LXR linux/drivers/virtio/virtio_pci

   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
  22static bool force_legacy = false;
  23
  24#if IS_ENABLED(CONFIG_VIRTIO_PCI_LEGACY)
  25module_param(force_legacy, bool, 0444);
  26MODULE_PARM_DESC(force_legacy,
  27                 "Force legacy mode for transitional virtio 1 devices");
  28#endif
  29
  30/* wait for pending irq handlers */
  31void 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 */
  44bool 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. */
  53static 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. */
  62static 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. */
  85static 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
 105static 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 flags = PCI_IRQ_MSIX;
 111        unsigned i, v;
 112        int err = -ENOMEM;
 113
 114        vp_dev->msix_vectors = nvectors;
 115
 116        vp_dev->msix_names = kmalloc(nvectors * sizeof *vp_dev->msix_names,
 117                                     GFP_KERNEL);
 118        if (!vp_dev->msix_names)
 119                goto error;
 120        vp_dev->msix_affinity_masks
 121                = kzalloc(nvectors * sizeof *vp_dev->msix_affinity_masks,
 122                          GFP_KERNEL);
 123        if (!vp_dev->msix_affinity_masks)
 124                goto error;
 125        for (i = 0; i < nvectors; ++i)
 126                if (!alloc_cpumask_var(&vp_dev->msix_affinity_masks[i],
 127                                        GFP_KERNEL))
 128                        goto error;
 129
 130        if (desc) {
 131                flags |= PCI_IRQ_AFFINITY;
 132                desc->pre_vectors++; /* virtio config vector */
 133        }
 134
 135        err = pci_alloc_irq_vectors_affinity(vp_dev->pci_dev, nvectors,
 136                                             nvectors, flags, desc);
 137        if (err < 0)
 138                goto error;
 139        vp_dev->msix_enabled = 1;
 140
 141        /* Set the vector used for configuration */
 142        v = vp_dev->msix_used_vectors;
 143        snprintf(vp_dev->msix_names[v], sizeof *vp_dev->msix_names,
 144                 "%s-config", name);
 145        err = request_irq(pci_irq_vector(vp_dev->pci_dev, v),
 146                          vp_config_changed, 0, vp_dev->msix_names[v],
 147                          vp_dev);
 148        if (err)
 149                goto error;
 150        ++vp_dev->msix_used_vectors;
 151
 152        v = vp_dev->config_vector(vp_dev, v);
 153        /* Verify we had enough resources to assign the vector */
 154        if (v == VIRTIO_MSI_NO_VECTOR) {
 155                err = -EBUSY;
 156                goto error;
 157        }
 158
 159        if (!per_vq_vectors) {
 160                /* Shared vector for all VQs */
 161                v = vp_dev->msix_used_vectors;
 162                snprintf(vp_dev->msix_names[v], sizeof *vp_dev->msix_names,
 163                         "%s-virtqueues", name);
 164                err = request_irq(pci_irq_vector(vp_dev->pci_dev, v),
 165                                  vp_vring_interrupt, 0, vp_dev->msix_names[v],
 166                                  vp_dev);
 167                if (err)
 168                        goto error;
 169                ++vp_dev->msix_used_vectors;
 170        }
 171        return 0;
 172error:
 173        return err;
 174}
 175
 176static struct virtqueue *vp_setup_vq(struct virtio_device *vdev, unsigned index,
 177                                     void (*callback)(struct virtqueue *vq),
 178                                     const char *name,
 179                                     bool ctx,
 180                                     u16 msix_vec)
 181{
 182        struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 183        struct virtio_pci_vq_info *info = kmalloc(sizeof *info, GFP_KERNEL);
 184        struct virtqueue *vq;
 185        unsigned long flags;
 186
 187        /* fill out our structure that represents an active queue */
 188        if (!info)
 189                return ERR_PTR(-ENOMEM);
 190
 191        vq = vp_dev->setup_vq(vp_dev, info, index, callback, name, ctx,
 192                              msix_vec);
 193        if (IS_ERR(vq))
 194                goto out_info;
 195
 196        info->vq = vq;
 197        if (callback) {
 198                spin_lock_irqsave(&vp_dev->lock, flags);
 199                list_add(&info->node, &vp_dev->virtqueues);
 200                spin_unlock_irqrestore(&vp_dev->lock, flags);
 201        } else {
 202                INIT_LIST_HEAD(&info->node);
 203        }
 204
 205        vp_dev->vqs[index] = info;
 206        return vq;
 207
 208out_info:
 209        kfree(info);
 210        return vq;
 211}
 212
 213static void vp_del_vq(struct virtqueue *vq)
 214{
 215        struct virtio_pci_device *vp_dev = to_vp_device(vq->vdev);
 216        struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index];
 217        unsigned long flags;
 218
 219        spin_lock_irqsave(&vp_dev->lock, flags);
 220        list_del(&info->node);
 221        spin_unlock_irqrestore(&vp_dev->lock, flags);
 222
 223        vp_dev->del_vq(info);
 224        kfree(info);
 225}
 226
 227/* the config->del_vqs() implementation */
 228void vp_del_vqs(struct virtio_device *vdev)
 229{
 230        struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 231        struct virtqueue *vq, *n;
 232        int i;
 233
 234        list_for_each_entry_safe(vq, n, &vdev->vqs, list) {
 235                if (vp_dev->per_vq_vectors) {
 236                        int v = vp_dev->vqs[vq->index]->msix_vector;
 237
 238                        if (v != VIRTIO_MSI_NO_VECTOR) {
 239                                int irq = pci_irq_vector(vp_dev->pci_dev, v);
 240
 241                                irq_set_affinity_hint(irq, NULL);
 242                                free_irq(irq, vq);
 243                        }
 244                }
 245                vp_del_vq(vq);
 246        }
 247        vp_dev->per_vq_vectors = false;
 248
 249        if (vp_dev->intx_enabled) {
 250                free_irq(vp_dev->pci_dev->irq, vp_dev);
 251                vp_dev->intx_enabled = 0;
 252        }
 253
 254        for (i = 0; i < vp_dev->msix_used_vectors; ++i)
 255                free_irq(pci_irq_vector(vp_dev->pci_dev, i), vp_dev);
 256
 257        for (i = 0; i < vp_dev->msix_vectors; i++)
 258                if (vp_dev->msix_affinity_masks[i])
 259                        free_cpumask_var(vp_dev->msix_affinity_masks[i]);
 260
 261        if (vp_dev->msix_enabled) {
 262                /* Disable the vector used for configuration */
 263                vp_dev->config_vector(vp_dev, VIRTIO_MSI_NO_VECTOR);
 264
 265                pci_free_irq_vectors(vp_dev->pci_dev);
 266                vp_dev->msix_enabled = 0;
 267        }
 268
 269        vp_dev->msix_vectors = 0;
 270        vp_dev->msix_used_vectors = 0;
 271        kfree(vp_dev->msix_names);
 272        vp_dev->msix_names = NULL;
 273        kfree(vp_dev->msix_affinity_masks);
 274        vp_dev->msix_affinity_masks = NULL;
 275        kfree(vp_dev->vqs);
 276        vp_dev->vqs = NULL;
 277}
 278
 279static int vp_find_vqs_msix(struct virtio_device *vdev, unsigned nvqs,
 280                struct virtqueue *vqs[], vq_callback_t *callbacks[],
 281                const char * const names[], bool per_vq_vectors,
 282                const bool *ctx,
 283                struct irq_affinity *desc)
 284{
 285        struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 286        u16 msix_vec;
 287        int i, err, nvectors, allocated_vectors;
 288
 289        vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL);
 290        if (!vp_dev->vqs)
 291                return -ENOMEM;
 292
 293        if (per_vq_vectors) {
 294                /* Best option: one for change interrupt, one per vq. */
 295                nvectors = 1;
 296                for (i = 0; i < nvqs; ++i)
 297                        if (callbacks[i])
 298                                ++nvectors;
 299        } else {
 300                /* Second best: one for change, shared for all vqs. */
 301                nvectors = 2;
 302        }
 303
 304        err = vp_request_msix_vectors(vdev, nvectors, per_vq_vectors,
 305                                      per_vq_vectors ? desc : NULL);
 306        if (err)
 307                goto error_find;
 308
 309        vp_dev->per_vq_vectors = per_vq_vectors;
 310        allocated_vectors = vp_dev->msix_used_vectors;
 311        for (i = 0; i < nvqs; ++i) {
 312                if (!names[i]) {
 313                        vqs[i] = NULL;
 314                        continue;
 315                }
 316
 317                if (!callbacks[i])
 318                        msix_vec = VIRTIO_MSI_NO_VECTOR;
 319                else if (vp_dev->per_vq_vectors)
 320                        msix_vec = allocated_vectors++;
 321                else
 322                        msix_vec = VP_MSIX_VQ_VECTOR;
 323                vqs[i] = vp_setup_vq(vdev, i, callbacks[i], names[i],
 324                                     ctx ? ctx[i] : false,
 325                                     msix_vec);
 326                if (IS_ERR(vqs[i])) {
 327                        err = PTR_ERR(vqs[i]);
 328                        goto error_find;
 329                }
 330
 331                if (!vp_dev->per_vq_vectors || msix_vec == VIRTIO_MSI_NO_VECTOR)
 332                        continue;
 333
 334                /* allocate per-vq irq if available and necessary */
 335                snprintf(vp_dev->msix_names[msix_vec],
 336                         sizeof *vp_dev->msix_names,
 337                         "%s-%s",
 338                         dev_name(&vp_dev->vdev.dev), names[i]);
 339                err = request_irq(pci_irq_vector(vp_dev->pci_dev, msix_vec),
 340                                  vring_interrupt, 0,
 341                                  vp_dev->msix_names[msix_vec],
 342                                  vqs[i]);
 343                if (err)
 344                        goto error_find;
 345        }
 346        return 0;
 347
 348error_find:
 349        vp_del_vqs(vdev);
 350        return err;
 351}
 352
 353static int vp_find_vqs_intx(struct virtio_device *vdev, unsigned nvqs,
 354                struct virtqueue *vqs[], vq_callback_t *callbacks[],
 355                const char * const names[], const bool *ctx)
 356{
 357        struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 358        int i, err;
 359
 360        vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL);
 361        if (!vp_dev->vqs)
 362                return -ENOMEM;
 363
 364        err = request_irq(vp_dev->pci_dev->irq, vp_interrupt, IRQF_SHARED,
 365                        dev_name(&vdev->dev), vp_dev);
 366        if (err)
 367                goto out_del_vqs;
 368
 369        vp_dev->intx_enabled = 1;
 370        vp_dev->per_vq_vectors = false;
 371        for (i = 0; i < nvqs; ++i) {
 372                if (!names[i]) {
 373                        vqs[i] = NULL;
 374                        continue;
 375                }
 376                vqs[i] = vp_setup_vq(vdev, i, callbacks[i], names[i],
 377                                     ctx ? ctx[i] : false,
 378                                     VIRTIO_MSI_NO_VECTOR);
 379                if (IS_ERR(vqs[i])) {
 380                        err = PTR_ERR(vqs[i]);
 381                        goto out_del_vqs;
 382                }
 383        }
 384
 385        return 0;
 386out_del_vqs:
 387        vp_del_vqs(vdev);
 388        return err;
 389}
 390
 391/* the config->find_vqs() implementation */
 392int vp_find_vqs(struct virtio_device *vdev, unsigned nvqs,
 393                struct virtqueue *vqs[], vq_callback_t *callbacks[],
 394                const char * const names[], const bool *ctx,
 395                struct irq_affinity *desc)
 396{
 397        int err;
 398
 399        /* Try MSI-X with one vector per queue. */
 400        err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, true, ctx, desc);
 401        if (!err)
 402                return 0;
 403        /* Fallback: MSI-X with one vector for config, one shared for queues. */
 404        err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, false, ctx, desc);
 405        if (!err)
 406                return 0;
 407        /* Finally fall back to regular interrupts. */
 408        return vp_find_vqs_intx(vdev, nvqs, vqs, callbacks, names, ctx);
 409}
 410
 411const char *vp_bus_name(struct virtio_device *vdev)
 412{
 413        struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 414
 415        return pci_name(vp_dev->pci_dev);
 416}
 417
 418/* Setup the affinity for a virtqueue:
 419 * - force the affinity for per vq vector
 420 * - OR over all affinities for shared MSI
 421 * - ignore the affinity request if we're using INTX
 422 */
 423int vp_set_vq_affinity(struct virtqueue *vq, int cpu)
 424{
 425        struct virtio_device *vdev = vq->vdev;
 426        struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 427        struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index];
 428        struct cpumask *mask;
 429        unsigned int irq;
 430
 431        if (!vq->callback)
 432                return -EINVAL;
 433
 434        if (vp_dev->msix_enabled) {
 435                mask = vp_dev->msix_affinity_masks[info->msix_vector];
 436                irq = pci_irq_vector(vp_dev->pci_dev, info->msix_vector);
 437                if (cpu == -1)
 438                        irq_set_affinity_hint(irq, NULL);
 439                else {
 440                        cpumask_clear(mask);
 441                        cpumask_set_cpu(cpu, mask);
 442                        irq_set_affinity_hint(irq, mask);
 443                }
 444        }
 445        return 0;
 446}
 447
 448const struct cpumask *vp_get_vq_affinity(struct virtio_device *vdev, int index)
 449{
 450        struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 451
 452        if (!vp_dev->per_vq_vectors ||
 453            vp_dev->vqs[index]->msix_vector == VIRTIO_MSI_NO_VECTOR)
 454                return NULL;
 455
 456        return pci_irq_get_affinity(vp_dev->pci_dev,
 457                                    vp_dev->vqs[index]->msix_vector);
 458}
 459
 460#ifdef CONFIG_PM_SLEEP
 461static int virtio_pci_freeze(struct device *dev)
 462{
 463        struct pci_dev *pci_dev = to_pci_dev(dev);
 464        struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
 465        int ret;
 466
 467        ret = virtio_device_freeze(&vp_dev->vdev);
 468
 469        if (!ret)
 470                pci_disable_device(pci_dev);
 471        return ret;
 472}
 473
 474static int virtio_pci_restore(struct device *dev)
 475{
 476        struct pci_dev *pci_dev = to_pci_dev(dev);
 477        struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
 478        int ret;
 479
 480        ret = pci_enable_device(pci_dev);
 481        if (ret)
 482                return ret;
 483
 484        pci_set_master(pci_dev);
 485        return virtio_device_restore(&vp_dev->vdev);
 486}
 487
 488static const struct dev_pm_ops virtio_pci_pm_ops = {
 489        SET_SYSTEM_SLEEP_PM_OPS(virtio_pci_freeze, virtio_pci_restore)
 490};
 491#endif
 492
 493
 494/* Qumranet donated their vendor ID for devices 0x1000 thru 0x10FF. */
 495static const struct pci_device_id virtio_pci_id_table[] = {
 496        { PCI_DEVICE(PCI_VENDOR_ID_REDHAT_QUMRANET, PCI_ANY_ID) },
 497        { 0 }
 498};
 499
 500MODULE_DEVICE_TABLE(pci, virtio_pci_id_table);
 501
 502static void virtio_pci_release_dev(struct device *_d)
 503{
 504        struct virtio_device *vdev = dev_to_virtio(_d);
 505        struct virtio_pci_device *vp_dev = to_vp_device(vdev);
 506
 507        /* As struct device is a kobject, it's not safe to
 508         * free the memory (including the reference counter itself)
 509         * until it's release callback. */
 510        kfree(vp_dev);
 511}
 512
 513static int virtio_pci_probe(struct pci_dev *pci_dev,
 514                            const struct pci_device_id *id)
 515{
 516        struct virtio_pci_device *vp_dev, *reg_dev = NULL;
 517        int rc;
 518
 519        /* allocate our structure and fill it out */
 520        vp_dev = kzalloc(sizeof(struct virtio_pci_device), GFP_KERNEL);
 521        if (!vp_dev)
 522                return -ENOMEM;
 523
 524        pci_set_drvdata(pci_dev, vp_dev);
 525        vp_dev->vdev.dev.parent = &pci_dev->dev;
 526        vp_dev->vdev.dev.release = virtio_pci_release_dev;
 527        vp_dev->pci_dev = pci_dev;
 528        INIT_LIST_HEAD(&vp_dev->virtqueues);
 529        spin_lock_init(&vp_dev->lock);
 530
 531        /* enable the device */
 532        rc = pci_enable_device(pci_dev);
 533        if (rc)
 534                goto err_enable_device;
 535
 536        if (force_legacy) {
 537                rc = virtio_pci_legacy_probe(vp_dev);
 538                /* Also try modern mode if we can't map BAR0 (no IO space). */
 539                if (rc == -ENODEV || rc == -ENOMEM)
 540                        rc = virtio_pci_modern_probe(vp_dev);
 541                if (rc)
 542                        goto err_probe;
 543        } else {
 544                rc = virtio_pci_modern_probe(vp_dev);
 545                if (rc == -ENODEV)
 546                        rc = virtio_pci_legacy_probe(vp_dev);
 547                if (rc)
 548                        goto err_probe;
 549        }
 550
 551        pci_set_master(pci_dev);
 552
 553        rc = register_virtio_device(&vp_dev->vdev);
 554        reg_dev = vp_dev;
 555        if (rc)
 556                goto err_register;
 557
 558        return 0;
 559
 560err_register:
 561        if (vp_dev->ioaddr)
 562             virtio_pci_legacy_remove(vp_dev);
 563        else
 564             virtio_pci_modern_remove(vp_dev);
 565err_probe:
 566        pci_disable_device(pci_dev);
 567err_enable_device:
 568        if (reg_dev)
 569                put_device(&vp_dev->vdev.dev);
 570        else
 571                kfree(vp_dev);
 572        return rc;
 573}
 574
 575static void virtio_pci_remove(struct pci_dev *pci_dev)
 576{
 577        struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
 578        struct device *dev = get_device(&vp_dev->vdev.dev);
 579
 580        unregister_virtio_device(&vp_dev->vdev);
 581
 582        if (vp_dev->ioaddr)
 583                virtio_pci_legacy_remove(vp_dev);
 584        else
 585                virtio_pci_modern_remove(vp_dev);
 586
 587        pci_disable_device(pci_dev);
 588        put_device(dev);
 589}
 590
 591static struct pci_driver virtio_pci_driver = {
 592        .name           = "virtio-pci",
 593        .id_table       = virtio_pci_id_table,
 594        .probe          = virtio_pci_probe,
 595        .remove         = virtio_pci_remove,
 596#ifdef CONFIG_PM_SLEEP
 597        .driver.pm      = &virtio_pci_pm_ops,
 598#endif
 599};
 600
 601module_pci_driver(virtio_pci_driver);
 602
 603MODULE_AUTHOR("Anthony Liguori <aliguori@us.ibm.com>");
 604MODULE_DESCRIPTION("virtio-pci");
 605MODULE_LICENSE("GPL");
 606MODULE_VERSION("1");
 607