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                                      u16 msix_vec)
 176 {
 177         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 178         struct virtio_pci_vq_info *info = kmalloc(sizeof *info, GFP_KERNEL);
 179         struct virtqueue *vq;
 180         unsigned long flags;
 181
 182         /* fill out our structure that represents an active queue */
 183         if (!info)
 184                 return ERR_PTR(-ENOMEM);
 185
 186         vq = vp_dev->setup_vq(vp_dev, info, index, callback, name,
 187                               msix_vec);
 188         if (IS_ERR(vq))
 189                 goto out_info;
 190
 191         info->vq = vq;
 192         if (callback) {
 193                 spin_lock_irqsave(&vp_dev->lock, flags);
 194                 list_add(&info->node, &vp_dev->virtqueues);
 195                 spin_unlock_irqrestore(&vp_dev->lock, flags);
 196         } else {
 197                 INIT_LIST_HEAD(&info->node);
 198         }
 199
 200         vp_dev->vqs[index] = info;
 201         return vq;
 202
 203 out_info:
 204         kfree(info);
 205         return vq;
 206 }
 207
 208 static void vp_del_vq(struct virtqueue *vq)
 209 {
 210         struct virtio_pci_device *vp_dev = to_vp_device(vq->vdev);
 211         struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index];
 212         unsigned long flags;
 213
 214         spin_lock_irqsave(&vp_dev->lock, flags);
 215         list_del(&info->node);
 216         spin_unlock_irqrestore(&vp_dev->lock, flags);
 217
 218         vp_dev->del_vq(info);
 219         kfree(info);
 220 }
 221
 222 /* the config->del_vqs() implementation */
 223 void vp_del_vqs(struct virtio_device *vdev)
 224 {
 225         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 226         struct virtqueue *vq, *n;
 227         int i;
 228
 229         list_for_each_entry_safe(vq, n, &vdev->vqs, list) {
 230                 if (vp_dev->per_vq_vectors) {
 231                         int v = vp_dev->vqs[vq->index]->msix_vector;
 232
 233                         if (v != VIRTIO_MSI_NO_VECTOR) {
 234                                 int irq = pci_irq_vector(vp_dev->pci_dev, v);
 235
 236                                 irq_set_affinity_hint(irq, NULL);
 237                                 free_irq(irq, vq);
 238                         }
 239                 }
 240                 vp_del_vq(vq);
 241         }
 242         vp_dev->per_vq_vectors = false;
 243
 244         if (vp_dev->intx_enabled) {
 245                 free_irq(vp_dev->pci_dev->irq, vp_dev);
 246                 vp_dev->intx_enabled = 0;
 247         }
 248
 249         for (i = 0; i < vp_dev->msix_used_vectors; ++i)
 250                 free_irq(pci_irq_vector(vp_dev->pci_dev, i), vp_dev);
 251
 252         for (i = 0; i < vp_dev->msix_vectors; i++)
 253                 if (vp_dev->msix_affinity_masks[i])
 254                         free_cpumask_var(vp_dev->msix_affinity_masks[i]);
 255
 256         if (vp_dev->msix_enabled) {
 257                 /* Disable the vector used for configuration */
 258                 vp_dev->config_vector(vp_dev, VIRTIO_MSI_NO_VECTOR);
 259
 260                 pci_free_irq_vectors(vp_dev->pci_dev);
 261                 vp_dev->msix_enabled = 0;
 262         }
 263
 264         vp_dev->msix_vectors = 0;
 265         vp_dev->msix_used_vectors = 0;
 266         kfree(vp_dev->msix_names);
 267         vp_dev->msix_names = NULL;
 268         kfree(vp_dev->msix_affinity_masks);
 269         vp_dev->msix_affinity_masks = NULL;
 270         kfree(vp_dev->vqs);
 271         vp_dev->vqs = NULL;
 272 }
 273
 274 static int vp_find_vqs_msix(struct virtio_device *vdev, unsigned nvqs,
 275                 struct virtqueue *vqs[], vq_callback_t *callbacks[],
 276                 const char * const names[], bool per_vq_vectors,
 277                 struct irq_affinity *desc)
 278 {
 279         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 280         u16 msix_vec;
 281         int i, err, nvectors, allocated_vectors;
 282
 283         vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL);
 284         if (!vp_dev->vqs)
 285                 return -ENOMEM;
 286
 287         if (per_vq_vectors) {
 288                 /* Best option: one for change interrupt, one per vq. */
 289                 nvectors = 1;
 290                 for (i = 0; i < nvqs; ++i)
 291                         if (callbacks[i])
 292                                 ++nvectors;
 293         } else {
 294                 /* Second best: one for change, shared for all vqs. */
 295                 nvectors = 2;
 296         }
 297
 298         err = vp_request_msix_vectors(vdev, nvectors, per_vq_vectors,
 299                                       per_vq_vectors ? desc : NULL);
 300         if (err)
 301                 goto error_find;
 302
 303         vp_dev->per_vq_vectors = per_vq_vectors;
 304         allocated_vectors = vp_dev->msix_used_vectors;
 305         for (i = 0; i < nvqs; ++i) {
 306                 if (!names[i]) {
 307                         vqs[i] = NULL;
 308                         continue;
 309                 }
 310
 311                 if (!callbacks[i])
 312                         msix_vec = VIRTIO_MSI_NO_VECTOR;
 313                 else if (vp_dev->per_vq_vectors)
 314                         msix_vec = allocated_vectors++;
 315                 else
 316                         msix_vec = VP_MSIX_VQ_VECTOR;
 317                 vqs[i] = vp_setup_vq(vdev, i, callbacks[i], names[i],
 318                                      msix_vec);
 319                 if (IS_ERR(vqs[i])) {
 320                         err = PTR_ERR(vqs[i]);
 321                         goto error_find;
 322                 }
 323
 324                 if (!vp_dev->per_vq_vectors || msix_vec == VIRTIO_MSI_NO_VECTOR)
 325                         continue;
 326
 327                 /* allocate per-vq irq if available and necessary */
 328                 snprintf(vp_dev->msix_names[msix_vec],
 329                          sizeof *vp_dev->msix_names,
 330                          "%s-%s",
 331                          dev_name(&vp_dev->vdev.dev), names[i]);
 332                 err = request_irq(pci_irq_vector(vp_dev->pci_dev, msix_vec),
 333                                   vring_interrupt, 0,
 334                                   vp_dev->msix_names[msix_vec],
 335                                   vqs[i]);
 336                 if (err)
 337                         goto error_find;
 338         }
 339         return 0;
 340
 341 error_find:
 342         vp_del_vqs(vdev);
 343         return err;
 344 }
 345
 346 static int vp_find_vqs_intx(struct virtio_device *vdev, unsigned nvqs,
 347                 struct virtqueue *vqs[], vq_callback_t *callbacks[],
 348                 const char * const names[])
 349 {
 350         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 351         int i, err;
 352
 353         vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL);
 354         if (!vp_dev->vqs)
 355                 return -ENOMEM;
 356
 357         err = request_irq(vp_dev->pci_dev->irq, vp_interrupt, IRQF_SHARED,
 358                         dev_name(&vdev->dev), vp_dev);
 359         if (err)
 360                 goto out_del_vqs;
 361
 362         vp_dev->intx_enabled = 1;
 363         vp_dev->per_vq_vectors = false;
 364         for (i = 0; i < nvqs; ++i) {
 365                 if (!names[i]) {
 366                         vqs[i] = NULL;
 367                         continue;
 368                 }
 369                 vqs[i] = vp_setup_vq(vdev, i, callbacks[i], names[i],
 370                                      VIRTIO_MSI_NO_VECTOR);
 371                 if (IS_ERR(vqs[i])) {
 372                         err = PTR_ERR(vqs[i]);
 373                         goto out_del_vqs;
 374                 }
 375         }
 376
 377         return 0;
 378 out_del_vqs:
 379         vp_del_vqs(vdev);
 380         return err;
 381 }
 382
 383 /* the config->find_vqs() implementation */
 384 int vp_find_vqs(struct virtio_device *vdev, unsigned nvqs,
 385                 struct virtqueue *vqs[], vq_callback_t *callbacks[],
 386                 const char * const names[], struct irq_affinity *desc)
 387 {
 388         int err;
 389
 390         /* Try MSI-X with one vector per queue. */
 391         err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, true, desc);
 392         if (!err)
 393                 return 0;
 394         /* Fallback: MSI-X with one vector for config, one shared for queues. */
 395         err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, false, desc);
 396         if (!err)
 397                 return 0;
 398         /* Finally fall back to regular interrupts. */
 399         return vp_find_vqs_intx(vdev, nvqs, vqs, callbacks, names);
 400 }
 401
 402 const char *vp_bus_name(struct virtio_device *vdev)
 403 {
 404         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 405
 406         return pci_name(vp_dev->pci_dev);
 407 }
 408
 409 /* Setup the affinity for a virtqueue:
 410  * - force the affinity for per vq vector
 411  * - OR over all affinities for shared MSI
 412  * - ignore the affinity request if we're using INTX
 413  */
 414 int vp_set_vq_affinity(struct virtqueue *vq, int cpu)
 415 {
 416         struct virtio_device *vdev = vq->vdev;
 417         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 418         struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index];
 419         struct cpumask *mask;
 420         unsigned int irq;
 421
 422         if (!vq->callback)
 423                 return -EINVAL;
 424
 425         if (vp_dev->msix_enabled) {
 426                 mask = vp_dev->msix_affinity_masks[info->msix_vector];
 427                 irq = pci_irq_vector(vp_dev->pci_dev, info->msix_vector);
 428                 if (cpu == -1)
 429                         irq_set_affinity_hint(irq, NULL);
 430                 else {
 431                         cpumask_clear(mask);
 432                         cpumask_set_cpu(cpu, mask);
 433                         irq_set_affinity_hint(irq, mask);
 434                 }
 435         }
 436         return 0;
 437 }
 438
 439 const struct cpumask *vp_get_vq_affinity(struct virtio_device *vdev, int index)
 440 {
 441         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 442
 443         if (!vp_dev->per_vq_vectors ||
 444             vp_dev->vqs[index]->msix_vector == VIRTIO_MSI_NO_VECTOR)
 445                 return NULL;
 446
 447         return pci_irq_get_affinity(vp_dev->pci_dev,
 448                                     vp_dev->vqs[index]->msix_vector);
 449 }
 450
 451 #ifdef CONFIG_PM_SLEEP
 452 static int virtio_pci_freeze(struct device *dev)
 453 {
 454         struct pci_dev *pci_dev = to_pci_dev(dev);
 455         struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
 456         int ret;
 457
 458         ret = virtio_device_freeze(&vp_dev->vdev);
 459
 460         if (!ret)
 461                 pci_disable_device(pci_dev);
 462         return ret;
 463 }
 464
 465 static int virtio_pci_restore(struct device *dev)
 466 {
 467         struct pci_dev *pci_dev = to_pci_dev(dev);
 468         struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
 469         int ret;
 470
 471         ret = pci_enable_device(pci_dev);
 472         if (ret)
 473                 return ret;
 474
 475         pci_set_master(pci_dev);
 476         return virtio_device_restore(&vp_dev->vdev);
 477 }
 478
 479 static const struct dev_pm_ops virtio_pci_pm_ops = {
 480         SET_SYSTEM_SLEEP_PM_OPS(virtio_pci_freeze, virtio_pci_restore)
 481 };
 482 #endif
 483
 484
 485 /* Qumranet donated their vendor ID for devices 0x1000 thru 0x10FF. */
 486 static const struct pci_device_id virtio_pci_id_table[] = {
 487         { PCI_DEVICE(PCI_VENDOR_ID_REDHAT_QUMRANET, PCI_ANY_ID) },
 488         { 0 }
 489 };
 490
 491 MODULE_DEVICE_TABLE(pci, virtio_pci_id_table);
 492
 493 static void virtio_pci_release_dev(struct device *_d)
 494 {
 495         struct virtio_device *vdev = dev_to_virtio(_d);
 496         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 497
 498         /* As struct device is a kobject, it's not safe to
 499          * free the memory (including the reference counter itself)
 500          * until it's release callback. */
 501         kfree(vp_dev);
 502 }
 503
 504 static int virtio_pci_probe(struct pci_dev *pci_dev,
 505                             const struct pci_device_id *id)
 506 {
 507         struct virtio_pci_device *vp_dev;
 508         int rc;
 509
 510         /* allocate our structure and fill it out */
 511         vp_dev = kzalloc(sizeof(struct virtio_pci_device), GFP_KERNEL);
 512         if (!vp_dev)
 513                 return -ENOMEM;
 514
 515         pci_set_drvdata(pci_dev, vp_dev);
 516         vp_dev->vdev.dev.parent = &pci_dev->dev;
 517         vp_dev->vdev.dev.release = virtio_pci_release_dev;
 518         vp_dev->pci_dev = pci_dev;
 519         INIT_LIST_HEAD(&vp_dev->virtqueues);
 520         spin_lock_init(&vp_dev->lock);
 521
 522         /* enable the device */
 523         rc = pci_enable_device(pci_dev);
 524         if (rc)
 525                 goto err_enable_device;
 526
 527         if (force_legacy) {
 528                 rc = virtio_pci_legacy_probe(vp_dev);
 529                 /* Also try modern mode if we can't map BAR0 (no IO space). */
 530                 if (rc == -ENODEV || rc == -ENOMEM)
 531                         rc = virtio_pci_modern_probe(vp_dev);
 532                 if (rc)
 533                         goto err_probe;
 534         } else {
 535                 rc = virtio_pci_modern_probe(vp_dev);
 536                 if (rc == -ENODEV)
 537                         rc = virtio_pci_legacy_probe(vp_dev);
 538                 if (rc)
 539                         goto err_probe;
 540         }
 541
 542         pci_set_master(pci_dev);
 543
 544         rc = register_virtio_device(&vp_dev->vdev);
 545         if (rc)
 546                 goto err_register;
 547
 548         return 0;
 549
 550 err_register:
 551         if (vp_dev->ioaddr)
 552              virtio_pci_legacy_remove(vp_dev);
 553         else
 554              virtio_pci_modern_remove(vp_dev);
 555 err_probe:
 556         pci_disable_device(pci_dev);
 557 err_enable_device:
 558         kfree(vp_dev);
 559         return rc;
 560 }
 561
 562 static void virtio_pci_remove(struct pci_dev *pci_dev)
 563 {
 564         struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
 565         struct device *dev = get_device(&vp_dev->vdev.dev);
 566
 567         unregister_virtio_device(&vp_dev->vdev);
 568
 569         if (vp_dev->ioaddr)
 570                 virtio_pci_legacy_remove(vp_dev);
 571         else
 572                 virtio_pci_modern_remove(vp_dev);
 573
 574         pci_disable_device(pci_dev);
 575         put_device(dev);
 576 }
 577
 578 static struct pci_driver virtio_pci_driver = {
 579         .name           = "virtio-pci",
 580         .id_table       = virtio_pci_id_table,
 581         .probe          = virtio_pci_probe,
 582         .remove         = virtio_pci_remove,
 583 #ifdef CONFIG_PM_SLEEP
 584         .driver.pm      = &virtio_pci_pm_ops,
 585 #endif
 586 };
 587
 588 module_pci_driver(virtio_pci_driver);
 589
 590 MODULE_AUTHOR("Anthony Liguori <aliguori@us.ibm.com>");
 591 MODULE_DESCRIPTION("virtio-pci");
 592 MODULE_LICENSE("GPL");
 593 MODULE_VERSION("1");