1 /*
2  * Virtio PCI driver - modern (virtio 1.0) device support
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 <linux/delay.h>
21 #define VIRTIO_PCI_NO_LEGACY
22 #include "virtio_pci_common.h"
23 
24 /*
25  * Type-safe wrappers for io accesses.
26  * Use these to enforce at compile time the following spec requirement:
27  *
28  * The driver MUST access each field using the “natural” access
29  * method, i.e. 32-bit accesses for 32-bit fields, 16-bit accesses
30  * for 16-bit fields and 8-bit accesses for 8-bit fields.
31  */
32 static inline u8 vp_ioread8(u8 __iomem *addr)
33 {
34 	return ioread8(addr);
35 }
36 static inline u16 vp_ioread16 (__le16 __iomem *addr)
37 {
38 	return ioread16(addr);
39 }
40 
41 static inline u32 vp_ioread32(__le32 __iomem *addr)
42 {
43 	return ioread32(addr);
44 }
45 
46 static inline void vp_iowrite8(u8 value, u8 __iomem *addr)
47 {
48 	iowrite8(value, addr);
49 }
50 
51 static inline void vp_iowrite16(u16 value, __le16 __iomem *addr)
52 {
53 	iowrite16(value, addr);
54 }
55 
56 static inline void vp_iowrite32(u32 value, __le32 __iomem *addr)
57 {
58 	iowrite32(value, addr);
59 }
60 
61 static void vp_iowrite64_twopart(u64 val,
62 				 __le32 __iomem *lo, __le32 __iomem *hi)
63 {
64 	vp_iowrite32((u32)val, lo);
65 	vp_iowrite32(val >> 32, hi);
66 }
67 
68 static void __iomem *map_capability(struct pci_dev *dev, int off,
69 				    size_t minlen,
70 				    u32 align,
71 				    u32 start, u32 size,
72 				    size_t *len)
73 {
74 	u8 bar;
75 	u32 offset, length;
76 	void __iomem *p;
77 
78 	pci_read_config_byte(dev, off + offsetof(struct virtio_pci_cap,
79 						 bar),
80 			     &bar);
81 	pci_read_config_dword(dev, off + offsetof(struct virtio_pci_cap, offset),
82 			     &offset);
83 	pci_read_config_dword(dev, off + offsetof(struct virtio_pci_cap, length),
84 			      &length);
85 
86 	if (length <= start) {
87 		dev_err(&dev->dev,
88 			"virtio_pci: bad capability len %u (>%u expected)\n",
89 			length, start);
90 		return NULL;
91 	}
92 
93 	if (length - start < minlen) {
94 		dev_err(&dev->dev,
95 			"virtio_pci: bad capability len %u (>=%zu expected)\n",
96 			length, minlen);
97 		return NULL;
98 	}
99 
100 	length -= start;
101 
102 	if (start + offset < offset) {
103 		dev_err(&dev->dev,
104 			"virtio_pci: map wrap-around %u+%u\n",
105 			start, offset);
106 		return NULL;
107 	}
108 
109 	offset += start;
110 
111 	if (offset & (align - 1)) {
112 		dev_err(&dev->dev,
113 			"virtio_pci: offset %u not aligned to %u\n",
114 			offset, align);
115 		return NULL;
116 	}
117 
118 	if (length > size)
119 		length = size;
120 
121 	if (len)
122 		*len = length;
123 
124 	if (minlen + offset < minlen ||
125 	    minlen + offset > pci_resource_len(dev, bar)) {
126 		dev_err(&dev->dev,
127 			"virtio_pci: map virtio %zu@%u "
128 			"out of range on bar %i length %lu\n",
129 			minlen, offset,
130 			bar, (unsigned long)pci_resource_len(dev, bar));
131 		return NULL;
132 	}
133 
134 	p = pci_iomap_range(dev, bar, offset, length);
135 	if (!p)
136 		dev_err(&dev->dev,
137 			"virtio_pci: unable to map virtio %u@%u on bar %i\n",
138 			length, offset, bar);
139 	return p;
140 }
141 
142 /* virtio config->get_features() implementation */
143 static u64 vp_get_features(struct virtio_device *vdev)
144 {
145 	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
146 	u64 features;
147 
148 	vp_iowrite32(0, &vp_dev->common->device_feature_select);
149 	features = vp_ioread32(&vp_dev->common->device_feature);
150 	vp_iowrite32(1, &vp_dev->common->device_feature_select);
151 	features |= ((u64)vp_ioread32(&vp_dev->common->device_feature) << 32);
152 
153 	return features;
154 }
155 
156 /* virtio config->finalize_features() implementation */
157 static int vp_finalize_features(struct virtio_device *vdev)
158 {
159 	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
160 
161 	/* Give virtio_ring a chance to accept features. */
162 	vring_transport_features(vdev);
163 
164 	if (!__virtio_test_bit(vdev, VIRTIO_F_VERSION_1)) {
165 		dev_err(&vdev->dev, "virtio: device uses modern interface "
166 			"but does not have VIRTIO_F_VERSION_1\n");
167 		return -EINVAL;
168 	}
169 
170 	vp_iowrite32(0, &vp_dev->common->guest_feature_select);
171 	vp_iowrite32((u32)vdev->features, &vp_dev->common->guest_feature);
172 	vp_iowrite32(1, &vp_dev->common->guest_feature_select);
173 	vp_iowrite32(vdev->features >> 32, &vp_dev->common->guest_feature);
174 
175 	return 0;
176 }
177 
178 /* virtio config->get() implementation */
179 static void vp_get(struct virtio_device *vdev, unsigned offset,
180 		   void *buf, unsigned len)
181 {
182 	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
183 	u8 b;
184 	__le16 w;
185 	__le32 l;
186 
187 	BUG_ON(offset + len > vp_dev->device_len);
188 
189 	switch (len) {
190 	case 1:
191 		b = ioread8(vp_dev->device + offset);
192 		memcpy(buf, &b, sizeof b);
193 		break;
194 	case 2:
195 		w = cpu_to_le16(ioread16(vp_dev->device + offset));
196 		memcpy(buf, &w, sizeof w);
197 		break;
198 	case 4:
199 		l = cpu_to_le32(ioread32(vp_dev->device + offset));
200 		memcpy(buf, &l, sizeof l);
201 		break;
202 	case 8:
203 		l = cpu_to_le32(ioread32(vp_dev->device + offset));
204 		memcpy(buf, &l, sizeof l);
205 		l = cpu_to_le32(ioread32(vp_dev->device + offset + sizeof l));
206 		memcpy(buf + sizeof l, &l, sizeof l);
207 		break;
208 	default:
209 		BUG();
210 	}
211 }
212 
213 /* the config->set() implementation.  it's symmetric to the config->get()
214  * implementation */
215 static void vp_set(struct virtio_device *vdev, unsigned offset,
216 		   const void *buf, unsigned len)
217 {
218 	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
219 	u8 b;
220 	__le16 w;
221 	__le32 l;
222 
223 	BUG_ON(offset + len > vp_dev->device_len);
224 
225 	switch (len) {
226 	case 1:
227 		memcpy(&b, buf, sizeof b);
228 		iowrite8(b, vp_dev->device + offset);
229 		break;
230 	case 2:
231 		memcpy(&w, buf, sizeof w);
232 		iowrite16(le16_to_cpu(w), vp_dev->device + offset);
233 		break;
234 	case 4:
235 		memcpy(&l, buf, sizeof l);
236 		iowrite32(le32_to_cpu(l), vp_dev->device + offset);
237 		break;
238 	case 8:
239 		memcpy(&l, buf, sizeof l);
240 		iowrite32(le32_to_cpu(l), vp_dev->device + offset);
241 		memcpy(&l, buf + sizeof l, sizeof l);
242 		iowrite32(le32_to_cpu(l), vp_dev->device + offset + sizeof l);
243 		break;
244 	default:
245 		BUG();
246 	}
247 }
248 
249 static u32 vp_generation(struct virtio_device *vdev)
250 {
251 	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
252 	return vp_ioread8(&vp_dev->common->config_generation);
253 }
254 
255 /* config->{get,set}_status() implementations */
256 static u8 vp_get_status(struct virtio_device *vdev)
257 {
258 	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
259 	return vp_ioread8(&vp_dev->common->device_status);
260 }
261 
262 static void vp_set_status(struct virtio_device *vdev, u8 status)
263 {
264 	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
265 	/* We should never be setting status to 0. */
266 	BUG_ON(status == 0);
267 	vp_iowrite8(status, &vp_dev->common->device_status);
268 }
269 
270 static void vp_reset(struct virtio_device *vdev)
271 {
272 	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
273 	/* 0 status means a reset. */
274 	vp_iowrite8(0, &vp_dev->common->device_status);
275 	/* After writing 0 to device_status, the driver MUST wait for a read of
276 	 * device_status to return 0 before reinitializing the device.
277 	 * This will flush out the status write, and flush in device writes,
278 	 * including MSI-X interrupts, if any.
279 	 */
280 	while (vp_ioread8(&vp_dev->common->device_status))
281 		msleep(1);
282 	/* Flush pending VQ/configuration callbacks. */
283 	vp_synchronize_vectors(vdev);
284 }
285 
286 static u16 vp_config_vector(struct virtio_pci_device *vp_dev, u16 vector)
287 {
288 	/* Setup the vector used for configuration events */
289 	vp_iowrite16(vector, &vp_dev->common->msix_config);
290 	/* Verify we had enough resources to assign the vector */
291 	/* Will also flush the write out to device */
292 	return vp_ioread16(&vp_dev->common->msix_config);
293 }
294 
295 static struct virtqueue *setup_vq(struct virtio_pci_device *vp_dev,
296 				  unsigned index,
297 				  void (*callback)(struct virtqueue *vq),
298 				  const char *name,
299 				  u16 msix_vec)
300 {
301 	struct virtio_pci_common_cfg __iomem *cfg = vp_dev->common;
302 	struct virtqueue *vq;
303 	u16 num, off;
304 	int err;
305 
306 	if (index >= vp_ioread16(&cfg->num_queues))
307 		return ERR_PTR(-ENOENT);
308 
309 	/* Select the queue we're interested in */
310 	vp_iowrite16(index, &cfg->queue_select);
311 
312 	/* Check if queue is either not available or already active. */
313 	num = vp_ioread16(&cfg->queue_size);
314 	if (!num || vp_ioread16(&cfg->queue_enable))
315 		return ERR_PTR(-ENOENT);
316 
317 	if (num & (num - 1)) {
318 		dev_warn(&vp_dev->pci_dev->dev, "bad queue size %u", num);
319 		return ERR_PTR(-EINVAL);
320 	}
321 
322 	/* get offset of notification word for this vq */
323 	off = vp_ioread16(&cfg->queue_notify_off);
324 
325 	/* create the vring */
326 	vq = vring_create_virtqueue(index, num,
327 				    SMP_CACHE_BYTES, &vp_dev->vdev,
328 				    true, true, vp_notify, callback, name);
329 	if (!vq)
330 		return ERR_PTR(-ENOMEM);
331 
332 	/* activate the queue */
333 	vp_iowrite16(virtqueue_get_vring_size(vq), &cfg->queue_size);
334 	vp_iowrite64_twopart(virtqueue_get_desc_addr(vq),
335 			     &cfg->queue_desc_lo, &cfg->queue_desc_hi);
336 	vp_iowrite64_twopart(virtqueue_get_avail_addr(vq),
337 			     &cfg->queue_avail_lo, &cfg->queue_avail_hi);
338 	vp_iowrite64_twopart(virtqueue_get_used_addr(vq),
339 			     &cfg->queue_used_lo, &cfg->queue_used_hi);
340 
341 	if (vp_dev->notify_base) {
342 		/* offset should not wrap */
343 		if ((u64)off * vp_dev->notify_offset_multiplier + 2
344 		    > vp_dev->notify_len) {
345 			dev_warn(&vp_dev->pci_dev->dev,
346 				 "bad notification offset %u (x %u) "
347 				 "for queue %u > %zd",
348 				 off, vp_dev->notify_offset_multiplier,
349 				 index, vp_dev->notify_len);
350 			err = -EINVAL;
351 			goto err_map_notify;
352 		}
353 		vq->priv = (void __force *)vp_dev->notify_base +
354 			off * vp_dev->notify_offset_multiplier;
355 	} else {
356 		vq->priv = (void __force *)map_capability(vp_dev->pci_dev,
357 					  vp_dev->notify_map_cap, 2, 2,
358 					  off * vp_dev->notify_offset_multiplier, 2,
359 					  NULL);
360 	}
361 
362 	if (!vq->priv) {
363 		err = -ENOMEM;
364 		goto err_map_notify;
365 	}
366 
367 	if (msix_vec != VIRTIO_MSI_NO_VECTOR) {
368 		vp_iowrite16(msix_vec, &cfg->queue_msix_vector);
369 		msix_vec = vp_ioread16(&cfg->queue_msix_vector);
370 		if (msix_vec == VIRTIO_MSI_NO_VECTOR) {
371 			err = -EBUSY;
372 			goto err_assign_vector;
373 		}
374 	}
375 
376 	return vq;
377 
378 err_assign_vector:
379 	if (!vp_dev->notify_base)
380 		pci_iounmap(vp_dev->pci_dev, (void __iomem __force *)vq->priv);
381 err_map_notify:
382 	vring_del_virtqueue(vq);
383 	return ERR_PTR(err);
384 }
385 
386 static int vp_modern_find_vqs(struct virtio_device *vdev, unsigned nvqs,
387 		struct virtqueue *vqs[], vq_callback_t *callbacks[],
388 		const char * const names[], struct irq_affinity *desc)
389 {
390 	struct virtio_pci_device *vp_dev = to_vp_device(vdev);
391 	struct virtqueue *vq;
392 	int rc = vp_find_vqs(vdev, nvqs, vqs, callbacks, names, desc);
393 
394 	if (rc)
395 		return rc;
396 
397 	/* Select and activate all queues. Has to be done last: once we do
398 	 * this, there's no way to go back except reset.
399 	 */
400 	list_for_each_entry(vq, &vdev->vqs, list) {
401 		vp_iowrite16(vq->index, &vp_dev->common->queue_select);
402 		vp_iowrite16(1, &vp_dev->common->queue_enable);
403 	}
404 
405 	return 0;
406 }
407 
408 static void del_vq(struct virtqueue *vq)
409 {
410 	struct virtio_pci_device *vp_dev = to_vp_device(vq->vdev);
411 
412 	vp_iowrite16(vq->index, &vp_dev->common->queue_select);
413 
414 	if (vp_dev->pci_dev->msix_enabled) {
415 		vp_iowrite16(VIRTIO_MSI_NO_VECTOR,
416 			     &vp_dev->common->queue_msix_vector);
417 		/* Flush the write out to device */
418 		vp_ioread16(&vp_dev->common->queue_msix_vector);
419 	}
420 
421 	if (!vp_dev->notify_base)
422 		pci_iounmap(vp_dev->pci_dev, (void __force __iomem *)vq->priv);
423 
424 	vring_del_virtqueue(vq);
425 }
426 
427 static const struct virtio_config_ops virtio_pci_config_nodev_ops = {
428 	.get		= NULL,
429 	.set		= NULL,
430 	.generation	= vp_generation,
431 	.get_status	= vp_get_status,
432 	.set_status	= vp_set_status,
433 	.reset		= vp_reset,
434 	.find_vqs	= vp_modern_find_vqs,
435 	.del_vqs	= vp_del_vqs,
436 	.get_features	= vp_get_features,
437 	.finalize_features = vp_finalize_features,
438 	.bus_name	= vp_bus_name,
439 	.set_vq_affinity = vp_set_vq_affinity,
440 	.get_vq_affinity = vp_get_vq_affinity,
441 };
442 
443 static const struct virtio_config_ops virtio_pci_config_ops = {
444 	.get		= vp_get,
445 	.set		= vp_set,
446 	.generation	= vp_generation,
447 	.get_status	= vp_get_status,
448 	.set_status	= vp_set_status,
449 	.reset		= vp_reset,
450 	.find_vqs	= vp_modern_find_vqs,
451 	.del_vqs	= vp_del_vqs,
452 	.get_features	= vp_get_features,
453 	.finalize_features = vp_finalize_features,
454 	.bus_name	= vp_bus_name,
455 	.set_vq_affinity = vp_set_vq_affinity,
456 	.get_vq_affinity = vp_get_vq_affinity,
457 };
458 
459 /**
460  * virtio_pci_find_capability - walk capabilities to find device info.
461  * @dev: the pci device
462  * @cfg_type: the VIRTIO_PCI_CAP_* value we seek
463  * @ioresource_types: IORESOURCE_MEM and/or IORESOURCE_IO.
464  *
465  * Returns offset of the capability, or 0.
466  */
467 static inline int virtio_pci_find_capability(struct pci_dev *dev, u8 cfg_type,
468 					     u32 ioresource_types, int *bars)
469 {
470 	int pos;
471 
472 	for (pos = pci_find_capability(dev, PCI_CAP_ID_VNDR);
473 	     pos > 0;
474 	     pos = pci_find_next_capability(dev, pos, PCI_CAP_ID_VNDR)) {
475 		u8 type, bar;
476 		pci_read_config_byte(dev, pos + offsetof(struct virtio_pci_cap,
477 							 cfg_type),
478 				     &type);
479 		pci_read_config_byte(dev, pos + offsetof(struct virtio_pci_cap,
480 							 bar),
481 				     &bar);
482 
483 		/* Ignore structures with reserved BAR values */
484 		if (bar > 0x5)
485 			continue;
486 
487 		if (type == cfg_type) {
488 			if (pci_resource_len(dev, bar) &&
489 			    pci_resource_flags(dev, bar) & ioresource_types) {
490 				*bars |= (1 << bar);
491 				return pos;
492 			}
493 		}
494 	}
495 	return 0;
496 }
497 
498 /* This is part of the ABI.  Don't screw with it. */
499 static inline void check_offsets(void)
500 {
501 	/* Note: disk space was harmed in compilation of this function. */
502 	BUILD_BUG_ON(VIRTIO_PCI_CAP_VNDR !=
503 		     offsetof(struct virtio_pci_cap, cap_vndr));
504 	BUILD_BUG_ON(VIRTIO_PCI_CAP_NEXT !=
505 		     offsetof(struct virtio_pci_cap, cap_next));
506 	BUILD_BUG_ON(VIRTIO_PCI_CAP_LEN !=
507 		     offsetof(struct virtio_pci_cap, cap_len));
508 	BUILD_BUG_ON(VIRTIO_PCI_CAP_CFG_TYPE !=
509 		     offsetof(struct virtio_pci_cap, cfg_type));
510 	BUILD_BUG_ON(VIRTIO_PCI_CAP_BAR !=
511 		     offsetof(struct virtio_pci_cap, bar));
512 	BUILD_BUG_ON(VIRTIO_PCI_CAP_OFFSET !=
513 		     offsetof(struct virtio_pci_cap, offset));
514 	BUILD_BUG_ON(VIRTIO_PCI_CAP_LENGTH !=
515 		     offsetof(struct virtio_pci_cap, length));
516 	BUILD_BUG_ON(VIRTIO_PCI_NOTIFY_CAP_MULT !=
517 		     offsetof(struct virtio_pci_notify_cap,
518 			      notify_off_multiplier));
519 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_DFSELECT !=
520 		     offsetof(struct virtio_pci_common_cfg,
521 			      device_feature_select));
522 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_DF !=
523 		     offsetof(struct virtio_pci_common_cfg, device_feature));
524 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_GFSELECT !=
525 		     offsetof(struct virtio_pci_common_cfg,
526 			      guest_feature_select));
527 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_GF !=
528 		     offsetof(struct virtio_pci_common_cfg, guest_feature));
529 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_MSIX !=
530 		     offsetof(struct virtio_pci_common_cfg, msix_config));
531 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_NUMQ !=
532 		     offsetof(struct virtio_pci_common_cfg, num_queues));
533 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_STATUS !=
534 		     offsetof(struct virtio_pci_common_cfg, device_status));
535 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_CFGGENERATION !=
536 		     offsetof(struct virtio_pci_common_cfg, config_generation));
537 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_Q_SELECT !=
538 		     offsetof(struct virtio_pci_common_cfg, queue_select));
539 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_Q_SIZE !=
540 		     offsetof(struct virtio_pci_common_cfg, queue_size));
541 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_Q_MSIX !=
542 		     offsetof(struct virtio_pci_common_cfg, queue_msix_vector));
543 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_Q_ENABLE !=
544 		     offsetof(struct virtio_pci_common_cfg, queue_enable));
545 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_Q_NOFF !=
546 		     offsetof(struct virtio_pci_common_cfg, queue_notify_off));
547 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_Q_DESCLO !=
548 		     offsetof(struct virtio_pci_common_cfg, queue_desc_lo));
549 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_Q_DESCHI !=
550 		     offsetof(struct virtio_pci_common_cfg, queue_desc_hi));
551 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_Q_AVAILLO !=
552 		     offsetof(struct virtio_pci_common_cfg, queue_avail_lo));
553 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_Q_AVAILHI !=
554 		     offsetof(struct virtio_pci_common_cfg, queue_avail_hi));
555 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_Q_USEDLO !=
556 		     offsetof(struct virtio_pci_common_cfg, queue_used_lo));
557 	BUILD_BUG_ON(VIRTIO_PCI_COMMON_Q_USEDHI !=
558 		     offsetof(struct virtio_pci_common_cfg, queue_used_hi));
559 }
560 
561 /* the PCI probing function */
562 int virtio_pci_modern_probe(struct virtio_pci_device *vp_dev)
563 {
564 	struct pci_dev *pci_dev = vp_dev->pci_dev;
565 	int err, common, isr, notify, device;
566 	u32 notify_length;
567 	u32 notify_offset;
568 
569 	check_offsets();
570 
571 	/* We only own devices >= 0x1000 and <= 0x107f: leave the rest. */
572 	if (pci_dev->device < 0x1000 || pci_dev->device > 0x107f)
573 		return -ENODEV;
574 
575 	if (pci_dev->device < 0x1040) {
576 		/* Transitional devices: use the PCI subsystem device id as
577 		 * virtio device id, same as legacy driver always did.
578 		 */
579 		vp_dev->vdev.id.device = pci_dev->subsystem_device;
580 	} else {
581 		/* Modern devices: simply use PCI device id, but start from 0x1040. */
582 		vp_dev->vdev.id.device = pci_dev->device - 0x1040;
583 	}
584 	vp_dev->vdev.id.vendor = pci_dev->subsystem_vendor;
585 
586 	/* check for a common config: if not, use legacy mode (bar 0). */
587 	common = virtio_pci_find_capability(pci_dev, VIRTIO_PCI_CAP_COMMON_CFG,
588 					    IORESOURCE_IO | IORESOURCE_MEM,
589 					    &vp_dev->modern_bars);
590 	if (!common) {
591 		dev_info(&pci_dev->dev,
592 			 "virtio_pci: leaving for legacy driver\n");
593 		return -ENODEV;
594 	}
595 
596 	/* If common is there, these should be too... */
597 	isr = virtio_pci_find_capability(pci_dev, VIRTIO_PCI_CAP_ISR_CFG,
598 					 IORESOURCE_IO | IORESOURCE_MEM,
599 					 &vp_dev->modern_bars);
600 	notify = virtio_pci_find_capability(pci_dev, VIRTIO_PCI_CAP_NOTIFY_CFG,
601 					    IORESOURCE_IO | IORESOURCE_MEM,
602 					    &vp_dev->modern_bars);
603 	if (!isr || !notify) {
604 		dev_err(&pci_dev->dev,
605 			"virtio_pci: missing capabilities %i/%i/%i\n",
606 			common, isr, notify);
607 		return -EINVAL;
608 	}
609 
610 	err = dma_set_mask_and_coherent(&pci_dev->dev, DMA_BIT_MASK(64));
611 	if (err)
612 		err = dma_set_mask_and_coherent(&pci_dev->dev,
613 						DMA_BIT_MASK(32));
614 	if (err)
615 		dev_warn(&pci_dev->dev, "Failed to enable 64-bit or 32-bit DMA.  Trying to continue, but this might not work.\n");
616 
617 	/* Device capability is only mandatory for devices that have
618 	 * device-specific configuration.
619 	 */
620 	device = virtio_pci_find_capability(pci_dev, VIRTIO_PCI_CAP_DEVICE_CFG,
621 					    IORESOURCE_IO | IORESOURCE_MEM,
622 					    &vp_dev->modern_bars);
623 
624 	err = pci_request_selected_regions(pci_dev, vp_dev->modern_bars,
625 					   "virtio-pci-modern");
626 	if (err)
627 		return err;
628 
629 	err = -EINVAL;
630 	vp_dev->common = map_capability(pci_dev, common,
631 					sizeof(struct virtio_pci_common_cfg), 4,
632 					0, sizeof(struct virtio_pci_common_cfg),
633 					NULL);
634 	if (!vp_dev->common)
635 		goto err_map_common;
636 	vp_dev->isr = map_capability(pci_dev, isr, sizeof(u8), 1,
637 				     0, 1,
638 				     NULL);
639 	if (!vp_dev->isr)
640 		goto err_map_isr;
641 
642 	/* Read notify_off_multiplier from config space. */
643 	pci_read_config_dword(pci_dev,
644 			      notify + offsetof(struct virtio_pci_notify_cap,
645 						notify_off_multiplier),
646 			      &vp_dev->notify_offset_multiplier);
647 	/* Read notify length and offset from config space. */
648 	pci_read_config_dword(pci_dev,
649 			      notify + offsetof(struct virtio_pci_notify_cap,
650 						cap.length),
651 			      &notify_length);
652 
653 	pci_read_config_dword(pci_dev,
654 			      notify + offsetof(struct virtio_pci_notify_cap,
655 						cap.offset),
656 			      &notify_offset);
657 
658 	/* We don't know how many VQs we'll map, ahead of the time.
659 	 * If notify length is small, map it all now.
660 	 * Otherwise, map each VQ individually later.
661 	 */
662 	if ((u64)notify_length + (notify_offset % PAGE_SIZE) <= PAGE_SIZE) {
663 		vp_dev->notify_base = map_capability(pci_dev, notify, 2, 2,
664 						     0, notify_length,
665 						     &vp_dev->notify_len);
666 		if (!vp_dev->notify_base)
667 			goto err_map_notify;
668 	} else {
669 		vp_dev->notify_map_cap = notify;
670 	}
671 
672 	/* Again, we don't know how much we should map, but PAGE_SIZE
673 	 * is more than enough for all existing devices.
674 	 */
675 	if (device) {
676 		vp_dev->device = map_capability(pci_dev, device, 0, 4,
677 						0, PAGE_SIZE,
678 						&vp_dev->device_len);
679 		if (!vp_dev->device)
680 			goto err_map_device;
681 
682 		vp_dev->vdev.config = &virtio_pci_config_ops;
683 	} else {
684 		vp_dev->vdev.config = &virtio_pci_config_nodev_ops;
685 	}
686 
687 	vp_dev->config_vector = vp_config_vector;
688 	vp_dev->setup_vq = setup_vq;
689 	vp_dev->del_vq = del_vq;
690 
691 	return 0;
692 
693 err_map_device:
694 	if (vp_dev->notify_base)
695 		pci_iounmap(pci_dev, vp_dev->notify_base);
696 err_map_notify:
697 	pci_iounmap(pci_dev, vp_dev->isr);
698 err_map_isr:
699 	pci_iounmap(pci_dev, vp_dev->common);
700 err_map_common:
701 	return err;
702 }
703 
704 void virtio_pci_modern_remove(struct virtio_pci_device *vp_dev)
705 {
706 	struct pci_dev *pci_dev = vp_dev->pci_dev;
707 
708 	if (vp_dev->device)
709 		pci_iounmap(pci_dev, vp_dev->device);
710 	if (vp_dev->notify_base)
711 		pci_iounmap(pci_dev, vp_dev->notify_base);
712 	pci_iounmap(pci_dev, vp_dev->isr);
713 	pci_iounmap(pci_dev, vp_dev->common);
714 	pci_release_selected_regions(pci_dev, vp_dev->modern_bars);
715 }
716