xref: /openbmc/linux/drivers/vfio/pci/vfio_pci.c (revision 2223cbec) 
1 /*
2  * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
3  *     Author: Alex Williamson <alex.williamson@redhat.com>
4  *
5  * This program is free software; you can redistribute it and/or modify
6  * it under the terms of the GNU General Public License version 2 as
7  * published by the Free Software Foundation.
8  *
9  * Derived from original vfio:
10  * Copyright 2010 Cisco Systems, Inc.  All rights reserved.
11  * Author: Tom Lyon, pugs@cisco.com
12  */
13 
14 #include <linux/device.h>
15 #include <linux/eventfd.h>
16 #include <linux/interrupt.h>
17 #include <linux/iommu.h>
18 #include <linux/module.h>
19 #include <linux/mutex.h>
20 #include <linux/notifier.h>
21 #include <linux/pci.h>
22 #include <linux/pm_runtime.h>
23 #include <linux/slab.h>
24 #include <linux/types.h>
25 #include <linux/uaccess.h>
26 #include <linux/vfio.h>
27 
28 #include "vfio_pci_private.h"
29 
30 #define DRIVER_VERSION  "0.2"
31 #define DRIVER_AUTHOR   "Alex Williamson <alex.williamson@redhat.com>"
32 #define DRIVER_DESC     "VFIO PCI - User Level meta-driver"
33 
34 static bool nointxmask;
35 module_param_named(nointxmask, nointxmask, bool, S_IRUGO | S_IWUSR);
36 MODULE_PARM_DESC(nointxmask,
37 		  "Disable support for PCI 2.3 style INTx masking.  If this resolves problems for specific devices, report lspci -vvvxxx to linux-pci@vger.kernel.org so the device can be fixed automatically via the broken_intx_masking flag.");
38 
39 static int vfio_pci_enable(struct vfio_pci_device *vdev)
40 {
41 	struct pci_dev *pdev = vdev->pdev;
42 	int ret;
43 	u16 cmd;
44 	u8 msix_pos;
45 
46 	vdev->reset_works = (pci_reset_function(pdev) == 0);
47 	pci_save_state(pdev);
48 	vdev->pci_saved_state = pci_store_saved_state(pdev);
49 	if (!vdev->pci_saved_state)
50 		pr_debug("%s: Couldn't store %s saved state\n",
51 			 __func__, dev_name(&pdev->dev));
52 
53 	ret = vfio_config_init(vdev);
54 	if (ret)
55 		goto out;
56 
57 	if (likely(!nointxmask))
58 		vdev->pci_2_3 = pci_intx_mask_supported(pdev);
59 
60 	pci_read_config_word(pdev, PCI_COMMAND, &cmd);
61 	if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
62 		cmd &= ~PCI_COMMAND_INTX_DISABLE;
63 		pci_write_config_word(pdev, PCI_COMMAND, cmd);
64 	}
65 
66 	msix_pos = pci_find_capability(pdev, PCI_CAP_ID_MSIX);
67 	if (msix_pos) {
68 		u16 flags;
69 		u32 table;
70 
71 		pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
72 		pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
73 
74 		vdev->msix_bar = table & PCI_MSIX_FLAGS_BIRMASK;
75 		vdev->msix_offset = table & ~PCI_MSIX_FLAGS_BIRMASK;
76 		vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
77 	} else
78 		vdev->msix_bar = 0xFF;
79 
80 	ret = pci_enable_device(pdev);
81 	if (ret)
82 		goto out;
83 
84 	return ret;
85 
86 out:
87 	kfree(vdev->pci_saved_state);
88 	vdev->pci_saved_state = NULL;
89 	vfio_config_free(vdev);
90 	return ret;
91 }
92 
93 static void vfio_pci_disable(struct vfio_pci_device *vdev)
94 {
95 	int bar;
96 
97 	pci_disable_device(vdev->pdev);
98 
99 	vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
100 				VFIO_IRQ_SET_ACTION_TRIGGER,
101 				vdev->irq_type, 0, 0, NULL);
102 
103 	vdev->virq_disabled = false;
104 
105 	vfio_config_free(vdev);
106 
107 	pci_reset_function(vdev->pdev);
108 
109 	if (pci_load_and_free_saved_state(vdev->pdev,
110 					  &vdev->pci_saved_state) == 0)
111 		pci_restore_state(vdev->pdev);
112 	else
113 		pr_info("%s: Couldn't reload %s saved state\n",
114 			__func__, dev_name(&vdev->pdev->dev));
115 
116 	for (bar = PCI_STD_RESOURCES; bar <= PCI_STD_RESOURCE_END; bar++) {
117 		if (!vdev->barmap[bar])
118 			continue;
119 		pci_iounmap(vdev->pdev, vdev->barmap[bar]);
120 		pci_release_selected_regions(vdev->pdev, 1 << bar);
121 		vdev->barmap[bar] = NULL;
122 	}
123 }
124 
125 static void vfio_pci_release(void *device_data)
126 {
127 	struct vfio_pci_device *vdev = device_data;
128 
129 	if (atomic_dec_and_test(&vdev->refcnt))
130 		vfio_pci_disable(vdev);
131 
132 	module_put(THIS_MODULE);
133 }
134 
135 static int vfio_pci_open(void *device_data)
136 {
137 	struct vfio_pci_device *vdev = device_data;
138 
139 	if (!try_module_get(THIS_MODULE))
140 		return -ENODEV;
141 
142 	if (atomic_inc_return(&vdev->refcnt) == 1) {
143 		int ret = vfio_pci_enable(vdev);
144 		if (ret) {
145 			module_put(THIS_MODULE);
146 			return ret;
147 		}
148 	}
149 
150 	return 0;
151 }
152 
153 static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
154 {
155 	if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
156 		u8 pin;
157 		pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
158 		if (pin)
159 			return 1;
160 
161 	} else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
162 		u8 pos;
163 		u16 flags;
164 
165 		pos = pci_find_capability(vdev->pdev, PCI_CAP_ID_MSI);
166 		if (pos) {
167 			pci_read_config_word(vdev->pdev,
168 					     pos + PCI_MSI_FLAGS, &flags);
169 
170 			return 1 << (flags & PCI_MSI_FLAGS_QMASK);
171 		}
172 	} else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
173 		u8 pos;
174 		u16 flags;
175 
176 		pos = pci_find_capability(vdev->pdev, PCI_CAP_ID_MSIX);
177 		if (pos) {
178 			pci_read_config_word(vdev->pdev,
179 					     pos + PCI_MSIX_FLAGS, &flags);
180 
181 			return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
182 		}
183 	}
184 
185 	return 0;
186 }
187 
188 static long vfio_pci_ioctl(void *device_data,
189 			   unsigned int cmd, unsigned long arg)
190 {
191 	struct vfio_pci_device *vdev = device_data;
192 	unsigned long minsz;
193 
194 	if (cmd == VFIO_DEVICE_GET_INFO) {
195 		struct vfio_device_info info;
196 
197 		minsz = offsetofend(struct vfio_device_info, num_irqs);
198 
199 		if (copy_from_user(&info, (void __user *)arg, minsz))
200 			return -EFAULT;
201 
202 		if (info.argsz < minsz)
203 			return -EINVAL;
204 
205 		info.flags = VFIO_DEVICE_FLAGS_PCI;
206 
207 		if (vdev->reset_works)
208 			info.flags |= VFIO_DEVICE_FLAGS_RESET;
209 
210 		info.num_regions = VFIO_PCI_NUM_REGIONS;
211 		info.num_irqs = VFIO_PCI_NUM_IRQS;
212 
213 		return copy_to_user((void __user *)arg, &info, minsz);
214 
215 	} else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
216 		struct pci_dev *pdev = vdev->pdev;
217 		struct vfio_region_info info;
218 
219 		minsz = offsetofend(struct vfio_region_info, offset);
220 
221 		if (copy_from_user(&info, (void __user *)arg, minsz))
222 			return -EFAULT;
223 
224 		if (info.argsz < minsz)
225 			return -EINVAL;
226 
227 		switch (info.index) {
228 		case VFIO_PCI_CONFIG_REGION_INDEX:
229 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
230 			info.size = pdev->cfg_size;
231 			info.flags = VFIO_REGION_INFO_FLAG_READ |
232 				     VFIO_REGION_INFO_FLAG_WRITE;
233 			break;
234 		case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
235 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
236 			info.size = pci_resource_len(pdev, info.index);
237 			if (!info.size) {
238 				info.flags = 0;
239 				break;
240 			}
241 
242 			info.flags = VFIO_REGION_INFO_FLAG_READ |
243 				     VFIO_REGION_INFO_FLAG_WRITE;
244 			if (pci_resource_flags(pdev, info.index) &
245 			    IORESOURCE_MEM && info.size >= PAGE_SIZE)
246 				info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
247 			break;
248 		case VFIO_PCI_ROM_REGION_INDEX:
249 		{
250 			void __iomem *io;
251 			size_t size;
252 
253 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
254 			info.flags = 0;
255 
256 			/* Report the BAR size, not the ROM size */
257 			info.size = pci_resource_len(pdev, info.index);
258 			if (!info.size)
259 				break;
260 
261 			/* Is it really there? */
262 			io = pci_map_rom(pdev, &size);
263 			if (!io || !size) {
264 				info.size = 0;
265 				break;
266 			}
267 			pci_unmap_rom(pdev, io);
268 
269 			info.flags = VFIO_REGION_INFO_FLAG_READ;
270 			break;
271 		}
272 		default:
273 			return -EINVAL;
274 		}
275 
276 		return copy_to_user((void __user *)arg, &info, minsz);
277 
278 	} else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
279 		struct vfio_irq_info info;
280 
281 		minsz = offsetofend(struct vfio_irq_info, count);
282 
283 		if (copy_from_user(&info, (void __user *)arg, minsz))
284 			return -EFAULT;
285 
286 		if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
287 			return -EINVAL;
288 
289 		info.flags = VFIO_IRQ_INFO_EVENTFD;
290 
291 		info.count = vfio_pci_get_irq_count(vdev, info.index);
292 
293 		if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
294 			info.flags |= (VFIO_IRQ_INFO_MASKABLE |
295 				       VFIO_IRQ_INFO_AUTOMASKED);
296 		else
297 			info.flags |= VFIO_IRQ_INFO_NORESIZE;
298 
299 		return copy_to_user((void __user *)arg, &info, minsz);
300 
301 	} else if (cmd == VFIO_DEVICE_SET_IRQS) {
302 		struct vfio_irq_set hdr;
303 		u8 *data = NULL;
304 		int ret = 0;
305 
306 		minsz = offsetofend(struct vfio_irq_set, count);
307 
308 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
309 			return -EFAULT;
310 
311 		if (hdr.argsz < minsz || hdr.index >= VFIO_PCI_NUM_IRQS ||
312 		    hdr.flags & ~(VFIO_IRQ_SET_DATA_TYPE_MASK |
313 				  VFIO_IRQ_SET_ACTION_TYPE_MASK))
314 			return -EINVAL;
315 
316 		if (!(hdr.flags & VFIO_IRQ_SET_DATA_NONE)) {
317 			size_t size;
318 
319 			if (hdr.flags & VFIO_IRQ_SET_DATA_BOOL)
320 				size = sizeof(uint8_t);
321 			else if (hdr.flags & VFIO_IRQ_SET_DATA_EVENTFD)
322 				size = sizeof(int32_t);
323 			else
324 				return -EINVAL;
325 
326 			if (hdr.argsz - minsz < hdr.count * size ||
327 			    hdr.count > vfio_pci_get_irq_count(vdev, hdr.index))
328 				return -EINVAL;
329 
330 			data = kmalloc(hdr.count * size, GFP_KERNEL);
331 			if (!data)
332 				return -ENOMEM;
333 
334 			if (copy_from_user(data, (void __user *)(arg + minsz),
335 					   hdr.count * size)) {
336 				kfree(data);
337 				return -EFAULT;
338 			}
339 		}
340 
341 		mutex_lock(&vdev->igate);
342 
343 		ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
344 					      hdr.start, hdr.count, data);
345 
346 		mutex_unlock(&vdev->igate);
347 		kfree(data);
348 
349 		return ret;
350 
351 	} else if (cmd == VFIO_DEVICE_RESET)
352 		return vdev->reset_works ?
353 			pci_reset_function(vdev->pdev) : -EINVAL;
354 
355 	return -ENOTTY;
356 }
357 
358 static ssize_t vfio_pci_read(void *device_data, char __user *buf,
359 			     size_t count, loff_t *ppos)
360 {
361 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
362 	struct vfio_pci_device *vdev = device_data;
363 	struct pci_dev *pdev = vdev->pdev;
364 
365 	if (index >= VFIO_PCI_NUM_REGIONS)
366 		return -EINVAL;
367 
368 	if (index == VFIO_PCI_CONFIG_REGION_INDEX)
369 		return vfio_pci_config_readwrite(vdev, buf, count, ppos, false);
370 	else if (index == VFIO_PCI_ROM_REGION_INDEX)
371 		return vfio_pci_mem_readwrite(vdev, buf, count, ppos, false);
372 	else if (pci_resource_flags(pdev, index) & IORESOURCE_IO)
373 		return vfio_pci_io_readwrite(vdev, buf, count, ppos, false);
374 	else if (pci_resource_flags(pdev, index) & IORESOURCE_MEM)
375 		return vfio_pci_mem_readwrite(vdev, buf, count, ppos, false);
376 
377 	return -EINVAL;
378 }
379 
380 static ssize_t vfio_pci_write(void *device_data, const char __user *buf,
381 			      size_t count, loff_t *ppos)
382 {
383 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
384 	struct vfio_pci_device *vdev = device_data;
385 	struct pci_dev *pdev = vdev->pdev;
386 
387 	if (index >= VFIO_PCI_NUM_REGIONS)
388 		return -EINVAL;
389 
390 	if (index == VFIO_PCI_CONFIG_REGION_INDEX)
391 		return vfio_pci_config_readwrite(vdev, (char __user *)buf,
392 						 count, ppos, true);
393 	else if (index == VFIO_PCI_ROM_REGION_INDEX)
394 		return -EINVAL;
395 	else if (pci_resource_flags(pdev, index) & IORESOURCE_IO)
396 		return vfio_pci_io_readwrite(vdev, (char __user *)buf,
397 					     count, ppos, true);
398 	else if (pci_resource_flags(pdev, index) & IORESOURCE_MEM) {
399 		return vfio_pci_mem_readwrite(vdev, (char __user *)buf,
400 					      count, ppos, true);
401 	}
402 
403 	return -EINVAL;
404 }
405 
406 static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma)
407 {
408 	struct vfio_pci_device *vdev = device_data;
409 	struct pci_dev *pdev = vdev->pdev;
410 	unsigned int index;
411 	u64 phys_len, req_len, pgoff, req_start;
412 	int ret;
413 
414 	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
415 
416 	if (vma->vm_end < vma->vm_start)
417 		return -EINVAL;
418 	if ((vma->vm_flags & VM_SHARED) == 0)
419 		return -EINVAL;
420 	if (index >= VFIO_PCI_ROM_REGION_INDEX)
421 		return -EINVAL;
422 	if (!(pci_resource_flags(pdev, index) & IORESOURCE_MEM))
423 		return -EINVAL;
424 
425 	phys_len = pci_resource_len(pdev, index);
426 	req_len = vma->vm_end - vma->vm_start;
427 	pgoff = vma->vm_pgoff &
428 		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
429 	req_start = pgoff << PAGE_SHIFT;
430 
431 	if (phys_len < PAGE_SIZE || req_start + req_len > phys_len)
432 		return -EINVAL;
433 
434 	if (index == vdev->msix_bar) {
435 		/*
436 		 * Disallow mmaps overlapping the MSI-X table; users don't
437 		 * get to touch this directly.  We could find somewhere
438 		 * else to map the overlap, but page granularity is only
439 		 * a recommendation, not a requirement, so the user needs
440 		 * to know which bits are real.  Requiring them to mmap
441 		 * around the table makes that clear.
442 		 */
443 
444 		/* If neither entirely above nor below, then it overlaps */
445 		if (!(req_start >= vdev->msix_offset + vdev->msix_size ||
446 		      req_start + req_len <= vdev->msix_offset))
447 			return -EINVAL;
448 	}
449 
450 	/*
451 	 * Even though we don't make use of the barmap for the mmap,
452 	 * we need to request the region and the barmap tracks that.
453 	 */
454 	if (!vdev->barmap[index]) {
455 		ret = pci_request_selected_regions(pdev,
456 						   1 << index, "vfio-pci");
457 		if (ret)
458 			return ret;
459 
460 		vdev->barmap[index] = pci_iomap(pdev, index, 0);
461 	}
462 
463 	vma->vm_private_data = vdev;
464 	vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
465 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
466 	vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
467 
468 	return remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
469 			       req_len, vma->vm_page_prot);
470 }
471 
472 static const struct vfio_device_ops vfio_pci_ops = {
473 	.name		= "vfio-pci",
474 	.open		= vfio_pci_open,
475 	.release	= vfio_pci_release,
476 	.ioctl		= vfio_pci_ioctl,
477 	.read		= vfio_pci_read,
478 	.write		= vfio_pci_write,
479 	.mmap		= vfio_pci_mmap,
480 };
481 
482 static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
483 {
484 	u8 type;
485 	struct vfio_pci_device *vdev;
486 	struct iommu_group *group;
487 	int ret;
488 
489 	pci_read_config_byte(pdev, PCI_HEADER_TYPE, &type);
490 	if ((type & PCI_HEADER_TYPE) != PCI_HEADER_TYPE_NORMAL)
491 		return -EINVAL;
492 
493 	group = iommu_group_get(&pdev->dev);
494 	if (!group)
495 		return -EINVAL;
496 
497 	vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
498 	if (!vdev) {
499 		iommu_group_put(group);
500 		return -ENOMEM;
501 	}
502 
503 	vdev->pdev = pdev;
504 	vdev->irq_type = VFIO_PCI_NUM_IRQS;
505 	mutex_init(&vdev->igate);
506 	spin_lock_init(&vdev->irqlock);
507 	atomic_set(&vdev->refcnt, 0);
508 
509 	ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
510 	if (ret) {
511 		iommu_group_put(group);
512 		kfree(vdev);
513 	}
514 
515 	return ret;
516 }
517 
518 static void vfio_pci_remove(struct pci_dev *pdev)
519 {
520 	struct vfio_pci_device *vdev;
521 
522 	vdev = vfio_del_group_dev(&pdev->dev);
523 	if (!vdev)
524 		return;
525 
526 	iommu_group_put(pdev->dev.iommu_group);
527 	kfree(vdev);
528 }
529 
530 static struct pci_driver vfio_pci_driver = {
531 	.name		= "vfio-pci",
532 	.id_table	= NULL, /* only dynamic ids */
533 	.probe		= vfio_pci_probe,
534 	.remove		= vfio_pci_remove,
535 };
536 
537 static void __exit vfio_pci_cleanup(void)
538 {
539 	pci_unregister_driver(&vfio_pci_driver);
540 	vfio_pci_virqfd_exit();
541 	vfio_pci_uninit_perm_bits();
542 }
543 
544 static int __init vfio_pci_init(void)
545 {
546 	int ret;
547 
548 	/* Allocate shared config space permision data used by all devices */
549 	ret = vfio_pci_init_perm_bits();
550 	if (ret)
551 		return ret;
552 
553 	/* Start the virqfd cleanup handler */
554 	ret = vfio_pci_virqfd_init();
555 	if (ret)
556 		goto out_virqfd;
557 
558 	/* Register and scan for devices */
559 	ret = pci_register_driver(&vfio_pci_driver);
560 	if (ret)
561 		goto out_driver;
562 
563 	return 0;
564 
565 out_virqfd:
566 	vfio_pci_virqfd_exit();
567 out_driver:
568 	vfio_pci_uninit_perm_bits();
569 	return ret;
570 }
571 
572 module_init(vfio_pci_init);
573 module_exit(vfio_pci_cleanup);
574 
575 MODULE_VERSION(DRIVER_VERSION);
576 MODULE_LICENSE("GPL v2");
577 MODULE_AUTHOR(DRIVER_AUTHOR);
578 MODULE_DESCRIPTION(DRIVER_DESC);
579