xref: /openbmc/linux/drivers/vfio/pci/vfio_pci.c (revision 6774def6)
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 DEFINE_MUTEX(driver_lock);
41 
42 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev);
43 
44 static int vfio_pci_enable(struct vfio_pci_device *vdev)
45 {
46 	struct pci_dev *pdev = vdev->pdev;
47 	int ret;
48 	u16 cmd;
49 	u8 msix_pos;
50 
51 	/* Don't allow our initial saved state to include busmaster */
52 	pci_clear_master(pdev);
53 
54 	ret = pci_enable_device(pdev);
55 	if (ret)
56 		return ret;
57 
58 	vdev->reset_works = (pci_reset_function(pdev) == 0);
59 	pci_save_state(pdev);
60 	vdev->pci_saved_state = pci_store_saved_state(pdev);
61 	if (!vdev->pci_saved_state)
62 		pr_debug("%s: Couldn't store %s saved state\n",
63 			 __func__, dev_name(&pdev->dev));
64 
65 	ret = vfio_config_init(vdev);
66 	if (ret) {
67 		kfree(vdev->pci_saved_state);
68 		vdev->pci_saved_state = NULL;
69 		pci_disable_device(pdev);
70 		return ret;
71 	}
72 
73 	if (likely(!nointxmask))
74 		vdev->pci_2_3 = pci_intx_mask_supported(pdev);
75 
76 	pci_read_config_word(pdev, PCI_COMMAND, &cmd);
77 	if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
78 		cmd &= ~PCI_COMMAND_INTX_DISABLE;
79 		pci_write_config_word(pdev, PCI_COMMAND, cmd);
80 	}
81 
82 	msix_pos = pdev->msix_cap;
83 	if (msix_pos) {
84 		u16 flags;
85 		u32 table;
86 
87 		pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
88 		pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
89 
90 		vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
91 		vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
92 		vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
93 	} else
94 		vdev->msix_bar = 0xFF;
95 
96 #ifdef CONFIG_VFIO_PCI_VGA
97 	if ((pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA)
98 		vdev->has_vga = true;
99 #endif
100 
101 	return 0;
102 }
103 
104 static void vfio_pci_disable(struct vfio_pci_device *vdev)
105 {
106 	struct pci_dev *pdev = vdev->pdev;
107 	int bar;
108 
109 	/* Stop the device from further DMA */
110 	pci_clear_master(pdev);
111 
112 	vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
113 				VFIO_IRQ_SET_ACTION_TRIGGER,
114 				vdev->irq_type, 0, 0, NULL);
115 
116 	vdev->virq_disabled = false;
117 
118 	vfio_config_free(vdev);
119 
120 	for (bar = PCI_STD_RESOURCES; bar <= PCI_STD_RESOURCE_END; bar++) {
121 		if (!vdev->barmap[bar])
122 			continue;
123 		pci_iounmap(pdev, vdev->barmap[bar]);
124 		pci_release_selected_regions(pdev, 1 << bar);
125 		vdev->barmap[bar] = NULL;
126 	}
127 
128 	vdev->needs_reset = true;
129 
130 	/*
131 	 * If we have saved state, restore it.  If we can reset the device,
132 	 * even better.  Resetting with current state seems better than
133 	 * nothing, but saving and restoring current state without reset
134 	 * is just busy work.
135 	 */
136 	if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
137 		pr_info("%s: Couldn't reload %s saved state\n",
138 			__func__, dev_name(&pdev->dev));
139 
140 		if (!vdev->reset_works)
141 			goto out;
142 
143 		pci_save_state(pdev);
144 	}
145 
146 	/*
147 	 * Disable INTx and MSI, presumably to avoid spurious interrupts
148 	 * during reset.  Stolen from pci_reset_function()
149 	 */
150 	pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
151 
152 	/*
153 	 * Try to reset the device.  The success of this is dependent on
154 	 * being able to lock the device, which is not always possible.
155 	 */
156 	if (vdev->reset_works) {
157 		int ret = pci_try_reset_function(pdev);
158 		if (ret)
159 			pr_warn("%s: Failed to reset device %s (%d)\n",
160 				__func__, dev_name(&pdev->dev), ret);
161 		else
162 			vdev->needs_reset = false;
163 	}
164 
165 	pci_restore_state(pdev);
166 out:
167 	pci_disable_device(pdev);
168 
169 	vfio_pci_try_bus_reset(vdev);
170 }
171 
172 static void vfio_pci_release(void *device_data)
173 {
174 	struct vfio_pci_device *vdev = device_data;
175 
176 	mutex_lock(&driver_lock);
177 
178 	if (!(--vdev->refcnt)) {
179 		vfio_spapr_pci_eeh_release(vdev->pdev);
180 		vfio_pci_disable(vdev);
181 	}
182 
183 	mutex_unlock(&driver_lock);
184 
185 	module_put(THIS_MODULE);
186 }
187 
188 static int vfio_pci_open(void *device_data)
189 {
190 	struct vfio_pci_device *vdev = device_data;
191 	int ret = 0;
192 
193 	if (!try_module_get(THIS_MODULE))
194 		return -ENODEV;
195 
196 	mutex_lock(&driver_lock);
197 
198 	if (!vdev->refcnt) {
199 		ret = vfio_pci_enable(vdev);
200 		if (ret)
201 			goto error;
202 
203 		vfio_spapr_pci_eeh_open(vdev->pdev);
204 	}
205 	vdev->refcnt++;
206 error:
207 	mutex_unlock(&driver_lock);
208 	if (ret)
209 		module_put(THIS_MODULE);
210 	return ret;
211 }
212 
213 static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
214 {
215 	if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
216 		u8 pin;
217 		pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
218 		if (pin)
219 			return 1;
220 
221 	} else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
222 		u8 pos;
223 		u16 flags;
224 
225 		pos = vdev->pdev->msi_cap;
226 		if (pos) {
227 			pci_read_config_word(vdev->pdev,
228 					     pos + PCI_MSI_FLAGS, &flags);
229 			return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
230 		}
231 	} else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
232 		u8 pos;
233 		u16 flags;
234 
235 		pos = vdev->pdev->msix_cap;
236 		if (pos) {
237 			pci_read_config_word(vdev->pdev,
238 					     pos + PCI_MSIX_FLAGS, &flags);
239 
240 			return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
241 		}
242 	} else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX)
243 		if (pci_is_pcie(vdev->pdev))
244 			return 1;
245 
246 	return 0;
247 }
248 
249 static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
250 {
251 	(*(int *)data)++;
252 	return 0;
253 }
254 
255 struct vfio_pci_fill_info {
256 	int max;
257 	int cur;
258 	struct vfio_pci_dependent_device *devices;
259 };
260 
261 static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
262 {
263 	struct vfio_pci_fill_info *fill = data;
264 	struct iommu_group *iommu_group;
265 
266 	if (fill->cur == fill->max)
267 		return -EAGAIN; /* Something changed, try again */
268 
269 	iommu_group = iommu_group_get(&pdev->dev);
270 	if (!iommu_group)
271 		return -EPERM; /* Cannot reset non-isolated devices */
272 
273 	fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
274 	fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
275 	fill->devices[fill->cur].bus = pdev->bus->number;
276 	fill->devices[fill->cur].devfn = pdev->devfn;
277 	fill->cur++;
278 	iommu_group_put(iommu_group);
279 	return 0;
280 }
281 
282 struct vfio_pci_group_entry {
283 	struct vfio_group *group;
284 	int id;
285 };
286 
287 struct vfio_pci_group_info {
288 	int count;
289 	struct vfio_pci_group_entry *groups;
290 };
291 
292 static int vfio_pci_validate_devs(struct pci_dev *pdev, void *data)
293 {
294 	struct vfio_pci_group_info *info = data;
295 	struct iommu_group *group;
296 	int id, i;
297 
298 	group = iommu_group_get(&pdev->dev);
299 	if (!group)
300 		return -EPERM;
301 
302 	id = iommu_group_id(group);
303 
304 	for (i = 0; i < info->count; i++)
305 		if (info->groups[i].id == id)
306 			break;
307 
308 	iommu_group_put(group);
309 
310 	return (i == info->count) ? -EINVAL : 0;
311 }
312 
313 static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
314 {
315 	for (; pdev; pdev = pdev->bus->self)
316 		if (pdev->bus == slot->bus)
317 			return (pdev->slot == slot);
318 	return false;
319 }
320 
321 struct vfio_pci_walk_info {
322 	int (*fn)(struct pci_dev *, void *data);
323 	void *data;
324 	struct pci_dev *pdev;
325 	bool slot;
326 	int ret;
327 };
328 
329 static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
330 {
331 	struct vfio_pci_walk_info *walk = data;
332 
333 	if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
334 		walk->ret = walk->fn(pdev, walk->data);
335 
336 	return walk->ret;
337 }
338 
339 static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
340 					 int (*fn)(struct pci_dev *,
341 						   void *data), void *data,
342 					 bool slot)
343 {
344 	struct vfio_pci_walk_info walk = {
345 		.fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
346 	};
347 
348 	pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
349 
350 	return walk.ret;
351 }
352 
353 static long vfio_pci_ioctl(void *device_data,
354 			   unsigned int cmd, unsigned long arg)
355 {
356 	struct vfio_pci_device *vdev = device_data;
357 	unsigned long minsz;
358 
359 	if (cmd == VFIO_DEVICE_GET_INFO) {
360 		struct vfio_device_info info;
361 
362 		minsz = offsetofend(struct vfio_device_info, num_irqs);
363 
364 		if (copy_from_user(&info, (void __user *)arg, minsz))
365 			return -EFAULT;
366 
367 		if (info.argsz < minsz)
368 			return -EINVAL;
369 
370 		info.flags = VFIO_DEVICE_FLAGS_PCI;
371 
372 		if (vdev->reset_works)
373 			info.flags |= VFIO_DEVICE_FLAGS_RESET;
374 
375 		info.num_regions = VFIO_PCI_NUM_REGIONS;
376 		info.num_irqs = VFIO_PCI_NUM_IRQS;
377 
378 		return copy_to_user((void __user *)arg, &info, minsz);
379 
380 	} else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
381 		struct pci_dev *pdev = vdev->pdev;
382 		struct vfio_region_info info;
383 
384 		minsz = offsetofend(struct vfio_region_info, offset);
385 
386 		if (copy_from_user(&info, (void __user *)arg, minsz))
387 			return -EFAULT;
388 
389 		if (info.argsz < minsz)
390 			return -EINVAL;
391 
392 		switch (info.index) {
393 		case VFIO_PCI_CONFIG_REGION_INDEX:
394 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
395 			info.size = pdev->cfg_size;
396 			info.flags = VFIO_REGION_INFO_FLAG_READ |
397 				     VFIO_REGION_INFO_FLAG_WRITE;
398 			break;
399 		case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
400 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
401 			info.size = pci_resource_len(pdev, info.index);
402 			if (!info.size) {
403 				info.flags = 0;
404 				break;
405 			}
406 
407 			info.flags = VFIO_REGION_INFO_FLAG_READ |
408 				     VFIO_REGION_INFO_FLAG_WRITE;
409 			if (pci_resource_flags(pdev, info.index) &
410 			    IORESOURCE_MEM && info.size >= PAGE_SIZE)
411 				info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
412 			break;
413 		case VFIO_PCI_ROM_REGION_INDEX:
414 		{
415 			void __iomem *io;
416 			size_t size;
417 
418 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
419 			info.flags = 0;
420 
421 			/* Report the BAR size, not the ROM size */
422 			info.size = pci_resource_len(pdev, info.index);
423 			if (!info.size)
424 				break;
425 
426 			/* Is it really there? */
427 			io = pci_map_rom(pdev, &size);
428 			if (!io || !size) {
429 				info.size = 0;
430 				break;
431 			}
432 			pci_unmap_rom(pdev, io);
433 
434 			info.flags = VFIO_REGION_INFO_FLAG_READ;
435 			break;
436 		}
437 		case VFIO_PCI_VGA_REGION_INDEX:
438 			if (!vdev->has_vga)
439 				return -EINVAL;
440 
441 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
442 			info.size = 0xc0000;
443 			info.flags = VFIO_REGION_INFO_FLAG_READ |
444 				     VFIO_REGION_INFO_FLAG_WRITE;
445 
446 			break;
447 		default:
448 			return -EINVAL;
449 		}
450 
451 		return copy_to_user((void __user *)arg, &info, minsz);
452 
453 	} else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
454 		struct vfio_irq_info info;
455 
456 		minsz = offsetofend(struct vfio_irq_info, count);
457 
458 		if (copy_from_user(&info, (void __user *)arg, minsz))
459 			return -EFAULT;
460 
461 		if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
462 			return -EINVAL;
463 
464 		switch (info.index) {
465 		case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
466 			break;
467 		case VFIO_PCI_ERR_IRQ_INDEX:
468 			if (pci_is_pcie(vdev->pdev))
469 				break;
470 		/* pass thru to return error */
471 		default:
472 			return -EINVAL;
473 		}
474 
475 		info.flags = VFIO_IRQ_INFO_EVENTFD;
476 
477 		info.count = vfio_pci_get_irq_count(vdev, info.index);
478 
479 		if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
480 			info.flags |= (VFIO_IRQ_INFO_MASKABLE |
481 				       VFIO_IRQ_INFO_AUTOMASKED);
482 		else
483 			info.flags |= VFIO_IRQ_INFO_NORESIZE;
484 
485 		return copy_to_user((void __user *)arg, &info, minsz);
486 
487 	} else if (cmd == VFIO_DEVICE_SET_IRQS) {
488 		struct vfio_irq_set hdr;
489 		u8 *data = NULL;
490 		int ret = 0;
491 
492 		minsz = offsetofend(struct vfio_irq_set, count);
493 
494 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
495 			return -EFAULT;
496 
497 		if (hdr.argsz < minsz || hdr.index >= VFIO_PCI_NUM_IRQS ||
498 		    hdr.flags & ~(VFIO_IRQ_SET_DATA_TYPE_MASK |
499 				  VFIO_IRQ_SET_ACTION_TYPE_MASK))
500 			return -EINVAL;
501 
502 		if (!(hdr.flags & VFIO_IRQ_SET_DATA_NONE)) {
503 			size_t size;
504 			int max = vfio_pci_get_irq_count(vdev, hdr.index);
505 
506 			if (hdr.flags & VFIO_IRQ_SET_DATA_BOOL)
507 				size = sizeof(uint8_t);
508 			else if (hdr.flags & VFIO_IRQ_SET_DATA_EVENTFD)
509 				size = sizeof(int32_t);
510 			else
511 				return -EINVAL;
512 
513 			if (hdr.argsz - minsz < hdr.count * size ||
514 			    hdr.start >= max || hdr.start + hdr.count > max)
515 				return -EINVAL;
516 
517 			data = memdup_user((void __user *)(arg + minsz),
518 					   hdr.count * size);
519 			if (IS_ERR(data))
520 				return PTR_ERR(data);
521 		}
522 
523 		mutex_lock(&vdev->igate);
524 
525 		ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
526 					      hdr.start, hdr.count, data);
527 
528 		mutex_unlock(&vdev->igate);
529 		kfree(data);
530 
531 		return ret;
532 
533 	} else if (cmd == VFIO_DEVICE_RESET) {
534 		return vdev->reset_works ?
535 			pci_try_reset_function(vdev->pdev) : -EINVAL;
536 
537 	} else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
538 		struct vfio_pci_hot_reset_info hdr;
539 		struct vfio_pci_fill_info fill = { 0 };
540 		struct vfio_pci_dependent_device *devices = NULL;
541 		bool slot = false;
542 		int ret = 0;
543 
544 		minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
545 
546 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
547 			return -EFAULT;
548 
549 		if (hdr.argsz < minsz)
550 			return -EINVAL;
551 
552 		hdr.flags = 0;
553 
554 		/* Can we do a slot or bus reset or neither? */
555 		if (!pci_probe_reset_slot(vdev->pdev->slot))
556 			slot = true;
557 		else if (pci_probe_reset_bus(vdev->pdev->bus))
558 			return -ENODEV;
559 
560 		/* How many devices are affected? */
561 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
562 						    vfio_pci_count_devs,
563 						    &fill.max, slot);
564 		if (ret)
565 			return ret;
566 
567 		WARN_ON(!fill.max); /* Should always be at least one */
568 
569 		/*
570 		 * If there's enough space, fill it now, otherwise return
571 		 * -ENOSPC and the number of devices affected.
572 		 */
573 		if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
574 			ret = -ENOSPC;
575 			hdr.count = fill.max;
576 			goto reset_info_exit;
577 		}
578 
579 		devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
580 		if (!devices)
581 			return -ENOMEM;
582 
583 		fill.devices = devices;
584 
585 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
586 						    vfio_pci_fill_devs,
587 						    &fill, slot);
588 
589 		/*
590 		 * If a device was removed between counting and filling,
591 		 * we may come up short of fill.max.  If a device was
592 		 * added, we'll have a return of -EAGAIN above.
593 		 */
594 		if (!ret)
595 			hdr.count = fill.cur;
596 
597 reset_info_exit:
598 		if (copy_to_user((void __user *)arg, &hdr, minsz))
599 			ret = -EFAULT;
600 
601 		if (!ret) {
602 			if (copy_to_user((void __user *)(arg + minsz), devices,
603 					 hdr.count * sizeof(*devices)))
604 				ret = -EFAULT;
605 		}
606 
607 		kfree(devices);
608 		return ret;
609 
610 	} else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
611 		struct vfio_pci_hot_reset hdr;
612 		int32_t *group_fds;
613 		struct vfio_pci_group_entry *groups;
614 		struct vfio_pci_group_info info;
615 		bool slot = false;
616 		int i, count = 0, ret = 0;
617 
618 		minsz = offsetofend(struct vfio_pci_hot_reset, count);
619 
620 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
621 			return -EFAULT;
622 
623 		if (hdr.argsz < minsz || hdr.flags)
624 			return -EINVAL;
625 
626 		/* Can we do a slot or bus reset or neither? */
627 		if (!pci_probe_reset_slot(vdev->pdev->slot))
628 			slot = true;
629 		else if (pci_probe_reset_bus(vdev->pdev->bus))
630 			return -ENODEV;
631 
632 		/*
633 		 * We can't let userspace give us an arbitrarily large
634 		 * buffer to copy, so verify how many we think there
635 		 * could be.  Note groups can have multiple devices so
636 		 * one group per device is the max.
637 		 */
638 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
639 						    vfio_pci_count_devs,
640 						    &count, slot);
641 		if (ret)
642 			return ret;
643 
644 		/* Somewhere between 1 and count is OK */
645 		if (!hdr.count || hdr.count > count)
646 			return -EINVAL;
647 
648 		group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
649 		groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
650 		if (!group_fds || !groups) {
651 			kfree(group_fds);
652 			kfree(groups);
653 			return -ENOMEM;
654 		}
655 
656 		if (copy_from_user(group_fds, (void __user *)(arg + minsz),
657 				   hdr.count * sizeof(*group_fds))) {
658 			kfree(group_fds);
659 			kfree(groups);
660 			return -EFAULT;
661 		}
662 
663 		/*
664 		 * For each group_fd, get the group through the vfio external
665 		 * user interface and store the group and iommu ID.  This
666 		 * ensures the group is held across the reset.
667 		 */
668 		for (i = 0; i < hdr.count; i++) {
669 			struct vfio_group *group;
670 			struct fd f = fdget(group_fds[i]);
671 			if (!f.file) {
672 				ret = -EBADF;
673 				break;
674 			}
675 
676 			group = vfio_group_get_external_user(f.file);
677 			fdput(f);
678 			if (IS_ERR(group)) {
679 				ret = PTR_ERR(group);
680 				break;
681 			}
682 
683 			groups[i].group = group;
684 			groups[i].id = vfio_external_user_iommu_id(group);
685 		}
686 
687 		kfree(group_fds);
688 
689 		/* release reference to groups on error */
690 		if (ret)
691 			goto hot_reset_release;
692 
693 		info.count = hdr.count;
694 		info.groups = groups;
695 
696 		/*
697 		 * Test whether all the affected devices are contained
698 		 * by the set of groups provided by the user.
699 		 */
700 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
701 						    vfio_pci_validate_devs,
702 						    &info, slot);
703 		if (!ret)
704 			/* User has access, do the reset */
705 			ret = slot ? pci_try_reset_slot(vdev->pdev->slot) :
706 				     pci_try_reset_bus(vdev->pdev->bus);
707 
708 hot_reset_release:
709 		for (i--; i >= 0; i--)
710 			vfio_group_put_external_user(groups[i].group);
711 
712 		kfree(groups);
713 		return ret;
714 	}
715 
716 	return -ENOTTY;
717 }
718 
719 static ssize_t vfio_pci_rw(void *device_data, char __user *buf,
720 			   size_t count, loff_t *ppos, bool iswrite)
721 {
722 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
723 	struct vfio_pci_device *vdev = device_data;
724 
725 	if (index >= VFIO_PCI_NUM_REGIONS)
726 		return -EINVAL;
727 
728 	switch (index) {
729 	case VFIO_PCI_CONFIG_REGION_INDEX:
730 		return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
731 
732 	case VFIO_PCI_ROM_REGION_INDEX:
733 		if (iswrite)
734 			return -EINVAL;
735 		return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
736 
737 	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
738 		return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
739 
740 	case VFIO_PCI_VGA_REGION_INDEX:
741 		return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
742 	}
743 
744 	return -EINVAL;
745 }
746 
747 static ssize_t vfio_pci_read(void *device_data, char __user *buf,
748 			     size_t count, loff_t *ppos)
749 {
750 	if (!count)
751 		return 0;
752 
753 	return vfio_pci_rw(device_data, buf, count, ppos, false);
754 }
755 
756 static ssize_t vfio_pci_write(void *device_data, const char __user *buf,
757 			      size_t count, loff_t *ppos)
758 {
759 	if (!count)
760 		return 0;
761 
762 	return vfio_pci_rw(device_data, (char __user *)buf, count, ppos, true);
763 }
764 
765 static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma)
766 {
767 	struct vfio_pci_device *vdev = device_data;
768 	struct pci_dev *pdev = vdev->pdev;
769 	unsigned int index;
770 	u64 phys_len, req_len, pgoff, req_start;
771 	int ret;
772 
773 	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
774 
775 	if (vma->vm_end < vma->vm_start)
776 		return -EINVAL;
777 	if ((vma->vm_flags & VM_SHARED) == 0)
778 		return -EINVAL;
779 	if (index >= VFIO_PCI_ROM_REGION_INDEX)
780 		return -EINVAL;
781 	if (!(pci_resource_flags(pdev, index) & IORESOURCE_MEM))
782 		return -EINVAL;
783 
784 	phys_len = pci_resource_len(pdev, index);
785 	req_len = vma->vm_end - vma->vm_start;
786 	pgoff = vma->vm_pgoff &
787 		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
788 	req_start = pgoff << PAGE_SHIFT;
789 
790 	if (phys_len < PAGE_SIZE || req_start + req_len > phys_len)
791 		return -EINVAL;
792 
793 	if (index == vdev->msix_bar) {
794 		/*
795 		 * Disallow mmaps overlapping the MSI-X table; users don't
796 		 * get to touch this directly.  We could find somewhere
797 		 * else to map the overlap, but page granularity is only
798 		 * a recommendation, not a requirement, so the user needs
799 		 * to know which bits are real.  Requiring them to mmap
800 		 * around the table makes that clear.
801 		 */
802 
803 		/* If neither entirely above nor below, then it overlaps */
804 		if (!(req_start >= vdev->msix_offset + vdev->msix_size ||
805 		      req_start + req_len <= vdev->msix_offset))
806 			return -EINVAL;
807 	}
808 
809 	/*
810 	 * Even though we don't make use of the barmap for the mmap,
811 	 * we need to request the region and the barmap tracks that.
812 	 */
813 	if (!vdev->barmap[index]) {
814 		ret = pci_request_selected_regions(pdev,
815 						   1 << index, "vfio-pci");
816 		if (ret)
817 			return ret;
818 
819 		vdev->barmap[index] = pci_iomap(pdev, index, 0);
820 	}
821 
822 	vma->vm_private_data = vdev;
823 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
824 	vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
825 
826 	return remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
827 			       req_len, vma->vm_page_prot);
828 }
829 
830 static const struct vfio_device_ops vfio_pci_ops = {
831 	.name		= "vfio-pci",
832 	.open		= vfio_pci_open,
833 	.release	= vfio_pci_release,
834 	.ioctl		= vfio_pci_ioctl,
835 	.read		= vfio_pci_read,
836 	.write		= vfio_pci_write,
837 	.mmap		= vfio_pci_mmap,
838 };
839 
840 static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
841 {
842 	u8 type;
843 	struct vfio_pci_device *vdev;
844 	struct iommu_group *group;
845 	int ret;
846 
847 	pci_read_config_byte(pdev, PCI_HEADER_TYPE, &type);
848 	if ((type & PCI_HEADER_TYPE) != PCI_HEADER_TYPE_NORMAL)
849 		return -EINVAL;
850 
851 	group = iommu_group_get(&pdev->dev);
852 	if (!group)
853 		return -EINVAL;
854 
855 	vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
856 	if (!vdev) {
857 		iommu_group_put(group);
858 		return -ENOMEM;
859 	}
860 
861 	vdev->pdev = pdev;
862 	vdev->irq_type = VFIO_PCI_NUM_IRQS;
863 	mutex_init(&vdev->igate);
864 	spin_lock_init(&vdev->irqlock);
865 
866 	ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
867 	if (ret) {
868 		iommu_group_put(group);
869 		kfree(vdev);
870 	}
871 
872 	return ret;
873 }
874 
875 static void vfio_pci_remove(struct pci_dev *pdev)
876 {
877 	struct vfio_pci_device *vdev;
878 
879 	vdev = vfio_del_group_dev(&pdev->dev);
880 	if (vdev) {
881 		iommu_group_put(pdev->dev.iommu_group);
882 		kfree(vdev);
883 	}
884 }
885 
886 static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
887 						  pci_channel_state_t state)
888 {
889 	struct vfio_pci_device *vdev;
890 	struct vfio_device *device;
891 
892 	device = vfio_device_get_from_dev(&pdev->dev);
893 	if (device == NULL)
894 		return PCI_ERS_RESULT_DISCONNECT;
895 
896 	vdev = vfio_device_data(device);
897 	if (vdev == NULL) {
898 		vfio_device_put(device);
899 		return PCI_ERS_RESULT_DISCONNECT;
900 	}
901 
902 	mutex_lock(&vdev->igate);
903 
904 	if (vdev->err_trigger)
905 		eventfd_signal(vdev->err_trigger, 1);
906 
907 	mutex_unlock(&vdev->igate);
908 
909 	vfio_device_put(device);
910 
911 	return PCI_ERS_RESULT_CAN_RECOVER;
912 }
913 
914 static struct pci_error_handlers vfio_err_handlers = {
915 	.error_detected = vfio_pci_aer_err_detected,
916 };
917 
918 static struct pci_driver vfio_pci_driver = {
919 	.name		= "vfio-pci",
920 	.id_table	= NULL, /* only dynamic ids */
921 	.probe		= vfio_pci_probe,
922 	.remove		= vfio_pci_remove,
923 	.err_handler	= &vfio_err_handlers,
924 };
925 
926 struct vfio_devices {
927 	struct vfio_device **devices;
928 	int cur_index;
929 	int max_index;
930 };
931 
932 static int vfio_pci_get_devs(struct pci_dev *pdev, void *data)
933 {
934 	struct vfio_devices *devs = data;
935 	struct pci_driver *pci_drv = ACCESS_ONCE(pdev->driver);
936 
937 	if (pci_drv != &vfio_pci_driver)
938 		return -EBUSY;
939 
940 	if (devs->cur_index == devs->max_index)
941 		return -ENOSPC;
942 
943 	devs->devices[devs->cur_index] = vfio_device_get_from_dev(&pdev->dev);
944 	if (!devs->devices[devs->cur_index])
945 		return -EINVAL;
946 
947 	devs->cur_index++;
948 	return 0;
949 }
950 
951 /*
952  * Attempt to do a bus/slot reset if there are devices affected by a reset for
953  * this device that are needs_reset and all of the affected devices are unused
954  * (!refcnt).  Callers are required to hold driver_lock when calling this to
955  * prevent device opens and concurrent bus reset attempts.  We prevent device
956  * unbinds by acquiring and holding a reference to the vfio_device.
957  *
958  * NB: vfio-core considers a group to be viable even if some devices are
959  * bound to drivers like pci-stub or pcieport.  Here we require all devices
960  * to be bound to vfio_pci since that's the only way we can be sure they
961  * stay put.
962  */
963 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev)
964 {
965 	struct vfio_devices devs = { .cur_index = 0 };
966 	int i = 0, ret = -EINVAL;
967 	bool needs_reset = false, slot = false;
968 	struct vfio_pci_device *tmp;
969 
970 	if (!pci_probe_reset_slot(vdev->pdev->slot))
971 		slot = true;
972 	else if (pci_probe_reset_bus(vdev->pdev->bus))
973 		return;
974 
975 	if (vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
976 					  &i, slot) || !i)
977 		return;
978 
979 	devs.max_index = i;
980 	devs.devices = kcalloc(i, sizeof(struct vfio_device *), GFP_KERNEL);
981 	if (!devs.devices)
982 		return;
983 
984 	if (vfio_pci_for_each_slot_or_bus(vdev->pdev,
985 					  vfio_pci_get_devs, &devs, slot))
986 		goto put_devs;
987 
988 	for (i = 0; i < devs.cur_index; i++) {
989 		tmp = vfio_device_data(devs.devices[i]);
990 		if (tmp->needs_reset)
991 			needs_reset = true;
992 		if (tmp->refcnt)
993 			goto put_devs;
994 	}
995 
996 	if (needs_reset)
997 		ret = slot ? pci_try_reset_slot(vdev->pdev->slot) :
998 			     pci_try_reset_bus(vdev->pdev->bus);
999 
1000 put_devs:
1001 	for (i = 0; i < devs.cur_index; i++) {
1002 		if (!ret) {
1003 			tmp = vfio_device_data(devs.devices[i]);
1004 			tmp->needs_reset = false;
1005 		}
1006 		vfio_device_put(devs.devices[i]);
1007 	}
1008 
1009 	kfree(devs.devices);
1010 }
1011 
1012 static void __exit vfio_pci_cleanup(void)
1013 {
1014 	pci_unregister_driver(&vfio_pci_driver);
1015 	vfio_pci_virqfd_exit();
1016 	vfio_pci_uninit_perm_bits();
1017 }
1018 
1019 static int __init vfio_pci_init(void)
1020 {
1021 	int ret;
1022 
1023 	/* Allocate shared config space permision data used by all devices */
1024 	ret = vfio_pci_init_perm_bits();
1025 	if (ret)
1026 		return ret;
1027 
1028 	/* Start the virqfd cleanup handler */
1029 	ret = vfio_pci_virqfd_init();
1030 	if (ret)
1031 		goto out_virqfd;
1032 
1033 	/* Register and scan for devices */
1034 	ret = pci_register_driver(&vfio_pci_driver);
1035 	if (ret)
1036 		goto out_driver;
1037 
1038 	return 0;
1039 
1040 out_driver:
1041 	vfio_pci_virqfd_exit();
1042 out_virqfd:
1043 	vfio_pci_uninit_perm_bits();
1044 	return ret;
1045 }
1046 
1047 module_init(vfio_pci_init);
1048 module_exit(vfio_pci_cleanup);
1049 
1050 MODULE_VERSION(DRIVER_VERSION);
1051 MODULE_LICENSE("GPL v2");
1052 MODULE_AUTHOR(DRIVER_AUTHOR);
1053 MODULE_DESCRIPTION(DRIVER_DESC);
1054