drivers/virtio/virtio_pci_common.c

   1 /*
   2  * Virtio PCI driver - common functionality for all device versions
   3  *
   4  * This module allows virtio devices to be used over a virtual PCI device.
   5  * This can be used with QEMU based VMMs like KVM or Xen.
   6  *
   7  * Copyright IBM Corp. 2007
   8  * Copyright Red Hat, Inc. 2014
   9  *
  10  * Authors:
  11  *  Anthony Liguori  <aliguori@us.ibm.com>
  12  *  Rusty Russell <rusty@rustcorp.com.au>
  13  *  Michael S. Tsirkin <mst@redhat.com>
  14  *
  15  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  16  * See the COPYING file in the top-level directory.
  17  *
  18  */
  19
  20 #include "virtio_pci_common.h"
  21
  22 static bool force_legacy = false;
  23
  24 #if IS_ENABLED(CONFIG_VIRTIO_PCI_LEGACY)
  25 module_param(force_legacy, bool, 0444);
  26 MODULE_PARM_DESC(force_legacy,
  27                  "Force legacy mode for transitional virtio 1 devices");
  28 #endif
  29
  30 /* wait for pending irq handlers */
  31 void vp_synchronize_vectors(struct virtio_device *vdev)
  32 {
  33         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
  34         int i;
  35
  36         if (vp_dev->intx_enabled)
  37                 synchronize_irq(vp_dev->pci_dev->irq);
  38
  39         for (i = 0; i < vp_dev->msix_vectors; ++i)
  40                 synchronize_irq(pci_irq_vector(vp_dev->pci_dev, i));
  41 }
  42
  43 /* the notify function used when creating a virt queue */
  44 bool vp_notify(struct virtqueue *vq)
  45 {
  46         /* we write the queue's selector into the notification register to
  47          * signal the other end */
  48         iowrite16(vq->index, (void __iomem *)vq->priv);
  49         return true;
  50 }
  51
  52 /* Handle a configuration change: Tell driver if it wants to know. */
  53 static irqreturn_t vp_config_changed(int irq, void *opaque)
  54 {
  55         struct virtio_pci_device *vp_dev = opaque;
  56
  57         virtio_config_changed(&vp_dev->vdev);
  58         return IRQ_HANDLED;
  59 }
  60
  61 /* Notify all virtqueues on an interrupt. */
  62 static irqreturn_t vp_vring_interrupt(int irq, void *opaque)
  63 {
  64         struct virtio_pci_device *vp_dev = opaque;
  65         struct virtio_pci_vq_info *info;
  66         irqreturn_t ret = IRQ_NONE;
  67         unsigned long flags;
  68
  69         spin_lock_irqsave(&vp_dev->lock, flags);
  70         list_for_each_entry(info, &vp_dev->virtqueues, node) {
  71                 if (vring_interrupt(irq, info->vq) == IRQ_HANDLED)
  72                         ret = IRQ_HANDLED;
  73         }
  74         spin_unlock_irqrestore(&vp_dev->lock, flags);
  75
  76         return ret;
  77 }
  78
  79 /* A small wrapper to also acknowledge the interrupt when it's handled.
  80  * I really need an EIO hook for the vring so I can ack the interrupt once we
  81  * know that we'll be handling the IRQ but before we invoke the callback since
  82  * the callback may notify the host which results in the host attempting to
  83  * raise an interrupt that we would then mask once we acknowledged the
  84  * interrupt. */
  85 static irqreturn_t vp_interrupt(int irq, void *opaque)
  86 {
  87         struct virtio_pci_device *vp_dev = opaque;
  88         u8 isr;
  89
  90         /* reading the ISR has the effect of also clearing it so it's very
  91          * important to save off the value. */
  92         isr = ioread8(vp_dev->isr);
  93
  94         /* It's definitely not us if the ISR was not high */
  95         if (!isr)
  96                 return IRQ_NONE;
  97
  98         /* Configuration change?  Tell driver if it wants to know. */
  99         if (isr & VIRTIO_PCI_ISR_CONFIG)
 100                 vp_config_changed(irq, opaque);
 101
 102         return vp_vring_interrupt(irq, opaque);
 103 }
 104
 105 static int vp_request_msix_vectors(struct virtio_device *vdev, int nvectors,
 106                                    bool per_vq_vectors, struct irq_affinity *desc)
 107 {
 108         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 109         const char *name = dev_name(&vp_dev->vdev.dev);
 110         unsigned i, v;
 111         int err = -ENOMEM;
 112
 113         vp_dev->msix_vectors = nvectors;
 114
 115         vp_dev->msix_names = kmalloc(nvectors * sizeof *vp_dev->msix_names,
 116                                      GFP_KERNEL);
 117         if (!vp_dev->msix_names)
 118                 goto error;
 119         vp_dev->msix_affinity_masks
 120                 = kzalloc(nvectors * sizeof *vp_dev->msix_affinity_masks,
 121                           GFP_KERNEL);
 122         if (!vp_dev->msix_affinity_masks)
 123                 goto error;
 124         for (i = 0; i < nvectors; ++i)
 125                 if (!alloc_cpumask_var(&vp_dev->msix_affinity_masks[i],
 126                                         GFP_KERNEL))
 127                         goto error;
 128
 129         err = pci_alloc_irq_vectors_affinity(vp_dev->pci_dev, nvectors,
 130                                              nvectors, PCI_IRQ_MSIX |
 131                                              (desc ? PCI_IRQ_AFFINITY : 0),
 132                                              desc);
 133         if (err < 0)
 134                 goto error;
 135         vp_dev->msix_enabled = 1;
 136
 137         /* Set the vector used for configuration */
 138         v = vp_dev->msix_used_vectors;
 139         snprintf(vp_dev->msix_names[v], sizeof *vp_dev->msix_names,
 140                  "%s-config", name);
 141         err = request_irq(pci_irq_vector(vp_dev->pci_dev, v),
 142                           vp_config_changed, 0, vp_dev->msix_names[v],
 143                           vp_dev);
 144         if (err)
 145                 goto error;
 146         ++vp_dev->msix_used_vectors;
 147
 148         v = vp_dev->config_vector(vp_dev, v);
 149         /* Verify we had enough resources to assign the vector */
 150         if (v == VIRTIO_MSI_NO_VECTOR) {
 151                 err = -EBUSY;
 152                 goto error;
 153         }
 154
 155         if (!per_vq_vectors) {
 156                 /* Shared vector for all VQs */
 157                 v = vp_dev->msix_used_vectors;
 158                 snprintf(vp_dev->msix_names[v], sizeof *vp_dev->msix_names,
 159                          "%s-virtqueues", name);
 160                 err = request_irq(pci_irq_vector(vp_dev->pci_dev, v),
 161                                   vp_vring_interrupt, 0, vp_dev->msix_names[v],
 162                                   vp_dev);
 163                 if (err)
 164                         goto error;
 165                 ++vp_dev->msix_used_vectors;
 166         }
 167         return 0;
 168 error:
 169         return err;
 170 }
 171
 172 static struct virtqueue *vp_setup_vq(struct virtio_device *vdev, unsigned index,
 173                                      void (*callback)(struct virtqueue *vq),
 174                                      const char *name,
 175                                      bool ctx,
 176                                      u16 msix_vec)
 177 {
 178         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 179         struct virtio_pci_vq_info *info = kmalloc(sizeof *info, GFP_KERNEL);
 180         struct virtqueue *vq;
 181         unsigned long flags;
 182
 183         /* fill out our structure that represents an active queue */
 184         if (!info)
 185                 return ERR_PTR(-ENOMEM);
 186
 187         vq = vp_dev->setup_vq(vp_dev, info, index, callback, name, ctx,
 188                               msix_vec);
 189         if (IS_ERR(vq))
 190                 goto out_info;
 191
 192         info->vq = vq;
 193         if (callback) {
 194                 spin_lock_irqsave(&vp_dev->lock, flags);
 195                 list_add(&info->node, &vp_dev->virtqueues);
 196                 spin_unlock_irqrestore(&vp_dev->lock, flags);
 197         } else {
 198                 INIT_LIST_HEAD(&info->node);
 199         }
 200
 201         vp_dev->vqs[index] = info;
 202         return vq;
 203
 204 out_info:
 205         kfree(info);
 206         return vq;
 207 }
 208
 209 static void vp_del_vq(struct virtqueue *vq)
 210 {
 211         struct virtio_pci_device *vp_dev = to_vp_device(vq->vdev);
 212         struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index];
 213         unsigned long flags;
 214
 215         spin_lock_irqsave(&vp_dev->lock, flags);
 216         list_del(&info->node);
 217         spin_unlock_irqrestore(&vp_dev->lock, flags);
 218
 219         vp_dev->del_vq(info);
 220         kfree(info);
 221 }
 222
 223 /* the config->del_vqs() implementation */
 224 void vp_del_vqs(struct virtio_device *vdev)
 225 {
 226         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 227         struct virtqueue *vq, *n;
 228         int i;
 229
 230         list_for_each_entry_safe(vq, n, &vdev->vqs, list) {
 231                 if (vp_dev->per_vq_vectors) {
 232                         int v = vp_dev->vqs[vq->index]->msix_vector;
 233
 234                         if (v != VIRTIO_MSI_NO_VECTOR) {
 235                                 int irq = pci_irq_vector(vp_dev->pci_dev, v);
 236
 237                                 irq_set_affinity_hint(irq, NULL);
 238                                 free_irq(irq, vq);
 239                         }
 240                 }
 241                 vp_del_vq(vq);
 242         }
 243         vp_dev->per_vq_vectors = false;
 244
 245         if (vp_dev->intx_enabled) {
 246                 free_irq(vp_dev->pci_dev->irq, vp_dev);
 247                 vp_dev->intx_enabled = 0;
 248         }
 249
 250         for (i = 0; i < vp_dev->msix_used_vectors; ++i)
 251                 free_irq(pci_irq_vector(vp_dev->pci_dev, i), vp_dev);
 252
 253         for (i = 0; i < vp_dev->msix_vectors; i++)
 254                 if (vp_dev->msix_affinity_masks[i])
 255                         free_cpumask_var(vp_dev->msix_affinity_masks[i]);
 256
 257         if (vp_dev->msix_enabled) {
 258                 /* Disable the vector used for configuration */
 259                 vp_dev->config_vector(vp_dev, VIRTIO_MSI_NO_VECTOR);
 260
 261                 pci_free_irq_vectors(vp_dev->pci_dev);
 262                 vp_dev->msix_enabled = 0;
 263         }
 264
 265         vp_dev->msix_vectors = 0;
 266         vp_dev->msix_used_vectors = 0;
 267         kfree(vp_dev->msix_names);
 268         vp_dev->msix_names = NULL;
 269         kfree(vp_dev->msix_affinity_masks);
 270         vp_dev->msix_affinity_masks = NULL;
 271         kfree(vp_dev->vqs);
 272         vp_dev->vqs = NULL;
 273 }
 274
 275 static int vp_find_vqs_msix(struct virtio_device *vdev, unsigned nvqs,
 276                 struct virtqueue *vqs[], vq_callback_t *callbacks[],
 277                 const char * const names[], bool per_vq_vectors,
 278                 const bool *ctx,
 279                 struct irq_affinity *desc)
 280 {
 281         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 282         u16 msix_vec;
 283         int i, err, nvectors, allocated_vectors;
 284
 285         vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL);
 286         if (!vp_dev->vqs)
 287                 return -ENOMEM;
 288
 289         if (per_vq_vectors) {
 290                 /* Best option: one for change interrupt, one per vq. */
 291                 nvectors = 1;
 292                 for (i = 0; i < nvqs; ++i)
 293                         if (callbacks[i])
 294                                 ++nvectors;
 295         } else {
 296                 /* Second best: one for change, shared for all vqs. */
 297                 nvectors = 2;
 298         }
 299
 300         err = vp_request_msix_vectors(vdev, nvectors, per_vq_vectors,
 301                                       per_vq_vectors ? desc : NULL);
 302         if (err)
 303                 goto error_find;
 304
 305         vp_dev->per_vq_vectors = per_vq_vectors;
 306         allocated_vectors = vp_dev->msix_used_vectors;
 307         for (i = 0; i < nvqs; ++i) {
 308                 if (!names[i]) {
 309                         vqs[i] = NULL;
 310                         continue;
 311                 }
 312
 313                 if (!callbacks[i])
 314                         msix_vec = VIRTIO_MSI_NO_VECTOR;
 315                 else if (vp_dev->per_vq_vectors)
 316                         msix_vec = allocated_vectors++;
 317                 else
 318                         msix_vec = VP_MSIX_VQ_VECTOR;
 319                 vqs[i] = vp_setup_vq(vdev, i, callbacks[i], names[i],
 320                                      ctx ? ctx[i] : false,
 321                                      msix_vec);
 322                 if (IS_ERR(vqs[i])) {
 323                         err = PTR_ERR(vqs[i]);
 324                         goto error_find;
 325                 }
 326
 327                 if (!vp_dev->per_vq_vectors || msix_vec == VIRTIO_MSI_NO_VECTOR)
 328                         continue;
 329
 330                 /* allocate per-vq irq if available and necessary */
 331                 snprintf(vp_dev->msix_names[msix_vec],
 332                          sizeof *vp_dev->msix_names,
 333                          "%s-%s",
 334                          dev_name(&vp_dev->vdev.dev), names[i]);
 335                 err = request_irq(pci_irq_vector(vp_dev->pci_dev, msix_vec),
 336                                   vring_interrupt, 0,
 337                                   vp_dev->msix_names[msix_vec],
 338                                   vqs[i]);
 339                 if (err)
 340                         goto error_find;
 341         }
 342         return 0;
 343
 344 error_find:
 345         vp_del_vqs(vdev);
 346         return err;
 347 }
 348
 349 static int vp_find_vqs_intx(struct virtio_device *vdev, unsigned nvqs,
 350                 struct virtqueue *vqs[], vq_callback_t *callbacks[],
 351                 const char * const names[], const bool *ctx)
 352 {
 353         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 354         int i, err;
 355
 356         vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL);
 357         if (!vp_dev->vqs)
 358                 return -ENOMEM;
 359
 360         err = request_irq(vp_dev->pci_dev->irq, vp_interrupt, IRQF_SHARED,
 361                         dev_name(&vdev->dev), vp_dev);
 362         if (err)
 363                 goto out_del_vqs;
 364
 365         vp_dev->intx_enabled = 1;
 366         vp_dev->per_vq_vectors = false;
 367         for (i = 0; i < nvqs; ++i) {
 368                 if (!names[i]) {
 369                         vqs[i] = NULL;
 370                         continue;
 371                 }
 372                 vqs[i] = vp_setup_vq(vdev, i, callbacks[i], names[i],
 373                                      ctx ? ctx[i] : false,
 374                                      VIRTIO_MSI_NO_VECTOR);
 375                 if (IS_ERR(vqs[i])) {
 376                         err = PTR_ERR(vqs[i]);
 377                         goto out_del_vqs;
 378                 }
 379         }
 380
 381         return 0;
 382 out_del_vqs:
 383         vp_del_vqs(vdev);
 384         return err;
 385 }
 386
 387 /* the config->find_vqs() implementation */
 388 int vp_find_vqs(struct virtio_device *vdev, unsigned nvqs,
 389                 struct virtqueue *vqs[], vq_callback_t *callbacks[],
 390                 const char * const names[], const bool *ctx,
 391                 struct irq_affinity *desc)
 392 {
 393         int err;
 394
 395         /* Try MSI-X with one vector per queue. */
 396         err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, true, ctx, desc);
 397         if (!err)
 398                 return 0;
 399         /* Fallback: MSI-X with one vector for config, one shared for queues. */
 400         err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, false, ctx, desc);
 401         if (!err)
 402                 return 0;
 403         /* Finally fall back to regular interrupts. */
 404         return vp_find_vqs_intx(vdev, nvqs, vqs, callbacks, names, ctx);
 405 }
 406
 407 const char *vp_bus_name(struct virtio_device *vdev)
 408 {
 409         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 410
 411         return pci_name(vp_dev->pci_dev);
 412 }
 413
 414 /* Setup the affinity for a virtqueue:
 415  * - force the affinity for per vq vector
 416  * - OR over all affinities for shared MSI
 417  * - ignore the affinity request if we're using INTX
 418  */
 419 int vp_set_vq_affinity(struct virtqueue *vq, int cpu)
 420 {
 421         struct virtio_device *vdev = vq->vdev;
 422         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 423         struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index];
 424         struct cpumask *mask;
 425         unsigned int irq;
 426
 427         if (!vq->callback)
 428                 return -EINVAL;
 429
 430         if (vp_dev->msix_enabled) {
 431                 mask = vp_dev->msix_affinity_masks[info->msix_vector];
 432                 irq = pci_irq_vector(vp_dev->pci_dev, info->msix_vector);
 433                 if (cpu == -1)
 434                         irq_set_affinity_hint(irq, NULL);
 435                 else {
 436                         cpumask_clear(mask);
 437                         cpumask_set_cpu(cpu, mask);
 438                         irq_set_affinity_hint(irq, mask);
 439                 }
 440         }
 441         return 0;
 442 }
 443
 444 const struct cpumask *vp_get_vq_affinity(struct virtio_device *vdev, int index)
 445 {
 446         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 447
 448         if (!vp_dev->per_vq_vectors ||
 449             vp_dev->vqs[index]->msix_vector == VIRTIO_MSI_NO_VECTOR)
 450                 return NULL;
 451
 452         return pci_irq_get_affinity(vp_dev->pci_dev,
 453                                     vp_dev->vqs[index]->msix_vector);
 454 }
 455
 456 #ifdef CONFIG_PM_SLEEP
 457 static int virtio_pci_freeze(struct device *dev)
 458 {
 459         struct pci_dev *pci_dev = to_pci_dev(dev);
 460         struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
 461         int ret;
 462
 463         ret = virtio_device_freeze(&vp_dev->vdev);
 464
 465         if (!ret)
 466                 pci_disable_device(pci_dev);
 467         return ret;
 468 }
 469
 470 static int virtio_pci_restore(struct device *dev)
 471 {
 472         struct pci_dev *pci_dev = to_pci_dev(dev);
 473         struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
 474         int ret;
 475
 476         ret = pci_enable_device(pci_dev);
 477         if (ret)
 478                 return ret;
 479
 480         pci_set_master(pci_dev);
 481         return virtio_device_restore(&vp_dev->vdev);
 482 }
 483
 484 static const struct dev_pm_ops virtio_pci_pm_ops = {
 485         SET_SYSTEM_SLEEP_PM_OPS(virtio_pci_freeze, virtio_pci_restore)
 486 };
 487 #endif
 488
 489
 490 /* Qumranet donated their vendor ID for devices 0x1000 thru 0x10FF. */
 491 static const struct pci_device_id virtio_pci_id_table[] = {
 492         { PCI_DEVICE(PCI_VENDOR_ID_REDHAT_QUMRANET, PCI_ANY_ID) },
 493         { 0 }
 494 };
 495
 496 MODULE_DEVICE_TABLE(pci, virtio_pci_id_table);
 497
 498 static void virtio_pci_release_dev(struct device *_d)
 499 {
 500         struct virtio_device *vdev = dev_to_virtio(_d);
 501         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 502
 503         /* As struct device is a kobject, it's not safe to
 504          * free the memory (including the reference counter itself)
 505          * until it's release callback. */
 506         kfree(vp_dev);
 507 }
 508
 509 static int virtio_pci_probe(struct pci_dev *pci_dev,
 510                             const struct pci_device_id *id)
 511 {
 512         struct virtio_pci_device *vp_dev;
 513         int rc;
 514
 515         /* allocate our structure and fill it out */
 516         vp_dev = kzalloc(sizeof(struct virtio_pci_device), GFP_KERNEL);
 517         if (!vp_dev)
 518                 return -ENOMEM;
 519
 520         pci_set_drvdata(pci_dev, vp_dev);
 521         vp_dev->vdev.dev.parent = &pci_dev->dev;
 522         vp_dev->vdev.dev.release = virtio_pci_release_dev;
 523         vp_dev->pci_dev = pci_dev;
 524         INIT_LIST_HEAD(&vp_dev->virtqueues);
 525         spin_lock_init(&vp_dev->lock);
 526
 527         /* enable the device */
 528         rc = pci_enable_device(pci_dev);
 529         if (rc)
 530                 goto err_enable_device;
 531
 532         if (force_legacy) {
 533                 rc = virtio_pci_legacy_probe(vp_dev);
 534                 /* Also try modern mode if we can't map BAR0 (no IO space). */
 535                 if (rc == -ENODEV || rc == -ENOMEM)
 536                         rc = virtio_pci_modern_probe(vp_dev);
 537                 if (rc)
 538                         goto err_probe;
 539         } else {
 540                 rc = virtio_pci_modern_probe(vp_dev);
 541                 if (rc == -ENODEV)
 542                         rc = virtio_pci_legacy_probe(vp_dev);
 543                 if (rc)
 544                         goto err_probe;
 545         }
 546
 547         pci_set_master(pci_dev);
 548
 549         rc = register_virtio_device(&vp_dev->vdev);
 550         if (rc)
 551                 goto err_register;
 552
 553         return 0;
 554
 555 err_register:
 556         if (vp_dev->ioaddr)
 557              virtio_pci_legacy_remove(vp_dev);
 558         else
 559              virtio_pci_modern_remove(vp_dev);
 560 err_probe:
 561         pci_disable_device(pci_dev);
 562 err_enable_device:
 563         kfree(vp_dev);
 564         return rc;
 565 }
 566
 567 static void virtio_pci_remove(struct pci_dev *pci_dev)
 568 {
 569         struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
 570         struct device *dev = get_device(&vp_dev->vdev.dev);
 571
 572         unregister_virtio_device(&vp_dev->vdev);
 573
 574         if (vp_dev->ioaddr)
 575                 virtio_pci_legacy_remove(vp_dev);
 576         else
 577                 virtio_pci_modern_remove(vp_dev);
 578
 579         pci_disable_device(pci_dev);
 580         put_device(dev);
 581 }
 582
 583 static struct pci_driver virtio_pci_driver = {
 584         .name           = "virtio-pci",
 585         .id_table       = virtio_pci_id_table,
 586         .probe          = virtio_pci_probe,
 587         .remove         = virtio_pci_remove,
 588 #ifdef CONFIG_PM_SLEEP
 589         .driver.pm      = &virtio_pci_pm_ops,
 590 #endif
 591 };
 592
 593 module_pci_driver(virtio_pci_driver);
 594
 595 MODULE_AUTHOR("Anthony Liguori <aliguori@us.ibm.com>");
 596 MODULE_DESCRIPTION("virtio-pci");
 597 MODULE_LICENSE("GPL");
 598 MODULE_VERSION("1");