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)
 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(vp_dev->pci_dev, nvectors, nvectors,
 130                         PCI_IRQ_MSIX);
 131         if (err < 0)
 132                 goto error;
 133         vp_dev->msix_enabled = 1;
 134
 135         /* Set the vector used for configuration */
 136         v = vp_dev->msix_used_vectors;
 137         snprintf(vp_dev->msix_names[v], sizeof *vp_dev->msix_names,
 138                  "%s-config", name);
 139         err = request_irq(pci_irq_vector(vp_dev->pci_dev, v),
 140                           vp_config_changed, 0, vp_dev->msix_names[v],
 141                           vp_dev);
 142         if (err)
 143                 goto error;
 144         ++vp_dev->msix_used_vectors;
 145
 146         v = vp_dev->config_vector(vp_dev, v);
 147         /* Verify we had enough resources to assign the vector */
 148         if (v == VIRTIO_MSI_NO_VECTOR) {
 149                 err = -EBUSY;
 150                 goto error;
 151         }
 152
 153         if (!per_vq_vectors) {
 154                 /* Shared vector for all VQs */
 155                 v = vp_dev->msix_used_vectors;
 156                 snprintf(vp_dev->msix_names[v], sizeof *vp_dev->msix_names,
 157                          "%s-virtqueues", name);
 158                 err = request_irq(pci_irq_vector(vp_dev->pci_dev, v),
 159                                   vp_vring_interrupt, 0, vp_dev->msix_names[v],
 160                                   vp_dev);
 161                 if (err)
 162                         goto error;
 163                 ++vp_dev->msix_used_vectors;
 164         }
 165         return 0;
 166 error:
 167         return err;
 168 }
 169
 170 static struct virtqueue *vp_setup_vq(struct virtio_device *vdev, unsigned index,
 171                                      void (*callback)(struct virtqueue *vq),
 172                                      const char *name,
 173                                      u16 msix_vec)
 174 {
 175         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 176         struct virtio_pci_vq_info *info = kmalloc(sizeof *info, GFP_KERNEL);
 177         struct virtqueue *vq;
 178         unsigned long flags;
 179
 180         /* fill out our structure that represents an active queue */
 181         if (!info)
 182                 return ERR_PTR(-ENOMEM);
 183
 184         vq = vp_dev->setup_vq(vp_dev, info, index, callback, name, msix_vec);
 185         if (IS_ERR(vq))
 186                 goto out_info;
 187
 188         info->vq = vq;
 189         if (callback) {
 190                 spin_lock_irqsave(&vp_dev->lock, flags);
 191                 list_add(&info->node, &vp_dev->virtqueues);
 192                 spin_unlock_irqrestore(&vp_dev->lock, flags);
 193         } else {
 194                 INIT_LIST_HEAD(&info->node);
 195         }
 196
 197         vp_dev->vqs[index] = info;
 198         return vq;
 199
 200 out_info:
 201         kfree(info);
 202         return vq;
 203 }
 204
 205 static void vp_del_vq(struct virtqueue *vq)
 206 {
 207         struct virtio_pci_device *vp_dev = to_vp_device(vq->vdev);
 208         struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index];
 209         unsigned long flags;
 210
 211         spin_lock_irqsave(&vp_dev->lock, flags);
 212         list_del(&info->node);
 213         spin_unlock_irqrestore(&vp_dev->lock, flags);
 214
 215         vp_dev->del_vq(info);
 216         kfree(info);
 217 }
 218
 219 /* the config->del_vqs() implementation */
 220 void vp_del_vqs(struct virtio_device *vdev)
 221 {
 222         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 223         struct virtqueue *vq, *n;
 224         int i;
 225
 226         list_for_each_entry_safe(vq, n, &vdev->vqs, list) {
 227                 if (vp_dev->per_vq_vectors) {
 228                         int v = vp_dev->vqs[vq->index]->msix_vector;
 229
 230                         if (v != VIRTIO_MSI_NO_VECTOR)
 231                                 free_irq(pci_irq_vector(vp_dev->pci_dev, v),
 232                                         vq);
 233                 }
 234                 vp_del_vq(vq);
 235         }
 236         vp_dev->per_vq_vectors = false;
 237
 238         if (vp_dev->intx_enabled) {
 239                 free_irq(vp_dev->pci_dev->irq, vp_dev);
 240                 vp_dev->intx_enabled = 0;
 241         }
 242
 243         for (i = 0; i < vp_dev->msix_used_vectors; ++i)
 244                 free_irq(pci_irq_vector(vp_dev->pci_dev, i), vp_dev);
 245
 246         for (i = 0; i < vp_dev->msix_vectors; i++)
 247                 if (vp_dev->msix_affinity_masks[i])
 248                         free_cpumask_var(vp_dev->msix_affinity_masks[i]);
 249
 250         if (vp_dev->msix_enabled) {
 251                 /* Disable the vector used for configuration */
 252                 vp_dev->config_vector(vp_dev, VIRTIO_MSI_NO_VECTOR);
 253
 254                 pci_free_irq_vectors(vp_dev->pci_dev);
 255                 vp_dev->msix_enabled = 0;
 256         }
 257
 258         vp_dev->msix_vectors = 0;
 259         vp_dev->msix_used_vectors = 0;
 260         kfree(vp_dev->msix_names);
 261         vp_dev->msix_names = NULL;
 262         kfree(vp_dev->msix_affinity_masks);
 263         vp_dev->msix_affinity_masks = NULL;
 264         kfree(vp_dev->vqs);
 265         vp_dev->vqs = NULL;
 266 }
 267
 268 static int vp_find_vqs_msix(struct virtio_device *vdev, unsigned nvqs,
 269                               struct virtqueue *vqs[],
 270                               vq_callback_t *callbacks[],
 271                               const char * const names[],
 272                               bool per_vq_vectors)
 273 {
 274         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 275         u16 msix_vec;
 276         int i, err, nvectors, allocated_vectors;
 277
 278         vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL);
 279         if (!vp_dev->vqs)
 280                 return -ENOMEM;
 281
 282         if (per_vq_vectors) {
 283                 /* Best option: one for change interrupt, one per vq. */
 284                 nvectors = 1;
 285                 for (i = 0; i < nvqs; ++i)
 286                         if (callbacks[i])
 287                                 ++nvectors;
 288         } else {
 289                 /* Second best: one for change, shared for all vqs. */
 290                 nvectors = 2;
 291         }
 292
 293         err = vp_request_msix_vectors(vdev, nvectors, per_vq_vectors);
 294         if (err)
 295                 goto error_find;
 296
 297         vp_dev->per_vq_vectors = per_vq_vectors;
 298         allocated_vectors = vp_dev->msix_used_vectors;
 299         for (i = 0; i < nvqs; ++i) {
 300                 if (!names[i]) {
 301                         vqs[i] = NULL;
 302                         continue;
 303                 }
 304
 305                 if (!callbacks[i])
 306                         msix_vec = VIRTIO_MSI_NO_VECTOR;
 307                 else if (vp_dev->per_vq_vectors)
 308                         msix_vec = allocated_vectors++;
 309                 else
 310                         msix_vec = VP_MSIX_VQ_VECTOR;
 311                 vqs[i] = vp_setup_vq(vdev, i, callbacks[i], names[i], msix_vec);
 312                 if (IS_ERR(vqs[i])) {
 313                         err = PTR_ERR(vqs[i]);
 314                         goto error_find;
 315                 }
 316
 317                 if (!vp_dev->per_vq_vectors || msix_vec == VIRTIO_MSI_NO_VECTOR)
 318                         continue;
 319
 320                 /* allocate per-vq irq if available and necessary */
 321                 snprintf(vp_dev->msix_names[msix_vec],
 322                          sizeof *vp_dev->msix_names,
 323                          "%s-%s",
 324                          dev_name(&vp_dev->vdev.dev), names[i]);
 325                 err = request_irq(pci_irq_vector(vp_dev->pci_dev, msix_vec),
 326                                   vring_interrupt, 0,
 327                                   vp_dev->msix_names[msix_vec],
 328                                   vqs[i]);
 329                 if (err)
 330                         goto error_find;
 331         }
 332         return 0;
 333
 334 error_find:
 335         vp_del_vqs(vdev);
 336         return err;
 337 }
 338
 339 static int vp_find_vqs_intx(struct virtio_device *vdev, unsigned nvqs,
 340                 struct virtqueue *vqs[], vq_callback_t *callbacks[],
 341                 const char * const names[])
 342 {
 343         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 344         int i, err;
 345
 346         vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL);
 347         if (!vp_dev->vqs)
 348                 return -ENOMEM;
 349
 350         err = request_irq(vp_dev->pci_dev->irq, vp_interrupt, IRQF_SHARED,
 351                         dev_name(&vdev->dev), vp_dev);
 352         if (err)
 353                 goto out_del_vqs;
 354
 355         vp_dev->intx_enabled = 1;
 356         vp_dev->per_vq_vectors = false;
 357         for (i = 0; i < nvqs; ++i) {
 358                 if (!names[i]) {
 359                         vqs[i] = NULL;
 360                         continue;
 361                 }
 362                 vqs[i] = vp_setup_vq(vdev, i, callbacks[i], names[i],
 363                                 VIRTIO_MSI_NO_VECTOR);
 364                 if (IS_ERR(vqs[i])) {
 365                         err = PTR_ERR(vqs[i]);
 366                         goto out_del_vqs;
 367                 }
 368         }
 369
 370         return 0;
 371 out_del_vqs:
 372         vp_del_vqs(vdev);
 373         return err;
 374 }
 375
 376 /* the config->find_vqs() implementation */
 377 int vp_find_vqs(struct virtio_device *vdev, unsigned nvqs,
 378                 struct virtqueue *vqs[],
 379                 vq_callback_t *callbacks[],
 380                 const char * const names[])
 381 {
 382         int err;
 383
 384         /* Try MSI-X with one vector per queue. */
 385         err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, true);
 386         if (!err)
 387                 return 0;
 388         /* Fallback: MSI-X with one vector for config, one shared for queues. */
 389         err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, false);
 390         if (!err)
 391                 return 0;
 392         /* Finally fall back to regular interrupts. */
 393         return vp_find_vqs_intx(vdev, nvqs, vqs, callbacks, names);
 394 }
 395
 396 const char *vp_bus_name(struct virtio_device *vdev)
 397 {
 398         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 399
 400         return pci_name(vp_dev->pci_dev);
 401 }
 402
 403 /* Setup the affinity for a virtqueue:
 404  * - force the affinity for per vq vector
 405  * - OR over all affinities for shared MSI
 406  * - ignore the affinity request if we're using INTX
 407  */
 408 int vp_set_vq_affinity(struct virtqueue *vq, int cpu)
 409 {
 410         struct virtio_device *vdev = vq->vdev;
 411         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 412         struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index];
 413         struct cpumask *mask;
 414         unsigned int irq;
 415
 416         if (!vq->callback)
 417                 return -EINVAL;
 418
 419         if (vp_dev->msix_enabled) {
 420                 mask = vp_dev->msix_affinity_masks[info->msix_vector];
 421                 irq = pci_irq_vector(vp_dev->pci_dev, info->msix_vector);
 422                 if (cpu == -1)
 423                         irq_set_affinity_hint(irq, NULL);
 424                 else {
 425                         cpumask_clear(mask);
 426                         cpumask_set_cpu(cpu, mask);
 427                         irq_set_affinity_hint(irq, mask);
 428                 }
 429         }
 430         return 0;
 431 }
 432
 433 #ifdef CONFIG_PM_SLEEP
 434 static int virtio_pci_freeze(struct device *dev)
 435 {
 436         struct pci_dev *pci_dev = to_pci_dev(dev);
 437         struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
 438         int ret;
 439
 440         ret = virtio_device_freeze(&vp_dev->vdev);
 441
 442         if (!ret)
 443                 pci_disable_device(pci_dev);
 444         return ret;
 445 }
 446
 447 static int virtio_pci_restore(struct device *dev)
 448 {
 449         struct pci_dev *pci_dev = to_pci_dev(dev);
 450         struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
 451         int ret;
 452
 453         ret = pci_enable_device(pci_dev);
 454         if (ret)
 455                 return ret;
 456
 457         pci_set_master(pci_dev);
 458         return virtio_device_restore(&vp_dev->vdev);
 459 }
 460
 461 static const struct dev_pm_ops virtio_pci_pm_ops = {
 462         SET_SYSTEM_SLEEP_PM_OPS(virtio_pci_freeze, virtio_pci_restore)
 463 };
 464 #endif
 465
 466
 467 /* Qumranet donated their vendor ID for devices 0x1000 thru 0x10FF. */
 468 static const struct pci_device_id virtio_pci_id_table[] = {
 469         { PCI_DEVICE(PCI_VENDOR_ID_REDHAT_QUMRANET, PCI_ANY_ID) },
 470         { 0 }
 471 };
 472
 473 MODULE_DEVICE_TABLE(pci, virtio_pci_id_table);
 474
 475 static void virtio_pci_release_dev(struct device *_d)
 476 {
 477         struct virtio_device *vdev = dev_to_virtio(_d);
 478         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 479
 480         /* As struct device is a kobject, it's not safe to
 481          * free the memory (including the reference counter itself)
 482          * until it's release callback. */
 483         kfree(vp_dev);
 484 }
 485
 486 static int virtio_pci_probe(struct pci_dev *pci_dev,
 487                             const struct pci_device_id *id)
 488 {
 489         struct virtio_pci_device *vp_dev;
 490         int rc;
 491
 492         /* allocate our structure and fill it out */
 493         vp_dev = kzalloc(sizeof(struct virtio_pci_device), GFP_KERNEL);
 494         if (!vp_dev)
 495                 return -ENOMEM;
 496
 497         pci_set_drvdata(pci_dev, vp_dev);
 498         vp_dev->vdev.dev.parent = &pci_dev->dev;
 499         vp_dev->vdev.dev.release = virtio_pci_release_dev;
 500         vp_dev->pci_dev = pci_dev;
 501         INIT_LIST_HEAD(&vp_dev->virtqueues);
 502         spin_lock_init(&vp_dev->lock);
 503
 504         /* enable the device */
 505         rc = pci_enable_device(pci_dev);
 506         if (rc)
 507                 goto err_enable_device;
 508
 509         if (force_legacy) {
 510                 rc = virtio_pci_legacy_probe(vp_dev);
 511                 /* Also try modern mode if we can't map BAR0 (no IO space). */
 512                 if (rc == -ENODEV || rc == -ENOMEM)
 513                         rc = virtio_pci_modern_probe(vp_dev);
 514                 if (rc)
 515                         goto err_probe;
 516         } else {
 517                 rc = virtio_pci_modern_probe(vp_dev);
 518                 if (rc == -ENODEV)
 519                         rc = virtio_pci_legacy_probe(vp_dev);
 520                 if (rc)
 521                         goto err_probe;
 522         }
 523
 524         pci_set_master(pci_dev);
 525
 526         rc = register_virtio_device(&vp_dev->vdev);
 527         if (rc)
 528                 goto err_register;
 529
 530         return 0;
 531
 532 err_register:
 533         if (vp_dev->ioaddr)
 534              virtio_pci_legacy_remove(vp_dev);
 535         else
 536              virtio_pci_modern_remove(vp_dev);
 537 err_probe:
 538         pci_disable_device(pci_dev);
 539 err_enable_device:
 540         kfree(vp_dev);
 541         return rc;
 542 }
 543
 544 static void virtio_pci_remove(struct pci_dev *pci_dev)
 545 {
 546         struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
 547         struct device *dev = get_device(&vp_dev->vdev.dev);
 548
 549         unregister_virtio_device(&vp_dev->vdev);
 550
 551         if (vp_dev->ioaddr)
 552                 virtio_pci_legacy_remove(vp_dev);
 553         else
 554                 virtio_pci_modern_remove(vp_dev);
 555
 556         pci_disable_device(pci_dev);
 557         put_device(dev);
 558 }
 559
 560 static struct pci_driver virtio_pci_driver = {
 561         .name           = "virtio-pci",
 562         .id_table       = virtio_pci_id_table,
 563         .probe          = virtio_pci_probe,
 564         .remove         = virtio_pci_remove,
 565 #ifdef CONFIG_PM_SLEEP
 566         .driver.pm      = &virtio_pci_pm_ops,
 567 #endif
 568 };
 569
 570 module_pci_driver(virtio_pci_driver);
 571
 572 MODULE_AUTHOR("Anthony Liguori <aliguori@us.ibm.com>");
 573 MODULE_DESCRIPTION("virtio-pci");
 574 MODULE_LICENSE("GPL");
 575 MODULE_VERSION("1");