xref: /linux/drivers/vfio/pci/vfio_pci.c (revision 4e0385dd7469d933c4adf84a617f872ca547aa07)
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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15 
16 #include <linux/device.h>
17 #include <linux/eventfd.h>
18 #include <linux/file.h>
19 #include <linux/interrupt.h>
20 #include <linux/iommu.h>
21 #include <linux/module.h>
22 #include <linux/mutex.h>
23 #include <linux/notifier.h>
24 #include <linux/pci.h>
25 #include <linux/pm_runtime.h>
26 #include <linux/slab.h>
27 #include <linux/types.h>
28 #include <linux/uaccess.h>
29 #include <linux/vfio.h>
30 #include <linux/vgaarb.h>
31 
32 #include "vfio_pci_private.h"
33 
34 #define DRIVER_VERSION  "0.2"
35 #define DRIVER_AUTHOR   "Alex Williamson <alex.williamson@redhat.com>"
36 #define DRIVER_DESC     "VFIO PCI - User Level meta-driver"
37 
38 static char ids[1024] __initdata;
39 module_param_string(ids, ids, sizeof(ids), 0);
40 MODULE_PARM_DESC(ids, "Initial PCI IDs to add to the vfio driver, format is \"vendor:device[:subvendor[:subdevice[:class[:class_mask]]]]\" and multiple comma separated entries can be specified");
41 
42 static bool nointxmask;
43 module_param_named(nointxmask, nointxmask, bool, S_IRUGO | S_IWUSR);
44 MODULE_PARM_DESC(nointxmask,
45 		  "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.");
46 
47 #ifdef CONFIG_VFIO_PCI_VGA
48 static bool disable_vga;
49 module_param(disable_vga, bool, S_IRUGO);
50 MODULE_PARM_DESC(disable_vga, "Disable VGA resource access through vfio-pci");
51 #endif
52 
53 static bool disable_idle_d3;
54 module_param(disable_idle_d3, bool, S_IRUGO | S_IWUSR);
55 MODULE_PARM_DESC(disable_idle_d3,
56 		 "Disable using the PCI D3 low power state for idle, unused devices");
57 
58 static DEFINE_MUTEX(driver_lock);
59 
60 static inline bool vfio_vga_disabled(void)
61 {
62 #ifdef CONFIG_VFIO_PCI_VGA
63 	return disable_vga;
64 #else
65 	return true;
66 #endif
67 }
68 
69 /*
70  * Our VGA arbiter participation is limited since we don't know anything
71  * about the device itself.  However, if the device is the only VGA device
72  * downstream of a bridge and VFIO VGA support is disabled, then we can
73  * safely return legacy VGA IO and memory as not decoded since the user
74  * has no way to get to it and routing can be disabled externally at the
75  * bridge.
76  */
77 static unsigned int vfio_pci_set_vga_decode(void *opaque, bool single_vga)
78 {
79 	struct vfio_pci_device *vdev = opaque;
80 	struct pci_dev *tmp = NULL, *pdev = vdev->pdev;
81 	unsigned char max_busnr;
82 	unsigned int decodes;
83 
84 	if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
85 		return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
86 		       VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
87 
88 	max_busnr = pci_bus_max_busnr(pdev->bus);
89 	decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
90 
91 	while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
92 		if (tmp == pdev ||
93 		    pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
94 		    pci_is_root_bus(tmp->bus))
95 			continue;
96 
97 		if (tmp->bus->number >= pdev->bus->number &&
98 		    tmp->bus->number <= max_busnr) {
99 			pci_dev_put(tmp);
100 			decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
101 			break;
102 		}
103 	}
104 
105 	return decodes;
106 }
107 
108 static inline bool vfio_pci_is_vga(struct pci_dev *pdev)
109 {
110 	return (pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA;
111 }
112 
113 static void vfio_pci_probe_mmaps(struct vfio_pci_device *vdev)
114 {
115 	struct resource *res;
116 	int bar;
117 	struct vfio_pci_dummy_resource *dummy_res;
118 
119 	INIT_LIST_HEAD(&vdev->dummy_resources_list);
120 
121 	for (bar = PCI_STD_RESOURCES; bar <= PCI_STD_RESOURCE_END; bar++) {
122 		res = vdev->pdev->resource + bar;
123 
124 		if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
125 			goto no_mmap;
126 
127 		if (!(res->flags & IORESOURCE_MEM))
128 			goto no_mmap;
129 
130 		/*
131 		 * The PCI core shouldn't set up a resource with a
132 		 * type but zero size. But there may be bugs that
133 		 * cause us to do that.
134 		 */
135 		if (!resource_size(res))
136 			goto no_mmap;
137 
138 		if (resource_size(res) >= PAGE_SIZE) {
139 			vdev->bar_mmap_supported[bar] = true;
140 			continue;
141 		}
142 
143 		if (!(res->start & ~PAGE_MASK)) {
144 			/*
145 			 * Add a dummy resource to reserve the remainder
146 			 * of the exclusive page in case that hot-add
147 			 * device's bar is assigned into it.
148 			 */
149 			dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
150 			if (dummy_res == NULL)
151 				goto no_mmap;
152 
153 			dummy_res->resource.name = "vfio sub-page reserved";
154 			dummy_res->resource.start = res->end + 1;
155 			dummy_res->resource.end = res->start + PAGE_SIZE - 1;
156 			dummy_res->resource.flags = res->flags;
157 			if (request_resource(res->parent,
158 						&dummy_res->resource)) {
159 				kfree(dummy_res);
160 				goto no_mmap;
161 			}
162 			dummy_res->index = bar;
163 			list_add(&dummy_res->res_next,
164 					&vdev->dummy_resources_list);
165 			vdev->bar_mmap_supported[bar] = true;
166 			continue;
167 		}
168 		/*
169 		 * Here we don't handle the case when the BAR is not page
170 		 * aligned because we can't expect the BAR will be
171 		 * assigned into the same location in a page in guest
172 		 * when we passthrough the BAR. And it's hard to access
173 		 * this BAR in userspace because we have no way to get
174 		 * the BAR's location in a page.
175 		 */
176 no_mmap:
177 		vdev->bar_mmap_supported[bar] = false;
178 	}
179 }
180 
181 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev);
182 static void vfio_pci_disable(struct vfio_pci_device *vdev);
183 
184 /*
185  * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
186  * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
187  * If a device implements the former but not the latter we would typically
188  * expect broken_intx_masking be set and require an exclusive interrupt.
189  * However since we do have control of the device's ability to assert INTx,
190  * we can instead pretend that the device does not implement INTx, virtualizing
191  * the pin register to report zero and maintaining DisINTx set on the host.
192  */
193 static bool vfio_pci_nointx(struct pci_dev *pdev)
194 {
195 	switch (pdev->vendor) {
196 	case PCI_VENDOR_ID_INTEL:
197 		switch (pdev->device) {
198 		/* All i40e (XL710/X710) 10/20/40GbE NICs */
199 		case 0x1572:
200 		case 0x1574:
201 		case 0x1580 ... 0x1581:
202 		case 0x1583 ... 0x1589:
203 		case 0x37d0 ... 0x37d2:
204 			return true;
205 		default:
206 			return false;
207 		}
208 	}
209 
210 	return false;
211 }
212 
213 static int vfio_pci_enable(struct vfio_pci_device *vdev)
214 {
215 	struct pci_dev *pdev = vdev->pdev;
216 	int ret;
217 	u16 cmd;
218 	u8 msix_pos;
219 
220 	pci_set_power_state(pdev, PCI_D0);
221 
222 	/* Don't allow our initial saved state to include busmaster */
223 	pci_clear_master(pdev);
224 
225 	ret = pci_enable_device(pdev);
226 	if (ret)
227 		return ret;
228 
229 	vdev->reset_works = (pci_reset_function(pdev) == 0);
230 	pci_save_state(pdev);
231 	vdev->pci_saved_state = pci_store_saved_state(pdev);
232 	if (!vdev->pci_saved_state)
233 		pr_debug("%s: Couldn't store %s saved state\n",
234 			 __func__, dev_name(&pdev->dev));
235 
236 	if (likely(!nointxmask)) {
237 		if (vfio_pci_nointx(pdev)) {
238 			dev_info(&pdev->dev, "Masking broken INTx support\n");
239 			vdev->nointx = true;
240 			pci_intx(pdev, 0);
241 		} else
242 			vdev->pci_2_3 = pci_intx_mask_supported(pdev);
243 	}
244 
245 	pci_read_config_word(pdev, PCI_COMMAND, &cmd);
246 	if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
247 		cmd &= ~PCI_COMMAND_INTX_DISABLE;
248 		pci_write_config_word(pdev, PCI_COMMAND, cmd);
249 	}
250 
251 	ret = vfio_config_init(vdev);
252 	if (ret) {
253 		kfree(vdev->pci_saved_state);
254 		vdev->pci_saved_state = NULL;
255 		pci_disable_device(pdev);
256 		return ret;
257 	}
258 
259 	msix_pos = pdev->msix_cap;
260 	if (msix_pos) {
261 		u16 flags;
262 		u32 table;
263 
264 		pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
265 		pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
266 
267 		vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
268 		vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
269 		vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
270 	} else
271 		vdev->msix_bar = 0xFF;
272 
273 	if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
274 		vdev->has_vga = true;
275 
276 
277 	if (vfio_pci_is_vga(pdev) &&
278 	    pdev->vendor == PCI_VENDOR_ID_INTEL &&
279 	    IS_ENABLED(CONFIG_VFIO_PCI_IGD)) {
280 		ret = vfio_pci_igd_init(vdev);
281 		if (ret) {
282 			dev_warn(&vdev->pdev->dev,
283 				 "Failed to setup Intel IGD regions\n");
284 			vfio_pci_disable(vdev);
285 			return ret;
286 		}
287 	}
288 
289 	vfio_pci_probe_mmaps(vdev);
290 
291 	return 0;
292 }
293 
294 static void vfio_pci_disable(struct vfio_pci_device *vdev)
295 {
296 	struct pci_dev *pdev = vdev->pdev;
297 	struct vfio_pci_dummy_resource *dummy_res, *tmp;
298 	int i, bar;
299 
300 	/* Stop the device from further DMA */
301 	pci_clear_master(pdev);
302 
303 	vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
304 				VFIO_IRQ_SET_ACTION_TRIGGER,
305 				vdev->irq_type, 0, 0, NULL);
306 
307 	vdev->virq_disabled = false;
308 
309 	for (i = 0; i < vdev->num_regions; i++)
310 		vdev->region[i].ops->release(vdev, &vdev->region[i]);
311 
312 	vdev->num_regions = 0;
313 	kfree(vdev->region);
314 	vdev->region = NULL; /* don't krealloc a freed pointer */
315 
316 	vfio_config_free(vdev);
317 
318 	for (bar = PCI_STD_RESOURCES; bar <= PCI_STD_RESOURCE_END; bar++) {
319 		if (!vdev->barmap[bar])
320 			continue;
321 		pci_iounmap(pdev, vdev->barmap[bar]);
322 		pci_release_selected_regions(pdev, 1 << bar);
323 		vdev->barmap[bar] = NULL;
324 	}
325 
326 	list_for_each_entry_safe(dummy_res, tmp,
327 				 &vdev->dummy_resources_list, res_next) {
328 		list_del(&dummy_res->res_next);
329 		release_resource(&dummy_res->resource);
330 		kfree(dummy_res);
331 	}
332 
333 	vdev->needs_reset = true;
334 
335 	/*
336 	 * If we have saved state, restore it.  If we can reset the device,
337 	 * even better.  Resetting with current state seems better than
338 	 * nothing, but saving and restoring current state without reset
339 	 * is just busy work.
340 	 */
341 	if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
342 		pr_info("%s: Couldn't reload %s saved state\n",
343 			__func__, dev_name(&pdev->dev));
344 
345 		if (!vdev->reset_works)
346 			goto out;
347 
348 		pci_save_state(pdev);
349 	}
350 
351 	/*
352 	 * Disable INTx and MSI, presumably to avoid spurious interrupts
353 	 * during reset.  Stolen from pci_reset_function()
354 	 */
355 	pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
356 
357 	/*
358 	 * Try to reset the device.  The success of this is dependent on
359 	 * being able to lock the device, which is not always possible.
360 	 */
361 	if (vdev->reset_works && !pci_try_reset_function(pdev))
362 		vdev->needs_reset = false;
363 
364 	pci_restore_state(pdev);
365 out:
366 	pci_disable_device(pdev);
367 
368 	vfio_pci_try_bus_reset(vdev);
369 
370 	if (!disable_idle_d3)
371 		pci_set_power_state(pdev, PCI_D3hot);
372 }
373 
374 static void vfio_pci_release(void *device_data)
375 {
376 	struct vfio_pci_device *vdev = device_data;
377 
378 	mutex_lock(&driver_lock);
379 
380 	if (!(--vdev->refcnt)) {
381 		vfio_spapr_pci_eeh_release(vdev->pdev);
382 		vfio_pci_disable(vdev);
383 	}
384 
385 	mutex_unlock(&driver_lock);
386 
387 	module_put(THIS_MODULE);
388 }
389 
390 static int vfio_pci_open(void *device_data)
391 {
392 	struct vfio_pci_device *vdev = device_data;
393 	int ret = 0;
394 
395 	if (!try_module_get(THIS_MODULE))
396 		return -ENODEV;
397 
398 	mutex_lock(&driver_lock);
399 
400 	if (!vdev->refcnt) {
401 		ret = vfio_pci_enable(vdev);
402 		if (ret)
403 			goto error;
404 
405 		vfio_spapr_pci_eeh_open(vdev->pdev);
406 	}
407 	vdev->refcnt++;
408 error:
409 	mutex_unlock(&driver_lock);
410 	if (ret)
411 		module_put(THIS_MODULE);
412 	return ret;
413 }
414 
415 static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
416 {
417 	if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
418 		u8 pin;
419 		pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
420 		if (IS_ENABLED(CONFIG_VFIO_PCI_INTX) && !vdev->nointx && pin)
421 			return 1;
422 
423 	} else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
424 		u8 pos;
425 		u16 flags;
426 
427 		pos = vdev->pdev->msi_cap;
428 		if (pos) {
429 			pci_read_config_word(vdev->pdev,
430 					     pos + PCI_MSI_FLAGS, &flags);
431 			return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
432 		}
433 	} else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
434 		u8 pos;
435 		u16 flags;
436 
437 		pos = vdev->pdev->msix_cap;
438 		if (pos) {
439 			pci_read_config_word(vdev->pdev,
440 					     pos + PCI_MSIX_FLAGS, &flags);
441 
442 			return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
443 		}
444 	} else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
445 		if (pci_is_pcie(vdev->pdev))
446 			return 1;
447 	} else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
448 		return 1;
449 	}
450 
451 	return 0;
452 }
453 
454 static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
455 {
456 	(*(int *)data)++;
457 	return 0;
458 }
459 
460 struct vfio_pci_fill_info {
461 	int max;
462 	int cur;
463 	struct vfio_pci_dependent_device *devices;
464 };
465 
466 static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
467 {
468 	struct vfio_pci_fill_info *fill = data;
469 	struct iommu_group *iommu_group;
470 
471 	if (fill->cur == fill->max)
472 		return -EAGAIN; /* Something changed, try again */
473 
474 	iommu_group = iommu_group_get(&pdev->dev);
475 	if (!iommu_group)
476 		return -EPERM; /* Cannot reset non-isolated devices */
477 
478 	fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
479 	fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
480 	fill->devices[fill->cur].bus = pdev->bus->number;
481 	fill->devices[fill->cur].devfn = pdev->devfn;
482 	fill->cur++;
483 	iommu_group_put(iommu_group);
484 	return 0;
485 }
486 
487 struct vfio_pci_group_entry {
488 	struct vfio_group *group;
489 	int id;
490 };
491 
492 struct vfio_pci_group_info {
493 	int count;
494 	struct vfio_pci_group_entry *groups;
495 };
496 
497 static int vfio_pci_validate_devs(struct pci_dev *pdev, void *data)
498 {
499 	struct vfio_pci_group_info *info = data;
500 	struct iommu_group *group;
501 	int id, i;
502 
503 	group = iommu_group_get(&pdev->dev);
504 	if (!group)
505 		return -EPERM;
506 
507 	id = iommu_group_id(group);
508 
509 	for (i = 0; i < info->count; i++)
510 		if (info->groups[i].id == id)
511 			break;
512 
513 	iommu_group_put(group);
514 
515 	return (i == info->count) ? -EINVAL : 0;
516 }
517 
518 static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
519 {
520 	for (; pdev; pdev = pdev->bus->self)
521 		if (pdev->bus == slot->bus)
522 			return (pdev->slot == slot);
523 	return false;
524 }
525 
526 struct vfio_pci_walk_info {
527 	int (*fn)(struct pci_dev *, void *data);
528 	void *data;
529 	struct pci_dev *pdev;
530 	bool slot;
531 	int ret;
532 };
533 
534 static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
535 {
536 	struct vfio_pci_walk_info *walk = data;
537 
538 	if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
539 		walk->ret = walk->fn(pdev, walk->data);
540 
541 	return walk->ret;
542 }
543 
544 static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
545 					 int (*fn)(struct pci_dev *,
546 						   void *data), void *data,
547 					 bool slot)
548 {
549 	struct vfio_pci_walk_info walk = {
550 		.fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
551 	};
552 
553 	pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
554 
555 	return walk.ret;
556 }
557 
558 static int msix_sparse_mmap_cap(struct vfio_pci_device *vdev,
559 				struct vfio_info_cap *caps)
560 {
561 	struct vfio_region_info_cap_sparse_mmap *sparse;
562 	size_t end, size;
563 	int nr_areas = 2, i = 0, ret;
564 
565 	end = pci_resource_len(vdev->pdev, vdev->msix_bar);
566 
567 	/* If MSI-X table is aligned to the start or end, only one area */
568 	if (((vdev->msix_offset & PAGE_MASK) == 0) ||
569 	    (PAGE_ALIGN(vdev->msix_offset + vdev->msix_size) >= end))
570 		nr_areas = 1;
571 
572 	size = sizeof(*sparse) + (nr_areas * sizeof(*sparse->areas));
573 
574 	sparse = kzalloc(size, GFP_KERNEL);
575 	if (!sparse)
576 		return -ENOMEM;
577 
578 	sparse->nr_areas = nr_areas;
579 
580 	if (vdev->msix_offset & PAGE_MASK) {
581 		sparse->areas[i].offset = 0;
582 		sparse->areas[i].size = vdev->msix_offset & PAGE_MASK;
583 		i++;
584 	}
585 
586 	if (PAGE_ALIGN(vdev->msix_offset + vdev->msix_size) < end) {
587 		sparse->areas[i].offset = PAGE_ALIGN(vdev->msix_offset +
588 						     vdev->msix_size);
589 		sparse->areas[i].size = end - sparse->areas[i].offset;
590 		i++;
591 	}
592 
593 	ret = vfio_info_add_capability(caps, VFIO_REGION_INFO_CAP_SPARSE_MMAP,
594 				       sparse);
595 	kfree(sparse);
596 
597 	return ret;
598 }
599 
600 int vfio_pci_register_dev_region(struct vfio_pci_device *vdev,
601 				 unsigned int type, unsigned int subtype,
602 				 const struct vfio_pci_regops *ops,
603 				 size_t size, u32 flags, void *data)
604 {
605 	struct vfio_pci_region *region;
606 
607 	region = krealloc(vdev->region,
608 			  (vdev->num_regions + 1) * sizeof(*region),
609 			  GFP_KERNEL);
610 	if (!region)
611 		return -ENOMEM;
612 
613 	vdev->region = region;
614 	vdev->region[vdev->num_regions].type = type;
615 	vdev->region[vdev->num_regions].subtype = subtype;
616 	vdev->region[vdev->num_regions].ops = ops;
617 	vdev->region[vdev->num_regions].size = size;
618 	vdev->region[vdev->num_regions].flags = flags;
619 	vdev->region[vdev->num_regions].data = data;
620 
621 	vdev->num_regions++;
622 
623 	return 0;
624 }
625 
626 static long vfio_pci_ioctl(void *device_data,
627 			   unsigned int cmd, unsigned long arg)
628 {
629 	struct vfio_pci_device *vdev = device_data;
630 	unsigned long minsz;
631 
632 	if (cmd == VFIO_DEVICE_GET_INFO) {
633 		struct vfio_device_info info;
634 
635 		minsz = offsetofend(struct vfio_device_info, num_irqs);
636 
637 		if (copy_from_user(&info, (void __user *)arg, minsz))
638 			return -EFAULT;
639 
640 		if (info.argsz < minsz)
641 			return -EINVAL;
642 
643 		info.flags = VFIO_DEVICE_FLAGS_PCI;
644 
645 		if (vdev->reset_works)
646 			info.flags |= VFIO_DEVICE_FLAGS_RESET;
647 
648 		info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
649 		info.num_irqs = VFIO_PCI_NUM_IRQS;
650 
651 		return copy_to_user((void __user *)arg, &info, minsz) ?
652 			-EFAULT : 0;
653 
654 	} else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
655 		struct pci_dev *pdev = vdev->pdev;
656 		struct vfio_region_info info;
657 		struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
658 		int i, ret;
659 
660 		minsz = offsetofend(struct vfio_region_info, offset);
661 
662 		if (copy_from_user(&info, (void __user *)arg, minsz))
663 			return -EFAULT;
664 
665 		if (info.argsz < minsz)
666 			return -EINVAL;
667 
668 		switch (info.index) {
669 		case VFIO_PCI_CONFIG_REGION_INDEX:
670 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
671 			info.size = pdev->cfg_size;
672 			info.flags = VFIO_REGION_INFO_FLAG_READ |
673 				     VFIO_REGION_INFO_FLAG_WRITE;
674 			break;
675 		case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
676 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
677 			info.size = pci_resource_len(pdev, info.index);
678 			if (!info.size) {
679 				info.flags = 0;
680 				break;
681 			}
682 
683 			info.flags = VFIO_REGION_INFO_FLAG_READ |
684 				     VFIO_REGION_INFO_FLAG_WRITE;
685 			if (vdev->bar_mmap_supported[info.index]) {
686 				info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
687 				if (info.index == vdev->msix_bar) {
688 					ret = msix_sparse_mmap_cap(vdev, &caps);
689 					if (ret)
690 						return ret;
691 				}
692 			}
693 
694 			break;
695 		case VFIO_PCI_ROM_REGION_INDEX:
696 		{
697 			void __iomem *io;
698 			size_t size;
699 
700 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
701 			info.flags = 0;
702 
703 			/* Report the BAR size, not the ROM size */
704 			info.size = pci_resource_len(pdev, info.index);
705 			if (!info.size) {
706 				/* Shadow ROMs appear as PCI option ROMs */
707 				if (pdev->resource[PCI_ROM_RESOURCE].flags &
708 							IORESOURCE_ROM_SHADOW)
709 					info.size = 0x20000;
710 				else
711 					break;
712 			}
713 
714 			/* Is it really there? */
715 			io = pci_map_rom(pdev, &size);
716 			if (!io || !size) {
717 				info.size = 0;
718 				break;
719 			}
720 			pci_unmap_rom(pdev, io);
721 
722 			info.flags = VFIO_REGION_INFO_FLAG_READ;
723 			break;
724 		}
725 		case VFIO_PCI_VGA_REGION_INDEX:
726 			if (!vdev->has_vga)
727 				return -EINVAL;
728 
729 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
730 			info.size = 0xc0000;
731 			info.flags = VFIO_REGION_INFO_FLAG_READ |
732 				     VFIO_REGION_INFO_FLAG_WRITE;
733 
734 			break;
735 		default:
736 		{
737 			struct vfio_region_info_cap_type cap_type;
738 
739 			if (info.index >=
740 			    VFIO_PCI_NUM_REGIONS + vdev->num_regions)
741 				return -EINVAL;
742 
743 			i = info.index - VFIO_PCI_NUM_REGIONS;
744 
745 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
746 			info.size = vdev->region[i].size;
747 			info.flags = vdev->region[i].flags;
748 
749 			cap_type.type = vdev->region[i].type;
750 			cap_type.subtype = vdev->region[i].subtype;
751 
752 			ret = vfio_info_add_capability(&caps,
753 						      VFIO_REGION_INFO_CAP_TYPE,
754 						      &cap_type);
755 			if (ret)
756 				return ret;
757 
758 		}
759 		}
760 
761 		if (caps.size) {
762 			info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
763 			if (info.argsz < sizeof(info) + caps.size) {
764 				info.argsz = sizeof(info) + caps.size;
765 				info.cap_offset = 0;
766 			} else {
767 				vfio_info_cap_shift(&caps, sizeof(info));
768 				if (copy_to_user((void __user *)arg +
769 						  sizeof(info), caps.buf,
770 						  caps.size)) {
771 					kfree(caps.buf);
772 					return -EFAULT;
773 				}
774 				info.cap_offset = sizeof(info);
775 			}
776 
777 			kfree(caps.buf);
778 		}
779 
780 		return copy_to_user((void __user *)arg, &info, minsz) ?
781 			-EFAULT : 0;
782 
783 	} else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
784 		struct vfio_irq_info info;
785 
786 		minsz = offsetofend(struct vfio_irq_info, count);
787 
788 		if (copy_from_user(&info, (void __user *)arg, minsz))
789 			return -EFAULT;
790 
791 		if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
792 			return -EINVAL;
793 
794 		switch (info.index) {
795 		case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
796 		case VFIO_PCI_REQ_IRQ_INDEX:
797 			break;
798 		case VFIO_PCI_ERR_IRQ_INDEX:
799 			if (pci_is_pcie(vdev->pdev))
800 				break;
801 		/* pass thru to return error */
802 		default:
803 			return -EINVAL;
804 		}
805 
806 		info.flags = VFIO_IRQ_INFO_EVENTFD;
807 
808 		info.count = vfio_pci_get_irq_count(vdev, info.index);
809 
810 		if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
811 			info.flags |= (VFIO_IRQ_INFO_MASKABLE |
812 				       VFIO_IRQ_INFO_AUTOMASKED);
813 		else
814 			info.flags |= VFIO_IRQ_INFO_NORESIZE;
815 
816 		return copy_to_user((void __user *)arg, &info, minsz) ?
817 			-EFAULT : 0;
818 
819 	} else if (cmd == VFIO_DEVICE_SET_IRQS) {
820 		struct vfio_irq_set hdr;
821 		u8 *data = NULL;
822 		int max, ret = 0;
823 		size_t data_size = 0;
824 
825 		minsz = offsetofend(struct vfio_irq_set, count);
826 
827 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
828 			return -EFAULT;
829 
830 		max = vfio_pci_get_irq_count(vdev, hdr.index);
831 
832 		ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
833 						 VFIO_PCI_NUM_IRQS, &data_size);
834 		if (ret)
835 			return ret;
836 
837 		if (data_size) {
838 			data = memdup_user((void __user *)(arg + minsz),
839 					    data_size);
840 			if (IS_ERR(data))
841 				return PTR_ERR(data);
842 		}
843 
844 		mutex_lock(&vdev->igate);
845 
846 		ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
847 					      hdr.start, hdr.count, data);
848 
849 		mutex_unlock(&vdev->igate);
850 		kfree(data);
851 
852 		return ret;
853 
854 	} else if (cmd == VFIO_DEVICE_RESET) {
855 		return vdev->reset_works ?
856 			pci_try_reset_function(vdev->pdev) : -EINVAL;
857 
858 	} else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
859 		struct vfio_pci_hot_reset_info hdr;
860 		struct vfio_pci_fill_info fill = { 0 };
861 		struct vfio_pci_dependent_device *devices = NULL;
862 		bool slot = false;
863 		int ret = 0;
864 
865 		minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
866 
867 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
868 			return -EFAULT;
869 
870 		if (hdr.argsz < minsz)
871 			return -EINVAL;
872 
873 		hdr.flags = 0;
874 
875 		/* Can we do a slot or bus reset or neither? */
876 		if (!pci_probe_reset_slot(vdev->pdev->slot))
877 			slot = true;
878 		else if (pci_probe_reset_bus(vdev->pdev->bus))
879 			return -ENODEV;
880 
881 		/* How many devices are affected? */
882 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
883 						    vfio_pci_count_devs,
884 						    &fill.max, slot);
885 		if (ret)
886 			return ret;
887 
888 		WARN_ON(!fill.max); /* Should always be at least one */
889 
890 		/*
891 		 * If there's enough space, fill it now, otherwise return
892 		 * -ENOSPC and the number of devices affected.
893 		 */
894 		if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
895 			ret = -ENOSPC;
896 			hdr.count = fill.max;
897 			goto reset_info_exit;
898 		}
899 
900 		devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
901 		if (!devices)
902 			return -ENOMEM;
903 
904 		fill.devices = devices;
905 
906 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
907 						    vfio_pci_fill_devs,
908 						    &fill, slot);
909 
910 		/*
911 		 * If a device was removed between counting and filling,
912 		 * we may come up short of fill.max.  If a device was
913 		 * added, we'll have a return of -EAGAIN above.
914 		 */
915 		if (!ret)
916 			hdr.count = fill.cur;
917 
918 reset_info_exit:
919 		if (copy_to_user((void __user *)arg, &hdr, minsz))
920 			ret = -EFAULT;
921 
922 		if (!ret) {
923 			if (copy_to_user((void __user *)(arg + minsz), devices,
924 					 hdr.count * sizeof(*devices)))
925 				ret = -EFAULT;
926 		}
927 
928 		kfree(devices);
929 		return ret;
930 
931 	} else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
932 		struct vfio_pci_hot_reset hdr;
933 		int32_t *group_fds;
934 		struct vfio_pci_group_entry *groups;
935 		struct vfio_pci_group_info info;
936 		bool slot = false;
937 		int i, count = 0, ret = 0;
938 
939 		minsz = offsetofend(struct vfio_pci_hot_reset, count);
940 
941 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
942 			return -EFAULT;
943 
944 		if (hdr.argsz < minsz || hdr.flags)
945 			return -EINVAL;
946 
947 		/* Can we do a slot or bus reset or neither? */
948 		if (!pci_probe_reset_slot(vdev->pdev->slot))
949 			slot = true;
950 		else if (pci_probe_reset_bus(vdev->pdev->bus))
951 			return -ENODEV;
952 
953 		/*
954 		 * We can't let userspace give us an arbitrarily large
955 		 * buffer to copy, so verify how many we think there
956 		 * could be.  Note groups can have multiple devices so
957 		 * one group per device is the max.
958 		 */
959 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
960 						    vfio_pci_count_devs,
961 						    &count, slot);
962 		if (ret)
963 			return ret;
964 
965 		/* Somewhere between 1 and count is OK */
966 		if (!hdr.count || hdr.count > count)
967 			return -EINVAL;
968 
969 		group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
970 		groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
971 		if (!group_fds || !groups) {
972 			kfree(group_fds);
973 			kfree(groups);
974 			return -ENOMEM;
975 		}
976 
977 		if (copy_from_user(group_fds, (void __user *)(arg + minsz),
978 				   hdr.count * sizeof(*group_fds))) {
979 			kfree(group_fds);
980 			kfree(groups);
981 			return -EFAULT;
982 		}
983 
984 		/*
985 		 * For each group_fd, get the group through the vfio external
986 		 * user interface and store the group and iommu ID.  This
987 		 * ensures the group is held across the reset.
988 		 */
989 		for (i = 0; i < hdr.count; i++) {
990 			struct vfio_group *group;
991 			struct fd f = fdget(group_fds[i]);
992 			if (!f.file) {
993 				ret = -EBADF;
994 				break;
995 			}
996 
997 			group = vfio_group_get_external_user(f.file);
998 			fdput(f);
999 			if (IS_ERR(group)) {
1000 				ret = PTR_ERR(group);
1001 				break;
1002 			}
1003 
1004 			groups[i].group = group;
1005 			groups[i].id = vfio_external_user_iommu_id(group);
1006 		}
1007 
1008 		kfree(group_fds);
1009 
1010 		/* release reference to groups on error */
1011 		if (ret)
1012 			goto hot_reset_release;
1013 
1014 		info.count = hdr.count;
1015 		info.groups = groups;
1016 
1017 		/*
1018 		 * Test whether all the affected devices are contained
1019 		 * by the set of groups provided by the user.
1020 		 */
1021 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1022 						    vfio_pci_validate_devs,
1023 						    &info, slot);
1024 		if (!ret)
1025 			/* User has access, do the reset */
1026 			ret = slot ? pci_try_reset_slot(vdev->pdev->slot) :
1027 				     pci_try_reset_bus(vdev->pdev->bus);
1028 
1029 hot_reset_release:
1030 		for (i--; i >= 0; i--)
1031 			vfio_group_put_external_user(groups[i].group);
1032 
1033 		kfree(groups);
1034 		return ret;
1035 	}
1036 
1037 	return -ENOTTY;
1038 }
1039 
1040 static ssize_t vfio_pci_rw(void *device_data, char __user *buf,
1041 			   size_t count, loff_t *ppos, bool iswrite)
1042 {
1043 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1044 	struct vfio_pci_device *vdev = device_data;
1045 
1046 	if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1047 		return -EINVAL;
1048 
1049 	switch (index) {
1050 	case VFIO_PCI_CONFIG_REGION_INDEX:
1051 		return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1052 
1053 	case VFIO_PCI_ROM_REGION_INDEX:
1054 		if (iswrite)
1055 			return -EINVAL;
1056 		return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1057 
1058 	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1059 		return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1060 
1061 	case VFIO_PCI_VGA_REGION_INDEX:
1062 		return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1063 	default:
1064 		index -= VFIO_PCI_NUM_REGIONS;
1065 		return vdev->region[index].ops->rw(vdev, buf,
1066 						   count, ppos, iswrite);
1067 	}
1068 
1069 	return -EINVAL;
1070 }
1071 
1072 static ssize_t vfio_pci_read(void *device_data, char __user *buf,
1073 			     size_t count, loff_t *ppos)
1074 {
1075 	if (!count)
1076 		return 0;
1077 
1078 	return vfio_pci_rw(device_data, buf, count, ppos, false);
1079 }
1080 
1081 static ssize_t vfio_pci_write(void *device_data, const char __user *buf,
1082 			      size_t count, loff_t *ppos)
1083 {
1084 	if (!count)
1085 		return 0;
1086 
1087 	return vfio_pci_rw(device_data, (char __user *)buf, count, ppos, true);
1088 }
1089 
1090 static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma)
1091 {
1092 	struct vfio_pci_device *vdev = device_data;
1093 	struct pci_dev *pdev = vdev->pdev;
1094 	unsigned int index;
1095 	u64 phys_len, req_len, pgoff, req_start;
1096 	int ret;
1097 
1098 	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1099 
1100 	if (vma->vm_end < vma->vm_start)
1101 		return -EINVAL;
1102 	if ((vma->vm_flags & VM_SHARED) == 0)
1103 		return -EINVAL;
1104 	if (index >= VFIO_PCI_ROM_REGION_INDEX)
1105 		return -EINVAL;
1106 	if (!vdev->bar_mmap_supported[index])
1107 		return -EINVAL;
1108 
1109 	phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1110 	req_len = vma->vm_end - vma->vm_start;
1111 	pgoff = vma->vm_pgoff &
1112 		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1113 	req_start = pgoff << PAGE_SHIFT;
1114 
1115 	if (req_start + req_len > phys_len)
1116 		return -EINVAL;
1117 
1118 	if (index == vdev->msix_bar) {
1119 		/*
1120 		 * Disallow mmaps overlapping the MSI-X table; users don't
1121 		 * get to touch this directly.  We could find somewhere
1122 		 * else to map the overlap, but page granularity is only
1123 		 * a recommendation, not a requirement, so the user needs
1124 		 * to know which bits are real.  Requiring them to mmap
1125 		 * around the table makes that clear.
1126 		 */
1127 
1128 		/* If neither entirely above nor below, then it overlaps */
1129 		if (!(req_start >= vdev->msix_offset + vdev->msix_size ||
1130 		      req_start + req_len <= vdev->msix_offset))
1131 			return -EINVAL;
1132 	}
1133 
1134 	/*
1135 	 * Even though we don't make use of the barmap for the mmap,
1136 	 * we need to request the region and the barmap tracks that.
1137 	 */
1138 	if (!vdev->barmap[index]) {
1139 		ret = pci_request_selected_regions(pdev,
1140 						   1 << index, "vfio-pci");
1141 		if (ret)
1142 			return ret;
1143 
1144 		vdev->barmap[index] = pci_iomap(pdev, index, 0);
1145 		if (!vdev->barmap[index]) {
1146 			pci_release_selected_regions(pdev, 1 << index);
1147 			return -ENOMEM;
1148 		}
1149 	}
1150 
1151 	vma->vm_private_data = vdev;
1152 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1153 	vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1154 
1155 	return remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1156 			       req_len, vma->vm_page_prot);
1157 }
1158 
1159 static void vfio_pci_request(void *device_data, unsigned int count)
1160 {
1161 	struct vfio_pci_device *vdev = device_data;
1162 
1163 	mutex_lock(&vdev->igate);
1164 
1165 	if (vdev->req_trigger) {
1166 		if (!(count % 10))
1167 			dev_notice_ratelimited(&vdev->pdev->dev,
1168 				"Relaying device request to user (#%u)\n",
1169 				count);
1170 		eventfd_signal(vdev->req_trigger, 1);
1171 	} else if (count == 0) {
1172 		dev_warn(&vdev->pdev->dev,
1173 			"No device request channel registered, blocked until released by user\n");
1174 	}
1175 
1176 	mutex_unlock(&vdev->igate);
1177 }
1178 
1179 static const struct vfio_device_ops vfio_pci_ops = {
1180 	.name		= "vfio-pci",
1181 	.open		= vfio_pci_open,
1182 	.release	= vfio_pci_release,
1183 	.ioctl		= vfio_pci_ioctl,
1184 	.read		= vfio_pci_read,
1185 	.write		= vfio_pci_write,
1186 	.mmap		= vfio_pci_mmap,
1187 	.request	= vfio_pci_request,
1188 };
1189 
1190 static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1191 {
1192 	struct vfio_pci_device *vdev;
1193 	struct iommu_group *group;
1194 	int ret;
1195 
1196 	if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1197 		return -EINVAL;
1198 
1199 	group = vfio_iommu_group_get(&pdev->dev);
1200 	if (!group)
1201 		return -EINVAL;
1202 
1203 	vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
1204 	if (!vdev) {
1205 		vfio_iommu_group_put(group, &pdev->dev);
1206 		return -ENOMEM;
1207 	}
1208 
1209 	vdev->pdev = pdev;
1210 	vdev->irq_type = VFIO_PCI_NUM_IRQS;
1211 	mutex_init(&vdev->igate);
1212 	spin_lock_init(&vdev->irqlock);
1213 
1214 	ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
1215 	if (ret) {
1216 		vfio_iommu_group_put(group, &pdev->dev);
1217 		kfree(vdev);
1218 		return ret;
1219 	}
1220 
1221 	if (vfio_pci_is_vga(pdev)) {
1222 		vga_client_register(pdev, vdev, NULL, vfio_pci_set_vga_decode);
1223 		vga_set_legacy_decoding(pdev,
1224 					vfio_pci_set_vga_decode(vdev, false));
1225 	}
1226 
1227 	if (!disable_idle_d3) {
1228 		/*
1229 		 * pci-core sets the device power state to an unknown value at
1230 		 * bootup and after being removed from a driver.  The only
1231 		 * transition it allows from this unknown state is to D0, which
1232 		 * typically happens when a driver calls pci_enable_device().
1233 		 * We're not ready to enable the device yet, but we do want to
1234 		 * be able to get to D3.  Therefore first do a D0 transition
1235 		 * before going to D3.
1236 		 */
1237 		pci_set_power_state(pdev, PCI_D0);
1238 		pci_set_power_state(pdev, PCI_D3hot);
1239 	}
1240 
1241 	return ret;
1242 }
1243 
1244 static void vfio_pci_remove(struct pci_dev *pdev)
1245 {
1246 	struct vfio_pci_device *vdev;
1247 
1248 	vdev = vfio_del_group_dev(&pdev->dev);
1249 	if (!vdev)
1250 		return;
1251 
1252 	vfio_iommu_group_put(pdev->dev.iommu_group, &pdev->dev);
1253 	kfree(vdev->region);
1254 	kfree(vdev);
1255 
1256 	if (vfio_pci_is_vga(pdev)) {
1257 		vga_client_register(pdev, NULL, NULL, NULL);
1258 		vga_set_legacy_decoding(pdev,
1259 				VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
1260 				VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM);
1261 	}
1262 
1263 	if (!disable_idle_d3)
1264 		pci_set_power_state(pdev, PCI_D0);
1265 }
1266 
1267 static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
1268 						  pci_channel_state_t state)
1269 {
1270 	struct vfio_pci_device *vdev;
1271 	struct vfio_device *device;
1272 
1273 	device = vfio_device_get_from_dev(&pdev->dev);
1274 	if (device == NULL)
1275 		return PCI_ERS_RESULT_DISCONNECT;
1276 
1277 	vdev = vfio_device_data(device);
1278 	if (vdev == NULL) {
1279 		vfio_device_put(device);
1280 		return PCI_ERS_RESULT_DISCONNECT;
1281 	}
1282 
1283 	mutex_lock(&vdev->igate);
1284 
1285 	if (vdev->err_trigger)
1286 		eventfd_signal(vdev->err_trigger, 1);
1287 
1288 	mutex_unlock(&vdev->igate);
1289 
1290 	vfio_device_put(device);
1291 
1292 	return PCI_ERS_RESULT_CAN_RECOVER;
1293 }
1294 
1295 static const struct pci_error_handlers vfio_err_handlers = {
1296 	.error_detected = vfio_pci_aer_err_detected,
1297 };
1298 
1299 static struct pci_driver vfio_pci_driver = {
1300 	.name		= "vfio-pci",
1301 	.id_table	= NULL, /* only dynamic ids */
1302 	.probe		= vfio_pci_probe,
1303 	.remove		= vfio_pci_remove,
1304 	.err_handler	= &vfio_err_handlers,
1305 };
1306 
1307 struct vfio_devices {
1308 	struct vfio_device **devices;
1309 	int cur_index;
1310 	int max_index;
1311 };
1312 
1313 static int vfio_pci_get_devs(struct pci_dev *pdev, void *data)
1314 {
1315 	struct vfio_devices *devs = data;
1316 	struct vfio_device *device;
1317 
1318 	if (devs->cur_index == devs->max_index)
1319 		return -ENOSPC;
1320 
1321 	device = vfio_device_get_from_dev(&pdev->dev);
1322 	if (!device)
1323 		return -EINVAL;
1324 
1325 	if (pci_dev_driver(pdev) != &vfio_pci_driver) {
1326 		vfio_device_put(device);
1327 		return -EBUSY;
1328 	}
1329 
1330 	devs->devices[devs->cur_index++] = device;
1331 	return 0;
1332 }
1333 
1334 /*
1335  * Attempt to do a bus/slot reset if there are devices affected by a reset for
1336  * this device that are needs_reset and all of the affected devices are unused
1337  * (!refcnt).  Callers are required to hold driver_lock when calling this to
1338  * prevent device opens and concurrent bus reset attempts.  We prevent device
1339  * unbinds by acquiring and holding a reference to the vfio_device.
1340  *
1341  * NB: vfio-core considers a group to be viable even if some devices are
1342  * bound to drivers like pci-stub or pcieport.  Here we require all devices
1343  * to be bound to vfio_pci since that's the only way we can be sure they
1344  * stay put.
1345  */
1346 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev)
1347 {
1348 	struct vfio_devices devs = { .cur_index = 0 };
1349 	int i = 0, ret = -EINVAL;
1350 	bool needs_reset = false, slot = false;
1351 	struct vfio_pci_device *tmp;
1352 
1353 	if (!pci_probe_reset_slot(vdev->pdev->slot))
1354 		slot = true;
1355 	else if (pci_probe_reset_bus(vdev->pdev->bus))
1356 		return;
1357 
1358 	if (vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
1359 					  &i, slot) || !i)
1360 		return;
1361 
1362 	devs.max_index = i;
1363 	devs.devices = kcalloc(i, sizeof(struct vfio_device *), GFP_KERNEL);
1364 	if (!devs.devices)
1365 		return;
1366 
1367 	if (vfio_pci_for_each_slot_or_bus(vdev->pdev,
1368 					  vfio_pci_get_devs, &devs, slot))
1369 		goto put_devs;
1370 
1371 	for (i = 0; i < devs.cur_index; i++) {
1372 		tmp = vfio_device_data(devs.devices[i]);
1373 		if (tmp->needs_reset)
1374 			needs_reset = true;
1375 		if (tmp->refcnt)
1376 			goto put_devs;
1377 	}
1378 
1379 	if (needs_reset)
1380 		ret = slot ? pci_try_reset_slot(vdev->pdev->slot) :
1381 			     pci_try_reset_bus(vdev->pdev->bus);
1382 
1383 put_devs:
1384 	for (i = 0; i < devs.cur_index; i++) {
1385 		tmp = vfio_device_data(devs.devices[i]);
1386 		if (!ret)
1387 			tmp->needs_reset = false;
1388 
1389 		if (!tmp->refcnt && !disable_idle_d3)
1390 			pci_set_power_state(tmp->pdev, PCI_D3hot);
1391 
1392 		vfio_device_put(devs.devices[i]);
1393 	}
1394 
1395 	kfree(devs.devices);
1396 }
1397 
1398 static void __exit vfio_pci_cleanup(void)
1399 {
1400 	pci_unregister_driver(&vfio_pci_driver);
1401 	vfio_pci_uninit_perm_bits();
1402 }
1403 
1404 static void __init vfio_pci_fill_ids(void)
1405 {
1406 	char *p, *id;
1407 	int rc;
1408 
1409 	/* no ids passed actually */
1410 	if (ids[0] == '\0')
1411 		return;
1412 
1413 	/* add ids specified in the module parameter */
1414 	p = ids;
1415 	while ((id = strsep(&p, ","))) {
1416 		unsigned int vendor, device, subvendor = PCI_ANY_ID,
1417 			subdevice = PCI_ANY_ID, class = 0, class_mask = 0;
1418 		int fields;
1419 
1420 		if (!strlen(id))
1421 			continue;
1422 
1423 		fields = sscanf(id, "%x:%x:%x:%x:%x:%x",
1424 				&vendor, &device, &subvendor, &subdevice,
1425 				&class, &class_mask);
1426 
1427 		if (fields < 2) {
1428 			pr_warn("invalid id string \"%s\"\n", id);
1429 			continue;
1430 		}
1431 
1432 		rc = pci_add_dynid(&vfio_pci_driver, vendor, device,
1433 				   subvendor, subdevice, class, class_mask, 0);
1434 		if (rc)
1435 			pr_warn("failed to add dynamic id [%04hx:%04hx[%04hx:%04hx]] class %#08x/%08x (%d)\n",
1436 				vendor, device, subvendor, subdevice,
1437 				class, class_mask, rc);
1438 		else
1439 			pr_info("add [%04hx:%04hx[%04hx:%04hx]] class %#08x/%08x\n",
1440 				vendor, device, subvendor, subdevice,
1441 				class, class_mask);
1442 	}
1443 }
1444 
1445 static int __init vfio_pci_init(void)
1446 {
1447 	int ret;
1448 
1449 	/* Allocate shared config space permision data used by all devices */
1450 	ret = vfio_pci_init_perm_bits();
1451 	if (ret)
1452 		return ret;
1453 
1454 	/* Register and scan for devices */
1455 	ret = pci_register_driver(&vfio_pci_driver);
1456 	if (ret)
1457 		goto out_driver;
1458 
1459 	vfio_pci_fill_ids();
1460 
1461 	return 0;
1462 
1463 out_driver:
1464 	vfio_pci_uninit_perm_bits();
1465 	return ret;
1466 }
1467 
1468 module_init(vfio_pci_init);
1469 module_exit(vfio_pci_cleanup);
1470 
1471 MODULE_VERSION(DRIVER_VERSION);
1472 MODULE_LICENSE("GPL v2");
1473 MODULE_AUTHOR(DRIVER_AUTHOR);
1474 MODULE_DESCRIPTION(DRIVER_DESC);
1475