {
 140                 ret = (int)pendb - (int)penda;
 141                 goto out;
 142         }
 143
 144         /* Both pending and enabled, sort by priority */
 145         ret = irqa->priority - irqb->priority;
 146 out:
 147         spin_unlock(&irqb->irq_lock);
 148         spin_unlock(&irqa->irq_lock);
 149         return ret;
 150 }
 151
 152 /* Must be called with the ap_list_lock held */
 153 static void vgic_sort_ap_list(struct kvm_vcpu *vcpu)
 154 {
 155         struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
 156
 157         DEBUG_SPINLOCK_BUG_ON(!spin_is_locked(&vgic_cpu->ap_list_lock));
 158
 159         list_sort(NULL, &vgic_cpu->ap_list_head, vgic_irq_cmp);
 160 }
 161
 162 /*
 163  * Only valid injection if changing level for level-triggered IRQs or for a
 164  * rising edge.
 165  */
 166 static bool vgic_validate_injection(struct vgic_irq *irq, bool level)
 167 {
 168         switch (irq->config) {
 169         case VGIC_CONFIG_LEVEL:
 170                 return irq->line_level != level;
 171         case VGIC_CONFIG_EDGE:
 172                 return level;
 173         }
 174
 175         return false;
 176 }
 177
 178 /*
 179  * Check whether an IRQ needs to (and can) be queued to a VCPU's ap list.
 180  * Do the queuing if necessary, taking the right locks in the right order.
 181  * Returns true when the IRQ was queued, false otherwise.
 182  *
 183  * Needs to be entered with the IRQ lock already held, but will return
 184  * with all locks dropped.
 185  */
 186 bool vgic_queue_irq_unlock(struct kvm *kvm, struct vgic_irq *irq)
 187 {
 188         struct kvm_vcpu *vcpu;
 189
 190         DEBUG_SPINLOCK_BUG_ON(!spin_is_locked(&irq->irq_lock));
 191
 192 retry:
 193         vcpu = vgic_target_oracle(irq);
 194         if (irq->vcpu || !vcpu) {
 195                 /*
 196                  * If this IRQ is already on a VCPU's ap_list, then it
 197                  * cannot be moved or modified and there is no more work for
 198                  * us to do.
 199                  *
 200                  * Otherwise, if the irq is not pending and enabled, it does
 201                  * not need to be inserted into an ap_list and there is also
 202                  * no more work for us to do.
 203                  */
 204                 spin_unlock(&irq->irq_lock);
 205                 return false;
 206         }
 207
 208         /*
 209          * We must unlock the irq lock to take the ap_list_lock where
 210          * we are going to insert this new pending interrupt.
 211          */
 212         spin_unlock(&irq->irq_lock);
 213
 214         /* someone can do stuff here, which we re-check below */
 215
 216         spin_lock(&vcpu->arch.vgic_cpu.ap_list_lock);
 217         spin_lock(&irq->irq_lock);
 218
 219         /*
 220          * Did something change behind our backs?
 221          *
 222          * There are two cases:
 223          * 1) The irq lost its pending state or was disabled behind our
 224          *    backs and/or it was queued to another VCPU's ap_list.
 225          * 2) Someone changed the affinity on this irq behind our
 226          *    backs and we are now holding the wrong ap_list_lock.
 227          *
 228          * In both cases, drop the locks and retry.
 229          */
 230
 231         if (unlikely(irq->vcpu || vcpu != vgic_target_oracle(irq))) {
 232                 spin_unlock(&irq->irq_lock);
 233                 spin_unlock(&vcpu->arch.vgic_cpu.ap_list_lock);
 234
 235                 spin_lock(&irq->irq_lock);
 236                 goto retry;
 237         }
 238
 239         list_add_tail(&irq->ap_list, &vcpu->arch.vgic_cpu.ap_list_head);
 240         irq->vcpu = vcpu;
 241
 242         spin_unlock(&irq->irq_lock);
 243         spin_unlock(&vcpu->arch.vgic_cpu.ap_list_lock);
 244
 245         kvm_vcpu_kick(vcpu);
 246
 247         return true;
 248 }
 249
 250 static int vgic_update_irq_pending(struct kvm *kvm, int cpuid,
 251                                    unsigned int intid, bool level,
 252                                    bool mapped_irq)
 253 {
 254         struct kvm_vcpu *vcpu;
 255         struct vgic_irq *irq;
 256         int ret;
 257
 258         trace_vgic_update_irq_pending(cpuid, intid, level);
 259
 260         ret = vgic_lazy_init(kvm);
 261         if (ret)
 262                 return ret;
 263
 264         vcpu = kvm_get_vcpu(kvm, cpuid);
 265         if (!vcpu && intid < VGIC_NR_PRIVATE_IRQS)
 266                 return -EINVAL;
 267
 268         irq = vgic_get_irq(kvm, vcpu, intid);
 269         if (!irq)
 270                 return -EINVAL;
 271
 272         if (irq->hw != mapped_irq)
 273                 return -EINVAL;
 274
 275         spin_lock(&irq->irq_lock);
 276
 277         if (!vgic_validate_injection(irq, level)) {
 278                 /* Nothing to see here, move along... */
 279                 spin_unlock(&irq->irq_lock);
 280                 return 0;
 281         }
 282
 283         if (irq->config == VGIC_CONFIG_LEVEL) {
 284                 irq->line_level = level;
 285                 irq->pending = level || irq->soft_pending;
 286         } else {
 287                 irq->pending = true;
 288         }
 289
 290         vgic_queue_irq_unlock(kvm, irq);
 291
 292         return 0;
 293 }
 294
 295 /**
 296  * kvm_vgic_inject_irq - Inject an IRQ from a device to the vgic
 297  * @kvm:     The VM structure pointer
 298  * @cpuid:   The CPU for PPIs
 299  * @intid:   The INTID to inject a new state to.
 300  * @level:   Edge-triggered:  true:  to trigger the interrupt
 301  *                            false: to ignore the call
 302  *           Level-sensitive  true:  raise the input signal
 303  *                            false: lower the input signal
 304  *
 305  * The VGIC is not concerned with devices being active-LOW or active-HIGH for
 306  * level-sensitive interrupts.  You can think of the level parameter as 1
 307  * being HIGH and 0 being LOW and all devices being active-HIGH.
 308  */
 309 int kvm_vgic_inject_irq(struct kvm *kvm, int cpuid, unsigned int intid,
 310                         bool level)
 311 {
 312         return vgic_update_irq_pending(kvm, cpuid, intid, level, false);
 313 }
 314
 315 /**
 316  * vgic_prune_ap_list - Remove non-relevant interrupts from the list
 317  *
 318  * @vcpu: The VCPU pointer
 319  *
 320  * Go over the list of "interesting" interrupts, and prune those that we
 321  * won't have to consider in the near future.
 322  */
 323 static void vgic_prune_ap_list(struct kvm_vcpu *vcpu)
 324 {
 325         struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
 326         struct vgic_irq *irq, *tmp;
 327
 328 retry:
 329         spin_lock(&vgic_cpu->ap_list_lock);
 330
 331         list_for_each_entry_safe(irq, tmp, &vgic_cpu->ap_list_head, ap_list) {
 332                 struct kvm_vcpu *target_vcpu, *vcpuA, *vcpuB;
 333
 334                 spin_lock(&irq->irq_lock);
 335
 336                 BUG_ON(vcpu != irq->vcpu);
 337
 338                 target_vcpu = vgic_target_oracle(irq);
 339
 340                 if (!target_vcpu) {
 341                         /*
 342                          * We don't need to process this interrupt any
 343                          * further, move it off the list.
 344                          */
 345                         list_del(&irq->ap_list);
 346                         irq->vcpu = NULL;
 347                         spin_unlock(&irq->irq_lock);
 348                         continue;
 349                 }
 350
 351                 if (target_vcpu == vcpu) {
 352                         /* We're on the right CPU */
 353                         spin_unlock(&irq->irq_lock);
 354                         continue;
 355                 }
 356
 357                 /* This interrupt looks like it has to be migrated. */
 358
 359                 spin_unlock(&irq->irq_lock);
 360                 spin_unlock(&vgic_cpu->ap_list_lock);
 361
 362                 /*
 363                  * Ensure locking order by always locking the smallest
 364                  * ID first.
 365                  */
 366                 if (vcpu->vcpu_id < target_vcpu->vcpu_id) {
 367                         vcpuA = vcpu;
 368                         vcpuB = target_vcpu;
 369                 } else {
 370                         vcpuA = target_vcpu;
 371                         vcpuB = vcpu;
 372                 }
 373
 374                 spin_lock(&vcpuA->arch.vgic_cpu.ap_list_lock);
 375                 spin_lock_nested(&vcpuB->arch.vgic_cpu.ap_list_lock,
 376                                  SINGLE_DEPTH_NESTING);
 377                 spin_lock(&irq->irq_lock);
 378
 379                 /*
 380                  * If the affinity has been preserved, move the
 381                  * interrupt around. Otherwise, it means things have
 382                  * changed while the interrupt was unlocked, and we
 383                  * need to replay this.
 384                  *
 385                  * In all cases, we cannot trust the list not to have
 386                  * changed, so we restart from the beginning.
 387                  */
 388                 if (target_vcpu == vgic_target_oracle(irq)) {
 389                         struct vgic_cpu *new_cpu = &target_vcpu->arch.vgic_cpu;
 390
 391                         list_del(&irq->ap_list);
 392                         irq->vcpu = target_vcpu;
 393                         list_add_tail(&irq->ap_list, &new_cpu->ap_list_head);
 394                 }
 395
 396                 spin_unlock(&irq->irq_lock);
 397                 spin_unlock(&vcpuB->arch.vgic_cpu.ap_list_lock);
 398                 spin_unlock(&vcpuA->arch.vgic_cpu.ap_list_lock);
 399                 goto retry;
 400         }
 401
 402         spin_unlock(&vgic_cpu->ap_list_lock);
 403 }
 404
 405 static inline void vgic_process_maintenance_interrupt(struct kvm_vcpu *vcpu)
 406 {
 407         if (kvm_vgic_global_state.type == VGIC_V2)
 408                 vgic_v2_process_maintenance(vcpu);
 409         else
 410                 vgic_v3_process_maintenance(vcpu);
 411 }
 412
 413 static inline void vgic_fold_lr_state(struct kvm_vcpu *vcpu)
 414 {
 415         if (kvm_vgic_global_state.type == VGIC_V2)
 416                 vgic_v2_fold_lr_state(vcpu);
 417         else
 418                 vgic_v3_fold_lr_state(vcpu);
 419 }
 420
 421 /* Requires the irq_lock to be held. */
 422 static inline void vgic_populate_lr(struct kvm_vcpu *vcpu,
 423                                     struct vgic_irq *irq, int lr)
 424 {
 425         DEBUG_SPINLOCK_BUG_ON(!spin_is_locked(&irq->irq_lock));
 426
 427         if (kvm_vgic_global_state.type == VGIC_V2)
 428                 vgic_v2_populate_lr(vcpu, irq, lr);
 429         else
 430                 vgic_v3_populate_lr(vcpu, irq, lr);
 431 }
 432
 433 static inline void vgic_clear_lr(struct kvm_vcpu *vcpu, int lr)
 434 {
 435         if (kvm_vgic_global_state.type == VGIC_V2)
 436                 vgic_v2_clear_lr(vcpu, lr);
 437         else
 438                 vgic_v3_clear_lr(vcpu, lr);
 439 }
 440
 441 static inline void vgic_set_underflow(struct kvm_vcpu *vcpu)
 442 {
 443         if (kvm_vgic_global_state.type == VGIC_V2)
 444                 vgic_v2_set_underflow(vcpu);
 445         else
 446                 vgic_v3_set_underflow(vcpu);
 447 }
 448
 449 /* Requires the ap_list_lock to be held. */
 450 static int compute_ap_list_depth(struct kvm_vcpu *vcpu)
 451 {
 452         struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
 453         struct vgic_irq *irq;
 454         int count = 0;
 455
 456         DEBUG_SPINLOCK_BUG_ON(!spin_is_locked(&vgic_cpu->ap_list_lock));
 457
 458         list_for_each_entry(irq, &vgic_cpu->ap_list_head, ap_list) {
 459                 spin_lock(&irq->irq_lock);
 460                 /* GICv2 SGIs can count for more than one... */
 461                 if (vgic_irq_is_sgi(irq->intid) && irq->source)
 462                         count += hweight8(irq->source);
 463                 else
 464                         count++;
 465                 spin_unlock(&irq->irq_lock);
 466         }
 467         return count;
 468 }
 469
 470 /* Requires the VCPU's ap_list_lock to be held. */
 471 static void vgic_flush_lr_state(struct kvm_vcpu *vcpu)
 472 {
 473         struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
 474         struct vgic_irq *irq;
 475         int count = 0;
 476
 477         DEBUG_SPINLOCK_BUG_ON(!spin_is_locked(&vgic_cpu->ap_list_lock));
 478
 479         if (compute_ap_list_depth(vcpu) > kvm_vgic_global_state.nr_lr) {
 480                 vgic_set_underflow(vcpu);
 481                 vgic_sort_ap_list(vcpu);
 482         }
 483
 484         list_for_each_entry(irq, &vgic_cpu->ap_list_head, ap_list) {
 485                 spin_lock(&irq->irq_lock);
 486
 487                 if (unlikely(vgic_target_oracle(irq) != vcpu))
 488                         goto next;
 489
 490                 /*
 491                  * If we get an SGI with multiple sources, try to get
 492                  * them in all at once.
 493                  */
 494                 do {
 495                         vgic_populate_lr(vcpu, irq, count++);
 496                 } while (irq->source && count < kvm_vgic_global_state.nr_lr);
 497
 498 next:
 499                 spin_unlock(&irq->irq_lock);
 500
 501                 if (count == kvm_vgic_global_state.nr_lr)
 502                         break;
 503         }
 504
 505         vcpu->arch.vgic_cpu.used_lrs = count;
 506
 507         /* Nuke remaining LRs */
 508         for ( ; count < kvm_vgic_global_state.nr_lr; count++)
 509                 vgic_clear_lr(vcpu, count);
 510 }
 511
 512 /* Sync back the hardware VGIC state into our emulation after a guest's run. */
 513 void kvm_vgic_sync_hwstate(struct kvm_vcpu *vcpu)
 514 {
 515         vgic_process_maintenance_interrupt(vcpu);
 516         vgic_fold_lr_state(vcpu);
 517         vgic_prune_ap_list(vcpu);
 518 }
 519
 520 /* Flush our emulation state into the GIC hardware before entering the guest. */
 521 void kvm_vgic_flush_hwstate(struct kvm_vcpu *vcpu)
 522 {
 523         spin_lock(&vcpu->arch.vgic_cpu.ap_list_lock);
 524         vgic_flush_lr_state(vcpu);
 525         spin_unlock(&vcpu->arch.vgic_cpu.ap_list_lock);
 526 }
 527
 528 int kvm_vgic_vcpu_pending_irq(struct kvm_vcpu *vcpu)
 529 {
 530         struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
 531         struct vgic_irq *irq;
 532         bool pending = false;
 533
 534         if (!vcpu->kvm->arch.vgic.enabled)
 535                 return false;
 536
 537         spin_lock(&vgic_cpu->ap_list_lock);
 538
 539         list_for_each_entry(irq, &vgic_cpu->ap_list_head, ap_list) {
 540                 spin_lock(&irq->irq_lock);
 541                 pending = irq->pending && irq->enabled;
 542                 spin_unlock(&irq->irq_lock);
 543
 544                 if (pending)
 545                         break;
 546         }
 547
 548         spin_unlock(&vgic_cpu->ap_list_lock);
 549
 550         return pending;
 551 }
 552
 553 void vgic_kick_vcpus(struct kvm *kvm)
 554 {
 555         struct kvm_vcpu *vcpu;
 556         int c;
 557
 558         /*
 559          * We've injected an interrupt, time to find out who deserves
 560          * a good kick...
 561          */
 562         kvm_for_each_vcpu(c, vcpu, kvm) {
 563                 if (kvm_vgic_vcpu_pending_irq(vcpu))
 564                         kvm_vcpu_kick(vcpu);
 565         }
 566 }