xref: /openbmc/linux/drivers/vfio/pci/vfio_pci.c (revision bc5aa3a0)
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_info_cap_header *header;
562 	struct vfio_region_info_cap_sparse_mmap *sparse;
563 	size_t end, size;
564 	int nr_areas = 2, i = 0;
565 
566 	end = pci_resource_len(vdev->pdev, vdev->msix_bar);
567 
568 	/* If MSI-X table is aligned to the start or end, only one area */
569 	if (((vdev->msix_offset & PAGE_MASK) == 0) ||
570 	    (PAGE_ALIGN(vdev->msix_offset + vdev->msix_size) >= end))
571 		nr_areas = 1;
572 
573 	size = sizeof(*sparse) + (nr_areas * sizeof(*sparse->areas));
574 
575 	header = vfio_info_cap_add(caps, size,
576 				   VFIO_REGION_INFO_CAP_SPARSE_MMAP, 1);
577 	if (IS_ERR(header))
578 		return PTR_ERR(header);
579 
580 	sparse = container_of(header,
581 			      struct vfio_region_info_cap_sparse_mmap, header);
582 	sparse->nr_areas = nr_areas;
583 
584 	if (vdev->msix_offset & PAGE_MASK) {
585 		sparse->areas[i].offset = 0;
586 		sparse->areas[i].size = vdev->msix_offset & PAGE_MASK;
587 		i++;
588 	}
589 
590 	if (PAGE_ALIGN(vdev->msix_offset + vdev->msix_size) < end) {
591 		sparse->areas[i].offset = PAGE_ALIGN(vdev->msix_offset +
592 						     vdev->msix_size);
593 		sparse->areas[i].size = end - sparse->areas[i].offset;
594 		i++;
595 	}
596 
597 	return 0;
598 }
599 
600 static int region_type_cap(struct vfio_pci_device *vdev,
601 			   struct vfio_info_cap *caps,
602 			   unsigned int type, unsigned int subtype)
603 {
604 	struct vfio_info_cap_header *header;
605 	struct vfio_region_info_cap_type *cap;
606 
607 	header = vfio_info_cap_add(caps, sizeof(*cap),
608 				   VFIO_REGION_INFO_CAP_TYPE, 1);
609 	if (IS_ERR(header))
610 		return PTR_ERR(header);
611 
612 	cap = container_of(header, struct vfio_region_info_cap_type, header);
613 	cap->type = type;
614 	cap->subtype = subtype;
615 
616 	return 0;
617 }
618 
619 int vfio_pci_register_dev_region(struct vfio_pci_device *vdev,
620 				 unsigned int type, unsigned int subtype,
621 				 const struct vfio_pci_regops *ops,
622 				 size_t size, u32 flags, void *data)
623 {
624 	struct vfio_pci_region *region;
625 
626 	region = krealloc(vdev->region,
627 			  (vdev->num_regions + 1) * sizeof(*region),
628 			  GFP_KERNEL);
629 	if (!region)
630 		return -ENOMEM;
631 
632 	vdev->region = region;
633 	vdev->region[vdev->num_regions].type = type;
634 	vdev->region[vdev->num_regions].subtype = subtype;
635 	vdev->region[vdev->num_regions].ops = ops;
636 	vdev->region[vdev->num_regions].size = size;
637 	vdev->region[vdev->num_regions].flags = flags;
638 	vdev->region[vdev->num_regions].data = data;
639 
640 	vdev->num_regions++;
641 
642 	return 0;
643 }
644 
645 static long vfio_pci_ioctl(void *device_data,
646 			   unsigned int cmd, unsigned long arg)
647 {
648 	struct vfio_pci_device *vdev = device_data;
649 	unsigned long minsz;
650 
651 	if (cmd == VFIO_DEVICE_GET_INFO) {
652 		struct vfio_device_info info;
653 
654 		minsz = offsetofend(struct vfio_device_info, num_irqs);
655 
656 		if (copy_from_user(&info, (void __user *)arg, minsz))
657 			return -EFAULT;
658 
659 		if (info.argsz < minsz)
660 			return -EINVAL;
661 
662 		info.flags = VFIO_DEVICE_FLAGS_PCI;
663 
664 		if (vdev->reset_works)
665 			info.flags |= VFIO_DEVICE_FLAGS_RESET;
666 
667 		info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
668 		info.num_irqs = VFIO_PCI_NUM_IRQS;
669 
670 		return copy_to_user((void __user *)arg, &info, minsz) ?
671 			-EFAULT : 0;
672 
673 	} else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
674 		struct pci_dev *pdev = vdev->pdev;
675 		struct vfio_region_info info;
676 		struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
677 		int i, ret;
678 
679 		minsz = offsetofend(struct vfio_region_info, offset);
680 
681 		if (copy_from_user(&info, (void __user *)arg, minsz))
682 			return -EFAULT;
683 
684 		if (info.argsz < minsz)
685 			return -EINVAL;
686 
687 		switch (info.index) {
688 		case VFIO_PCI_CONFIG_REGION_INDEX:
689 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
690 			info.size = pdev->cfg_size;
691 			info.flags = VFIO_REGION_INFO_FLAG_READ |
692 				     VFIO_REGION_INFO_FLAG_WRITE;
693 			break;
694 		case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
695 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
696 			info.size = pci_resource_len(pdev, info.index);
697 			if (!info.size) {
698 				info.flags = 0;
699 				break;
700 			}
701 
702 			info.flags = VFIO_REGION_INFO_FLAG_READ |
703 				     VFIO_REGION_INFO_FLAG_WRITE;
704 			if (vdev->bar_mmap_supported[info.index]) {
705 				info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
706 				if (info.index == vdev->msix_bar) {
707 					ret = msix_sparse_mmap_cap(vdev, &caps);
708 					if (ret)
709 						return ret;
710 				}
711 			}
712 
713 			break;
714 		case VFIO_PCI_ROM_REGION_INDEX:
715 		{
716 			void __iomem *io;
717 			size_t size;
718 
719 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
720 			info.flags = 0;
721 
722 			/* Report the BAR size, not the ROM size */
723 			info.size = pci_resource_len(pdev, info.index);
724 			if (!info.size) {
725 				/* Shadow ROMs appear as PCI option ROMs */
726 				if (pdev->resource[PCI_ROM_RESOURCE].flags &
727 							IORESOURCE_ROM_SHADOW)
728 					info.size = 0x20000;
729 				else
730 					break;
731 			}
732 
733 			/* Is it really there? */
734 			io = pci_map_rom(pdev, &size);
735 			if (!io || !size) {
736 				info.size = 0;
737 				break;
738 			}
739 			pci_unmap_rom(pdev, io);
740 
741 			info.flags = VFIO_REGION_INFO_FLAG_READ;
742 			break;
743 		}
744 		case VFIO_PCI_VGA_REGION_INDEX:
745 			if (!vdev->has_vga)
746 				return -EINVAL;
747 
748 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
749 			info.size = 0xc0000;
750 			info.flags = VFIO_REGION_INFO_FLAG_READ |
751 				     VFIO_REGION_INFO_FLAG_WRITE;
752 
753 			break;
754 		default:
755 			if (info.index >=
756 			    VFIO_PCI_NUM_REGIONS + vdev->num_regions)
757 				return -EINVAL;
758 
759 			i = info.index - VFIO_PCI_NUM_REGIONS;
760 
761 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
762 			info.size = vdev->region[i].size;
763 			info.flags = vdev->region[i].flags;
764 
765 			ret = region_type_cap(vdev, &caps,
766 					      vdev->region[i].type,
767 					      vdev->region[i].subtype);
768 			if (ret)
769 				return ret;
770 		}
771 
772 		if (caps.size) {
773 			info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
774 			if (info.argsz < sizeof(info) + caps.size) {
775 				info.argsz = sizeof(info) + caps.size;
776 				info.cap_offset = 0;
777 			} else {
778 				vfio_info_cap_shift(&caps, sizeof(info));
779 				if (copy_to_user((void __user *)arg +
780 						  sizeof(info), caps.buf,
781 						  caps.size)) {
782 					kfree(caps.buf);
783 					return -EFAULT;
784 				}
785 				info.cap_offset = sizeof(info);
786 			}
787 
788 			kfree(caps.buf);
789 		}
790 
791 		return copy_to_user((void __user *)arg, &info, minsz) ?
792 			-EFAULT : 0;
793 
794 	} else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
795 		struct vfio_irq_info info;
796 
797 		minsz = offsetofend(struct vfio_irq_info, count);
798 
799 		if (copy_from_user(&info, (void __user *)arg, minsz))
800 			return -EFAULT;
801 
802 		if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
803 			return -EINVAL;
804 
805 		switch (info.index) {
806 		case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
807 		case VFIO_PCI_REQ_IRQ_INDEX:
808 			break;
809 		case VFIO_PCI_ERR_IRQ_INDEX:
810 			if (pci_is_pcie(vdev->pdev))
811 				break;
812 		/* pass thru to return error */
813 		default:
814 			return -EINVAL;
815 		}
816 
817 		info.flags = VFIO_IRQ_INFO_EVENTFD;
818 
819 		info.count = vfio_pci_get_irq_count(vdev, info.index);
820 
821 		if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
822 			info.flags |= (VFIO_IRQ_INFO_MASKABLE |
823 				       VFIO_IRQ_INFO_AUTOMASKED);
824 		else
825 			info.flags |= VFIO_IRQ_INFO_NORESIZE;
826 
827 		return copy_to_user((void __user *)arg, &info, minsz) ?
828 			-EFAULT : 0;
829 
830 	} else if (cmd == VFIO_DEVICE_SET_IRQS) {
831 		struct vfio_irq_set hdr;
832 		u8 *data = NULL;
833 		int ret = 0;
834 
835 		minsz = offsetofend(struct vfio_irq_set, count);
836 
837 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
838 			return -EFAULT;
839 
840 		if (hdr.argsz < minsz || hdr.index >= VFIO_PCI_NUM_IRQS ||
841 		    hdr.flags & ~(VFIO_IRQ_SET_DATA_TYPE_MASK |
842 				  VFIO_IRQ_SET_ACTION_TYPE_MASK))
843 			return -EINVAL;
844 
845 		if (!(hdr.flags & VFIO_IRQ_SET_DATA_NONE)) {
846 			size_t size;
847 			int max = vfio_pci_get_irq_count(vdev, hdr.index);
848 
849 			if (hdr.flags & VFIO_IRQ_SET_DATA_BOOL)
850 				size = sizeof(uint8_t);
851 			else if (hdr.flags & VFIO_IRQ_SET_DATA_EVENTFD)
852 				size = sizeof(int32_t);
853 			else
854 				return -EINVAL;
855 
856 			if (hdr.argsz - minsz < hdr.count * size ||
857 			    hdr.start >= max || hdr.start + hdr.count > max)
858 				return -EINVAL;
859 
860 			data = memdup_user((void __user *)(arg + minsz),
861 					   hdr.count * size);
862 			if (IS_ERR(data))
863 				return PTR_ERR(data);
864 		}
865 
866 		mutex_lock(&vdev->igate);
867 
868 		ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
869 					      hdr.start, hdr.count, data);
870 
871 		mutex_unlock(&vdev->igate);
872 		kfree(data);
873 
874 		return ret;
875 
876 	} else if (cmd == VFIO_DEVICE_RESET) {
877 		return vdev->reset_works ?
878 			pci_try_reset_function(vdev->pdev) : -EINVAL;
879 
880 	} else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
881 		struct vfio_pci_hot_reset_info hdr;
882 		struct vfio_pci_fill_info fill = { 0 };
883 		struct vfio_pci_dependent_device *devices = NULL;
884 		bool slot = false;
885 		int ret = 0;
886 
887 		minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
888 
889 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
890 			return -EFAULT;
891 
892 		if (hdr.argsz < minsz)
893 			return -EINVAL;
894 
895 		hdr.flags = 0;
896 
897 		/* Can we do a slot or bus reset or neither? */
898 		if (!pci_probe_reset_slot(vdev->pdev->slot))
899 			slot = true;
900 		else if (pci_probe_reset_bus(vdev->pdev->bus))
901 			return -ENODEV;
902 
903 		/* How many devices are affected? */
904 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
905 						    vfio_pci_count_devs,
906 						    &fill.max, slot);
907 		if (ret)
908 			return ret;
909 
910 		WARN_ON(!fill.max); /* Should always be at least one */
911 
912 		/*
913 		 * If there's enough space, fill it now, otherwise return
914 		 * -ENOSPC and the number of devices affected.
915 		 */
916 		if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
917 			ret = -ENOSPC;
918 			hdr.count = fill.max;
919 			goto reset_info_exit;
920 		}
921 
922 		devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
923 		if (!devices)
924 			return -ENOMEM;
925 
926 		fill.devices = devices;
927 
928 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
929 						    vfio_pci_fill_devs,
930 						    &fill, slot);
931 
932 		/*
933 		 * If a device was removed between counting and filling,
934 		 * we may come up short of fill.max.  If a device was
935 		 * added, we'll have a return of -EAGAIN above.
936 		 */
937 		if (!ret)
938 			hdr.count = fill.cur;
939 
940 reset_info_exit:
941 		if (copy_to_user((void __user *)arg, &hdr, minsz))
942 			ret = -EFAULT;
943 
944 		if (!ret) {
945 			if (copy_to_user((void __user *)(arg + minsz), devices,
946 					 hdr.count * sizeof(*devices)))
947 				ret = -EFAULT;
948 		}
949 
950 		kfree(devices);
951 		return ret;
952 
953 	} else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
954 		struct vfio_pci_hot_reset hdr;
955 		int32_t *group_fds;
956 		struct vfio_pci_group_entry *groups;
957 		struct vfio_pci_group_info info;
958 		bool slot = false;
959 		int i, count = 0, ret = 0;
960 
961 		minsz = offsetofend(struct vfio_pci_hot_reset, count);
962 
963 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
964 			return -EFAULT;
965 
966 		if (hdr.argsz < minsz || hdr.flags)
967 			return -EINVAL;
968 
969 		/* Can we do a slot or bus reset or neither? */
970 		if (!pci_probe_reset_slot(vdev->pdev->slot))
971 			slot = true;
972 		else if (pci_probe_reset_bus(vdev->pdev->bus))
973 			return -ENODEV;
974 
975 		/*
976 		 * We can't let userspace give us an arbitrarily large
977 		 * buffer to copy, so verify how many we think there
978 		 * could be.  Note groups can have multiple devices so
979 		 * one group per device is the max.
980 		 */
981 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
982 						    vfio_pci_count_devs,
983 						    &count, slot);
984 		if (ret)
985 			return ret;
986 
987 		/* Somewhere between 1 and count is OK */
988 		if (!hdr.count || hdr.count > count)
989 			return -EINVAL;
990 
991 		group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
992 		groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
993 		if (!group_fds || !groups) {
994 			kfree(group_fds);
995 			kfree(groups);
996 			return -ENOMEM;
997 		}
998 
999 		if (copy_from_user(group_fds, (void __user *)(arg + minsz),
1000 				   hdr.count * sizeof(*group_fds))) {
1001 			kfree(group_fds);
1002 			kfree(groups);
1003 			return -EFAULT;
1004 		}
1005 
1006 		/*
1007 		 * For each group_fd, get the group through the vfio external
1008 		 * user interface and store the group and iommu ID.  This
1009 		 * ensures the group is held across the reset.
1010 		 */
1011 		for (i = 0; i < hdr.count; i++) {
1012 			struct vfio_group *group;
1013 			struct fd f = fdget(group_fds[i]);
1014 			if (!f.file) {
1015 				ret = -EBADF;
1016 				break;
1017 			}
1018 
1019 			group = vfio_group_get_external_user(f.file);
1020 			fdput(f);
1021 			if (IS_ERR(group)) {
1022 				ret = PTR_ERR(group);
1023 				break;
1024 			}
1025 
1026 			groups[i].group = group;
1027 			groups[i].id = vfio_external_user_iommu_id(group);
1028 		}
1029 
1030 		kfree(group_fds);
1031 
1032 		/* release reference to groups on error */
1033 		if (ret)
1034 			goto hot_reset_release;
1035 
1036 		info.count = hdr.count;
1037 		info.groups = groups;
1038 
1039 		/*
1040 		 * Test whether all the affected devices are contained
1041 		 * by the set of groups provided by the user.
1042 		 */
1043 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1044 						    vfio_pci_validate_devs,
1045 						    &info, slot);
1046 		if (!ret)
1047 			/* User has access, do the reset */
1048 			ret = slot ? pci_try_reset_slot(vdev->pdev->slot) :
1049 				     pci_try_reset_bus(vdev->pdev->bus);
1050 
1051 hot_reset_release:
1052 		for (i--; i >= 0; i--)
1053 			vfio_group_put_external_user(groups[i].group);
1054 
1055 		kfree(groups);
1056 		return ret;
1057 	}
1058 
1059 	return -ENOTTY;
1060 }
1061 
1062 static ssize_t vfio_pci_rw(void *device_data, char __user *buf,
1063 			   size_t count, loff_t *ppos, bool iswrite)
1064 {
1065 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1066 	struct vfio_pci_device *vdev = device_data;
1067 
1068 	if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1069 		return -EINVAL;
1070 
1071 	switch (index) {
1072 	case VFIO_PCI_CONFIG_REGION_INDEX:
1073 		return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1074 
1075 	case VFIO_PCI_ROM_REGION_INDEX:
1076 		if (iswrite)
1077 			return -EINVAL;
1078 		return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1079 
1080 	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1081 		return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1082 
1083 	case VFIO_PCI_VGA_REGION_INDEX:
1084 		return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1085 	default:
1086 		index -= VFIO_PCI_NUM_REGIONS;
1087 		return vdev->region[index].ops->rw(vdev, buf,
1088 						   count, ppos, iswrite);
1089 	}
1090 
1091 	return -EINVAL;
1092 }
1093 
1094 static ssize_t vfio_pci_read(void *device_data, char __user *buf,
1095 			     size_t count, loff_t *ppos)
1096 {
1097 	if (!count)
1098 		return 0;
1099 
1100 	return vfio_pci_rw(device_data, buf, count, ppos, false);
1101 }
1102 
1103 static ssize_t vfio_pci_write(void *device_data, const char __user *buf,
1104 			      size_t count, loff_t *ppos)
1105 {
1106 	if (!count)
1107 		return 0;
1108 
1109 	return vfio_pci_rw(device_data, (char __user *)buf, count, ppos, true);
1110 }
1111 
1112 static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma)
1113 {
1114 	struct vfio_pci_device *vdev = device_data;
1115 	struct pci_dev *pdev = vdev->pdev;
1116 	unsigned int index;
1117 	u64 phys_len, req_len, pgoff, req_start;
1118 	int ret;
1119 
1120 	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1121 
1122 	if (vma->vm_end < vma->vm_start)
1123 		return -EINVAL;
1124 	if ((vma->vm_flags & VM_SHARED) == 0)
1125 		return -EINVAL;
1126 	if (index >= VFIO_PCI_ROM_REGION_INDEX)
1127 		return -EINVAL;
1128 	if (!vdev->bar_mmap_supported[index])
1129 		return -EINVAL;
1130 
1131 	phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1132 	req_len = vma->vm_end - vma->vm_start;
1133 	pgoff = vma->vm_pgoff &
1134 		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1135 	req_start = pgoff << PAGE_SHIFT;
1136 
1137 	if (req_start + req_len > phys_len)
1138 		return -EINVAL;
1139 
1140 	if (index == vdev->msix_bar) {
1141 		/*
1142 		 * Disallow mmaps overlapping the MSI-X table; users don't
1143 		 * get to touch this directly.  We could find somewhere
1144 		 * else to map the overlap, but page granularity is only
1145 		 * a recommendation, not a requirement, so the user needs
1146 		 * to know which bits are real.  Requiring them to mmap
1147 		 * around the table makes that clear.
1148 		 */
1149 
1150 		/* If neither entirely above nor below, then it overlaps */
1151 		if (!(req_start >= vdev->msix_offset + vdev->msix_size ||
1152 		      req_start + req_len <= vdev->msix_offset))
1153 			return -EINVAL;
1154 	}
1155 
1156 	/*
1157 	 * Even though we don't make use of the barmap for the mmap,
1158 	 * we need to request the region and the barmap tracks that.
1159 	 */
1160 	if (!vdev->barmap[index]) {
1161 		ret = pci_request_selected_regions(pdev,
1162 						   1 << index, "vfio-pci");
1163 		if (ret)
1164 			return ret;
1165 
1166 		vdev->barmap[index] = pci_iomap(pdev, index, 0);
1167 	}
1168 
1169 	vma->vm_private_data = vdev;
1170 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1171 	vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1172 
1173 	return remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1174 			       req_len, vma->vm_page_prot);
1175 }
1176 
1177 static void vfio_pci_request(void *device_data, unsigned int count)
1178 {
1179 	struct vfio_pci_device *vdev = device_data;
1180 
1181 	mutex_lock(&vdev->igate);
1182 
1183 	if (vdev->req_trigger) {
1184 		if (!(count % 10))
1185 			dev_notice_ratelimited(&vdev->pdev->dev,
1186 				"Relaying device request to user (#%u)\n",
1187 				count);
1188 		eventfd_signal(vdev->req_trigger, 1);
1189 	} else if (count == 0) {
1190 		dev_warn(&vdev->pdev->dev,
1191 			"No device request channel registered, blocked until released by user\n");
1192 	}
1193 
1194 	mutex_unlock(&vdev->igate);
1195 }
1196 
1197 static const struct vfio_device_ops vfio_pci_ops = {
1198 	.name		= "vfio-pci",
1199 	.open		= vfio_pci_open,
1200 	.release	= vfio_pci_release,
1201 	.ioctl		= vfio_pci_ioctl,
1202 	.read		= vfio_pci_read,
1203 	.write		= vfio_pci_write,
1204 	.mmap		= vfio_pci_mmap,
1205 	.request	= vfio_pci_request,
1206 };
1207 
1208 static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1209 {
1210 	struct vfio_pci_device *vdev;
1211 	struct iommu_group *group;
1212 	int ret;
1213 
1214 	if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1215 		return -EINVAL;
1216 
1217 	group = vfio_iommu_group_get(&pdev->dev);
1218 	if (!group)
1219 		return -EINVAL;
1220 
1221 	vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
1222 	if (!vdev) {
1223 		vfio_iommu_group_put(group, &pdev->dev);
1224 		return -ENOMEM;
1225 	}
1226 
1227 	vdev->pdev = pdev;
1228 	vdev->irq_type = VFIO_PCI_NUM_IRQS;
1229 	mutex_init(&vdev->igate);
1230 	spin_lock_init(&vdev->irqlock);
1231 
1232 	ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
1233 	if (ret) {
1234 		vfio_iommu_group_put(group, &pdev->dev);
1235 		kfree(vdev);
1236 		return ret;
1237 	}
1238 
1239 	if (vfio_pci_is_vga(pdev)) {
1240 		vga_client_register(pdev, vdev, NULL, vfio_pci_set_vga_decode);
1241 		vga_set_legacy_decoding(pdev,
1242 					vfio_pci_set_vga_decode(vdev, false));
1243 	}
1244 
1245 	if (!disable_idle_d3) {
1246 		/*
1247 		 * pci-core sets the device power state to an unknown value at
1248 		 * bootup and after being removed from a driver.  The only
1249 		 * transition it allows from this unknown state is to D0, which
1250 		 * typically happens when a driver calls pci_enable_device().
1251 		 * We're not ready to enable the device yet, but we do want to
1252 		 * be able to get to D3.  Therefore first do a D0 transition
1253 		 * before going to D3.
1254 		 */
1255 		pci_set_power_state(pdev, PCI_D0);
1256 		pci_set_power_state(pdev, PCI_D3hot);
1257 	}
1258 
1259 	return ret;
1260 }
1261 
1262 static void vfio_pci_remove(struct pci_dev *pdev)
1263 {
1264 	struct vfio_pci_device *vdev;
1265 
1266 	vdev = vfio_del_group_dev(&pdev->dev);
1267 	if (!vdev)
1268 		return;
1269 
1270 	vfio_iommu_group_put(pdev->dev.iommu_group, &pdev->dev);
1271 	kfree(vdev->region);
1272 	kfree(vdev);
1273 
1274 	if (vfio_pci_is_vga(pdev)) {
1275 		vga_client_register(pdev, NULL, NULL, NULL);
1276 		vga_set_legacy_decoding(pdev,
1277 				VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
1278 				VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM);
1279 	}
1280 
1281 	if (!disable_idle_d3)
1282 		pci_set_power_state(pdev, PCI_D0);
1283 }
1284 
1285 static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
1286 						  pci_channel_state_t state)
1287 {
1288 	struct vfio_pci_device *vdev;
1289 	struct vfio_device *device;
1290 
1291 	device = vfio_device_get_from_dev(&pdev->dev);
1292 	if (device == NULL)
1293 		return PCI_ERS_RESULT_DISCONNECT;
1294 
1295 	vdev = vfio_device_data(device);
1296 	if (vdev == NULL) {
1297 		vfio_device_put(device);
1298 		return PCI_ERS_RESULT_DISCONNECT;
1299 	}
1300 
1301 	mutex_lock(&vdev->igate);
1302 
1303 	if (vdev->err_trigger)
1304 		eventfd_signal(vdev->err_trigger, 1);
1305 
1306 	mutex_unlock(&vdev->igate);
1307 
1308 	vfio_device_put(device);
1309 
1310 	return PCI_ERS_RESULT_CAN_RECOVER;
1311 }
1312 
1313 static const struct pci_error_handlers vfio_err_handlers = {
1314 	.error_detected = vfio_pci_aer_err_detected,
1315 };
1316 
1317 static struct pci_driver vfio_pci_driver = {
1318 	.name		= "vfio-pci",
1319 	.id_table	= NULL, /* only dynamic ids */
1320 	.probe		= vfio_pci_probe,
1321 	.remove		= vfio_pci_remove,
1322 	.err_handler	= &vfio_err_handlers,
1323 };
1324 
1325 struct vfio_devices {
1326 	struct vfio_device **devices;
1327 	int cur_index;
1328 	int max_index;
1329 };
1330 
1331 static int vfio_pci_get_devs(struct pci_dev *pdev, void *data)
1332 {
1333 	struct vfio_devices *devs = data;
1334 	struct vfio_device *device;
1335 
1336 	if (devs->cur_index == devs->max_index)
1337 		return -ENOSPC;
1338 
1339 	device = vfio_device_get_from_dev(&pdev->dev);
1340 	if (!device)
1341 		return -EINVAL;
1342 
1343 	if (pci_dev_driver(pdev) != &vfio_pci_driver) {
1344 		vfio_device_put(device);
1345 		return -EBUSY;
1346 	}
1347 
1348 	devs->devices[devs->cur_index++] = device;
1349 	return 0;
1350 }
1351 
1352 /*
1353  * Attempt to do a bus/slot reset if there are devices affected by a reset for
1354  * this device that are needs_reset and all of the affected devices are unused
1355  * (!refcnt).  Callers are required to hold driver_lock when calling this to
1356  * prevent device opens and concurrent bus reset attempts.  We prevent device
1357  * unbinds by acquiring and holding a reference to the vfio_device.
1358  *
1359  * NB: vfio-core considers a group to be viable even if some devices are
1360  * bound to drivers like pci-stub or pcieport.  Here we require all devices
1361  * to be bound to vfio_pci since that's the only way we can be sure they
1362  * stay put.
1363  */
1364 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev)
1365 {
1366 	struct vfio_devices devs = { .cur_index = 0 };
1367 	int i = 0, ret = -EINVAL;
1368 	bool needs_reset = false, slot = false;
1369 	struct vfio_pci_device *tmp;
1370 
1371 	if (!pci_probe_reset_slot(vdev->pdev->slot))
1372 		slot = true;
1373 	else if (pci_probe_reset_bus(vdev->pdev->bus))
1374 		return;
1375 
1376 	if (vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
1377 					  &i, slot) || !i)
1378 		return;
1379 
1380 	devs.max_index = i;
1381 	devs.devices = kcalloc(i, sizeof(struct vfio_device *), GFP_KERNEL);
1382 	if (!devs.devices)
1383 		return;
1384 
1385 	if (vfio_pci_for_each_slot_or_bus(vdev->pdev,
1386 					  vfio_pci_get_devs, &devs, slot))
1387 		goto put_devs;
1388 
1389 	for (i = 0; i < devs.cur_index; i++) {
1390 		tmp = vfio_device_data(devs.devices[i]);
1391 		if (tmp->needs_reset)
1392 			needs_reset = true;
1393 		if (tmp->refcnt)
1394 			goto put_devs;
1395 	}
1396 
1397 	if (needs_reset)
1398 		ret = slot ? pci_try_reset_slot(vdev->pdev->slot) :
1399 			     pci_try_reset_bus(vdev->pdev->bus);
1400 
1401 put_devs:
1402 	for (i = 0; i < devs.cur_index; i++) {
1403 		tmp = vfio_device_data(devs.devices[i]);
1404 		if (!ret)
1405 			tmp->needs_reset = false;
1406 
1407 		if (!tmp->refcnt && !disable_idle_d3)
1408 			pci_set_power_state(tmp->pdev, PCI_D3hot);
1409 
1410 		vfio_device_put(devs.devices[i]);
1411 	}
1412 
1413 	kfree(devs.devices);
1414 }
1415 
1416 static void __exit vfio_pci_cleanup(void)
1417 {
1418 	pci_unregister_driver(&vfio_pci_driver);
1419 	vfio_pci_uninit_perm_bits();
1420 }
1421 
1422 static void __init vfio_pci_fill_ids(void)
1423 {
1424 	char *p, *id;
1425 	int rc;
1426 
1427 	/* no ids passed actually */
1428 	if (ids[0] == '\0')
1429 		return;
1430 
1431 	/* add ids specified in the module parameter */
1432 	p = ids;
1433 	while ((id = strsep(&p, ","))) {
1434 		unsigned int vendor, device, subvendor = PCI_ANY_ID,
1435 			subdevice = PCI_ANY_ID, class = 0, class_mask = 0;
1436 		int fields;
1437 
1438 		if (!strlen(id))
1439 			continue;
1440 
1441 		fields = sscanf(id, "%x:%x:%x:%x:%x:%x",
1442 				&vendor, &device, &subvendor, &subdevice,
1443 				&class, &class_mask);
1444 
1445 		if (fields < 2) {
1446 			pr_warn("invalid id string \"%s\"\n", id);
1447 			continue;
1448 		}
1449 
1450 		rc = pci_add_dynid(&vfio_pci_driver, vendor, device,
1451 				   subvendor, subdevice, class, class_mask, 0);
1452 		if (rc)
1453 			pr_warn("failed to add dynamic id [%04hx:%04hx[%04hx:%04hx]] class %#08x/%08x (%d)\n",
1454 				vendor, device, subvendor, subdevice,
1455 				class, class_mask, rc);
1456 		else
1457 			pr_info("add [%04hx:%04hx[%04hx:%04hx]] class %#08x/%08x\n",
1458 				vendor, device, subvendor, subdevice,
1459 				class, class_mask);
1460 	}
1461 }
1462 
1463 static int __init vfio_pci_init(void)
1464 {
1465 	int ret;
1466 
1467 	/* Allocate shared config space permision data used by all devices */
1468 	ret = vfio_pci_init_perm_bits();
1469 	if (ret)
1470 		return ret;
1471 
1472 	/* Register and scan for devices */
1473 	ret = pci_register_driver(&vfio_pci_driver);
1474 	if (ret)
1475 		goto out_driver;
1476 
1477 	vfio_pci_fill_ids();
1478 
1479 	return 0;
1480 
1481 out_driver:
1482 	vfio_pci_uninit_perm_bits();
1483 	return ret;
1484 }
1485 
1486 module_init(vfio_pci_init);
1487 module_exit(vfio_pci_cleanup);
1488 
1489 MODULE_VERSION(DRIVER_VERSION);
1490 MODULE_LICENSE("GPL v2");
1491 MODULE_AUTHOR(DRIVER_AUTHOR);
1492 MODULE_DESCRIPTION(DRIVER_DESC);
1493