drivers/virtio/virtio_pci.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_legacy.c"
  21
  22 /* wait for pending irq handlers */
  23 void vp_synchronize_vectors(struct virtio_device *vdev)
  24 {
  25         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
  26         int i;
  27
  28         if (vp_dev->intx_enabled)
  29                 synchronize_irq(vp_dev->pci_dev->irq);
  30
  31         for (i = 0; i < vp_dev->msix_vectors; ++i)
  32                 synchronize_irq(vp_dev->msix_entries[i].vector);
  33 }
  34
  35 /* the notify function used when creating a virt queue */
  36 bool vp_notify(struct virtqueue *vq)
  37 {
  38         /* we write the queue's selector into the notification register to
  39          * signal the other end */
  40         iowrite16(vq->index, (void __iomem *)vq->priv);
  41         return true;
  42 }
  43
  44 /* Handle a configuration change: Tell driver if it wants to know. */
  45 static irqreturn_t vp_config_changed(int irq, void *opaque)
  46 {
  47         struct virtio_pci_device *vp_dev = opaque;
  48
  49         virtio_config_changed(&vp_dev->vdev);
  50         return IRQ_HANDLED;
  51 }
  52
  53 /* Notify all virtqueues on an interrupt. */
  54 static irqreturn_t vp_vring_interrupt(int irq, void *opaque)
  55 {
  56         struct virtio_pci_device *vp_dev = opaque;
  57         struct virtio_pci_vq_info *info;
  58         irqreturn_t ret = IRQ_NONE;
  59         unsigned long flags;
  60
  61         spin_lock_irqsave(&vp_dev->lock, flags);
  62         list_for_each_entry(info, &vp_dev->virtqueues, node) {
  63                 if (vring_interrupt(irq, info->vq) == IRQ_HANDLED)
  64                         ret = IRQ_HANDLED;
  65         }
  66         spin_unlock_irqrestore(&vp_dev->lock, flags);
  67
  68         return ret;
  69 }
  70
  71 /* A small wrapper to also acknowledge the interrupt when it's handled.
  72  * I really need an EIO hook for the vring so I can ack the interrupt once we
  73  * know that we'll be handling the IRQ but before we invoke the callback since
  74  * the callback may notify the host which results in the host attempting to
  75  * raise an interrupt that we would then mask once we acknowledged the
  76  * interrupt. */
  77 static irqreturn_t vp_interrupt(int irq, void *opaque)
  78 {
  79         struct virtio_pci_device *vp_dev = opaque;
  80         u8 isr;
  81
  82         /* reading the ISR has the effect of also clearing it so it's very
  83          * important to save off the value. */
  84         isr = ioread8(vp_dev->isr);
  85
  86         /* It's definitely not us if the ISR was not high */
  87         if (!isr)
  88                 return IRQ_NONE;
  89
  90         /* Configuration change?  Tell driver if it wants to know. */
  91         if (isr & VIRTIO_PCI_ISR_CONFIG)
  92                 vp_config_changed(irq, opaque);
  93
  94         return vp_vring_interrupt(irq, opaque);
  95 }
  96
  97 static void vp_free_vectors(struct virtio_device *vdev)
  98 {
  99         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 100         int i;
 101
 102         if (vp_dev->intx_enabled) {
 103                 free_irq(vp_dev->pci_dev->irq, vp_dev);
 104                 vp_dev->intx_enabled = 0;
 105         }
 106
 107         for (i = 0; i < vp_dev->msix_used_vectors; ++i)
 108                 free_irq(vp_dev->msix_entries[i].vector, vp_dev);
 109
 110         for (i = 0; i < vp_dev->msix_vectors; i++)
 111                 if (vp_dev->msix_affinity_masks[i])
 112                         free_cpumask_var(vp_dev->msix_affinity_masks[i]);
 113
 114         if (vp_dev->msix_enabled) {
 115                 /* Disable the vector used for configuration */
 116                 vp_dev->config_vector(vp_dev, VIRTIO_MSI_NO_VECTOR);
 117
 118                 pci_disable_msix(vp_dev->pci_dev);
 119                 vp_dev->msix_enabled = 0;
 120         }
 121
 122         vp_dev->msix_vectors = 0;
 123         vp_dev->msix_used_vectors = 0;
 124         kfree(vp_dev->msix_names);
 125         vp_dev->msix_names = NULL;
 126         kfree(vp_dev->msix_entries);
 127         vp_dev->msix_entries = NULL;
 128         kfree(vp_dev->msix_affinity_masks);
 129         vp_dev->msix_affinity_masks = NULL;
 130 }
 131
 132 static int vp_request_msix_vectors(struct virtio_device *vdev, int nvectors,
 133                                    bool per_vq_vectors)
 134 {
 135         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 136         const char *name = dev_name(&vp_dev->vdev.dev);
 137         unsigned i, v;
 138         int err = -ENOMEM;
 139
 140         vp_dev->msix_vectors = nvectors;
 141
 142         vp_dev->msix_entries = kmalloc(nvectors * sizeof *vp_dev->msix_entries,
 143                                        GFP_KERNEL);
 144         if (!vp_dev->msix_entries)
 145                 goto error;
 146         vp_dev->msix_names = kmalloc(nvectors * sizeof *vp_dev->msix_names,
 147                                      GFP_KERNEL);
 148         if (!vp_dev->msix_names)
 149                 goto error;
 150         vp_dev->msix_affinity_masks
 151                 = kzalloc(nvectors * sizeof *vp_dev->msix_affinity_masks,
 152                           GFP_KERNEL);
 153         if (!vp_dev->msix_affinity_masks)
 154                 goto error;
 155         for (i = 0; i < nvectors; ++i)
 156                 if (!alloc_cpumask_var(&vp_dev->msix_affinity_masks[i],
 157                                         GFP_KERNEL))
 158                         goto error;
 159
 160         for (i = 0; i < nvectors; ++i)
 161                 vp_dev->msix_entries[i].entry = i;
 162
 163         err = pci_enable_msix_exact(vp_dev->pci_dev,
 164                                     vp_dev->msix_entries, nvectors);
 165         if (err)
 166                 goto error;
 167         vp_dev->msix_enabled = 1;
 168
 169         /* Set the vector used for configuration */
 170         v = vp_dev->msix_used_vectors;
 171         snprintf(vp_dev->msix_names[v], sizeof *vp_dev->msix_names,
 172                  "%s-config", name);
 173         err = request_irq(vp_dev->msix_entries[v].vector,
 174                           vp_config_changed, 0, vp_dev->msix_names[v],
 175                           vp_dev);
 176         if (err)
 177                 goto error;
 178         ++vp_dev->msix_used_vectors;
 179
 180         v = vp_dev->config_vector(vp_dev, v);
 181         /* Verify we had enough resources to assign the vector */
 182         if (v == VIRTIO_MSI_NO_VECTOR) {
 183                 err = -EBUSY;
 184                 goto error;
 185         }
 186
 187         if (!per_vq_vectors) {
 188                 /* Shared vector for all VQs */
 189                 v = vp_dev->msix_used_vectors;
 190                 snprintf(vp_dev->msix_names[v], sizeof *vp_dev->msix_names,
 191                          "%s-virtqueues", name);
 192                 err = request_irq(vp_dev->msix_entries[v].vector,
 193                                   vp_vring_interrupt, 0, vp_dev->msix_names[v],
 194                                   vp_dev);
 195                 if (err)
 196                         goto error;
 197                 ++vp_dev->msix_used_vectors;
 198         }
 199         return 0;
 200 error:
 201         vp_free_vectors(vdev);
 202         return err;
 203 }
 204
 205 static int vp_request_intx(struct virtio_device *vdev)
 206 {
 207         int err;
 208         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 209
 210         err = request_irq(vp_dev->pci_dev->irq, vp_interrupt,
 211                           IRQF_SHARED, dev_name(&vdev->dev), vp_dev);
 212         if (!err)
 213                 vp_dev->intx_enabled = 1;
 214         return err;
 215 }
 216
 217 static struct virtqueue *vp_setup_vq(struct virtio_device *vdev, unsigned index,
 218                                      void (*callback)(struct virtqueue *vq),
 219                                      const char *name,
 220                                      u16 msix_vec)
 221 {
 222         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 223         struct virtio_pci_vq_info *info = kmalloc(sizeof *info, GFP_KERNEL);
 224         struct virtqueue *vq;
 225         unsigned long flags;
 226
 227         /* fill out our structure that represents an active queue */
 228         if (!info)
 229                 return ERR_PTR(-ENOMEM);
 230
 231         vq = vp_dev->setup_vq(vp_dev, info, index, callback, name, msix_vec);
 232         if (IS_ERR(vq))
 233                 goto out_info;
 234
 235         info->vq = vq;
 236         if (callback) {
 237                 spin_lock_irqsave(&vp_dev->lock, flags);
 238                 list_add(&info->node, &vp_dev->virtqueues);
 239                 spin_unlock_irqrestore(&vp_dev->lock, flags);
 240         } else {
 241                 INIT_LIST_HEAD(&info->node);
 242         }
 243
 244         vp_dev->vqs[index] = info;
 245         return vq;
 246
 247 out_info:
 248         kfree(info);
 249         return vq;
 250 }
 251
 252 static void vp_del_vq(struct virtqueue *vq)
 253 {
 254         struct virtio_pci_device *vp_dev = to_vp_device(vq->vdev);
 255         struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index];
 256         unsigned long flags;
 257
 258         spin_lock_irqsave(&vp_dev->lock, flags);
 259         list_del(&info->node);
 260         spin_unlock_irqrestore(&vp_dev->lock, flags);
 261
 262         vp_dev->del_vq(info);
 263         kfree(info);
 264 }
 265
 266 /* the config->del_vqs() implementation */
 267 void vp_del_vqs(struct virtio_device *vdev)
 268 {
 269         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 270         struct virtqueue *vq, *n;
 271         struct virtio_pci_vq_info *info;
 272
 273         list_for_each_entry_safe(vq, n, &vdev->vqs, list) {
 274                 info = vp_dev->vqs[vq->index];
 275                 if (vp_dev->per_vq_vectors &&
 276                         info->msix_vector != VIRTIO_MSI_NO_VECTOR)
 277                         free_irq(vp_dev->msix_entries[info->msix_vector].vector,
 278                                  vq);
 279                 vp_del_vq(vq);
 280         }
 281         vp_dev->per_vq_vectors = false;
 282
 283         vp_free_vectors(vdev);
 284         kfree(vp_dev->vqs);
 285 }
 286
 287 static int vp_try_to_find_vqs(struct virtio_device *vdev, unsigned nvqs,
 288                               struct virtqueue *vqs[],
 289                               vq_callback_t *callbacks[],
 290                               const char *names[],
 291                               bool use_msix,
 292                               bool per_vq_vectors)
 293 {
 294         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 295         u16 msix_vec;
 296         int i, err, nvectors, allocated_vectors;
 297
 298         vp_dev->vqs = kmalloc(nvqs * sizeof *vp_dev->vqs, GFP_KERNEL);
 299         if (!vp_dev->vqs)
 300                 return -ENOMEM;
 301
 302         if (!use_msix) {
 303                 /* Old style: one normal interrupt for change and all vqs. */
 304                 err = vp_request_intx(vdev);
 305                 if (err)
 306                         goto error_find;
 307         } else {
 308                 if (per_vq_vectors) {
 309                         /* Best option: one for change interrupt, one per vq. */
 310                         nvectors = 1;
 311                         for (i = 0; i < nvqs; ++i)
 312                                 if (callbacks[i])
 313                                         ++nvectors;
 314                 } else {
 315                         /* Second best: one for change, shared for all vqs. */
 316                         nvectors = 2;
 317                 }
 318
 319                 err = vp_request_msix_vectors(vdev, nvectors, per_vq_vectors);
 320                 if (err)
 321                         goto error_find;
 322         }
 323
 324         vp_dev->per_vq_vectors = per_vq_vectors;
 325         allocated_vectors = vp_dev->msix_used_vectors;
 326         for (i = 0; i < nvqs; ++i) {
 327                 if (!names[i]) {
 328                         vqs[i] = NULL;
 329                         continue;
 330                 } else if (!callbacks[i] || !vp_dev->msix_enabled)
 331                         msix_vec = VIRTIO_MSI_NO_VECTOR;
 332                 else if (vp_dev->per_vq_vectors)
 333                         msix_vec = allocated_vectors++;
 334                 else
 335                         msix_vec = VP_MSIX_VQ_VECTOR;
 336                 vqs[i] = vp_setup_vq(vdev, i, callbacks[i], names[i], msix_vec);
 337                 if (IS_ERR(vqs[i])) {
 338                         err = PTR_ERR(vqs[i]);
 339                         goto error_find;
 340                 }
 341
 342                 if (!vp_dev->per_vq_vectors || msix_vec == VIRTIO_MSI_NO_VECTOR)
 343                         continue;
 344
 345                 /* allocate per-vq irq if available and necessary */
 346                 snprintf(vp_dev->msix_names[msix_vec],
 347                          sizeof *vp_dev->msix_names,
 348                          "%s-%s",
 349                          dev_name(&vp_dev->vdev.dev), names[i]);
 350                 err = request_irq(vp_dev->msix_entries[msix_vec].vector,
 351                                   vring_interrupt, 0,
 352                                   vp_dev->msix_names[msix_vec],
 353                                   vqs[i]);
 354                 if (err) {
 355                         vp_del_vq(vqs[i]);
 356                         goto error_find;
 357                 }
 358         }
 359         return 0;
 360
 361 error_find:
 362         vp_del_vqs(vdev);
 363         return err;
 364 }
 365
 366 /* the config->find_vqs() implementation */
 367 int vp_find_vqs(struct virtio_device *vdev, unsigned nvqs,
 368                 struct virtqueue *vqs[],
 369                 vq_callback_t *callbacks[],
 370                 const char *names[])
 371 {
 372         int err;
 373
 374         /* Try MSI-X with one vector per queue. */
 375         err = vp_try_to_find_vqs(vdev, nvqs, vqs, callbacks, names, true, true);
 376         if (!err)
 377                 return 0;
 378         /* Fallback: MSI-X with one vector for config, one shared for queues. */
 379         err = vp_try_to_find_vqs(vdev, nvqs, vqs, callbacks, names,
 380                                  true, false);
 381         if (!err)
 382                 return 0;
 383         /* Finally fall back to regular interrupts. */
 384         return vp_try_to_find_vqs(vdev, nvqs, vqs, callbacks, names,
 385                                   false, false);
 386 }
 387
 388 const char *vp_bus_name(struct virtio_device *vdev)
 389 {
 390         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 391
 392         return pci_name(vp_dev->pci_dev);
 393 }
 394
 395 /* Setup the affinity for a virtqueue:
 396  * - force the affinity for per vq vector
 397  * - OR over all affinities for shared MSI
 398  * - ignore the affinity request if we're using INTX
 399  */
 400 int vp_set_vq_affinity(struct virtqueue *vq, int cpu)
 401 {
 402         struct virtio_device *vdev = vq->vdev;
 403         struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 404         struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index];
 405         struct cpumask *mask;
 406         unsigned int irq;
 407
 408         if (!vq->callback)
 409                 return -EINVAL;
 410
 411         if (vp_dev->msix_enabled) {
 412                 mask = vp_dev->msix_affinity_masks[info->msix_vector];
 413                 irq = vp_dev->msix_entries[info->msix_vector].vector;
 414                 if (cpu == -1)
 415                         irq_set_affinity_hint(irq, NULL);
 416                 else {
 417                         cpumask_set_cpu(cpu, mask);
 418                         irq_set_affinity_hint(irq, mask);
 419                 }
 420         }
 421         return 0;
 422 }
 423
 424 void virtio_pci_release_dev(struct device *_d)
 425 {
 426         /*
 427          * No need for a release method as we allocate/free
 428          * all devices together with the pci devices.
 429          * Provide an empty one to avoid getting a warning from core.
 430          */
 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 const struct dev_pm_ops virtio_pci_pm_ops = {
 462         SET_SYSTEM_SLEEP_PM_OPS(virtio_pci_freeze, virtio_pci_restore)
 463 };
 464 #endif