xref: /openbmc/linux/drivers/vfio/pci/vfio_pci.c (revision f7777dcc)
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/file.h>
17 #include <linux/interrupt.h>
18 #include <linux/iommu.h>
19 #include <linux/module.h>
20 #include <linux/mutex.h>
21 #include <linux/notifier.h>
22 #include <linux/pci.h>
23 #include <linux/pm_runtime.h>
24 #include <linux/slab.h>
25 #include <linux/types.h>
26 #include <linux/uaccess.h>
27 #include <linux/vfio.h>
28 
29 #include "vfio_pci_private.h"
30 
31 #define DRIVER_VERSION  "0.2"
32 #define DRIVER_AUTHOR   "Alex Williamson <alex.williamson@redhat.com>"
33 #define DRIVER_DESC     "VFIO PCI - User Level meta-driver"
34 
35 static bool nointxmask;
36 module_param_named(nointxmask, nointxmask, bool, S_IRUGO | S_IWUSR);
37 MODULE_PARM_DESC(nointxmask,
38 		  "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.");
39 
40 static int vfio_pci_enable(struct vfio_pci_device *vdev)
41 {
42 	struct pci_dev *pdev = vdev->pdev;
43 	int ret;
44 	u16 cmd;
45 	u8 msix_pos;
46 
47 	ret = pci_enable_device(pdev);
48 	if (ret)
49 		return ret;
50 
51 	vdev->reset_works = (pci_reset_function(pdev) == 0);
52 	pci_save_state(pdev);
53 	vdev->pci_saved_state = pci_store_saved_state(pdev);
54 	if (!vdev->pci_saved_state)
55 		pr_debug("%s: Couldn't store %s saved state\n",
56 			 __func__, dev_name(&pdev->dev));
57 
58 	ret = vfio_config_init(vdev);
59 	if (ret) {
60 		pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state);
61 		pci_disable_device(pdev);
62 		return ret;
63 	}
64 
65 	if (likely(!nointxmask))
66 		vdev->pci_2_3 = pci_intx_mask_supported(pdev);
67 
68 	pci_read_config_word(pdev, PCI_COMMAND, &cmd);
69 	if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
70 		cmd &= ~PCI_COMMAND_INTX_DISABLE;
71 		pci_write_config_word(pdev, PCI_COMMAND, cmd);
72 	}
73 
74 	msix_pos = pdev->msix_cap;
75 	if (msix_pos) {
76 		u16 flags;
77 		u32 table;
78 
79 		pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
80 		pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
81 
82 		vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
83 		vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
84 		vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
85 	} else
86 		vdev->msix_bar = 0xFF;
87 
88 #ifdef CONFIG_VFIO_PCI_VGA
89 	if ((pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA)
90 		vdev->has_vga = true;
91 #endif
92 
93 	return 0;
94 }
95 
96 static void vfio_pci_disable(struct vfio_pci_device *vdev)
97 {
98 	struct pci_dev *pdev = vdev->pdev;
99 	int bar;
100 
101 	pci_disable_device(pdev);
102 
103 	vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
104 				VFIO_IRQ_SET_ACTION_TRIGGER,
105 				vdev->irq_type, 0, 0, NULL);
106 
107 	vdev->virq_disabled = false;
108 
109 	vfio_config_free(vdev);
110 
111 	for (bar = PCI_STD_RESOURCES; bar <= PCI_STD_RESOURCE_END; bar++) {
112 		if (!vdev->barmap[bar])
113 			continue;
114 		pci_iounmap(pdev, vdev->barmap[bar]);
115 		pci_release_selected_regions(pdev, 1 << bar);
116 		vdev->barmap[bar] = NULL;
117 	}
118 
119 	/*
120 	 * If we have saved state, restore it.  If we can reset the device,
121 	 * even better.  Resetting with current state seems better than
122 	 * nothing, but saving and restoring current state without reset
123 	 * is just busy work.
124 	 */
125 	if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
126 		pr_info("%s: Couldn't reload %s saved state\n",
127 			__func__, dev_name(&pdev->dev));
128 
129 		if (!vdev->reset_works)
130 			return;
131 
132 		pci_save_state(pdev);
133 	}
134 
135 	/*
136 	 * Disable INTx and MSI, presumably to avoid spurious interrupts
137 	 * during reset.  Stolen from pci_reset_function()
138 	 */
139 	pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
140 
141 	/*
142 	 * Careful, device_lock may already be held.  This is the case if
143 	 * a driver unbind is blocked.  Try to get the locks ourselves to
144 	 * prevent a deadlock.
145 	 */
146 	if (vdev->reset_works) {
147 		bool reset_done = false;
148 
149 		if (pci_cfg_access_trylock(pdev)) {
150 			if (device_trylock(&pdev->dev)) {
151 				__pci_reset_function_locked(pdev);
152 				reset_done = true;
153 				device_unlock(&pdev->dev);
154 			}
155 			pci_cfg_access_unlock(pdev);
156 		}
157 
158 		if (!reset_done)
159 			pr_warn("%s: Unable to acquire locks for reset of %s\n",
160 				__func__, dev_name(&pdev->dev));
161 	}
162 
163 	pci_restore_state(pdev);
164 }
165 
166 static void vfio_pci_release(void *device_data)
167 {
168 	struct vfio_pci_device *vdev = device_data;
169 
170 	if (atomic_dec_and_test(&vdev->refcnt))
171 		vfio_pci_disable(vdev);
172 
173 	module_put(THIS_MODULE);
174 }
175 
176 static int vfio_pci_open(void *device_data)
177 {
178 	struct vfio_pci_device *vdev = device_data;
179 
180 	if (!try_module_get(THIS_MODULE))
181 		return -ENODEV;
182 
183 	if (atomic_inc_return(&vdev->refcnt) == 1) {
184 		int ret = vfio_pci_enable(vdev);
185 		if (ret) {
186 			module_put(THIS_MODULE);
187 			return ret;
188 		}
189 	}
190 
191 	return 0;
192 }
193 
194 static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
195 {
196 	if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
197 		u8 pin;
198 		pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
199 		if (pin)
200 			return 1;
201 
202 	} else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
203 		u8 pos;
204 		u16 flags;
205 
206 		pos = vdev->pdev->msi_cap;
207 		if (pos) {
208 			pci_read_config_word(vdev->pdev,
209 					     pos + PCI_MSI_FLAGS, &flags);
210 
211 			return 1 << (flags & PCI_MSI_FLAGS_QMASK);
212 		}
213 	} else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
214 		u8 pos;
215 		u16 flags;
216 
217 		pos = vdev->pdev->msix_cap;
218 		if (pos) {
219 			pci_read_config_word(vdev->pdev,
220 					     pos + PCI_MSIX_FLAGS, &flags);
221 
222 			return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
223 		}
224 	} else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX)
225 		if (pci_is_pcie(vdev->pdev))
226 			return 1;
227 
228 	return 0;
229 }
230 
231 static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
232 {
233 	(*(int *)data)++;
234 	return 0;
235 }
236 
237 struct vfio_pci_fill_info {
238 	int max;
239 	int cur;
240 	struct vfio_pci_dependent_device *devices;
241 };
242 
243 static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
244 {
245 	struct vfio_pci_fill_info *fill = data;
246 	struct iommu_group *iommu_group;
247 
248 	if (fill->cur == fill->max)
249 		return -EAGAIN; /* Something changed, try again */
250 
251 	iommu_group = iommu_group_get(&pdev->dev);
252 	if (!iommu_group)
253 		return -EPERM; /* Cannot reset non-isolated devices */
254 
255 	fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
256 	fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
257 	fill->devices[fill->cur].bus = pdev->bus->number;
258 	fill->devices[fill->cur].devfn = pdev->devfn;
259 	fill->cur++;
260 	iommu_group_put(iommu_group);
261 	return 0;
262 }
263 
264 struct vfio_pci_group_entry {
265 	struct vfio_group *group;
266 	int id;
267 };
268 
269 struct vfio_pci_group_info {
270 	int count;
271 	struct vfio_pci_group_entry *groups;
272 };
273 
274 static int vfio_pci_validate_devs(struct pci_dev *pdev, void *data)
275 {
276 	struct vfio_pci_group_info *info = data;
277 	struct iommu_group *group;
278 	int id, i;
279 
280 	group = iommu_group_get(&pdev->dev);
281 	if (!group)
282 		return -EPERM;
283 
284 	id = iommu_group_id(group);
285 
286 	for (i = 0; i < info->count; i++)
287 		if (info->groups[i].id == id)
288 			break;
289 
290 	iommu_group_put(group);
291 
292 	return (i == info->count) ? -EINVAL : 0;
293 }
294 
295 static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
296 {
297 	for (; pdev; pdev = pdev->bus->self)
298 		if (pdev->bus == slot->bus)
299 			return (pdev->slot == slot);
300 	return false;
301 }
302 
303 struct vfio_pci_walk_info {
304 	int (*fn)(struct pci_dev *, void *data);
305 	void *data;
306 	struct pci_dev *pdev;
307 	bool slot;
308 	int ret;
309 };
310 
311 static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
312 {
313 	struct vfio_pci_walk_info *walk = data;
314 
315 	if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
316 		walk->ret = walk->fn(pdev, walk->data);
317 
318 	return walk->ret;
319 }
320 
321 static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
322 					 int (*fn)(struct pci_dev *,
323 						   void *data), void *data,
324 					 bool slot)
325 {
326 	struct vfio_pci_walk_info walk = {
327 		.fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
328 	};
329 
330 	pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
331 
332 	return walk.ret;
333 }
334 
335 static long vfio_pci_ioctl(void *device_data,
336 			   unsigned int cmd, unsigned long arg)
337 {
338 	struct vfio_pci_device *vdev = device_data;
339 	unsigned long minsz;
340 
341 	if (cmd == VFIO_DEVICE_GET_INFO) {
342 		struct vfio_device_info info;
343 
344 		minsz = offsetofend(struct vfio_device_info, num_irqs);
345 
346 		if (copy_from_user(&info, (void __user *)arg, minsz))
347 			return -EFAULT;
348 
349 		if (info.argsz < minsz)
350 			return -EINVAL;
351 
352 		info.flags = VFIO_DEVICE_FLAGS_PCI;
353 
354 		if (vdev->reset_works)
355 			info.flags |= VFIO_DEVICE_FLAGS_RESET;
356 
357 		info.num_regions = VFIO_PCI_NUM_REGIONS;
358 		info.num_irqs = VFIO_PCI_NUM_IRQS;
359 
360 		return copy_to_user((void __user *)arg, &info, minsz);
361 
362 	} else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
363 		struct pci_dev *pdev = vdev->pdev;
364 		struct vfio_region_info info;
365 
366 		minsz = offsetofend(struct vfio_region_info, offset);
367 
368 		if (copy_from_user(&info, (void __user *)arg, minsz))
369 			return -EFAULT;
370 
371 		if (info.argsz < minsz)
372 			return -EINVAL;
373 
374 		switch (info.index) {
375 		case VFIO_PCI_CONFIG_REGION_INDEX:
376 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
377 			info.size = pdev->cfg_size;
378 			info.flags = VFIO_REGION_INFO_FLAG_READ |
379 				     VFIO_REGION_INFO_FLAG_WRITE;
380 			break;
381 		case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
382 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
383 			info.size = pci_resource_len(pdev, info.index);
384 			if (!info.size) {
385 				info.flags = 0;
386 				break;
387 			}
388 
389 			info.flags = VFIO_REGION_INFO_FLAG_READ |
390 				     VFIO_REGION_INFO_FLAG_WRITE;
391 			if (pci_resource_flags(pdev, info.index) &
392 			    IORESOURCE_MEM && info.size >= PAGE_SIZE)
393 				info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
394 			break;
395 		case VFIO_PCI_ROM_REGION_INDEX:
396 		{
397 			void __iomem *io;
398 			size_t size;
399 
400 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
401 			info.flags = 0;
402 
403 			/* Report the BAR size, not the ROM size */
404 			info.size = pci_resource_len(pdev, info.index);
405 			if (!info.size)
406 				break;
407 
408 			/* Is it really there? */
409 			io = pci_map_rom(pdev, &size);
410 			if (!io || !size) {
411 				info.size = 0;
412 				break;
413 			}
414 			pci_unmap_rom(pdev, io);
415 
416 			info.flags = VFIO_REGION_INFO_FLAG_READ;
417 			break;
418 		}
419 		case VFIO_PCI_VGA_REGION_INDEX:
420 			if (!vdev->has_vga)
421 				return -EINVAL;
422 
423 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
424 			info.size = 0xc0000;
425 			info.flags = VFIO_REGION_INFO_FLAG_READ |
426 				     VFIO_REGION_INFO_FLAG_WRITE;
427 
428 			break;
429 		default:
430 			return -EINVAL;
431 		}
432 
433 		return copy_to_user((void __user *)arg, &info, minsz);
434 
435 	} else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
436 		struct vfio_irq_info info;
437 
438 		minsz = offsetofend(struct vfio_irq_info, count);
439 
440 		if (copy_from_user(&info, (void __user *)arg, minsz))
441 			return -EFAULT;
442 
443 		if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
444 			return -EINVAL;
445 
446 		switch (info.index) {
447 		case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
448 			break;
449 		case VFIO_PCI_ERR_IRQ_INDEX:
450 			if (pci_is_pcie(vdev->pdev))
451 				break;
452 		/* pass thru to return error */
453 		default:
454 			return -EINVAL;
455 		}
456 
457 		info.flags = VFIO_IRQ_INFO_EVENTFD;
458 
459 		info.count = vfio_pci_get_irq_count(vdev, info.index);
460 
461 		if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
462 			info.flags |= (VFIO_IRQ_INFO_MASKABLE |
463 				       VFIO_IRQ_INFO_AUTOMASKED);
464 		else
465 			info.flags |= VFIO_IRQ_INFO_NORESIZE;
466 
467 		return copy_to_user((void __user *)arg, &info, minsz);
468 
469 	} else if (cmd == VFIO_DEVICE_SET_IRQS) {
470 		struct vfio_irq_set hdr;
471 		u8 *data = NULL;
472 		int ret = 0;
473 
474 		minsz = offsetofend(struct vfio_irq_set, count);
475 
476 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
477 			return -EFAULT;
478 
479 		if (hdr.argsz < minsz || hdr.index >= VFIO_PCI_NUM_IRQS ||
480 		    hdr.flags & ~(VFIO_IRQ_SET_DATA_TYPE_MASK |
481 				  VFIO_IRQ_SET_ACTION_TYPE_MASK))
482 			return -EINVAL;
483 
484 		if (!(hdr.flags & VFIO_IRQ_SET_DATA_NONE)) {
485 			size_t size;
486 			int max = vfio_pci_get_irq_count(vdev, hdr.index);
487 
488 			if (hdr.flags & VFIO_IRQ_SET_DATA_BOOL)
489 				size = sizeof(uint8_t);
490 			else if (hdr.flags & VFIO_IRQ_SET_DATA_EVENTFD)
491 				size = sizeof(int32_t);
492 			else
493 				return -EINVAL;
494 
495 			if (hdr.argsz - minsz < hdr.count * size ||
496 			    hdr.start >= max || hdr.start + hdr.count > max)
497 				return -EINVAL;
498 
499 			data = memdup_user((void __user *)(arg + minsz),
500 					   hdr.count * size);
501 			if (IS_ERR(data))
502 				return PTR_ERR(data);
503 		}
504 
505 		mutex_lock(&vdev->igate);
506 
507 		ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
508 					      hdr.start, hdr.count, data);
509 
510 		mutex_unlock(&vdev->igate);
511 		kfree(data);
512 
513 		return ret;
514 
515 	} else if (cmd == VFIO_DEVICE_RESET) {
516 		return vdev->reset_works ?
517 			pci_reset_function(vdev->pdev) : -EINVAL;
518 
519 	} else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
520 		struct vfio_pci_hot_reset_info hdr;
521 		struct vfio_pci_fill_info fill = { 0 };
522 		struct vfio_pci_dependent_device *devices = NULL;
523 		bool slot = false;
524 		int ret = 0;
525 
526 		minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
527 
528 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
529 			return -EFAULT;
530 
531 		if (hdr.argsz < minsz)
532 			return -EINVAL;
533 
534 		hdr.flags = 0;
535 
536 		/* Can we do a slot or bus reset or neither? */
537 		if (!pci_probe_reset_slot(vdev->pdev->slot))
538 			slot = true;
539 		else if (pci_probe_reset_bus(vdev->pdev->bus))
540 			return -ENODEV;
541 
542 		/* How many devices are affected? */
543 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
544 						    vfio_pci_count_devs,
545 						    &fill.max, slot);
546 		if (ret)
547 			return ret;
548 
549 		WARN_ON(!fill.max); /* Should always be at least one */
550 
551 		/*
552 		 * If there's enough space, fill it now, otherwise return
553 		 * -ENOSPC and the number of devices affected.
554 		 */
555 		if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
556 			ret = -ENOSPC;
557 			hdr.count = fill.max;
558 			goto reset_info_exit;
559 		}
560 
561 		devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
562 		if (!devices)
563 			return -ENOMEM;
564 
565 		fill.devices = devices;
566 
567 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
568 						    vfio_pci_fill_devs,
569 						    &fill, slot);
570 
571 		/*
572 		 * If a device was removed between counting and filling,
573 		 * we may come up short of fill.max.  If a device was
574 		 * added, we'll have a return of -EAGAIN above.
575 		 */
576 		if (!ret)
577 			hdr.count = fill.cur;
578 
579 reset_info_exit:
580 		if (copy_to_user((void __user *)arg, &hdr, minsz))
581 			ret = -EFAULT;
582 
583 		if (!ret) {
584 			if (copy_to_user((void __user *)(arg + minsz), devices,
585 					 hdr.count * sizeof(*devices)))
586 				ret = -EFAULT;
587 		}
588 
589 		kfree(devices);
590 		return ret;
591 
592 	} else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
593 		struct vfio_pci_hot_reset hdr;
594 		int32_t *group_fds;
595 		struct vfio_pci_group_entry *groups;
596 		struct vfio_pci_group_info info;
597 		bool slot = false;
598 		int i, count = 0, ret = 0;
599 
600 		minsz = offsetofend(struct vfio_pci_hot_reset, count);
601 
602 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
603 			return -EFAULT;
604 
605 		if (hdr.argsz < minsz || hdr.flags)
606 			return -EINVAL;
607 
608 		/* Can we do a slot or bus reset or neither? */
609 		if (!pci_probe_reset_slot(vdev->pdev->slot))
610 			slot = true;
611 		else if (pci_probe_reset_bus(vdev->pdev->bus))
612 			return -ENODEV;
613 
614 		/*
615 		 * We can't let userspace give us an arbitrarily large
616 		 * buffer to copy, so verify how many we think there
617 		 * could be.  Note groups can have multiple devices so
618 		 * one group per device is the max.
619 		 */
620 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
621 						    vfio_pci_count_devs,
622 						    &count, slot);
623 		if (ret)
624 			return ret;
625 
626 		/* Somewhere between 1 and count is OK */
627 		if (!hdr.count || hdr.count > count)
628 			return -EINVAL;
629 
630 		group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
631 		groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
632 		if (!group_fds || !groups) {
633 			kfree(group_fds);
634 			kfree(groups);
635 			return -ENOMEM;
636 		}
637 
638 		if (copy_from_user(group_fds, (void __user *)(arg + minsz),
639 				   hdr.count * sizeof(*group_fds))) {
640 			kfree(group_fds);
641 			kfree(groups);
642 			return -EFAULT;
643 		}
644 
645 		/*
646 		 * For each group_fd, get the group through the vfio external
647 		 * user interface and store the group and iommu ID.  This
648 		 * ensures the group is held across the reset.
649 		 */
650 		for (i = 0; i < hdr.count; i++) {
651 			struct vfio_group *group;
652 			struct fd f = fdget(group_fds[i]);
653 			if (!f.file) {
654 				ret = -EBADF;
655 				break;
656 			}
657 
658 			group = vfio_group_get_external_user(f.file);
659 			fdput(f);
660 			if (IS_ERR(group)) {
661 				ret = PTR_ERR(group);
662 				break;
663 			}
664 
665 			groups[i].group = group;
666 			groups[i].id = vfio_external_user_iommu_id(group);
667 		}
668 
669 		kfree(group_fds);
670 
671 		/* release reference to groups on error */
672 		if (ret)
673 			goto hot_reset_release;
674 
675 		info.count = hdr.count;
676 		info.groups = groups;
677 
678 		/*
679 		 * Test whether all the affected devices are contained
680 		 * by the set of groups provided by the user.
681 		 */
682 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
683 						    vfio_pci_validate_devs,
684 						    &info, slot);
685 		if (!ret)
686 			/* User has access, do the reset */
687 			ret = slot ? pci_reset_slot(vdev->pdev->slot) :
688 				     pci_reset_bus(vdev->pdev->bus);
689 
690 hot_reset_release:
691 		for (i--; i >= 0; i--)
692 			vfio_group_put_external_user(groups[i].group);
693 
694 		kfree(groups);
695 		return ret;
696 	}
697 
698 	return -ENOTTY;
699 }
700 
701 static ssize_t vfio_pci_rw(void *device_data, char __user *buf,
702 			   size_t count, loff_t *ppos, bool iswrite)
703 {
704 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
705 	struct vfio_pci_device *vdev = device_data;
706 
707 	if (index >= VFIO_PCI_NUM_REGIONS)
708 		return -EINVAL;
709 
710 	switch (index) {
711 	case VFIO_PCI_CONFIG_REGION_INDEX:
712 		return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
713 
714 	case VFIO_PCI_ROM_REGION_INDEX:
715 		if (iswrite)
716 			return -EINVAL;
717 		return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
718 
719 	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
720 		return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
721 
722 	case VFIO_PCI_VGA_REGION_INDEX:
723 		return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
724 	}
725 
726 	return -EINVAL;
727 }
728 
729 static ssize_t vfio_pci_read(void *device_data, char __user *buf,
730 			     size_t count, loff_t *ppos)
731 {
732 	if (!count)
733 		return 0;
734 
735 	return vfio_pci_rw(device_data, buf, count, ppos, false);
736 }
737 
738 static ssize_t vfio_pci_write(void *device_data, const char __user *buf,
739 			      size_t count, loff_t *ppos)
740 {
741 	if (!count)
742 		return 0;
743 
744 	return vfio_pci_rw(device_data, (char __user *)buf, count, ppos, true);
745 }
746 
747 static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma)
748 {
749 	struct vfio_pci_device *vdev = device_data;
750 	struct pci_dev *pdev = vdev->pdev;
751 	unsigned int index;
752 	u64 phys_len, req_len, pgoff, req_start;
753 	int ret;
754 
755 	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
756 
757 	if (vma->vm_end < vma->vm_start)
758 		return -EINVAL;
759 	if ((vma->vm_flags & VM_SHARED) == 0)
760 		return -EINVAL;
761 	if (index >= VFIO_PCI_ROM_REGION_INDEX)
762 		return -EINVAL;
763 	if (!(pci_resource_flags(pdev, index) & IORESOURCE_MEM))
764 		return -EINVAL;
765 
766 	phys_len = pci_resource_len(pdev, index);
767 	req_len = vma->vm_end - vma->vm_start;
768 	pgoff = vma->vm_pgoff &
769 		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
770 	req_start = pgoff << PAGE_SHIFT;
771 
772 	if (phys_len < PAGE_SIZE || req_start + req_len > phys_len)
773 		return -EINVAL;
774 
775 	if (index == vdev->msix_bar) {
776 		/*
777 		 * Disallow mmaps overlapping the MSI-X table; users don't
778 		 * get to touch this directly.  We could find somewhere
779 		 * else to map the overlap, but page granularity is only
780 		 * a recommendation, not a requirement, so the user needs
781 		 * to know which bits are real.  Requiring them to mmap
782 		 * around the table makes that clear.
783 		 */
784 
785 		/* If neither entirely above nor below, then it overlaps */
786 		if (!(req_start >= vdev->msix_offset + vdev->msix_size ||
787 		      req_start + req_len <= vdev->msix_offset))
788 			return -EINVAL;
789 	}
790 
791 	/*
792 	 * Even though we don't make use of the barmap for the mmap,
793 	 * we need to request the region and the barmap tracks that.
794 	 */
795 	if (!vdev->barmap[index]) {
796 		ret = pci_request_selected_regions(pdev,
797 						   1 << index, "vfio-pci");
798 		if (ret)
799 			return ret;
800 
801 		vdev->barmap[index] = pci_iomap(pdev, index, 0);
802 	}
803 
804 	vma->vm_private_data = vdev;
805 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
806 	vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
807 
808 	return remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
809 			       req_len, vma->vm_page_prot);
810 }
811 
812 static const struct vfio_device_ops vfio_pci_ops = {
813 	.name		= "vfio-pci",
814 	.open		= vfio_pci_open,
815 	.release	= vfio_pci_release,
816 	.ioctl		= vfio_pci_ioctl,
817 	.read		= vfio_pci_read,
818 	.write		= vfio_pci_write,
819 	.mmap		= vfio_pci_mmap,
820 };
821 
822 static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
823 {
824 	u8 type;
825 	struct vfio_pci_device *vdev;
826 	struct iommu_group *group;
827 	int ret;
828 
829 	pci_read_config_byte(pdev, PCI_HEADER_TYPE, &type);
830 	if ((type & PCI_HEADER_TYPE) != PCI_HEADER_TYPE_NORMAL)
831 		return -EINVAL;
832 
833 	group = iommu_group_get(&pdev->dev);
834 	if (!group)
835 		return -EINVAL;
836 
837 	vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
838 	if (!vdev) {
839 		iommu_group_put(group);
840 		return -ENOMEM;
841 	}
842 
843 	vdev->pdev = pdev;
844 	vdev->irq_type = VFIO_PCI_NUM_IRQS;
845 	mutex_init(&vdev->igate);
846 	spin_lock_init(&vdev->irqlock);
847 	atomic_set(&vdev->refcnt, 0);
848 
849 	ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
850 	if (ret) {
851 		iommu_group_put(group);
852 		kfree(vdev);
853 	}
854 
855 	return ret;
856 }
857 
858 static void vfio_pci_remove(struct pci_dev *pdev)
859 {
860 	struct vfio_pci_device *vdev;
861 
862 	vdev = vfio_del_group_dev(&pdev->dev);
863 	if (!vdev)
864 		return;
865 
866 	iommu_group_put(pdev->dev.iommu_group);
867 	kfree(vdev);
868 }
869 
870 static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
871 						  pci_channel_state_t state)
872 {
873 	struct vfio_pci_device *vdev;
874 	struct vfio_device *device;
875 
876 	device = vfio_device_get_from_dev(&pdev->dev);
877 	if (device == NULL)
878 		return PCI_ERS_RESULT_DISCONNECT;
879 
880 	vdev = vfio_device_data(device);
881 	if (vdev == NULL) {
882 		vfio_device_put(device);
883 		return PCI_ERS_RESULT_DISCONNECT;
884 	}
885 
886 	if (vdev->err_trigger)
887 		eventfd_signal(vdev->err_trigger, 1);
888 
889 	vfio_device_put(device);
890 
891 	return PCI_ERS_RESULT_CAN_RECOVER;
892 }
893 
894 static struct pci_error_handlers vfio_err_handlers = {
895 	.error_detected = vfio_pci_aer_err_detected,
896 };
897 
898 static struct pci_driver vfio_pci_driver = {
899 	.name		= "vfio-pci",
900 	.id_table	= NULL, /* only dynamic ids */
901 	.probe		= vfio_pci_probe,
902 	.remove		= vfio_pci_remove,
903 	.err_handler	= &vfio_err_handlers,
904 };
905 
906 static void __exit vfio_pci_cleanup(void)
907 {
908 	pci_unregister_driver(&vfio_pci_driver);
909 	vfio_pci_virqfd_exit();
910 	vfio_pci_uninit_perm_bits();
911 }
912 
913 static int __init vfio_pci_init(void)
914 {
915 	int ret;
916 
917 	/* Allocate shared config space permision data used by all devices */
918 	ret = vfio_pci_init_perm_bits();
919 	if (ret)
920 		return ret;
921 
922 	/* Start the virqfd cleanup handler */
923 	ret = vfio_pci_virqfd_init();
924 	if (ret)
925 		goto out_virqfd;
926 
927 	/* Register and scan for devices */
928 	ret = pci_register_driver(&vfio_pci_driver);
929 	if (ret)
930 		goto out_driver;
931 
932 	return 0;
933 
934 out_driver:
935 	vfio_pci_virqfd_exit();
936 out_virqfd:
937 	vfio_pci_uninit_perm_bits();
938 	return ret;
939 }
940 
941 module_init(vfio_pci_init);
942 module_exit(vfio_pci_cleanup);
943 
944 MODULE_VERSION(DRIVER_VERSION);
945 MODULE_LICENSE("GPL v2");
946 MODULE_AUTHOR(DRIVER_AUTHOR);
947 MODULE_DESCRIPTION(DRIVER_DESC);
948