Linux Audio

Check our new training course

Loading...
v3.1
  1/*
  2 * Virtio PCI driver
  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 *
  9 * Authors:
 10 *  Anthony Liguori  <aliguori@us.ibm.com>
 11 *
 12 * This work is licensed under the terms of the GNU GPL, version 2 or later.
 13 * See the COPYING file in the top-level directory.
 14 *
 15 */
 16
 17#include <linux/module.h>
 18#include <linux/list.h>
 19#include <linux/pci.h>
 20#include <linux/slab.h>
 21#include <linux/interrupt.h>
 22#include <linux/virtio.h>
 23#include <linux/virtio_config.h>
 24#include <linux/virtio_ring.h>
 25#include <linux/virtio_pci.h>
 26#include <linux/highmem.h>
 27#include <linux/spinlock.h>
 28
 29MODULE_AUTHOR("Anthony Liguori <aliguori@us.ibm.com>");
 30MODULE_DESCRIPTION("virtio-pci");
 31MODULE_LICENSE("GPL");
 32MODULE_VERSION("1");
 33
 34/* Our device structure */
 35struct virtio_pci_device
 36{
 37	struct virtio_device vdev;
 38	struct pci_dev *pci_dev;
 39
 40	/* the IO mapping for the PCI config space */
 41	void __iomem *ioaddr;
 42
 43	/* a list of queues so we can dispatch IRQs */
 44	spinlock_t lock;
 45	struct list_head virtqueues;
 46
 47	/* MSI-X support */
 48	int msix_enabled;
 49	int intx_enabled;
 50	struct msix_entry *msix_entries;
 
 51	/* Name strings for interrupts. This size should be enough,
 52	 * and I'm too lazy to allocate each name separately. */
 53	char (*msix_names)[256];
 54	/* Number of available vectors */
 55	unsigned msix_vectors;
 56	/* Vectors allocated, excluding per-vq vectors if any */
 57	unsigned msix_used_vectors;
 
 
 
 
 58	/* Whether we have vector per vq */
 59	bool per_vq_vectors;
 60};
 61
 62/* Constants for MSI-X */
 63/* Use first vector for configuration changes, second and the rest for
 64 * virtqueues Thus, we need at least 2 vectors for MSI. */
 65enum {
 66	VP_MSIX_CONFIG_VECTOR = 0,
 67	VP_MSIX_VQ_VECTOR = 1,
 68};
 69
 70struct virtio_pci_vq_info
 71{
 72	/* the actual virtqueue */
 73	struct virtqueue *vq;
 74
 75	/* the number of entries in the queue */
 76	int num;
 77
 78	/* the index of the queue */
 79	int queue_index;
 80
 81	/* the virtual address of the ring queue */
 82	void *queue;
 83
 84	/* the list node for the virtqueues list */
 85	struct list_head node;
 86
 87	/* MSI-X vector (or none) */
 88	unsigned msix_vector;
 89};
 90
 91/* Qumranet donated their vendor ID for devices 0x1000 thru 0x10FF. */
 92static struct pci_device_id virtio_pci_id_table[] = {
 93	{ 0x1af4, PCI_ANY_ID, PCI_ANY_ID, PCI_ANY_ID, 0, 0, 0 },
 94	{ 0 },
 95};
 96
 97MODULE_DEVICE_TABLE(pci, virtio_pci_id_table);
 98
 99/* Convert a generic virtio device to our structure */
100static struct virtio_pci_device *to_vp_device(struct virtio_device *vdev)
101{
102	return container_of(vdev, struct virtio_pci_device, vdev);
103}
104
105/* virtio config->get_features() implementation */
106static u32 vp_get_features(struct virtio_device *vdev)
107{
108	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
109
110	/* When someone needs more than 32 feature bits, we'll need to
111	 * steal a bit to indicate that the rest are somewhere else. */
112	return ioread32(vp_dev->ioaddr + VIRTIO_PCI_HOST_FEATURES);
113}
114
115/* virtio config->finalize_features() implementation */
116static void vp_finalize_features(struct virtio_device *vdev)
117{
118	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
119
120	/* Give virtio_ring a chance to accept features. */
121	vring_transport_features(vdev);
122
123	/* We only support 32 feature bits. */
124	BUILD_BUG_ON(ARRAY_SIZE(vdev->features) != 1);
125	iowrite32(vdev->features[0], vp_dev->ioaddr+VIRTIO_PCI_GUEST_FEATURES);
126}
127
128/* virtio config->get() implementation */
129static void vp_get(struct virtio_device *vdev, unsigned offset,
130		   void *buf, unsigned len)
131{
132	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
133	void __iomem *ioaddr = vp_dev->ioaddr +
134				VIRTIO_PCI_CONFIG(vp_dev) + offset;
135	u8 *ptr = buf;
136	int i;
137
138	for (i = 0; i < len; i++)
139		ptr[i] = ioread8(ioaddr + i);
140}
141
142/* the config->set() implementation.  it's symmetric to the config->get()
143 * implementation */
144static void vp_set(struct virtio_device *vdev, unsigned offset,
145		   const void *buf, unsigned len)
146{
147	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
148	void __iomem *ioaddr = vp_dev->ioaddr +
149				VIRTIO_PCI_CONFIG(vp_dev) + offset;
150	const u8 *ptr = buf;
151	int i;
152
153	for (i = 0; i < len; i++)
154		iowrite8(ptr[i], ioaddr + i);
155}
156
157/* config->{get,set}_status() implementations */
158static u8 vp_get_status(struct virtio_device *vdev)
159{
160	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
161	return ioread8(vp_dev->ioaddr + VIRTIO_PCI_STATUS);
162}
163
164static void vp_set_status(struct virtio_device *vdev, u8 status)
165{
166	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
167	/* We should never be setting status to 0. */
168	BUG_ON(status == 0);
169	iowrite8(status, vp_dev->ioaddr + VIRTIO_PCI_STATUS);
170}
171
 
 
 
 
 
 
 
 
 
 
 
 
 
172static void vp_reset(struct virtio_device *vdev)
173{
174	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
175	/* 0 status means a reset. */
176	iowrite8(0, vp_dev->ioaddr + VIRTIO_PCI_STATUS);
 
 
 
 
 
177}
178
179/* the notify function used when creating a virt queue */
180static void vp_notify(struct virtqueue *vq)
181{
182	struct virtio_pci_device *vp_dev = to_vp_device(vq->vdev);
183	struct virtio_pci_vq_info *info = vq->priv;
184
185	/* we write the queue's selector into the notification register to
186	 * signal the other end */
187	iowrite16(info->queue_index, vp_dev->ioaddr + VIRTIO_PCI_QUEUE_NOTIFY);
 
188}
189
190/* Handle a configuration change: Tell driver if it wants to know. */
191static irqreturn_t vp_config_changed(int irq, void *opaque)
192{
193	struct virtio_pci_device *vp_dev = opaque;
194	struct virtio_driver *drv;
195	drv = container_of(vp_dev->vdev.dev.driver,
196			   struct virtio_driver, driver);
197
198	if (drv && drv->config_changed)
199		drv->config_changed(&vp_dev->vdev);
200	return IRQ_HANDLED;
201}
202
203/* Notify all virtqueues on an interrupt. */
204static irqreturn_t vp_vring_interrupt(int irq, void *opaque)
205{
206	struct virtio_pci_device *vp_dev = opaque;
207	struct virtio_pci_vq_info *info;
208	irqreturn_t ret = IRQ_NONE;
209	unsigned long flags;
210
211	spin_lock_irqsave(&vp_dev->lock, flags);
212	list_for_each_entry(info, &vp_dev->virtqueues, node) {
213		if (vring_interrupt(irq, info->vq) == IRQ_HANDLED)
214			ret = IRQ_HANDLED;
215	}
216	spin_unlock_irqrestore(&vp_dev->lock, flags);
217
218	return ret;
219}
220
221/* A small wrapper to also acknowledge the interrupt when it's handled.
222 * I really need an EIO hook for the vring so I can ack the interrupt once we
223 * know that we'll be handling the IRQ but before we invoke the callback since
224 * the callback may notify the host which results in the host attempting to
225 * raise an interrupt that we would then mask once we acknowledged the
226 * interrupt. */
227static irqreturn_t vp_interrupt(int irq, void *opaque)
228{
229	struct virtio_pci_device *vp_dev = opaque;
230	u8 isr;
231
232	/* reading the ISR has the effect of also clearing it so it's very
233	 * important to save off the value. */
234	isr = ioread8(vp_dev->ioaddr + VIRTIO_PCI_ISR);
235
236	/* It's definitely not us if the ISR was not high */
237	if (!isr)
238		return IRQ_NONE;
239
240	/* Configuration change?  Tell driver if it wants to know. */
241	if (isr & VIRTIO_PCI_ISR_CONFIG)
242		vp_config_changed(irq, opaque);
243
244	return vp_vring_interrupt(irq, opaque);
245}
246
247static void vp_free_vectors(struct virtio_device *vdev)
248{
249	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
250	int i;
251
252	if (vp_dev->intx_enabled) {
253		free_irq(vp_dev->pci_dev->irq, vp_dev);
254		vp_dev->intx_enabled = 0;
255	}
256
257	for (i = 0; i < vp_dev->msix_used_vectors; ++i)
258		free_irq(vp_dev->msix_entries[i].vector, vp_dev);
259
 
 
 
 
260	if (vp_dev->msix_enabled) {
261		/* Disable the vector used for configuration */
262		iowrite16(VIRTIO_MSI_NO_VECTOR,
263			  vp_dev->ioaddr + VIRTIO_MSI_CONFIG_VECTOR);
264		/* Flush the write out to device */
265		ioread16(vp_dev->ioaddr + VIRTIO_MSI_CONFIG_VECTOR);
266
267		pci_disable_msix(vp_dev->pci_dev);
268		vp_dev->msix_enabled = 0;
269		vp_dev->msix_vectors = 0;
270	}
271
 
272	vp_dev->msix_used_vectors = 0;
273	kfree(vp_dev->msix_names);
274	vp_dev->msix_names = NULL;
275	kfree(vp_dev->msix_entries);
276	vp_dev->msix_entries = NULL;
 
 
277}
278
279static int vp_request_msix_vectors(struct virtio_device *vdev, int nvectors,
280				   bool per_vq_vectors)
281{
282	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
283	const char *name = dev_name(&vp_dev->vdev.dev);
284	unsigned i, v;
285	int err = -ENOMEM;
286
 
 
287	vp_dev->msix_entries = kmalloc(nvectors * sizeof *vp_dev->msix_entries,
288				       GFP_KERNEL);
289	if (!vp_dev->msix_entries)
290		goto error;
291	vp_dev->msix_names = kmalloc(nvectors * sizeof *vp_dev->msix_names,
292				     GFP_KERNEL);
293	if (!vp_dev->msix_names)
294		goto error;
 
 
 
 
 
 
 
 
 
295
296	for (i = 0; i < nvectors; ++i)
297		vp_dev->msix_entries[i].entry = i;
298
299	/* pci_enable_msix returns positive if we can't get this many. */
300	err = pci_enable_msix(vp_dev->pci_dev, vp_dev->msix_entries, nvectors);
301	if (err > 0)
302		err = -ENOSPC;
303	if (err)
304		goto error;
305	vp_dev->msix_vectors = nvectors;
306	vp_dev->msix_enabled = 1;
307
308	/* Set the vector used for configuration */
309	v = vp_dev->msix_used_vectors;
310	snprintf(vp_dev->msix_names[v], sizeof *vp_dev->msix_names,
311		 "%s-config", name);
312	err = request_irq(vp_dev->msix_entries[v].vector,
313			  vp_config_changed, 0, vp_dev->msix_names[v],
314			  vp_dev);
315	if (err)
316		goto error;
317	++vp_dev->msix_used_vectors;
318
319	iowrite16(v, vp_dev->ioaddr + VIRTIO_MSI_CONFIG_VECTOR);
320	/* Verify we had enough resources to assign the vector */
321	v = ioread16(vp_dev->ioaddr + VIRTIO_MSI_CONFIG_VECTOR);
322	if (v == VIRTIO_MSI_NO_VECTOR) {
323		err = -EBUSY;
324		goto error;
325	}
326
327	if (!per_vq_vectors) {
328		/* Shared vector for all VQs */
329		v = vp_dev->msix_used_vectors;
330		snprintf(vp_dev->msix_names[v], sizeof *vp_dev->msix_names,
331			 "%s-virtqueues", name);
332		err = request_irq(vp_dev->msix_entries[v].vector,
333				  vp_vring_interrupt, 0, vp_dev->msix_names[v],
334				  vp_dev);
335		if (err)
336			goto error;
337		++vp_dev->msix_used_vectors;
338	}
339	return 0;
340error:
341	vp_free_vectors(vdev);
342	return err;
343}
344
345static int vp_request_intx(struct virtio_device *vdev)
346{
347	int err;
348	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
349
350	err = request_irq(vp_dev->pci_dev->irq, vp_interrupt,
351			  IRQF_SHARED, dev_name(&vdev->dev), vp_dev);
352	if (!err)
353		vp_dev->intx_enabled = 1;
354	return err;
355}
356
357static struct virtqueue *setup_vq(struct virtio_device *vdev, unsigned index,
358				  void (*callback)(struct virtqueue *vq),
359				  const char *name,
360				  u16 msix_vec)
361{
362	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
363	struct virtio_pci_vq_info *info;
364	struct virtqueue *vq;
365	unsigned long flags, size;
366	u16 num;
367	int err;
368
369	/* Select the queue we're interested in */
370	iowrite16(index, vp_dev->ioaddr + VIRTIO_PCI_QUEUE_SEL);
371
372	/* Check if queue is either not available or already active. */
373	num = ioread16(vp_dev->ioaddr + VIRTIO_PCI_QUEUE_NUM);
374	if (!num || ioread32(vp_dev->ioaddr + VIRTIO_PCI_QUEUE_PFN))
375		return ERR_PTR(-ENOENT);
376
377	/* allocate and fill out our structure the represents an active
378	 * queue */
379	info = kmalloc(sizeof(struct virtio_pci_vq_info), GFP_KERNEL);
380	if (!info)
381		return ERR_PTR(-ENOMEM);
382
383	info->queue_index = index;
384	info->num = num;
385	info->msix_vector = msix_vec;
386
387	size = PAGE_ALIGN(vring_size(num, VIRTIO_PCI_VRING_ALIGN));
388	info->queue = alloc_pages_exact(size, GFP_KERNEL|__GFP_ZERO);
389	if (info->queue == NULL) {
390		err = -ENOMEM;
391		goto out_info;
392	}
393
394	/* activate the queue */
395	iowrite32(virt_to_phys(info->queue) >> VIRTIO_PCI_QUEUE_ADDR_SHIFT,
396		  vp_dev->ioaddr + VIRTIO_PCI_QUEUE_PFN);
397
398	/* create the vring */
399	vq = vring_new_virtqueue(info->num, VIRTIO_PCI_VRING_ALIGN,
400				 vdev, info->queue, vp_notify, callback, name);
401	if (!vq) {
402		err = -ENOMEM;
403		goto out_activate_queue;
404	}
405
406	vq->priv = info;
407	info->vq = vq;
408
409	if (msix_vec != VIRTIO_MSI_NO_VECTOR) {
410		iowrite16(msix_vec, vp_dev->ioaddr + VIRTIO_MSI_QUEUE_VECTOR);
411		msix_vec = ioread16(vp_dev->ioaddr + VIRTIO_MSI_QUEUE_VECTOR);
412		if (msix_vec == VIRTIO_MSI_NO_VECTOR) {
413			err = -EBUSY;
414			goto out_assign;
415		}
416	}
417
418	spin_lock_irqsave(&vp_dev->lock, flags);
419	list_add(&info->node, &vp_dev->virtqueues);
420	spin_unlock_irqrestore(&vp_dev->lock, flags);
 
 
 
 
421
422	return vq;
423
424out_assign:
425	vring_del_virtqueue(vq);
426out_activate_queue:
427	iowrite32(0, vp_dev->ioaddr + VIRTIO_PCI_QUEUE_PFN);
428	free_pages_exact(info->queue, size);
429out_info:
430	kfree(info);
431	return ERR_PTR(err);
432}
433
434static void vp_del_vq(struct virtqueue *vq)
435{
436	struct virtio_pci_device *vp_dev = to_vp_device(vq->vdev);
437	struct virtio_pci_vq_info *info = vq->priv;
438	unsigned long flags, size;
439
440	spin_lock_irqsave(&vp_dev->lock, flags);
441	list_del(&info->node);
442	spin_unlock_irqrestore(&vp_dev->lock, flags);
443
444	iowrite16(info->queue_index, vp_dev->ioaddr + VIRTIO_PCI_QUEUE_SEL);
445
446	if (vp_dev->msix_enabled) {
447		iowrite16(VIRTIO_MSI_NO_VECTOR,
448			  vp_dev->ioaddr + VIRTIO_MSI_QUEUE_VECTOR);
449		/* Flush the write out to device */
450		ioread8(vp_dev->ioaddr + VIRTIO_PCI_ISR);
451	}
452
453	vring_del_virtqueue(vq);
454
455	/* Select and deactivate the queue */
456	iowrite32(0, vp_dev->ioaddr + VIRTIO_PCI_QUEUE_PFN);
457
458	size = PAGE_ALIGN(vring_size(info->num, VIRTIO_PCI_VRING_ALIGN));
459	free_pages_exact(info->queue, size);
460	kfree(info);
461}
462
463/* the config->del_vqs() implementation */
464static void vp_del_vqs(struct virtio_device *vdev)
465{
466	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
467	struct virtqueue *vq, *n;
468	struct virtio_pci_vq_info *info;
469
470	list_for_each_entry_safe(vq, n, &vdev->vqs, list) {
471		info = vq->priv;
472		if (vp_dev->per_vq_vectors &&
473			info->msix_vector != VIRTIO_MSI_NO_VECTOR)
474			free_irq(vp_dev->msix_entries[info->msix_vector].vector,
475				 vq);
476		vp_del_vq(vq);
477	}
478	vp_dev->per_vq_vectors = false;
479
480	vp_free_vectors(vdev);
481}
482
483static int vp_try_to_find_vqs(struct virtio_device *vdev, unsigned nvqs,
484			      struct virtqueue *vqs[],
485			      vq_callback_t *callbacks[],
486			      const char *names[],
487			      bool use_msix,
488			      bool per_vq_vectors)
489{
490	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
491	u16 msix_vec;
492	int i, err, nvectors, allocated_vectors;
493
494	if (!use_msix) {
495		/* Old style: one normal interrupt for change and all vqs. */
496		err = vp_request_intx(vdev);
497		if (err)
498			goto error_request;
499	} else {
500		if (per_vq_vectors) {
501			/* Best option: one for change interrupt, one per vq. */
502			nvectors = 1;
503			for (i = 0; i < nvqs; ++i)
504				if (callbacks[i])
505					++nvectors;
506		} else {
507			/* Second best: one for change, shared for all vqs. */
508			nvectors = 2;
509		}
510
511		err = vp_request_msix_vectors(vdev, nvectors, per_vq_vectors);
512		if (err)
513			goto error_request;
514	}
515
516	vp_dev->per_vq_vectors = per_vq_vectors;
517	allocated_vectors = vp_dev->msix_used_vectors;
518	for (i = 0; i < nvqs; ++i) {
519		if (!callbacks[i] || !vp_dev->msix_enabled)
 
 
 
520			msix_vec = VIRTIO_MSI_NO_VECTOR;
521		else if (vp_dev->per_vq_vectors)
522			msix_vec = allocated_vectors++;
523		else
524			msix_vec = VP_MSIX_VQ_VECTOR;
525		vqs[i] = setup_vq(vdev, i, callbacks[i], names[i], msix_vec);
526		if (IS_ERR(vqs[i])) {
527			err = PTR_ERR(vqs[i]);
528			goto error_find;
529		}
530
531		if (!vp_dev->per_vq_vectors || msix_vec == VIRTIO_MSI_NO_VECTOR)
532			continue;
533
534		/* allocate per-vq irq if available and necessary */
535		snprintf(vp_dev->msix_names[msix_vec],
536			 sizeof *vp_dev->msix_names,
537			 "%s-%s",
538			 dev_name(&vp_dev->vdev.dev), names[i]);
539		err = request_irq(vp_dev->msix_entries[msix_vec].vector,
540				  vring_interrupt, 0,
541				  vp_dev->msix_names[msix_vec],
542				  vqs[i]);
543		if (err) {
544			vp_del_vq(vqs[i]);
545			goto error_find;
546		}
547	}
548	return 0;
549
550error_find:
551	vp_del_vqs(vdev);
552
553error_request:
554	return err;
555}
556
557/* the config->find_vqs() implementation */
558static int vp_find_vqs(struct virtio_device *vdev, unsigned nvqs,
559		       struct virtqueue *vqs[],
560		       vq_callback_t *callbacks[],
561		       const char *names[])
562{
563	int err;
564
565	/* Try MSI-X with one vector per queue. */
566	err = vp_try_to_find_vqs(vdev, nvqs, vqs, callbacks, names, true, true);
567	if (!err)
568		return 0;
569	/* Fallback: MSI-X with one vector for config, one shared for queues. */
570	err = vp_try_to_find_vqs(vdev, nvqs, vqs, callbacks, names,
571				 true, false);
572	if (!err)
573		return 0;
574	/* Finally fall back to regular interrupts. */
575	return vp_try_to_find_vqs(vdev, nvqs, vqs, callbacks, names,
576				  false, false);
577}
578
579static struct virtio_config_ops virtio_pci_config_ops = {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
580	.get		= vp_get,
581	.set		= vp_set,
582	.get_status	= vp_get_status,
583	.set_status	= vp_set_status,
584	.reset		= vp_reset,
585	.find_vqs	= vp_find_vqs,
586	.del_vqs	= vp_del_vqs,
587	.get_features	= vp_get_features,
588	.finalize_features = vp_finalize_features,
 
 
589};
590
591static void virtio_pci_release_dev(struct device *_d)
592{
593	struct virtio_device *dev = container_of(_d, struct virtio_device,
594						 dev);
595	struct virtio_pci_device *vp_dev = to_vp_device(dev);
596
597	kfree(vp_dev);
598}
599
600/* the PCI probing function */
601static int __devinit virtio_pci_probe(struct pci_dev *pci_dev,
602				      const struct pci_device_id *id)
603{
604	struct virtio_pci_device *vp_dev;
605	int err;
606
607	/* We only own devices >= 0x1000 and <= 0x103f: leave the rest. */
608	if (pci_dev->device < 0x1000 || pci_dev->device > 0x103f)
609		return -ENODEV;
610
611	if (pci_dev->revision != VIRTIO_PCI_ABI_VERSION) {
612		printk(KERN_ERR "virtio_pci: expected ABI version %d, got %d\n",
613		       VIRTIO_PCI_ABI_VERSION, pci_dev->revision);
614		return -ENODEV;
615	}
616
617	/* allocate our structure and fill it out */
618	vp_dev = kzalloc(sizeof(struct virtio_pci_device), GFP_KERNEL);
619	if (vp_dev == NULL)
620		return -ENOMEM;
621
622	vp_dev->vdev.dev.parent = &pci_dev->dev;
623	vp_dev->vdev.dev.release = virtio_pci_release_dev;
624	vp_dev->vdev.config = &virtio_pci_config_ops;
625	vp_dev->pci_dev = pci_dev;
626	INIT_LIST_HEAD(&vp_dev->virtqueues);
627	spin_lock_init(&vp_dev->lock);
628
629	/* Disable MSI/MSIX to bring device to a known good state. */
630	pci_msi_off(pci_dev);
631
632	/* enable the device */
633	err = pci_enable_device(pci_dev);
634	if (err)
635		goto out;
636
637	err = pci_request_regions(pci_dev, "virtio-pci");
638	if (err)
639		goto out_enable_device;
640
641	vp_dev->ioaddr = pci_iomap(pci_dev, 0, 0);
642	if (vp_dev->ioaddr == NULL)
 
643		goto out_req_regions;
 
644
645	pci_set_drvdata(pci_dev, vp_dev);
646	pci_set_master(pci_dev);
647
648	/* we use the subsystem vendor/device id as the virtio vendor/device
649	 * id.  this allows us to use the same PCI vendor/device id for all
650	 * virtio devices and to identify the particular virtio driver by
651	 * the subsystem ids */
652	vp_dev->vdev.id.vendor = pci_dev->subsystem_vendor;
653	vp_dev->vdev.id.device = pci_dev->subsystem_device;
654
655	/* finally register the virtio device */
656	err = register_virtio_device(&vp_dev->vdev);
657	if (err)
658		goto out_set_drvdata;
659
660	return 0;
661
662out_set_drvdata:
663	pci_set_drvdata(pci_dev, NULL);
664	pci_iounmap(pci_dev, vp_dev->ioaddr);
665out_req_regions:
666	pci_release_regions(pci_dev);
667out_enable_device:
668	pci_disable_device(pci_dev);
669out:
670	kfree(vp_dev);
671	return err;
672}
673
674static void __devexit virtio_pci_remove(struct pci_dev *pci_dev)
675{
676	struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
677
678	unregister_virtio_device(&vp_dev->vdev);
679
680	vp_del_vqs(&vp_dev->vdev);
681	pci_set_drvdata(pci_dev, NULL);
682	pci_iounmap(pci_dev, vp_dev->ioaddr);
683	pci_release_regions(pci_dev);
684	pci_disable_device(pci_dev);
 
685}
686
687#ifdef CONFIG_PM
688static int virtio_pci_suspend(struct pci_dev *pci_dev, pm_message_t state)
689{
690	pci_save_state(pci_dev);
691	pci_set_power_state(pci_dev, PCI_D3hot);
692	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
693}
694
695static int virtio_pci_resume(struct pci_dev *pci_dev)
696{
697	pci_restore_state(pci_dev);
698	pci_set_power_state(pci_dev, PCI_D0);
699	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
700}
 
 
 
 
701#endif
702
703static struct pci_driver virtio_pci_driver = {
704	.name		= "virtio-pci",
705	.id_table	= virtio_pci_id_table,
706	.probe		= virtio_pci_probe,
707	.remove		= __devexit_p(virtio_pci_remove),
708#ifdef CONFIG_PM
709	.suspend	= virtio_pci_suspend,
710	.resume		= virtio_pci_resume,
711#endif
712};
713
714static int __init virtio_pci_init(void)
715{
716	return pci_register_driver(&virtio_pci_driver);
717}
718
719module_init(virtio_pci_init);
720
721static void __exit virtio_pci_exit(void)
722{
723	pci_unregister_driver(&virtio_pci_driver);
724}
725
726module_exit(virtio_pci_exit);
v3.15
  1/*
  2 * Virtio PCI driver
  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 *
  9 * Authors:
 10 *  Anthony Liguori  <aliguori@us.ibm.com>
 11 *
 12 * This work is licensed under the terms of the GNU GPL, version 2 or later.
 13 * See the COPYING file in the top-level directory.
 14 *
 15 */
 16
 17#include <linux/module.h>
 18#include <linux/list.h>
 19#include <linux/pci.h>
 20#include <linux/slab.h>
 21#include <linux/interrupt.h>
 22#include <linux/virtio.h>
 23#include <linux/virtio_config.h>
 24#include <linux/virtio_ring.h>
 25#include <linux/virtio_pci.h>
 26#include <linux/highmem.h>
 27#include <linux/spinlock.h>
 28
 29MODULE_AUTHOR("Anthony Liguori <aliguori@us.ibm.com>");
 30MODULE_DESCRIPTION("virtio-pci");
 31MODULE_LICENSE("GPL");
 32MODULE_VERSION("1");
 33
 34/* Our device structure */
 35struct virtio_pci_device
 36{
 37	struct virtio_device vdev;
 38	struct pci_dev *pci_dev;
 39
 40	/* the IO mapping for the PCI config space */
 41	void __iomem *ioaddr;
 42
 43	/* a list of queues so we can dispatch IRQs */
 44	spinlock_t lock;
 45	struct list_head virtqueues;
 46
 47	/* MSI-X support */
 48	int msix_enabled;
 49	int intx_enabled;
 50	struct msix_entry *msix_entries;
 51	cpumask_var_t *msix_affinity_masks;
 52	/* Name strings for interrupts. This size should be enough,
 53	 * and I'm too lazy to allocate each name separately. */
 54	char (*msix_names)[256];
 55	/* Number of available vectors */
 56	unsigned msix_vectors;
 57	/* Vectors allocated, excluding per-vq vectors if any */
 58	unsigned msix_used_vectors;
 59
 60	/* Status saved during hibernate/restore */
 61	u8 saved_status;
 62
 63	/* Whether we have vector per vq */
 64	bool per_vq_vectors;
 65};
 66
 67/* Constants for MSI-X */
 68/* Use first vector for configuration changes, second and the rest for
 69 * virtqueues Thus, we need at least 2 vectors for MSI. */
 70enum {
 71	VP_MSIX_CONFIG_VECTOR = 0,
 72	VP_MSIX_VQ_VECTOR = 1,
 73};
 74
 75struct virtio_pci_vq_info
 76{
 77	/* the actual virtqueue */
 78	struct virtqueue *vq;
 79
 80	/* the number of entries in the queue */
 81	int num;
 82
 
 
 
 83	/* the virtual address of the ring queue */
 84	void *queue;
 85
 86	/* the list node for the virtqueues list */
 87	struct list_head node;
 88
 89	/* MSI-X vector (or none) */
 90	unsigned msix_vector;
 91};
 92
 93/* Qumranet donated their vendor ID for devices 0x1000 thru 0x10FF. */
 94static DEFINE_PCI_DEVICE_TABLE(virtio_pci_id_table) = {
 95	{ PCI_DEVICE(0x1af4, PCI_ANY_ID) },
 96	{ 0 }
 97};
 98
 99MODULE_DEVICE_TABLE(pci, virtio_pci_id_table);
100
101/* Convert a generic virtio device to our structure */
102static struct virtio_pci_device *to_vp_device(struct virtio_device *vdev)
103{
104	return container_of(vdev, struct virtio_pci_device, vdev);
105}
106
107/* virtio config->get_features() implementation */
108static u32 vp_get_features(struct virtio_device *vdev)
109{
110	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
111
112	/* When someone needs more than 32 feature bits, we'll need to
113	 * steal a bit to indicate that the rest are somewhere else. */
114	return ioread32(vp_dev->ioaddr + VIRTIO_PCI_HOST_FEATURES);
115}
116
117/* virtio config->finalize_features() implementation */
118static void vp_finalize_features(struct virtio_device *vdev)
119{
120	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
121
122	/* Give virtio_ring a chance to accept features. */
123	vring_transport_features(vdev);
124
125	/* We only support 32 feature bits. */
126	BUILD_BUG_ON(ARRAY_SIZE(vdev->features) != 1);
127	iowrite32(vdev->features[0], vp_dev->ioaddr+VIRTIO_PCI_GUEST_FEATURES);
128}
129
130/* virtio config->get() implementation */
131static void vp_get(struct virtio_device *vdev, unsigned offset,
132		   void *buf, unsigned len)
133{
134	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
135	void __iomem *ioaddr = vp_dev->ioaddr +
136				VIRTIO_PCI_CONFIG(vp_dev) + offset;
137	u8 *ptr = buf;
138	int i;
139
140	for (i = 0; i < len; i++)
141		ptr[i] = ioread8(ioaddr + i);
142}
143
144/* the config->set() implementation.  it's symmetric to the config->get()
145 * implementation */
146static void vp_set(struct virtio_device *vdev, unsigned offset,
147		   const void *buf, unsigned len)
148{
149	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
150	void __iomem *ioaddr = vp_dev->ioaddr +
151				VIRTIO_PCI_CONFIG(vp_dev) + offset;
152	const u8 *ptr = buf;
153	int i;
154
155	for (i = 0; i < len; i++)
156		iowrite8(ptr[i], ioaddr + i);
157}
158
159/* config->{get,set}_status() implementations */
160static u8 vp_get_status(struct virtio_device *vdev)
161{
162	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
163	return ioread8(vp_dev->ioaddr + VIRTIO_PCI_STATUS);
164}
165
166static void vp_set_status(struct virtio_device *vdev, u8 status)
167{
168	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
169	/* We should never be setting status to 0. */
170	BUG_ON(status == 0);
171	iowrite8(status, vp_dev->ioaddr + VIRTIO_PCI_STATUS);
172}
173
174/* wait for pending irq handlers */
175static void vp_synchronize_vectors(struct virtio_device *vdev)
176{
177	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
178	int i;
179
180	if (vp_dev->intx_enabled)
181		synchronize_irq(vp_dev->pci_dev->irq);
182
183	for (i = 0; i < vp_dev->msix_vectors; ++i)
184		synchronize_irq(vp_dev->msix_entries[i].vector);
185}
186
187static void vp_reset(struct virtio_device *vdev)
188{
189	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
190	/* 0 status means a reset. */
191	iowrite8(0, vp_dev->ioaddr + VIRTIO_PCI_STATUS);
192	/* Flush out the status write, and flush in device writes,
193	 * including MSi-X interrupts, if any. */
194	ioread8(vp_dev->ioaddr + VIRTIO_PCI_STATUS);
195	/* Flush pending VQ/configuration callbacks. */
196	vp_synchronize_vectors(vdev);
197}
198
199/* the notify function used when creating a virt queue */
200static bool vp_notify(struct virtqueue *vq)
201{
202	struct virtio_pci_device *vp_dev = to_vp_device(vq->vdev);
 
203
204	/* we write the queue's selector into the notification register to
205	 * signal the other end */
206	iowrite16(vq->index, vp_dev->ioaddr + VIRTIO_PCI_QUEUE_NOTIFY);
207	return true;
208}
209
210/* Handle a configuration change: Tell driver if it wants to know. */
211static irqreturn_t vp_config_changed(int irq, void *opaque)
212{
213	struct virtio_pci_device *vp_dev = opaque;
214	struct virtio_driver *drv;
215	drv = container_of(vp_dev->vdev.dev.driver,
216			   struct virtio_driver, driver);
217
218	if (drv && drv->config_changed)
219		drv->config_changed(&vp_dev->vdev);
220	return IRQ_HANDLED;
221}
222
223/* Notify all virtqueues on an interrupt. */
224static irqreturn_t vp_vring_interrupt(int irq, void *opaque)
225{
226	struct virtio_pci_device *vp_dev = opaque;
227	struct virtio_pci_vq_info *info;
228	irqreturn_t ret = IRQ_NONE;
229	unsigned long flags;
230
231	spin_lock_irqsave(&vp_dev->lock, flags);
232	list_for_each_entry(info, &vp_dev->virtqueues, node) {
233		if (vring_interrupt(irq, info->vq) == IRQ_HANDLED)
234			ret = IRQ_HANDLED;
235	}
236	spin_unlock_irqrestore(&vp_dev->lock, flags);
237
238	return ret;
239}
240
241/* A small wrapper to also acknowledge the interrupt when it's handled.
242 * I really need an EIO hook for the vring so I can ack the interrupt once we
243 * know that we'll be handling the IRQ but before we invoke the callback since
244 * the callback may notify the host which results in the host attempting to
245 * raise an interrupt that we would then mask once we acknowledged the
246 * interrupt. */
247static irqreturn_t vp_interrupt(int irq, void *opaque)
248{
249	struct virtio_pci_device *vp_dev = opaque;
250	u8 isr;
251
252	/* reading the ISR has the effect of also clearing it so it's very
253	 * important to save off the value. */
254	isr = ioread8(vp_dev->ioaddr + VIRTIO_PCI_ISR);
255
256	/* It's definitely not us if the ISR was not high */
257	if (!isr)
258		return IRQ_NONE;
259
260	/* Configuration change?  Tell driver if it wants to know. */
261	if (isr & VIRTIO_PCI_ISR_CONFIG)
262		vp_config_changed(irq, opaque);
263
264	return vp_vring_interrupt(irq, opaque);
265}
266
267static void vp_free_vectors(struct virtio_device *vdev)
268{
269	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
270	int i;
271
272	if (vp_dev->intx_enabled) {
273		free_irq(vp_dev->pci_dev->irq, vp_dev);
274		vp_dev->intx_enabled = 0;
275	}
276
277	for (i = 0; i < vp_dev->msix_used_vectors; ++i)
278		free_irq(vp_dev->msix_entries[i].vector, vp_dev);
279
280	for (i = 0; i < vp_dev->msix_vectors; i++)
281		if (vp_dev->msix_affinity_masks[i])
282			free_cpumask_var(vp_dev->msix_affinity_masks[i]);
283
284	if (vp_dev->msix_enabled) {
285		/* Disable the vector used for configuration */
286		iowrite16(VIRTIO_MSI_NO_VECTOR,
287			  vp_dev->ioaddr + VIRTIO_MSI_CONFIG_VECTOR);
288		/* Flush the write out to device */
289		ioread16(vp_dev->ioaddr + VIRTIO_MSI_CONFIG_VECTOR);
290
291		pci_disable_msix(vp_dev->pci_dev);
292		vp_dev->msix_enabled = 0;
 
293	}
294
295	vp_dev->msix_vectors = 0;
296	vp_dev->msix_used_vectors = 0;
297	kfree(vp_dev->msix_names);
298	vp_dev->msix_names = NULL;
299	kfree(vp_dev->msix_entries);
300	vp_dev->msix_entries = NULL;
301	kfree(vp_dev->msix_affinity_masks);
302	vp_dev->msix_affinity_masks = NULL;
303}
304
305static int vp_request_msix_vectors(struct virtio_device *vdev, int nvectors,
306				   bool per_vq_vectors)
307{
308	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
309	const char *name = dev_name(&vp_dev->vdev.dev);
310	unsigned i, v;
311	int err = -ENOMEM;
312
313	vp_dev->msix_vectors = nvectors;
314
315	vp_dev->msix_entries = kmalloc(nvectors * sizeof *vp_dev->msix_entries,
316				       GFP_KERNEL);
317	if (!vp_dev->msix_entries)
318		goto error;
319	vp_dev->msix_names = kmalloc(nvectors * sizeof *vp_dev->msix_names,
320				     GFP_KERNEL);
321	if (!vp_dev->msix_names)
322		goto error;
323	vp_dev->msix_affinity_masks
324		= kzalloc(nvectors * sizeof *vp_dev->msix_affinity_masks,
325			  GFP_KERNEL);
326	if (!vp_dev->msix_affinity_masks)
327		goto error;
328	for (i = 0; i < nvectors; ++i)
329		if (!alloc_cpumask_var(&vp_dev->msix_affinity_masks[i],
330					GFP_KERNEL))
331			goto error;
332
333	for (i = 0; i < nvectors; ++i)
334		vp_dev->msix_entries[i].entry = i;
335
336	err = pci_enable_msix_exact(vp_dev->pci_dev,
337				    vp_dev->msix_entries, nvectors);
 
 
338	if (err)
339		goto error;
 
340	vp_dev->msix_enabled = 1;
341
342	/* Set the vector used for configuration */
343	v = vp_dev->msix_used_vectors;
344	snprintf(vp_dev->msix_names[v], sizeof *vp_dev->msix_names,
345		 "%s-config", name);
346	err = request_irq(vp_dev->msix_entries[v].vector,
347			  vp_config_changed, 0, vp_dev->msix_names[v],
348			  vp_dev);
349	if (err)
350		goto error;
351	++vp_dev->msix_used_vectors;
352
353	iowrite16(v, vp_dev->ioaddr + VIRTIO_MSI_CONFIG_VECTOR);
354	/* Verify we had enough resources to assign the vector */
355	v = ioread16(vp_dev->ioaddr + VIRTIO_MSI_CONFIG_VECTOR);
356	if (v == VIRTIO_MSI_NO_VECTOR) {
357		err = -EBUSY;
358		goto error;
359	}
360
361	if (!per_vq_vectors) {
362		/* Shared vector for all VQs */
363		v = vp_dev->msix_used_vectors;
364		snprintf(vp_dev->msix_names[v], sizeof *vp_dev->msix_names,
365			 "%s-virtqueues", name);
366		err = request_irq(vp_dev->msix_entries[v].vector,
367				  vp_vring_interrupt, 0, vp_dev->msix_names[v],
368				  vp_dev);
369		if (err)
370			goto error;
371		++vp_dev->msix_used_vectors;
372	}
373	return 0;
374error:
375	vp_free_vectors(vdev);
376	return err;
377}
378
379static int vp_request_intx(struct virtio_device *vdev)
380{
381	int err;
382	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
383
384	err = request_irq(vp_dev->pci_dev->irq, vp_interrupt,
385			  IRQF_SHARED, dev_name(&vdev->dev), vp_dev);
386	if (!err)
387		vp_dev->intx_enabled = 1;
388	return err;
389}
390
391static struct virtqueue *setup_vq(struct virtio_device *vdev, unsigned index,
392				  void (*callback)(struct virtqueue *vq),
393				  const char *name,
394				  u16 msix_vec)
395{
396	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
397	struct virtio_pci_vq_info *info;
398	struct virtqueue *vq;
399	unsigned long flags, size;
400	u16 num;
401	int err;
402
403	/* Select the queue we're interested in */
404	iowrite16(index, vp_dev->ioaddr + VIRTIO_PCI_QUEUE_SEL);
405
406	/* Check if queue is either not available or already active. */
407	num = ioread16(vp_dev->ioaddr + VIRTIO_PCI_QUEUE_NUM);
408	if (!num || ioread32(vp_dev->ioaddr + VIRTIO_PCI_QUEUE_PFN))
409		return ERR_PTR(-ENOENT);
410
411	/* allocate and fill out our structure the represents an active
412	 * queue */
413	info = kmalloc(sizeof(struct virtio_pci_vq_info), GFP_KERNEL);
414	if (!info)
415		return ERR_PTR(-ENOMEM);
416
 
417	info->num = num;
418	info->msix_vector = msix_vec;
419
420	size = PAGE_ALIGN(vring_size(num, VIRTIO_PCI_VRING_ALIGN));
421	info->queue = alloc_pages_exact(size, GFP_KERNEL|__GFP_ZERO);
422	if (info->queue == NULL) {
423		err = -ENOMEM;
424		goto out_info;
425	}
426
427	/* activate the queue */
428	iowrite32(virt_to_phys(info->queue) >> VIRTIO_PCI_QUEUE_ADDR_SHIFT,
429		  vp_dev->ioaddr + VIRTIO_PCI_QUEUE_PFN);
430
431	/* create the vring */
432	vq = vring_new_virtqueue(index, info->num, VIRTIO_PCI_VRING_ALIGN, vdev,
433				 true, info->queue, vp_notify, callback, name);
434	if (!vq) {
435		err = -ENOMEM;
436		goto out_activate_queue;
437	}
438
439	vq->priv = info;
440	info->vq = vq;
441
442	if (msix_vec != VIRTIO_MSI_NO_VECTOR) {
443		iowrite16(msix_vec, vp_dev->ioaddr + VIRTIO_MSI_QUEUE_VECTOR);
444		msix_vec = ioread16(vp_dev->ioaddr + VIRTIO_MSI_QUEUE_VECTOR);
445		if (msix_vec == VIRTIO_MSI_NO_VECTOR) {
446			err = -EBUSY;
447			goto out_assign;
448		}
449	}
450
451	if (callback) {
452		spin_lock_irqsave(&vp_dev->lock, flags);
453		list_add(&info->node, &vp_dev->virtqueues);
454		spin_unlock_irqrestore(&vp_dev->lock, flags);
455	} else {
456		INIT_LIST_HEAD(&info->node);
457	}
458
459	return vq;
460
461out_assign:
462	vring_del_virtqueue(vq);
463out_activate_queue:
464	iowrite32(0, vp_dev->ioaddr + VIRTIO_PCI_QUEUE_PFN);
465	free_pages_exact(info->queue, size);
466out_info:
467	kfree(info);
468	return ERR_PTR(err);
469}
470
471static void vp_del_vq(struct virtqueue *vq)
472{
473	struct virtio_pci_device *vp_dev = to_vp_device(vq->vdev);
474	struct virtio_pci_vq_info *info = vq->priv;
475	unsigned long flags, size;
476
477	spin_lock_irqsave(&vp_dev->lock, flags);
478	list_del(&info->node);
479	spin_unlock_irqrestore(&vp_dev->lock, flags);
480
481	iowrite16(vq->index, vp_dev->ioaddr + VIRTIO_PCI_QUEUE_SEL);
482
483	if (vp_dev->msix_enabled) {
484		iowrite16(VIRTIO_MSI_NO_VECTOR,
485			  vp_dev->ioaddr + VIRTIO_MSI_QUEUE_VECTOR);
486		/* Flush the write out to device */
487		ioread8(vp_dev->ioaddr + VIRTIO_PCI_ISR);
488	}
489
490	vring_del_virtqueue(vq);
491
492	/* Select and deactivate the queue */
493	iowrite32(0, vp_dev->ioaddr + VIRTIO_PCI_QUEUE_PFN);
494
495	size = PAGE_ALIGN(vring_size(info->num, VIRTIO_PCI_VRING_ALIGN));
496	free_pages_exact(info->queue, size);
497	kfree(info);
498}
499
500/* the config->del_vqs() implementation */
501static void vp_del_vqs(struct virtio_device *vdev)
502{
503	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
504	struct virtqueue *vq, *n;
505	struct virtio_pci_vq_info *info;
506
507	list_for_each_entry_safe(vq, n, &vdev->vqs, list) {
508		info = vq->priv;
509		if (vp_dev->per_vq_vectors &&
510			info->msix_vector != VIRTIO_MSI_NO_VECTOR)
511			free_irq(vp_dev->msix_entries[info->msix_vector].vector,
512				 vq);
513		vp_del_vq(vq);
514	}
515	vp_dev->per_vq_vectors = false;
516
517	vp_free_vectors(vdev);
518}
519
520static int vp_try_to_find_vqs(struct virtio_device *vdev, unsigned nvqs,
521			      struct virtqueue *vqs[],
522			      vq_callback_t *callbacks[],
523			      const char *names[],
524			      bool use_msix,
525			      bool per_vq_vectors)
526{
527	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
528	u16 msix_vec;
529	int i, err, nvectors, allocated_vectors;
530
531	if (!use_msix) {
532		/* Old style: one normal interrupt for change and all vqs. */
533		err = vp_request_intx(vdev);
534		if (err)
535			goto error_request;
536	} else {
537		if (per_vq_vectors) {
538			/* Best option: one for change interrupt, one per vq. */
539			nvectors = 1;
540			for (i = 0; i < nvqs; ++i)
541				if (callbacks[i])
542					++nvectors;
543		} else {
544			/* Second best: one for change, shared for all vqs. */
545			nvectors = 2;
546		}
547
548		err = vp_request_msix_vectors(vdev, nvectors, per_vq_vectors);
549		if (err)
550			goto error_request;
551	}
552
553	vp_dev->per_vq_vectors = per_vq_vectors;
554	allocated_vectors = vp_dev->msix_used_vectors;
555	for (i = 0; i < nvqs; ++i) {
556		if (!names[i]) {
557			vqs[i] = NULL;
558			continue;
559		} else if (!callbacks[i] || !vp_dev->msix_enabled)
560			msix_vec = VIRTIO_MSI_NO_VECTOR;
561		else if (vp_dev->per_vq_vectors)
562			msix_vec = allocated_vectors++;
563		else
564			msix_vec = VP_MSIX_VQ_VECTOR;
565		vqs[i] = setup_vq(vdev, i, callbacks[i], names[i], msix_vec);
566		if (IS_ERR(vqs[i])) {
567			err = PTR_ERR(vqs[i]);
568			goto error_find;
569		}
570
571		if (!vp_dev->per_vq_vectors || msix_vec == VIRTIO_MSI_NO_VECTOR)
572			continue;
573
574		/* allocate per-vq irq if available and necessary */
575		snprintf(vp_dev->msix_names[msix_vec],
576			 sizeof *vp_dev->msix_names,
577			 "%s-%s",
578			 dev_name(&vp_dev->vdev.dev), names[i]);
579		err = request_irq(vp_dev->msix_entries[msix_vec].vector,
580				  vring_interrupt, 0,
581				  vp_dev->msix_names[msix_vec],
582				  vqs[i]);
583		if (err) {
584			vp_del_vq(vqs[i]);
585			goto error_find;
586		}
587	}
588	return 0;
589
590error_find:
591	vp_del_vqs(vdev);
592
593error_request:
594	return err;
595}
596
597/* the config->find_vqs() implementation */
598static int vp_find_vqs(struct virtio_device *vdev, unsigned nvqs,
599		       struct virtqueue *vqs[],
600		       vq_callback_t *callbacks[],
601		       const char *names[])
602{
603	int err;
604
605	/* Try MSI-X with one vector per queue. */
606	err = vp_try_to_find_vqs(vdev, nvqs, vqs, callbacks, names, true, true);
607	if (!err)
608		return 0;
609	/* Fallback: MSI-X with one vector for config, one shared for queues. */
610	err = vp_try_to_find_vqs(vdev, nvqs, vqs, callbacks, names,
611				 true, false);
612	if (!err)
613		return 0;
614	/* Finally fall back to regular interrupts. */
615	return vp_try_to_find_vqs(vdev, nvqs, vqs, callbacks, names,
616				  false, false);
617}
618
619static const char *vp_bus_name(struct virtio_device *vdev)
620{
621	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
622
623	return pci_name(vp_dev->pci_dev);
624}
625
626/* Setup the affinity for a virtqueue:
627 * - force the affinity for per vq vector
628 * - OR over all affinities for shared MSI
629 * - ignore the affinity request if we're using INTX
630 */
631static int vp_set_vq_affinity(struct virtqueue *vq, int cpu)
632{
633	struct virtio_device *vdev = vq->vdev;
634	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
635	struct virtio_pci_vq_info *info = vq->priv;
636	struct cpumask *mask;
637	unsigned int irq;
638
639	if (!vq->callback)
640		return -EINVAL;
641
642	if (vp_dev->msix_enabled) {
643		mask = vp_dev->msix_affinity_masks[info->msix_vector];
644		irq = vp_dev->msix_entries[info->msix_vector].vector;
645		if (cpu == -1)
646			irq_set_affinity_hint(irq, NULL);
647		else {
648			cpumask_set_cpu(cpu, mask);
649			irq_set_affinity_hint(irq, mask);
650		}
651	}
652	return 0;
653}
654
655static const struct virtio_config_ops virtio_pci_config_ops = {
656	.get		= vp_get,
657	.set		= vp_set,
658	.get_status	= vp_get_status,
659	.set_status	= vp_set_status,
660	.reset		= vp_reset,
661	.find_vqs	= vp_find_vqs,
662	.del_vqs	= vp_del_vqs,
663	.get_features	= vp_get_features,
664	.finalize_features = vp_finalize_features,
665	.bus_name	= vp_bus_name,
666	.set_vq_affinity = vp_set_vq_affinity,
667};
668
669static void virtio_pci_release_dev(struct device *_d)
670{
671	/*
672	 * No need for a release method as we allocate/free
673	 * all devices together with the pci devices.
674	 * Provide an empty one to avoid getting a warning from core.
675	 */
676}
677
678/* the PCI probing function */
679static int virtio_pci_probe(struct pci_dev *pci_dev,
680			    const struct pci_device_id *id)
681{
682	struct virtio_pci_device *vp_dev;
683	int err;
684
685	/* We only own devices >= 0x1000 and <= 0x103f: leave the rest. */
686	if (pci_dev->device < 0x1000 || pci_dev->device > 0x103f)
687		return -ENODEV;
688
689	if (pci_dev->revision != VIRTIO_PCI_ABI_VERSION) {
690		printk(KERN_ERR "virtio_pci: expected ABI version %d, got %d\n",
691		       VIRTIO_PCI_ABI_VERSION, pci_dev->revision);
692		return -ENODEV;
693	}
694
695	/* allocate our structure and fill it out */
696	vp_dev = kzalloc(sizeof(struct virtio_pci_device), GFP_KERNEL);
697	if (vp_dev == NULL)
698		return -ENOMEM;
699
700	vp_dev->vdev.dev.parent = &pci_dev->dev;
701	vp_dev->vdev.dev.release = virtio_pci_release_dev;
702	vp_dev->vdev.config = &virtio_pci_config_ops;
703	vp_dev->pci_dev = pci_dev;
704	INIT_LIST_HEAD(&vp_dev->virtqueues);
705	spin_lock_init(&vp_dev->lock);
706
707	/* Disable MSI/MSIX to bring device to a known good state. */
708	pci_msi_off(pci_dev);
709
710	/* enable the device */
711	err = pci_enable_device(pci_dev);
712	if (err)
713		goto out;
714
715	err = pci_request_regions(pci_dev, "virtio-pci");
716	if (err)
717		goto out_enable_device;
718
719	vp_dev->ioaddr = pci_iomap(pci_dev, 0, 0);
720	if (vp_dev->ioaddr == NULL) {
721		err = -ENOMEM;
722		goto out_req_regions;
723	}
724
725	pci_set_drvdata(pci_dev, vp_dev);
726	pci_set_master(pci_dev);
727
728	/* we use the subsystem vendor/device id as the virtio vendor/device
729	 * id.  this allows us to use the same PCI vendor/device id for all
730	 * virtio devices and to identify the particular virtio driver by
731	 * the subsystem ids */
732	vp_dev->vdev.id.vendor = pci_dev->subsystem_vendor;
733	vp_dev->vdev.id.device = pci_dev->subsystem_device;
734
735	/* finally register the virtio device */
736	err = register_virtio_device(&vp_dev->vdev);
737	if (err)
738		goto out_set_drvdata;
739
740	return 0;
741
742out_set_drvdata:
 
743	pci_iounmap(pci_dev, vp_dev->ioaddr);
744out_req_regions:
745	pci_release_regions(pci_dev);
746out_enable_device:
747	pci_disable_device(pci_dev);
748out:
749	kfree(vp_dev);
750	return err;
751}
752
753static void virtio_pci_remove(struct pci_dev *pci_dev)
754{
755	struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
756
757	unregister_virtio_device(&vp_dev->vdev);
758
759	vp_del_vqs(&vp_dev->vdev);
 
760	pci_iounmap(pci_dev, vp_dev->ioaddr);
761	pci_release_regions(pci_dev);
762	pci_disable_device(pci_dev);
763	kfree(vp_dev);
764}
765
766#ifdef CONFIG_PM_SLEEP
767static int virtio_pci_freeze(struct device *dev)
768{
769	struct pci_dev *pci_dev = to_pci_dev(dev);
770	struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
771	struct virtio_driver *drv;
772	int ret;
773
774	drv = container_of(vp_dev->vdev.dev.driver,
775			   struct virtio_driver, driver);
776
777	ret = 0;
778	vp_dev->saved_status = vp_get_status(&vp_dev->vdev);
779	if (drv && drv->freeze)
780		ret = drv->freeze(&vp_dev->vdev);
781
782	if (!ret)
783		pci_disable_device(pci_dev);
784	return ret;
785}
786
787static int virtio_pci_restore(struct device *dev)
788{
789	struct pci_dev *pci_dev = to_pci_dev(dev);
790	struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev);
791	struct virtio_driver *drv;
792	int ret;
793
794	drv = container_of(vp_dev->vdev.dev.driver,
795			   struct virtio_driver, driver);
796
797	ret = pci_enable_device(pci_dev);
798	if (ret)
799		return ret;
800
801	pci_set_master(pci_dev);
802	vp_finalize_features(&vp_dev->vdev);
803
804	if (drv && drv->restore)
805		ret = drv->restore(&vp_dev->vdev);
806
807	/* Finally, tell the device we're all set */
808	if (!ret)
809		vp_set_status(&vp_dev->vdev, vp_dev->saved_status);
810
811	return ret;
812}
813
814static const struct dev_pm_ops virtio_pci_pm_ops = {
815	SET_SYSTEM_SLEEP_PM_OPS(virtio_pci_freeze, virtio_pci_restore)
816};
817#endif
818
819static struct pci_driver virtio_pci_driver = {
820	.name		= "virtio-pci",
821	.id_table	= virtio_pci_id_table,
822	.probe		= virtio_pci_probe,
823	.remove		= virtio_pci_remove,
824#ifdef CONFIG_PM_SLEEP
825	.driver.pm	= &virtio_pci_pm_ops,
 
826#endif
827};
828
829module_pci_driver(virtio_pci_driver);