xref: /openbmc/linux/drivers/vfio/pci/vfio_pci.c (revision 0c874100)
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 #include <linux/nospec.h>
32 
33 #include "vfio_pci_private.h"
34 
35 #define DRIVER_VERSION  "0.2"
36 #define DRIVER_AUTHOR   "Alex Williamson <alex.williamson@redhat.com>"
37 #define DRIVER_DESC     "VFIO PCI - User Level meta-driver"
38 
39 static char ids[1024] __initdata;
40 module_param_string(ids, ids, sizeof(ids), 0);
41 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");
42 
43 static bool nointxmask;
44 module_param_named(nointxmask, nointxmask, bool, S_IRUGO | S_IWUSR);
45 MODULE_PARM_DESC(nointxmask,
46 		  "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.");
47 
48 #ifdef CONFIG_VFIO_PCI_VGA
49 static bool disable_vga;
50 module_param(disable_vga, bool, S_IRUGO);
51 MODULE_PARM_DESC(disable_vga, "Disable VGA resource access through vfio-pci");
52 #endif
53 
54 static bool disable_idle_d3;
55 module_param(disable_idle_d3, bool, S_IRUGO | S_IWUSR);
56 MODULE_PARM_DESC(disable_idle_d3,
57 		 "Disable using the PCI D3 low power state for idle, unused devices");
58 
59 static DEFINE_MUTEX(driver_lock);
60 
61 static inline bool vfio_vga_disabled(void)
62 {
63 #ifdef CONFIG_VFIO_PCI_VGA
64 	return disable_vga;
65 #else
66 	return true;
67 #endif
68 }
69 
70 /*
71  * Our VGA arbiter participation is limited since we don't know anything
72  * about the device itself.  However, if the device is the only VGA device
73  * downstream of a bridge and VFIO VGA support is disabled, then we can
74  * safely return legacy VGA IO and memory as not decoded since the user
75  * has no way to get to it and routing can be disabled externally at the
76  * bridge.
77  */
78 static unsigned int vfio_pci_set_vga_decode(void *opaque, bool single_vga)
79 {
80 	struct vfio_pci_device *vdev = opaque;
81 	struct pci_dev *tmp = NULL, *pdev = vdev->pdev;
82 	unsigned char max_busnr;
83 	unsigned int decodes;
84 
85 	if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
86 		return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
87 		       VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
88 
89 	max_busnr = pci_bus_max_busnr(pdev->bus);
90 	decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
91 
92 	while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
93 		if (tmp == pdev ||
94 		    pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
95 		    pci_is_root_bus(tmp->bus))
96 			continue;
97 
98 		if (tmp->bus->number >= pdev->bus->number &&
99 		    tmp->bus->number <= max_busnr) {
100 			pci_dev_put(tmp);
101 			decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
102 			break;
103 		}
104 	}
105 
106 	return decodes;
107 }
108 
109 static inline bool vfio_pci_is_vga(struct pci_dev *pdev)
110 {
111 	return (pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA;
112 }
113 
114 static void vfio_pci_probe_mmaps(struct vfio_pci_device *vdev)
115 {
116 	struct resource *res;
117 	int bar;
118 	struct vfio_pci_dummy_resource *dummy_res;
119 
120 	INIT_LIST_HEAD(&vdev->dummy_resources_list);
121 
122 	for (bar = PCI_STD_RESOURCES; bar <= PCI_STD_RESOURCE_END; bar++) {
123 		res = vdev->pdev->resource + bar;
124 
125 		if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
126 			goto no_mmap;
127 
128 		if (!(res->flags & IORESOURCE_MEM))
129 			goto no_mmap;
130 
131 		/*
132 		 * The PCI core shouldn't set up a resource with a
133 		 * type but zero size. But there may be bugs that
134 		 * cause us to do that.
135 		 */
136 		if (!resource_size(res))
137 			goto no_mmap;
138 
139 		if (resource_size(res) >= PAGE_SIZE) {
140 			vdev->bar_mmap_supported[bar] = true;
141 			continue;
142 		}
143 
144 		if (!(res->start & ~PAGE_MASK)) {
145 			/*
146 			 * Add a dummy resource to reserve the remainder
147 			 * of the exclusive page in case that hot-add
148 			 * device's bar is assigned into it.
149 			 */
150 			dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
151 			if (dummy_res == NULL)
152 				goto no_mmap;
153 
154 			dummy_res->resource.name = "vfio sub-page reserved";
155 			dummy_res->resource.start = res->end + 1;
156 			dummy_res->resource.end = res->start + PAGE_SIZE - 1;
157 			dummy_res->resource.flags = res->flags;
158 			if (request_resource(res->parent,
159 						&dummy_res->resource)) {
160 				kfree(dummy_res);
161 				goto no_mmap;
162 			}
163 			dummy_res->index = bar;
164 			list_add(&dummy_res->res_next,
165 					&vdev->dummy_resources_list);
166 			vdev->bar_mmap_supported[bar] = true;
167 			continue;
168 		}
169 		/*
170 		 * Here we don't handle the case when the BAR is not page
171 		 * aligned because we can't expect the BAR will be
172 		 * assigned into the same location in a page in guest
173 		 * when we passthrough the BAR. And it's hard to access
174 		 * this BAR in userspace because we have no way to get
175 		 * the BAR's location in a page.
176 		 */
177 no_mmap:
178 		vdev->bar_mmap_supported[bar] = false;
179 	}
180 }
181 
182 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev);
183 static void vfio_pci_disable(struct vfio_pci_device *vdev);
184 
185 /*
186  * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
187  * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
188  * If a device implements the former but not the latter we would typically
189  * expect broken_intx_masking be set and require an exclusive interrupt.
190  * However since we do have control of the device's ability to assert INTx,
191  * we can instead pretend that the device does not implement INTx, virtualizing
192  * the pin register to report zero and maintaining DisINTx set on the host.
193  */
194 static bool vfio_pci_nointx(struct pci_dev *pdev)
195 {
196 	switch (pdev->vendor) {
197 	case PCI_VENDOR_ID_INTEL:
198 		switch (pdev->device) {
199 		/* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
200 		case 0x1572:
201 		case 0x1574:
202 		case 0x1580 ... 0x1581:
203 		case 0x1583 ... 0x158b:
204 		case 0x37d0 ... 0x37d2:
205 			return true;
206 		default:
207 			return false;
208 		}
209 	}
210 
211 	return false;
212 }
213 
214 static int vfio_pci_enable(struct vfio_pci_device *vdev)
215 {
216 	struct pci_dev *pdev = vdev->pdev;
217 	int ret;
218 	u16 cmd;
219 	u8 msix_pos;
220 
221 	pci_set_power_state(pdev, PCI_D0);
222 
223 	/* Don't allow our initial saved state to include busmaster */
224 	pci_clear_master(pdev);
225 
226 	ret = pci_enable_device(pdev);
227 	if (ret)
228 		return ret;
229 
230 	/* If reset fails because of the device lock, fail this path entirely */
231 	ret = pci_try_reset_function(pdev);
232 	if (ret == -EAGAIN) {
233 		pci_disable_device(pdev);
234 		return ret;
235 	}
236 
237 	vdev->reset_works = !ret;
238 	pci_save_state(pdev);
239 	vdev->pci_saved_state = pci_store_saved_state(pdev);
240 	if (!vdev->pci_saved_state)
241 		pr_debug("%s: Couldn't store %s saved state\n",
242 			 __func__, dev_name(&pdev->dev));
243 
244 	if (likely(!nointxmask)) {
245 		if (vfio_pci_nointx(pdev)) {
246 			dev_info(&pdev->dev, "Masking broken INTx support\n");
247 			vdev->nointx = true;
248 			pci_intx(pdev, 0);
249 		} else
250 			vdev->pci_2_3 = pci_intx_mask_supported(pdev);
251 	}
252 
253 	pci_read_config_word(pdev, PCI_COMMAND, &cmd);
254 	if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
255 		cmd &= ~PCI_COMMAND_INTX_DISABLE;
256 		pci_write_config_word(pdev, PCI_COMMAND, cmd);
257 	}
258 
259 	ret = vfio_config_init(vdev);
260 	if (ret) {
261 		kfree(vdev->pci_saved_state);
262 		vdev->pci_saved_state = NULL;
263 		pci_disable_device(pdev);
264 		return ret;
265 	}
266 
267 	msix_pos = pdev->msix_cap;
268 	if (msix_pos) {
269 		u16 flags;
270 		u32 table;
271 
272 		pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
273 		pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
274 
275 		vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
276 		vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
277 		vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
278 	} else
279 		vdev->msix_bar = 0xFF;
280 
281 	if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
282 		vdev->has_vga = true;
283 
284 
285 	if (vfio_pci_is_vga(pdev) &&
286 	    pdev->vendor == PCI_VENDOR_ID_INTEL &&
287 	    IS_ENABLED(CONFIG_VFIO_PCI_IGD)) {
288 		ret = vfio_pci_igd_init(vdev);
289 		if (ret) {
290 			dev_warn(&vdev->pdev->dev,
291 				 "Failed to setup Intel IGD regions\n");
292 			vfio_pci_disable(vdev);
293 			return ret;
294 		}
295 	}
296 
297 	vfio_pci_probe_mmaps(vdev);
298 
299 	return 0;
300 }
301 
302 static void vfio_pci_disable(struct vfio_pci_device *vdev)
303 {
304 	struct pci_dev *pdev = vdev->pdev;
305 	struct vfio_pci_dummy_resource *dummy_res, *tmp;
306 	struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
307 	int i, bar;
308 
309 	/* Stop the device from further DMA */
310 	pci_clear_master(pdev);
311 
312 	vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
313 				VFIO_IRQ_SET_ACTION_TRIGGER,
314 				vdev->irq_type, 0, 0, NULL);
315 
316 	/* Device closed, don't need mutex here */
317 	list_for_each_entry_safe(ioeventfd, ioeventfd_tmp,
318 				 &vdev->ioeventfds_list, next) {
319 		vfio_virqfd_disable(&ioeventfd->virqfd);
320 		list_del(&ioeventfd->next);
321 		kfree(ioeventfd);
322 	}
323 	vdev->ioeventfds_nr = 0;
324 
325 	vdev->virq_disabled = false;
326 
327 	for (i = 0; i < vdev->num_regions; i++)
328 		vdev->region[i].ops->release(vdev, &vdev->region[i]);
329 
330 	vdev->num_regions = 0;
331 	kfree(vdev->region);
332 	vdev->region = NULL; /* don't krealloc a freed pointer */
333 
334 	vfio_config_free(vdev);
335 
336 	for (bar = PCI_STD_RESOURCES; bar <= PCI_STD_RESOURCE_END; bar++) {
337 		if (!vdev->barmap[bar])
338 			continue;
339 		pci_iounmap(pdev, vdev->barmap[bar]);
340 		pci_release_selected_regions(pdev, 1 << bar);
341 		vdev->barmap[bar] = NULL;
342 	}
343 
344 	list_for_each_entry_safe(dummy_res, tmp,
345 				 &vdev->dummy_resources_list, res_next) {
346 		list_del(&dummy_res->res_next);
347 		release_resource(&dummy_res->resource);
348 		kfree(dummy_res);
349 	}
350 
351 	vdev->needs_reset = true;
352 
353 	/*
354 	 * If we have saved state, restore it.  If we can reset the device,
355 	 * even better.  Resetting with current state seems better than
356 	 * nothing, but saving and restoring current state without reset
357 	 * is just busy work.
358 	 */
359 	if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
360 		pr_info("%s: Couldn't reload %s saved state\n",
361 			__func__, dev_name(&pdev->dev));
362 
363 		if (!vdev->reset_works)
364 			goto out;
365 
366 		pci_save_state(pdev);
367 	}
368 
369 	/*
370 	 * Disable INTx and MSI, presumably to avoid spurious interrupts
371 	 * during reset.  Stolen from pci_reset_function()
372 	 */
373 	pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
374 
375 	/*
376 	 * Try to reset the device.  The success of this is dependent on
377 	 * being able to lock the device, which is not always possible.
378 	 */
379 	if (vdev->reset_works && !pci_try_reset_function(pdev))
380 		vdev->needs_reset = false;
381 
382 	pci_restore_state(pdev);
383 out:
384 	pci_disable_device(pdev);
385 
386 	vfio_pci_try_bus_reset(vdev);
387 
388 	if (!disable_idle_d3)
389 		pci_set_power_state(pdev, PCI_D3hot);
390 }
391 
392 static void vfio_pci_release(void *device_data)
393 {
394 	struct vfio_pci_device *vdev = device_data;
395 
396 	mutex_lock(&driver_lock);
397 
398 	if (!(--vdev->refcnt)) {
399 		vfio_spapr_pci_eeh_release(vdev->pdev);
400 		vfio_pci_disable(vdev);
401 	}
402 
403 	mutex_unlock(&driver_lock);
404 
405 	module_put(THIS_MODULE);
406 }
407 
408 static int vfio_pci_open(void *device_data)
409 {
410 	struct vfio_pci_device *vdev = device_data;
411 	int ret = 0;
412 
413 	if (!try_module_get(THIS_MODULE))
414 		return -ENODEV;
415 
416 	mutex_lock(&driver_lock);
417 
418 	if (!vdev->refcnt) {
419 		ret = vfio_pci_enable(vdev);
420 		if (ret)
421 			goto error;
422 
423 		vfio_spapr_pci_eeh_open(vdev->pdev);
424 	}
425 	vdev->refcnt++;
426 error:
427 	mutex_unlock(&driver_lock);
428 	if (ret)
429 		module_put(THIS_MODULE);
430 	return ret;
431 }
432 
433 static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
434 {
435 	if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
436 		u8 pin;
437 
438 		if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
439 		    vdev->nointx || vdev->pdev->is_virtfn)
440 			return 0;
441 
442 		pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
443 
444 		return pin ? 1 : 0;
445 	} else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
446 		u8 pos;
447 		u16 flags;
448 
449 		pos = vdev->pdev->msi_cap;
450 		if (pos) {
451 			pci_read_config_word(vdev->pdev,
452 					     pos + PCI_MSI_FLAGS, &flags);
453 			return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
454 		}
455 	} else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
456 		u8 pos;
457 		u16 flags;
458 
459 		pos = vdev->pdev->msix_cap;
460 		if (pos) {
461 			pci_read_config_word(vdev->pdev,
462 					     pos + PCI_MSIX_FLAGS, &flags);
463 
464 			return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
465 		}
466 	} else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
467 		if (pci_is_pcie(vdev->pdev))
468 			return 1;
469 	} else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
470 		return 1;
471 	}
472 
473 	return 0;
474 }
475 
476 static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
477 {
478 	(*(int *)data)++;
479 	return 0;
480 }
481 
482 struct vfio_pci_fill_info {
483 	int max;
484 	int cur;
485 	struct vfio_pci_dependent_device *devices;
486 };
487 
488 static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
489 {
490 	struct vfio_pci_fill_info *fill = data;
491 	struct iommu_group *iommu_group;
492 
493 	if (fill->cur == fill->max)
494 		return -EAGAIN; /* Something changed, try again */
495 
496 	iommu_group = iommu_group_get(&pdev->dev);
497 	if (!iommu_group)
498 		return -EPERM; /* Cannot reset non-isolated devices */
499 
500 	fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
501 	fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
502 	fill->devices[fill->cur].bus = pdev->bus->number;
503 	fill->devices[fill->cur].devfn = pdev->devfn;
504 	fill->cur++;
505 	iommu_group_put(iommu_group);
506 	return 0;
507 }
508 
509 struct vfio_pci_group_entry {
510 	struct vfio_group *group;
511 	int id;
512 };
513 
514 struct vfio_pci_group_info {
515 	int count;
516 	struct vfio_pci_group_entry *groups;
517 };
518 
519 static int vfio_pci_validate_devs(struct pci_dev *pdev, void *data)
520 {
521 	struct vfio_pci_group_info *info = data;
522 	struct iommu_group *group;
523 	int id, i;
524 
525 	group = iommu_group_get(&pdev->dev);
526 	if (!group)
527 		return -EPERM;
528 
529 	id = iommu_group_id(group);
530 
531 	for (i = 0; i < info->count; i++)
532 		if (info->groups[i].id == id)
533 			break;
534 
535 	iommu_group_put(group);
536 
537 	return (i == info->count) ? -EINVAL : 0;
538 }
539 
540 static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
541 {
542 	for (; pdev; pdev = pdev->bus->self)
543 		if (pdev->bus == slot->bus)
544 			return (pdev->slot == slot);
545 	return false;
546 }
547 
548 struct vfio_pci_walk_info {
549 	int (*fn)(struct pci_dev *, void *data);
550 	void *data;
551 	struct pci_dev *pdev;
552 	bool slot;
553 	int ret;
554 };
555 
556 static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
557 {
558 	struct vfio_pci_walk_info *walk = data;
559 
560 	if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
561 		walk->ret = walk->fn(pdev, walk->data);
562 
563 	return walk->ret;
564 }
565 
566 static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
567 					 int (*fn)(struct pci_dev *,
568 						   void *data), void *data,
569 					 bool slot)
570 {
571 	struct vfio_pci_walk_info walk = {
572 		.fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
573 	};
574 
575 	pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
576 
577 	return walk.ret;
578 }
579 
580 static int msix_mmappable_cap(struct vfio_pci_device *vdev,
581 			      struct vfio_info_cap *caps)
582 {
583 	struct vfio_info_cap_header header = {
584 		.id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
585 		.version = 1
586 	};
587 
588 	return vfio_info_add_capability(caps, &header, sizeof(header));
589 }
590 
591 int vfio_pci_register_dev_region(struct vfio_pci_device *vdev,
592 				 unsigned int type, unsigned int subtype,
593 				 const struct vfio_pci_regops *ops,
594 				 size_t size, u32 flags, void *data)
595 {
596 	struct vfio_pci_region *region;
597 
598 	region = krealloc(vdev->region,
599 			  (vdev->num_regions + 1) * sizeof(*region),
600 			  GFP_KERNEL);
601 	if (!region)
602 		return -ENOMEM;
603 
604 	vdev->region = region;
605 	vdev->region[vdev->num_regions].type = type;
606 	vdev->region[vdev->num_regions].subtype = subtype;
607 	vdev->region[vdev->num_regions].ops = ops;
608 	vdev->region[vdev->num_regions].size = size;
609 	vdev->region[vdev->num_regions].flags = flags;
610 	vdev->region[vdev->num_regions].data = data;
611 
612 	vdev->num_regions++;
613 
614 	return 0;
615 }
616 
617 static long vfio_pci_ioctl(void *device_data,
618 			   unsigned int cmd, unsigned long arg)
619 {
620 	struct vfio_pci_device *vdev = device_data;
621 	unsigned long minsz;
622 
623 	if (cmd == VFIO_DEVICE_GET_INFO) {
624 		struct vfio_device_info info;
625 
626 		minsz = offsetofend(struct vfio_device_info, num_irqs);
627 
628 		if (copy_from_user(&info, (void __user *)arg, minsz))
629 			return -EFAULT;
630 
631 		if (info.argsz < minsz)
632 			return -EINVAL;
633 
634 		info.flags = VFIO_DEVICE_FLAGS_PCI;
635 
636 		if (vdev->reset_works)
637 			info.flags |= VFIO_DEVICE_FLAGS_RESET;
638 
639 		info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
640 		info.num_irqs = VFIO_PCI_NUM_IRQS;
641 
642 		return copy_to_user((void __user *)arg, &info, minsz) ?
643 			-EFAULT : 0;
644 
645 	} else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
646 		struct pci_dev *pdev = vdev->pdev;
647 		struct vfio_region_info info;
648 		struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
649 		int i, ret;
650 
651 		minsz = offsetofend(struct vfio_region_info, offset);
652 
653 		if (copy_from_user(&info, (void __user *)arg, minsz))
654 			return -EFAULT;
655 
656 		if (info.argsz < minsz)
657 			return -EINVAL;
658 
659 		switch (info.index) {
660 		case VFIO_PCI_CONFIG_REGION_INDEX:
661 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
662 			info.size = pdev->cfg_size;
663 			info.flags = VFIO_REGION_INFO_FLAG_READ |
664 				     VFIO_REGION_INFO_FLAG_WRITE;
665 			break;
666 		case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
667 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
668 			info.size = pci_resource_len(pdev, info.index);
669 			if (!info.size) {
670 				info.flags = 0;
671 				break;
672 			}
673 
674 			info.flags = VFIO_REGION_INFO_FLAG_READ |
675 				     VFIO_REGION_INFO_FLAG_WRITE;
676 			if (vdev->bar_mmap_supported[info.index]) {
677 				info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
678 				if (info.index == vdev->msix_bar) {
679 					ret = msix_mmappable_cap(vdev, &caps);
680 					if (ret)
681 						return ret;
682 				}
683 			}
684 
685 			break;
686 		case VFIO_PCI_ROM_REGION_INDEX:
687 		{
688 			void __iomem *io;
689 			size_t size;
690 
691 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
692 			info.flags = 0;
693 
694 			/* Report the BAR size, not the ROM size */
695 			info.size = pci_resource_len(pdev, info.index);
696 			if (!info.size) {
697 				/* Shadow ROMs appear as PCI option ROMs */
698 				if (pdev->resource[PCI_ROM_RESOURCE].flags &
699 							IORESOURCE_ROM_SHADOW)
700 					info.size = 0x20000;
701 				else
702 					break;
703 			}
704 
705 			/* Is it really there? */
706 			io = pci_map_rom(pdev, &size);
707 			if (!io || !size) {
708 				info.size = 0;
709 				break;
710 			}
711 			pci_unmap_rom(pdev, io);
712 
713 			info.flags = VFIO_REGION_INFO_FLAG_READ;
714 			break;
715 		}
716 		case VFIO_PCI_VGA_REGION_INDEX:
717 			if (!vdev->has_vga)
718 				return -EINVAL;
719 
720 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
721 			info.size = 0xc0000;
722 			info.flags = VFIO_REGION_INFO_FLAG_READ |
723 				     VFIO_REGION_INFO_FLAG_WRITE;
724 
725 			break;
726 		default:
727 		{
728 			struct vfio_region_info_cap_type cap_type = {
729 					.header.id = VFIO_REGION_INFO_CAP_TYPE,
730 					.header.version = 1 };
731 
732 			if (info.index >=
733 			    VFIO_PCI_NUM_REGIONS + vdev->num_regions)
734 				return -EINVAL;
735 			info.index = array_index_nospec(info.index,
736 							VFIO_PCI_NUM_REGIONS +
737 							vdev->num_regions);
738 
739 			i = info.index - VFIO_PCI_NUM_REGIONS;
740 
741 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
742 			info.size = vdev->region[i].size;
743 			info.flags = vdev->region[i].flags;
744 
745 			cap_type.type = vdev->region[i].type;
746 			cap_type.subtype = vdev->region[i].subtype;
747 
748 			ret = vfio_info_add_capability(&caps, &cap_type.header,
749 						       sizeof(cap_type));
750 			if (ret)
751 				return ret;
752 
753 		}
754 		}
755 
756 		if (caps.size) {
757 			info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
758 			if (info.argsz < sizeof(info) + caps.size) {
759 				info.argsz = sizeof(info) + caps.size;
760 				info.cap_offset = 0;
761 			} else {
762 				vfio_info_cap_shift(&caps, sizeof(info));
763 				if (copy_to_user((void __user *)arg +
764 						  sizeof(info), caps.buf,
765 						  caps.size)) {
766 					kfree(caps.buf);
767 					return -EFAULT;
768 				}
769 				info.cap_offset = sizeof(info);
770 			}
771 
772 			kfree(caps.buf);
773 		}
774 
775 		return copy_to_user((void __user *)arg, &info, minsz) ?
776 			-EFAULT : 0;
777 
778 	} else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
779 		struct vfio_irq_info info;
780 
781 		minsz = offsetofend(struct vfio_irq_info, count);
782 
783 		if (copy_from_user(&info, (void __user *)arg, minsz))
784 			return -EFAULT;
785 
786 		if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
787 			return -EINVAL;
788 
789 		switch (info.index) {
790 		case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
791 		case VFIO_PCI_REQ_IRQ_INDEX:
792 			break;
793 		case VFIO_PCI_ERR_IRQ_INDEX:
794 			if (pci_is_pcie(vdev->pdev))
795 				break;
796 		/* fall through */
797 		default:
798 			return -EINVAL;
799 		}
800 
801 		info.flags = VFIO_IRQ_INFO_EVENTFD;
802 
803 		info.count = vfio_pci_get_irq_count(vdev, info.index);
804 
805 		if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
806 			info.flags |= (VFIO_IRQ_INFO_MASKABLE |
807 				       VFIO_IRQ_INFO_AUTOMASKED);
808 		else
809 			info.flags |= VFIO_IRQ_INFO_NORESIZE;
810 
811 		return copy_to_user((void __user *)arg, &info, minsz) ?
812 			-EFAULT : 0;
813 
814 	} else if (cmd == VFIO_DEVICE_SET_IRQS) {
815 		struct vfio_irq_set hdr;
816 		u8 *data = NULL;
817 		int max, ret = 0;
818 		size_t data_size = 0;
819 
820 		minsz = offsetofend(struct vfio_irq_set, count);
821 
822 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
823 			return -EFAULT;
824 
825 		max = vfio_pci_get_irq_count(vdev, hdr.index);
826 
827 		ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
828 						 VFIO_PCI_NUM_IRQS, &data_size);
829 		if (ret)
830 			return ret;
831 
832 		if (data_size) {
833 			data = memdup_user((void __user *)(arg + minsz),
834 					    data_size);
835 			if (IS_ERR(data))
836 				return PTR_ERR(data);
837 		}
838 
839 		mutex_lock(&vdev->igate);
840 
841 		ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
842 					      hdr.start, hdr.count, data);
843 
844 		mutex_unlock(&vdev->igate);
845 		kfree(data);
846 
847 		return ret;
848 
849 	} else if (cmd == VFIO_DEVICE_RESET) {
850 		return vdev->reset_works ?
851 			pci_try_reset_function(vdev->pdev) : -EINVAL;
852 
853 	} else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
854 		struct vfio_pci_hot_reset_info hdr;
855 		struct vfio_pci_fill_info fill = { 0 };
856 		struct vfio_pci_dependent_device *devices = NULL;
857 		bool slot = false;
858 		int ret = 0;
859 
860 		minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
861 
862 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
863 			return -EFAULT;
864 
865 		if (hdr.argsz < minsz)
866 			return -EINVAL;
867 
868 		hdr.flags = 0;
869 
870 		/* Can we do a slot or bus reset or neither? */
871 		if (!pci_probe_reset_slot(vdev->pdev->slot))
872 			slot = true;
873 		else if (pci_probe_reset_bus(vdev->pdev->bus))
874 			return -ENODEV;
875 
876 		/* How many devices are affected? */
877 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
878 						    vfio_pci_count_devs,
879 						    &fill.max, slot);
880 		if (ret)
881 			return ret;
882 
883 		WARN_ON(!fill.max); /* Should always be at least one */
884 
885 		/*
886 		 * If there's enough space, fill it now, otherwise return
887 		 * -ENOSPC and the number of devices affected.
888 		 */
889 		if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
890 			ret = -ENOSPC;
891 			hdr.count = fill.max;
892 			goto reset_info_exit;
893 		}
894 
895 		devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
896 		if (!devices)
897 			return -ENOMEM;
898 
899 		fill.devices = devices;
900 
901 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
902 						    vfio_pci_fill_devs,
903 						    &fill, slot);
904 
905 		/*
906 		 * If a device was removed between counting and filling,
907 		 * we may come up short of fill.max.  If a device was
908 		 * added, we'll have a return of -EAGAIN above.
909 		 */
910 		if (!ret)
911 			hdr.count = fill.cur;
912 
913 reset_info_exit:
914 		if (copy_to_user((void __user *)arg, &hdr, minsz))
915 			ret = -EFAULT;
916 
917 		if (!ret) {
918 			if (copy_to_user((void __user *)(arg + minsz), devices,
919 					 hdr.count * sizeof(*devices)))
920 				ret = -EFAULT;
921 		}
922 
923 		kfree(devices);
924 		return ret;
925 
926 	} else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
927 		struct vfio_pci_hot_reset hdr;
928 		int32_t *group_fds;
929 		struct vfio_pci_group_entry *groups;
930 		struct vfio_pci_group_info info;
931 		bool slot = false;
932 		int i, count = 0, ret = 0;
933 
934 		minsz = offsetofend(struct vfio_pci_hot_reset, count);
935 
936 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
937 			return -EFAULT;
938 
939 		if (hdr.argsz < minsz || hdr.flags)
940 			return -EINVAL;
941 
942 		/* Can we do a slot or bus reset or neither? */
943 		if (!pci_probe_reset_slot(vdev->pdev->slot))
944 			slot = true;
945 		else if (pci_probe_reset_bus(vdev->pdev->bus))
946 			return -ENODEV;
947 
948 		/*
949 		 * We can't let userspace give us an arbitrarily large
950 		 * buffer to copy, so verify how many we think there
951 		 * could be.  Note groups can have multiple devices so
952 		 * one group per device is the max.
953 		 */
954 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
955 						    vfio_pci_count_devs,
956 						    &count, slot);
957 		if (ret)
958 			return ret;
959 
960 		/* Somewhere between 1 and count is OK */
961 		if (!hdr.count || hdr.count > count)
962 			return -EINVAL;
963 
964 		group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
965 		groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
966 		if (!group_fds || !groups) {
967 			kfree(group_fds);
968 			kfree(groups);
969 			return -ENOMEM;
970 		}
971 
972 		if (copy_from_user(group_fds, (void __user *)(arg + minsz),
973 				   hdr.count * sizeof(*group_fds))) {
974 			kfree(group_fds);
975 			kfree(groups);
976 			return -EFAULT;
977 		}
978 
979 		/*
980 		 * For each group_fd, get the group through the vfio external
981 		 * user interface and store the group and iommu ID.  This
982 		 * ensures the group is held across the reset.
983 		 */
984 		for (i = 0; i < hdr.count; i++) {
985 			struct vfio_group *group;
986 			struct fd f = fdget(group_fds[i]);
987 			if (!f.file) {
988 				ret = -EBADF;
989 				break;
990 			}
991 
992 			group = vfio_group_get_external_user(f.file);
993 			fdput(f);
994 			if (IS_ERR(group)) {
995 				ret = PTR_ERR(group);
996 				break;
997 			}
998 
999 			groups[i].group = group;
1000 			groups[i].id = vfio_external_user_iommu_id(group);
1001 		}
1002 
1003 		kfree(group_fds);
1004 
1005 		/* release reference to groups on error */
1006 		if (ret)
1007 			goto hot_reset_release;
1008 
1009 		info.count = hdr.count;
1010 		info.groups = groups;
1011 
1012 		/*
1013 		 * Test whether all the affected devices are contained
1014 		 * by the set of groups provided by the user.
1015 		 */
1016 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1017 						    vfio_pci_validate_devs,
1018 						    &info, slot);
1019 		if (!ret)
1020 			/* User has access, do the reset */
1021 			ret = pci_reset_bus(vdev->pdev);
1022 
1023 hot_reset_release:
1024 		for (i--; i >= 0; i--)
1025 			vfio_group_put_external_user(groups[i].group);
1026 
1027 		kfree(groups);
1028 		return ret;
1029 	} else if (cmd == VFIO_DEVICE_IOEVENTFD) {
1030 		struct vfio_device_ioeventfd ioeventfd;
1031 		int count;
1032 
1033 		minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1034 
1035 		if (copy_from_user(&ioeventfd, (void __user *)arg, minsz))
1036 			return -EFAULT;
1037 
1038 		if (ioeventfd.argsz < minsz)
1039 			return -EINVAL;
1040 
1041 		if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1042 			return -EINVAL;
1043 
1044 		count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1045 
1046 		if (hweight8(count) != 1 || ioeventfd.fd < -1)
1047 			return -EINVAL;
1048 
1049 		return vfio_pci_ioeventfd(vdev, ioeventfd.offset,
1050 					  ioeventfd.data, count, ioeventfd.fd);
1051 	}
1052 
1053 	return -ENOTTY;
1054 }
1055 
1056 static ssize_t vfio_pci_rw(void *device_data, char __user *buf,
1057 			   size_t count, loff_t *ppos, bool iswrite)
1058 {
1059 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1060 	struct vfio_pci_device *vdev = device_data;
1061 
1062 	if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1063 		return -EINVAL;
1064 
1065 	switch (index) {
1066 	case VFIO_PCI_CONFIG_REGION_INDEX:
1067 		return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1068 
1069 	case VFIO_PCI_ROM_REGION_INDEX:
1070 		if (iswrite)
1071 			return -EINVAL;
1072 		return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1073 
1074 	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1075 		return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1076 
1077 	case VFIO_PCI_VGA_REGION_INDEX:
1078 		return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1079 	default:
1080 		index -= VFIO_PCI_NUM_REGIONS;
1081 		return vdev->region[index].ops->rw(vdev, buf,
1082 						   count, ppos, iswrite);
1083 	}
1084 
1085 	return -EINVAL;
1086 }
1087 
1088 static ssize_t vfio_pci_read(void *device_data, char __user *buf,
1089 			     size_t count, loff_t *ppos)
1090 {
1091 	if (!count)
1092 		return 0;
1093 
1094 	return vfio_pci_rw(device_data, buf, count, ppos, false);
1095 }
1096 
1097 static ssize_t vfio_pci_write(void *device_data, const char __user *buf,
1098 			      size_t count, loff_t *ppos)
1099 {
1100 	if (!count)
1101 		return 0;
1102 
1103 	return vfio_pci_rw(device_data, (char __user *)buf, count, ppos, true);
1104 }
1105 
1106 static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma)
1107 {
1108 	struct vfio_pci_device *vdev = device_data;
1109 	struct pci_dev *pdev = vdev->pdev;
1110 	unsigned int index;
1111 	u64 phys_len, req_len, pgoff, req_start;
1112 	int ret;
1113 
1114 	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1115 
1116 	if (vma->vm_end < vma->vm_start)
1117 		return -EINVAL;
1118 	if ((vma->vm_flags & VM_SHARED) == 0)
1119 		return -EINVAL;
1120 	if (index >= VFIO_PCI_ROM_REGION_INDEX)
1121 		return -EINVAL;
1122 	if (!vdev->bar_mmap_supported[index])
1123 		return -EINVAL;
1124 
1125 	phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1126 	req_len = vma->vm_end - vma->vm_start;
1127 	pgoff = vma->vm_pgoff &
1128 		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1129 	req_start = pgoff << PAGE_SHIFT;
1130 
1131 	if (req_start + req_len > phys_len)
1132 		return -EINVAL;
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 	/*
1200 	 * Prevent binding to PFs with VFs enabled, this too easily allows
1201 	 * userspace instance with VFs and PFs from the same device, which
1202 	 * cannot work.  Disabling SR-IOV here would initiate removing the
1203 	 * VFs, which would unbind the driver, which is prone to blocking
1204 	 * if that VF is also in use by vfio-pci.  Just reject these PFs
1205 	 * and let the user sort it out.
1206 	 */
1207 	if (pci_num_vf(pdev)) {
1208 		pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
1209 		return -EBUSY;
1210 	}
1211 
1212 	group = vfio_iommu_group_get(&pdev->dev);
1213 	if (!group)
1214 		return -EINVAL;
1215 
1216 	vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
1217 	if (!vdev) {
1218 		vfio_iommu_group_put(group, &pdev->dev);
1219 		return -ENOMEM;
1220 	}
1221 
1222 	vdev->pdev = pdev;
1223 	vdev->irq_type = VFIO_PCI_NUM_IRQS;
1224 	mutex_init(&vdev->igate);
1225 	spin_lock_init(&vdev->irqlock);
1226 	mutex_init(&vdev->ioeventfds_lock);
1227 	INIT_LIST_HEAD(&vdev->ioeventfds_list);
1228 
1229 	ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
1230 	if (ret) {
1231 		vfio_iommu_group_put(group, &pdev->dev);
1232 		kfree(vdev);
1233 		return ret;
1234 	}
1235 
1236 	if (vfio_pci_is_vga(pdev)) {
1237 		vga_client_register(pdev, vdev, NULL, vfio_pci_set_vga_decode);
1238 		vga_set_legacy_decoding(pdev,
1239 					vfio_pci_set_vga_decode(vdev, false));
1240 	}
1241 
1242 	if (!disable_idle_d3) {
1243 		/*
1244 		 * pci-core sets the device power state to an unknown value at
1245 		 * bootup and after being removed from a driver.  The only
1246 		 * transition it allows from this unknown state is to D0, which
1247 		 * typically happens when a driver calls pci_enable_device().
1248 		 * We're not ready to enable the device yet, but we do want to
1249 		 * be able to get to D3.  Therefore first do a D0 transition
1250 		 * before going to D3.
1251 		 */
1252 		pci_set_power_state(pdev, PCI_D0);
1253 		pci_set_power_state(pdev, PCI_D3hot);
1254 	}
1255 
1256 	return ret;
1257 }
1258 
1259 static void vfio_pci_remove(struct pci_dev *pdev)
1260 {
1261 	struct vfio_pci_device *vdev;
1262 
1263 	vdev = vfio_del_group_dev(&pdev->dev);
1264 	if (!vdev)
1265 		return;
1266 
1267 	vfio_iommu_group_put(pdev->dev.iommu_group, &pdev->dev);
1268 	kfree(vdev->region);
1269 	mutex_destroy(&vdev->ioeventfds_lock);
1270 	kfree(vdev);
1271 
1272 	if (vfio_pci_is_vga(pdev)) {
1273 		vga_client_register(pdev, NULL, NULL, NULL);
1274 		vga_set_legacy_decoding(pdev,
1275 				VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
1276 				VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM);
1277 	}
1278 
1279 	if (!disable_idle_d3)
1280 		pci_set_power_state(pdev, PCI_D0);
1281 }
1282 
1283 static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
1284 						  pci_channel_state_t state)
1285 {
1286 	struct vfio_pci_device *vdev;
1287 	struct vfio_device *device;
1288 
1289 	device = vfio_device_get_from_dev(&pdev->dev);
1290 	if (device == NULL)
1291 		return PCI_ERS_RESULT_DISCONNECT;
1292 
1293 	vdev = vfio_device_data(device);
1294 	if (vdev == NULL) {
1295 		vfio_device_put(device);
1296 		return PCI_ERS_RESULT_DISCONNECT;
1297 	}
1298 
1299 	mutex_lock(&vdev->igate);
1300 
1301 	if (vdev->err_trigger)
1302 		eventfd_signal(vdev->err_trigger, 1);
1303 
1304 	mutex_unlock(&vdev->igate);
1305 
1306 	vfio_device_put(device);
1307 
1308 	return PCI_ERS_RESULT_CAN_RECOVER;
1309 }
1310 
1311 static const struct pci_error_handlers vfio_err_handlers = {
1312 	.error_detected = vfio_pci_aer_err_detected,
1313 };
1314 
1315 static struct pci_driver vfio_pci_driver = {
1316 	.name		= "vfio-pci",
1317 	.id_table	= NULL, /* only dynamic ids */
1318 	.probe		= vfio_pci_probe,
1319 	.remove		= vfio_pci_remove,
1320 	.err_handler	= &vfio_err_handlers,
1321 };
1322 
1323 struct vfio_devices {
1324 	struct vfio_device **devices;
1325 	int cur_index;
1326 	int max_index;
1327 };
1328 
1329 static int vfio_pci_get_devs(struct pci_dev *pdev, void *data)
1330 {
1331 	struct vfio_devices *devs = data;
1332 	struct vfio_device *device;
1333 
1334 	if (devs->cur_index == devs->max_index)
1335 		return -ENOSPC;
1336 
1337 	device = vfio_device_get_from_dev(&pdev->dev);
1338 	if (!device)
1339 		return -EINVAL;
1340 
1341 	if (pci_dev_driver(pdev) != &vfio_pci_driver) {
1342 		vfio_device_put(device);
1343 		return -EBUSY;
1344 	}
1345 
1346 	devs->devices[devs->cur_index++] = device;
1347 	return 0;
1348 }
1349 
1350 /*
1351  * Attempt to do a bus/slot reset if there are devices affected by a reset for
1352  * this device that are needs_reset and all of the affected devices are unused
1353  * (!refcnt).  Callers are required to hold driver_lock when calling this to
1354  * prevent device opens and concurrent bus reset attempts.  We prevent device
1355  * unbinds by acquiring and holding a reference to the vfio_device.
1356  *
1357  * NB: vfio-core considers a group to be viable even if some devices are
1358  * bound to drivers like pci-stub or pcieport.  Here we require all devices
1359  * to be bound to vfio_pci since that's the only way we can be sure they
1360  * stay put.
1361  */
1362 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev)
1363 {
1364 	struct vfio_devices devs = { .cur_index = 0 };
1365 	int i = 0, ret = -EINVAL;
1366 	bool needs_reset = false, slot = false;
1367 	struct vfio_pci_device *tmp;
1368 
1369 	if (!pci_probe_reset_slot(vdev->pdev->slot))
1370 		slot = true;
1371 	else if (pci_probe_reset_bus(vdev->pdev->bus))
1372 		return;
1373 
1374 	if (vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
1375 					  &i, slot) || !i)
1376 		return;
1377 
1378 	devs.max_index = i;
1379 	devs.devices = kcalloc(i, sizeof(struct vfio_device *), GFP_KERNEL);
1380 	if (!devs.devices)
1381 		return;
1382 
1383 	if (vfio_pci_for_each_slot_or_bus(vdev->pdev,
1384 					  vfio_pci_get_devs, &devs, slot))
1385 		goto put_devs;
1386 
1387 	for (i = 0; i < devs.cur_index; i++) {
1388 		tmp = vfio_device_data(devs.devices[i]);
1389 		if (tmp->needs_reset)
1390 			needs_reset = true;
1391 		if (tmp->refcnt)
1392 			goto put_devs;
1393 	}
1394 
1395 	if (needs_reset)
1396 		ret = pci_reset_bus(vdev->pdev);
1397 
1398 put_devs:
1399 	for (i = 0; i < devs.cur_index; i++) {
1400 		tmp = vfio_device_data(devs.devices[i]);
1401 		if (!ret)
1402 			tmp->needs_reset = false;
1403 
1404 		if (!tmp->refcnt && !disable_idle_d3)
1405 			pci_set_power_state(tmp->pdev, PCI_D3hot);
1406 
1407 		vfio_device_put(devs.devices[i]);
1408 	}
1409 
1410 	kfree(devs.devices);
1411 }
1412 
1413 static void __exit vfio_pci_cleanup(void)
1414 {
1415 	pci_unregister_driver(&vfio_pci_driver);
1416 	vfio_pci_uninit_perm_bits();
1417 }
1418 
1419 static void __init vfio_pci_fill_ids(void)
1420 {
1421 	char *p, *id;
1422 	int rc;
1423 
1424 	/* no ids passed actually */
1425 	if (ids[0] == '\0')
1426 		return;
1427 
1428 	/* add ids specified in the module parameter */
1429 	p = ids;
1430 	while ((id = strsep(&p, ","))) {
1431 		unsigned int vendor, device, subvendor = PCI_ANY_ID,
1432 			subdevice = PCI_ANY_ID, class = 0, class_mask = 0;
1433 		int fields;
1434 
1435 		if (!strlen(id))
1436 			continue;
1437 
1438 		fields = sscanf(id, "%x:%x:%x:%x:%x:%x",
1439 				&vendor, &device, &subvendor, &subdevice,
1440 				&class, &class_mask);
1441 
1442 		if (fields < 2) {
1443 			pr_warn("invalid id string \"%s\"\n", id);
1444 			continue;
1445 		}
1446 
1447 		rc = pci_add_dynid(&vfio_pci_driver, vendor, device,
1448 				   subvendor, subdevice, class, class_mask, 0);
1449 		if (rc)
1450 			pr_warn("failed to add dynamic id [%04hx:%04hx[%04hx:%04hx]] class %#08x/%08x (%d)\n",
1451 				vendor, device, subvendor, subdevice,
1452 				class, class_mask, rc);
1453 		else
1454 			pr_info("add [%04hx:%04hx[%04hx:%04hx]] class %#08x/%08x\n",
1455 				vendor, device, subvendor, subdevice,
1456 				class, class_mask);
1457 	}
1458 }
1459 
1460 static int __init vfio_pci_init(void)
1461 {
1462 	int ret;
1463 
1464 	/* Allocate shared config space permision data used by all devices */
1465 	ret = vfio_pci_init_perm_bits();
1466 	if (ret)
1467 		return ret;
1468 
1469 	/* Register and scan for devices */
1470 	ret = pci_register_driver(&vfio_pci_driver);
1471 	if (ret)
1472 		goto out_driver;
1473 
1474 	vfio_pci_fill_ids();
1475 
1476 	return 0;
1477 
1478 out_driver:
1479 	vfio_pci_uninit_perm_bits();
1480 	return ret;
1481 }
1482 
1483 module_init(vfio_pci_init);
1484 module_exit(vfio_pci_cleanup);
1485 
1486 MODULE_VERSION(DRIVER_VERSION);
1487 MODULE_LICENSE("GPL v2");
1488 MODULE_AUTHOR(DRIVER_AUTHOR);
1489 MODULE_DESCRIPTION(DRIVER_DESC);
1490