xref: /openbmc/linux/drivers/vfio/pci/vfio_pci.c (revision 2fa49589)
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 inline bool vfio_vga_disabled(void)
60 {
61 #ifdef CONFIG_VFIO_PCI_VGA
62 	return disable_vga;
63 #else
64 	return true;
65 #endif
66 }
67 
68 /*
69  * Our VGA arbiter participation is limited since we don't know anything
70  * about the device itself.  However, if the device is the only VGA device
71  * downstream of a bridge and VFIO VGA support is disabled, then we can
72  * safely return legacy VGA IO and memory as not decoded since the user
73  * has no way to get to it and routing can be disabled externally at the
74  * bridge.
75  */
76 static unsigned int vfio_pci_set_vga_decode(void *opaque, bool single_vga)
77 {
78 	struct vfio_pci_device *vdev = opaque;
79 	struct pci_dev *tmp = NULL, *pdev = vdev->pdev;
80 	unsigned char max_busnr;
81 	unsigned int decodes;
82 
83 	if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
84 		return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
85 		       VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
86 
87 	max_busnr = pci_bus_max_busnr(pdev->bus);
88 	decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
89 
90 	while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
91 		if (tmp == pdev ||
92 		    pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
93 		    pci_is_root_bus(tmp->bus))
94 			continue;
95 
96 		if (tmp->bus->number >= pdev->bus->number &&
97 		    tmp->bus->number <= max_busnr) {
98 			pci_dev_put(tmp);
99 			decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
100 			break;
101 		}
102 	}
103 
104 	return decodes;
105 }
106 
107 static inline bool vfio_pci_is_vga(struct pci_dev *pdev)
108 {
109 	return (pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA;
110 }
111 
112 static void vfio_pci_probe_mmaps(struct vfio_pci_device *vdev)
113 {
114 	struct resource *res;
115 	int bar;
116 	struct vfio_pci_dummy_resource *dummy_res;
117 
118 	INIT_LIST_HEAD(&vdev->dummy_resources_list);
119 
120 	for (bar = PCI_STD_RESOURCES; bar <= PCI_STD_RESOURCE_END; bar++) {
121 		res = vdev->pdev->resource + bar;
122 
123 		if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
124 			goto no_mmap;
125 
126 		if (!(res->flags & IORESOURCE_MEM))
127 			goto no_mmap;
128 
129 		/*
130 		 * The PCI core shouldn't set up a resource with a
131 		 * type but zero size. But there may be bugs that
132 		 * cause us to do that.
133 		 */
134 		if (!resource_size(res))
135 			goto no_mmap;
136 
137 		if (resource_size(res) >= PAGE_SIZE) {
138 			vdev->bar_mmap_supported[bar] = true;
139 			continue;
140 		}
141 
142 		if (!(res->start & ~PAGE_MASK)) {
143 			/*
144 			 * Add a dummy resource to reserve the remainder
145 			 * of the exclusive page in case that hot-add
146 			 * device's bar is assigned into it.
147 			 */
148 			dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
149 			if (dummy_res == NULL)
150 				goto no_mmap;
151 
152 			dummy_res->resource.name = "vfio sub-page reserved";
153 			dummy_res->resource.start = res->end + 1;
154 			dummy_res->resource.end = res->start + PAGE_SIZE - 1;
155 			dummy_res->resource.flags = res->flags;
156 			if (request_resource(res->parent,
157 						&dummy_res->resource)) {
158 				kfree(dummy_res);
159 				goto no_mmap;
160 			}
161 			dummy_res->index = bar;
162 			list_add(&dummy_res->res_next,
163 					&vdev->dummy_resources_list);
164 			vdev->bar_mmap_supported[bar] = true;
165 			continue;
166 		}
167 		/*
168 		 * Here we don't handle the case when the BAR is not page
169 		 * aligned because we can't expect the BAR will be
170 		 * assigned into the same location in a page in guest
171 		 * when we passthrough the BAR. And it's hard to access
172 		 * this BAR in userspace because we have no way to get
173 		 * the BAR's location in a page.
174 		 */
175 no_mmap:
176 		vdev->bar_mmap_supported[bar] = false;
177 	}
178 }
179 
180 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev);
181 static void vfio_pci_disable(struct vfio_pci_device *vdev);
182 
183 /*
184  * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
185  * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
186  * If a device implements the former but not the latter we would typically
187  * expect broken_intx_masking be set and require an exclusive interrupt.
188  * However since we do have control of the device's ability to assert INTx,
189  * we can instead pretend that the device does not implement INTx, virtualizing
190  * the pin register to report zero and maintaining DisINTx set on the host.
191  */
192 static bool vfio_pci_nointx(struct pci_dev *pdev)
193 {
194 	switch (pdev->vendor) {
195 	case PCI_VENDOR_ID_INTEL:
196 		switch (pdev->device) {
197 		/* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
198 		case 0x1572:
199 		case 0x1574:
200 		case 0x1580 ... 0x1581:
201 		case 0x1583 ... 0x158b:
202 		case 0x37d0 ... 0x37d2:
203 			return true;
204 		default:
205 			return false;
206 		}
207 	}
208 
209 	return false;
210 }
211 
212 static int vfio_pci_enable(struct vfio_pci_device *vdev)
213 {
214 	struct pci_dev *pdev = vdev->pdev;
215 	int ret;
216 	u16 cmd;
217 	u8 msix_pos;
218 
219 	pci_set_power_state(pdev, PCI_D0);
220 
221 	/* Don't allow our initial saved state to include busmaster */
222 	pci_clear_master(pdev);
223 
224 	ret = pci_enable_device(pdev);
225 	if (ret)
226 		return ret;
227 
228 	/* If reset fails because of the device lock, fail this path entirely */
229 	ret = pci_try_reset_function(pdev);
230 	if (ret == -EAGAIN) {
231 		pci_disable_device(pdev);
232 		return ret;
233 	}
234 
235 	vdev->reset_works = !ret;
236 	pci_save_state(pdev);
237 	vdev->pci_saved_state = pci_store_saved_state(pdev);
238 	if (!vdev->pci_saved_state)
239 		pr_debug("%s: Couldn't store %s saved state\n",
240 			 __func__, dev_name(&pdev->dev));
241 
242 	if (likely(!nointxmask)) {
243 		if (vfio_pci_nointx(pdev)) {
244 			dev_info(&pdev->dev, "Masking broken INTx support\n");
245 			vdev->nointx = true;
246 			pci_intx(pdev, 0);
247 		} else
248 			vdev->pci_2_3 = pci_intx_mask_supported(pdev);
249 	}
250 
251 	pci_read_config_word(pdev, PCI_COMMAND, &cmd);
252 	if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
253 		cmd &= ~PCI_COMMAND_INTX_DISABLE;
254 		pci_write_config_word(pdev, PCI_COMMAND, cmd);
255 	}
256 
257 	ret = vfio_config_init(vdev);
258 	if (ret) {
259 		kfree(vdev->pci_saved_state);
260 		vdev->pci_saved_state = NULL;
261 		pci_disable_device(pdev);
262 		return ret;
263 	}
264 
265 	msix_pos = pdev->msix_cap;
266 	if (msix_pos) {
267 		u16 flags;
268 		u32 table;
269 
270 		pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
271 		pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
272 
273 		vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
274 		vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
275 		vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
276 	} else
277 		vdev->msix_bar = 0xFF;
278 
279 	if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
280 		vdev->has_vga = true;
281 
282 
283 	if (vfio_pci_is_vga(pdev) &&
284 	    pdev->vendor == PCI_VENDOR_ID_INTEL &&
285 	    IS_ENABLED(CONFIG_VFIO_PCI_IGD)) {
286 		ret = vfio_pci_igd_init(vdev);
287 		if (ret) {
288 			dev_warn(&vdev->pdev->dev,
289 				 "Failed to setup Intel IGD regions\n");
290 			goto disable_exit;
291 		}
292 	}
293 
294 	if (pdev->vendor == PCI_VENDOR_ID_NVIDIA &&
295 	    IS_ENABLED(CONFIG_VFIO_PCI_NVLINK2)) {
296 		ret = vfio_pci_nvdia_v100_nvlink2_init(vdev);
297 		if (ret && ret != -ENODEV) {
298 			dev_warn(&vdev->pdev->dev,
299 				 "Failed to setup NVIDIA NV2 RAM region\n");
300 			goto disable_exit;
301 		}
302 	}
303 
304 	if (pdev->vendor == PCI_VENDOR_ID_IBM &&
305 	    IS_ENABLED(CONFIG_VFIO_PCI_NVLINK2)) {
306 		ret = vfio_pci_ibm_npu2_init(vdev);
307 		if (ret && ret != -ENODEV) {
308 			dev_warn(&vdev->pdev->dev,
309 					"Failed to setup NVIDIA NV2 ATSD region\n");
310 			goto disable_exit;
311 		}
312 	}
313 
314 	vfio_pci_probe_mmaps(vdev);
315 
316 	return 0;
317 
318 disable_exit:
319 	vfio_pci_disable(vdev);
320 	return ret;
321 }
322 
323 static void vfio_pci_disable(struct vfio_pci_device *vdev)
324 {
325 	struct pci_dev *pdev = vdev->pdev;
326 	struct vfio_pci_dummy_resource *dummy_res, *tmp;
327 	struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
328 	int i, bar;
329 
330 	/* Stop the device from further DMA */
331 	pci_clear_master(pdev);
332 
333 	vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
334 				VFIO_IRQ_SET_ACTION_TRIGGER,
335 				vdev->irq_type, 0, 0, NULL);
336 
337 	/* Device closed, don't need mutex here */
338 	list_for_each_entry_safe(ioeventfd, ioeventfd_tmp,
339 				 &vdev->ioeventfds_list, next) {
340 		vfio_virqfd_disable(&ioeventfd->virqfd);
341 		list_del(&ioeventfd->next);
342 		kfree(ioeventfd);
343 	}
344 	vdev->ioeventfds_nr = 0;
345 
346 	vdev->virq_disabled = false;
347 
348 	for (i = 0; i < vdev->num_regions; i++)
349 		vdev->region[i].ops->release(vdev, &vdev->region[i]);
350 
351 	vdev->num_regions = 0;
352 	kfree(vdev->region);
353 	vdev->region = NULL; /* don't krealloc a freed pointer */
354 
355 	vfio_config_free(vdev);
356 
357 	for (bar = PCI_STD_RESOURCES; bar <= PCI_STD_RESOURCE_END; bar++) {
358 		if (!vdev->barmap[bar])
359 			continue;
360 		pci_iounmap(pdev, vdev->barmap[bar]);
361 		pci_release_selected_regions(pdev, 1 << bar);
362 		vdev->barmap[bar] = NULL;
363 	}
364 
365 	list_for_each_entry_safe(dummy_res, tmp,
366 				 &vdev->dummy_resources_list, res_next) {
367 		list_del(&dummy_res->res_next);
368 		release_resource(&dummy_res->resource);
369 		kfree(dummy_res);
370 	}
371 
372 	vdev->needs_reset = true;
373 
374 	/*
375 	 * If we have saved state, restore it.  If we can reset the device,
376 	 * even better.  Resetting with current state seems better than
377 	 * nothing, but saving and restoring current state without reset
378 	 * is just busy work.
379 	 */
380 	if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
381 		pr_info("%s: Couldn't reload %s saved state\n",
382 			__func__, dev_name(&pdev->dev));
383 
384 		if (!vdev->reset_works)
385 			goto out;
386 
387 		pci_save_state(pdev);
388 	}
389 
390 	/*
391 	 * Disable INTx and MSI, presumably to avoid spurious interrupts
392 	 * during reset.  Stolen from pci_reset_function()
393 	 */
394 	pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
395 
396 	/*
397 	 * Try to reset the device.  The success of this is dependent on
398 	 * being able to lock the device, which is not always possible.
399 	 */
400 	if (vdev->reset_works && !pci_try_reset_function(pdev))
401 		vdev->needs_reset = false;
402 
403 	pci_restore_state(pdev);
404 out:
405 	pci_disable_device(pdev);
406 
407 	vfio_pci_try_bus_reset(vdev);
408 
409 	if (!disable_idle_d3)
410 		pci_set_power_state(pdev, PCI_D3hot);
411 }
412 
413 static void vfio_pci_release(void *device_data)
414 {
415 	struct vfio_pci_device *vdev = device_data;
416 
417 	mutex_lock(&vdev->reflck->lock);
418 
419 	if (!(--vdev->refcnt)) {
420 		vfio_spapr_pci_eeh_release(vdev->pdev);
421 		vfio_pci_disable(vdev);
422 	}
423 
424 	mutex_unlock(&vdev->reflck->lock);
425 
426 	module_put(THIS_MODULE);
427 }
428 
429 static int vfio_pci_open(void *device_data)
430 {
431 	struct vfio_pci_device *vdev = device_data;
432 	int ret = 0;
433 
434 	if (!try_module_get(THIS_MODULE))
435 		return -ENODEV;
436 
437 	mutex_lock(&vdev->reflck->lock);
438 
439 	if (!vdev->refcnt) {
440 		ret = vfio_pci_enable(vdev);
441 		if (ret)
442 			goto error;
443 
444 		vfio_spapr_pci_eeh_open(vdev->pdev);
445 	}
446 	vdev->refcnt++;
447 error:
448 	mutex_unlock(&vdev->reflck->lock);
449 	if (ret)
450 		module_put(THIS_MODULE);
451 	return ret;
452 }
453 
454 static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
455 {
456 	if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
457 		u8 pin;
458 
459 		if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
460 		    vdev->nointx || vdev->pdev->is_virtfn)
461 			return 0;
462 
463 		pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
464 
465 		return pin ? 1 : 0;
466 	} else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
467 		u8 pos;
468 		u16 flags;
469 
470 		pos = vdev->pdev->msi_cap;
471 		if (pos) {
472 			pci_read_config_word(vdev->pdev,
473 					     pos + PCI_MSI_FLAGS, &flags);
474 			return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
475 		}
476 	} else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
477 		u8 pos;
478 		u16 flags;
479 
480 		pos = vdev->pdev->msix_cap;
481 		if (pos) {
482 			pci_read_config_word(vdev->pdev,
483 					     pos + PCI_MSIX_FLAGS, &flags);
484 
485 			return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
486 		}
487 	} else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
488 		if (pci_is_pcie(vdev->pdev))
489 			return 1;
490 	} else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
491 		return 1;
492 	}
493 
494 	return 0;
495 }
496 
497 static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
498 {
499 	(*(int *)data)++;
500 	return 0;
501 }
502 
503 struct vfio_pci_fill_info {
504 	int max;
505 	int cur;
506 	struct vfio_pci_dependent_device *devices;
507 };
508 
509 static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
510 {
511 	struct vfio_pci_fill_info *fill = data;
512 	struct iommu_group *iommu_group;
513 
514 	if (fill->cur == fill->max)
515 		return -EAGAIN; /* Something changed, try again */
516 
517 	iommu_group = iommu_group_get(&pdev->dev);
518 	if (!iommu_group)
519 		return -EPERM; /* Cannot reset non-isolated devices */
520 
521 	fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
522 	fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
523 	fill->devices[fill->cur].bus = pdev->bus->number;
524 	fill->devices[fill->cur].devfn = pdev->devfn;
525 	fill->cur++;
526 	iommu_group_put(iommu_group);
527 	return 0;
528 }
529 
530 struct vfio_pci_group_entry {
531 	struct vfio_group *group;
532 	int id;
533 };
534 
535 struct vfio_pci_group_info {
536 	int count;
537 	struct vfio_pci_group_entry *groups;
538 };
539 
540 static int vfio_pci_validate_devs(struct pci_dev *pdev, void *data)
541 {
542 	struct vfio_pci_group_info *info = data;
543 	struct iommu_group *group;
544 	int id, i;
545 
546 	group = iommu_group_get(&pdev->dev);
547 	if (!group)
548 		return -EPERM;
549 
550 	id = iommu_group_id(group);
551 
552 	for (i = 0; i < info->count; i++)
553 		if (info->groups[i].id == id)
554 			break;
555 
556 	iommu_group_put(group);
557 
558 	return (i == info->count) ? -EINVAL : 0;
559 }
560 
561 static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
562 {
563 	for (; pdev; pdev = pdev->bus->self)
564 		if (pdev->bus == slot->bus)
565 			return (pdev->slot == slot);
566 	return false;
567 }
568 
569 struct vfio_pci_walk_info {
570 	int (*fn)(struct pci_dev *, void *data);
571 	void *data;
572 	struct pci_dev *pdev;
573 	bool slot;
574 	int ret;
575 };
576 
577 static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
578 {
579 	struct vfio_pci_walk_info *walk = data;
580 
581 	if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
582 		walk->ret = walk->fn(pdev, walk->data);
583 
584 	return walk->ret;
585 }
586 
587 static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
588 					 int (*fn)(struct pci_dev *,
589 						   void *data), void *data,
590 					 bool slot)
591 {
592 	struct vfio_pci_walk_info walk = {
593 		.fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
594 	};
595 
596 	pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
597 
598 	return walk.ret;
599 }
600 
601 static int msix_mmappable_cap(struct vfio_pci_device *vdev,
602 			      struct vfio_info_cap *caps)
603 {
604 	struct vfio_info_cap_header header = {
605 		.id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
606 		.version = 1
607 	};
608 
609 	return vfio_info_add_capability(caps, &header, sizeof(header));
610 }
611 
612 int vfio_pci_register_dev_region(struct vfio_pci_device *vdev,
613 				 unsigned int type, unsigned int subtype,
614 				 const struct vfio_pci_regops *ops,
615 				 size_t size, u32 flags, void *data)
616 {
617 	struct vfio_pci_region *region;
618 
619 	region = krealloc(vdev->region,
620 			  (vdev->num_regions + 1) * sizeof(*region),
621 			  GFP_KERNEL);
622 	if (!region)
623 		return -ENOMEM;
624 
625 	vdev->region = region;
626 	vdev->region[vdev->num_regions].type = type;
627 	vdev->region[vdev->num_regions].subtype = subtype;
628 	vdev->region[vdev->num_regions].ops = ops;
629 	vdev->region[vdev->num_regions].size = size;
630 	vdev->region[vdev->num_regions].flags = flags;
631 	vdev->region[vdev->num_regions].data = data;
632 
633 	vdev->num_regions++;
634 
635 	return 0;
636 }
637 
638 static long vfio_pci_ioctl(void *device_data,
639 			   unsigned int cmd, unsigned long arg)
640 {
641 	struct vfio_pci_device *vdev = device_data;
642 	unsigned long minsz;
643 
644 	if (cmd == VFIO_DEVICE_GET_INFO) {
645 		struct vfio_device_info info;
646 
647 		minsz = offsetofend(struct vfio_device_info, num_irqs);
648 
649 		if (copy_from_user(&info, (void __user *)arg, minsz))
650 			return -EFAULT;
651 
652 		if (info.argsz < minsz)
653 			return -EINVAL;
654 
655 		info.flags = VFIO_DEVICE_FLAGS_PCI;
656 
657 		if (vdev->reset_works)
658 			info.flags |= VFIO_DEVICE_FLAGS_RESET;
659 
660 		info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
661 		info.num_irqs = VFIO_PCI_NUM_IRQS;
662 
663 		return copy_to_user((void __user *)arg, &info, minsz) ?
664 			-EFAULT : 0;
665 
666 	} else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
667 		struct pci_dev *pdev = vdev->pdev;
668 		struct vfio_region_info info;
669 		struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
670 		int i, ret;
671 
672 		minsz = offsetofend(struct vfio_region_info, offset);
673 
674 		if (copy_from_user(&info, (void __user *)arg, minsz))
675 			return -EFAULT;
676 
677 		if (info.argsz < minsz)
678 			return -EINVAL;
679 
680 		switch (info.index) {
681 		case VFIO_PCI_CONFIG_REGION_INDEX:
682 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
683 			info.size = pdev->cfg_size;
684 			info.flags = VFIO_REGION_INFO_FLAG_READ |
685 				     VFIO_REGION_INFO_FLAG_WRITE;
686 			break;
687 		case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
688 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
689 			info.size = pci_resource_len(pdev, info.index);
690 			if (!info.size) {
691 				info.flags = 0;
692 				break;
693 			}
694 
695 			info.flags = VFIO_REGION_INFO_FLAG_READ |
696 				     VFIO_REGION_INFO_FLAG_WRITE;
697 			if (vdev->bar_mmap_supported[info.index]) {
698 				info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
699 				if (info.index == vdev->msix_bar) {
700 					ret = msix_mmappable_cap(vdev, &caps);
701 					if (ret)
702 						return ret;
703 				}
704 			}
705 
706 			break;
707 		case VFIO_PCI_ROM_REGION_INDEX:
708 		{
709 			void __iomem *io;
710 			size_t size;
711 
712 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
713 			info.flags = 0;
714 
715 			/* Report the BAR size, not the ROM size */
716 			info.size = pci_resource_len(pdev, info.index);
717 			if (!info.size) {
718 				/* Shadow ROMs appear as PCI option ROMs */
719 				if (pdev->resource[PCI_ROM_RESOURCE].flags &
720 							IORESOURCE_ROM_SHADOW)
721 					info.size = 0x20000;
722 				else
723 					break;
724 			}
725 
726 			/* Is it really there? */
727 			io = pci_map_rom(pdev, &size);
728 			if (!io || !size) {
729 				info.size = 0;
730 				break;
731 			}
732 			pci_unmap_rom(pdev, io);
733 
734 			info.flags = VFIO_REGION_INFO_FLAG_READ;
735 			break;
736 		}
737 		case VFIO_PCI_VGA_REGION_INDEX:
738 			if (!vdev->has_vga)
739 				return -EINVAL;
740 
741 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
742 			info.size = 0xc0000;
743 			info.flags = VFIO_REGION_INFO_FLAG_READ |
744 				     VFIO_REGION_INFO_FLAG_WRITE;
745 
746 			break;
747 		default:
748 		{
749 			struct vfio_region_info_cap_type cap_type = {
750 					.header.id = VFIO_REGION_INFO_CAP_TYPE,
751 					.header.version = 1 };
752 
753 			if (info.index >=
754 			    VFIO_PCI_NUM_REGIONS + vdev->num_regions)
755 				return -EINVAL;
756 			info.index = array_index_nospec(info.index,
757 							VFIO_PCI_NUM_REGIONS +
758 							vdev->num_regions);
759 
760 			i = info.index - VFIO_PCI_NUM_REGIONS;
761 
762 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
763 			info.size = vdev->region[i].size;
764 			info.flags = vdev->region[i].flags;
765 
766 			cap_type.type = vdev->region[i].type;
767 			cap_type.subtype = vdev->region[i].subtype;
768 
769 			ret = vfio_info_add_capability(&caps, &cap_type.header,
770 						       sizeof(cap_type));
771 			if (ret)
772 				return ret;
773 
774 			if (vdev->region[i].ops->add_capability) {
775 				ret = vdev->region[i].ops->add_capability(vdev,
776 						&vdev->region[i], &caps);
777 				if (ret)
778 					return ret;
779 			}
780 		}
781 		}
782 
783 		if (caps.size) {
784 			info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
785 			if (info.argsz < sizeof(info) + caps.size) {
786 				info.argsz = sizeof(info) + caps.size;
787 				info.cap_offset = 0;
788 			} else {
789 				vfio_info_cap_shift(&caps, sizeof(info));
790 				if (copy_to_user((void __user *)arg +
791 						  sizeof(info), caps.buf,
792 						  caps.size)) {
793 					kfree(caps.buf);
794 					return -EFAULT;
795 				}
796 				info.cap_offset = sizeof(info);
797 			}
798 
799 			kfree(caps.buf);
800 		}
801 
802 		return copy_to_user((void __user *)arg, &info, minsz) ?
803 			-EFAULT : 0;
804 
805 	} else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
806 		struct vfio_irq_info info;
807 
808 		minsz = offsetofend(struct vfio_irq_info, count);
809 
810 		if (copy_from_user(&info, (void __user *)arg, minsz))
811 			return -EFAULT;
812 
813 		if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
814 			return -EINVAL;
815 
816 		switch (info.index) {
817 		case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
818 		case VFIO_PCI_REQ_IRQ_INDEX:
819 			break;
820 		case VFIO_PCI_ERR_IRQ_INDEX:
821 			if (pci_is_pcie(vdev->pdev))
822 				break;
823 		/* fall through */
824 		default:
825 			return -EINVAL;
826 		}
827 
828 		info.flags = VFIO_IRQ_INFO_EVENTFD;
829 
830 		info.count = vfio_pci_get_irq_count(vdev, info.index);
831 
832 		if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
833 			info.flags |= (VFIO_IRQ_INFO_MASKABLE |
834 				       VFIO_IRQ_INFO_AUTOMASKED);
835 		else
836 			info.flags |= VFIO_IRQ_INFO_NORESIZE;
837 
838 		return copy_to_user((void __user *)arg, &info, minsz) ?
839 			-EFAULT : 0;
840 
841 	} else if (cmd == VFIO_DEVICE_SET_IRQS) {
842 		struct vfio_irq_set hdr;
843 		u8 *data = NULL;
844 		int max, ret = 0;
845 		size_t data_size = 0;
846 
847 		minsz = offsetofend(struct vfio_irq_set, count);
848 
849 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
850 			return -EFAULT;
851 
852 		max = vfio_pci_get_irq_count(vdev, hdr.index);
853 
854 		ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
855 						 VFIO_PCI_NUM_IRQS, &data_size);
856 		if (ret)
857 			return ret;
858 
859 		if (data_size) {
860 			data = memdup_user((void __user *)(arg + minsz),
861 					    data_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 = pci_reset_bus(vdev->pdev);
1049 
1050 hot_reset_release:
1051 		for (i--; i >= 0; i--)
1052 			vfio_group_put_external_user(groups[i].group);
1053 
1054 		kfree(groups);
1055 		return ret;
1056 	} else if (cmd == VFIO_DEVICE_IOEVENTFD) {
1057 		struct vfio_device_ioeventfd ioeventfd;
1058 		int count;
1059 
1060 		minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1061 
1062 		if (copy_from_user(&ioeventfd, (void __user *)arg, minsz))
1063 			return -EFAULT;
1064 
1065 		if (ioeventfd.argsz < minsz)
1066 			return -EINVAL;
1067 
1068 		if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1069 			return -EINVAL;
1070 
1071 		count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1072 
1073 		if (hweight8(count) != 1 || ioeventfd.fd < -1)
1074 			return -EINVAL;
1075 
1076 		return vfio_pci_ioeventfd(vdev, ioeventfd.offset,
1077 					  ioeventfd.data, count, ioeventfd.fd);
1078 	}
1079 
1080 	return -ENOTTY;
1081 }
1082 
1083 static ssize_t vfio_pci_rw(void *device_data, char __user *buf,
1084 			   size_t count, loff_t *ppos, bool iswrite)
1085 {
1086 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1087 	struct vfio_pci_device *vdev = device_data;
1088 
1089 	if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1090 		return -EINVAL;
1091 
1092 	switch (index) {
1093 	case VFIO_PCI_CONFIG_REGION_INDEX:
1094 		return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1095 
1096 	case VFIO_PCI_ROM_REGION_INDEX:
1097 		if (iswrite)
1098 			return -EINVAL;
1099 		return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1100 
1101 	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1102 		return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1103 
1104 	case VFIO_PCI_VGA_REGION_INDEX:
1105 		return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1106 	default:
1107 		index -= VFIO_PCI_NUM_REGIONS;
1108 		return vdev->region[index].ops->rw(vdev, buf,
1109 						   count, ppos, iswrite);
1110 	}
1111 
1112 	return -EINVAL;
1113 }
1114 
1115 static ssize_t vfio_pci_read(void *device_data, char __user *buf,
1116 			     size_t count, loff_t *ppos)
1117 {
1118 	if (!count)
1119 		return 0;
1120 
1121 	return vfio_pci_rw(device_data, buf, count, ppos, false);
1122 }
1123 
1124 static ssize_t vfio_pci_write(void *device_data, const char __user *buf,
1125 			      size_t count, loff_t *ppos)
1126 {
1127 	if (!count)
1128 		return 0;
1129 
1130 	return vfio_pci_rw(device_data, (char __user *)buf, count, ppos, true);
1131 }
1132 
1133 static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma)
1134 {
1135 	struct vfio_pci_device *vdev = device_data;
1136 	struct pci_dev *pdev = vdev->pdev;
1137 	unsigned int index;
1138 	u64 phys_len, req_len, pgoff, req_start;
1139 	int ret;
1140 
1141 	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1142 
1143 	if (vma->vm_end < vma->vm_start)
1144 		return -EINVAL;
1145 	if ((vma->vm_flags & VM_SHARED) == 0)
1146 		return -EINVAL;
1147 	if (index >= VFIO_PCI_NUM_REGIONS) {
1148 		int regnum = index - VFIO_PCI_NUM_REGIONS;
1149 		struct vfio_pci_region *region = vdev->region + regnum;
1150 
1151 		if (region && region->ops && region->ops->mmap &&
1152 		    (region->flags & VFIO_REGION_INFO_FLAG_MMAP))
1153 			return region->ops->mmap(vdev, region, vma);
1154 		return -EINVAL;
1155 	}
1156 	if (index >= VFIO_PCI_ROM_REGION_INDEX)
1157 		return -EINVAL;
1158 	if (!vdev->bar_mmap_supported[index])
1159 		return -EINVAL;
1160 
1161 	phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1162 	req_len = vma->vm_end - vma->vm_start;
1163 	pgoff = vma->vm_pgoff &
1164 		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1165 	req_start = pgoff << PAGE_SHIFT;
1166 
1167 	if (req_start + req_len > phys_len)
1168 		return -EINVAL;
1169 
1170 	/*
1171 	 * Even though we don't make use of the barmap for the mmap,
1172 	 * we need to request the region and the barmap tracks that.
1173 	 */
1174 	if (!vdev->barmap[index]) {
1175 		ret = pci_request_selected_regions(pdev,
1176 						   1 << index, "vfio-pci");
1177 		if (ret)
1178 			return ret;
1179 
1180 		vdev->barmap[index] = pci_iomap(pdev, index, 0);
1181 		if (!vdev->barmap[index]) {
1182 			pci_release_selected_regions(pdev, 1 << index);
1183 			return -ENOMEM;
1184 		}
1185 	}
1186 
1187 	vma->vm_private_data = vdev;
1188 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1189 	vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1190 
1191 	return remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1192 			       req_len, vma->vm_page_prot);
1193 }
1194 
1195 static void vfio_pci_request(void *device_data, unsigned int count)
1196 {
1197 	struct vfio_pci_device *vdev = device_data;
1198 
1199 	mutex_lock(&vdev->igate);
1200 
1201 	if (vdev->req_trigger) {
1202 		if (!(count % 10))
1203 			dev_notice_ratelimited(&vdev->pdev->dev,
1204 				"Relaying device request to user (#%u)\n",
1205 				count);
1206 		eventfd_signal(vdev->req_trigger, 1);
1207 	} else if (count == 0) {
1208 		dev_warn(&vdev->pdev->dev,
1209 			"No device request channel registered, blocked until released by user\n");
1210 	}
1211 
1212 	mutex_unlock(&vdev->igate);
1213 }
1214 
1215 static const struct vfio_device_ops vfio_pci_ops = {
1216 	.name		= "vfio-pci",
1217 	.open		= vfio_pci_open,
1218 	.release	= vfio_pci_release,
1219 	.ioctl		= vfio_pci_ioctl,
1220 	.read		= vfio_pci_read,
1221 	.write		= vfio_pci_write,
1222 	.mmap		= vfio_pci_mmap,
1223 	.request	= vfio_pci_request,
1224 };
1225 
1226 static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev);
1227 static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck);
1228 
1229 static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1230 {
1231 	struct vfio_pci_device *vdev;
1232 	struct iommu_group *group;
1233 	int ret;
1234 
1235 	if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1236 		return -EINVAL;
1237 
1238 	/*
1239 	 * Prevent binding to PFs with VFs enabled, this too easily allows
1240 	 * userspace instance with VFs and PFs from the same device, which
1241 	 * cannot work.  Disabling SR-IOV here would initiate removing the
1242 	 * VFs, which would unbind the driver, which is prone to blocking
1243 	 * if that VF is also in use by vfio-pci.  Just reject these PFs
1244 	 * and let the user sort it out.
1245 	 */
1246 	if (pci_num_vf(pdev)) {
1247 		pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
1248 		return -EBUSY;
1249 	}
1250 
1251 	group = vfio_iommu_group_get(&pdev->dev);
1252 	if (!group)
1253 		return -EINVAL;
1254 
1255 	vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
1256 	if (!vdev) {
1257 		vfio_iommu_group_put(group, &pdev->dev);
1258 		return -ENOMEM;
1259 	}
1260 
1261 	vdev->pdev = pdev;
1262 	vdev->irq_type = VFIO_PCI_NUM_IRQS;
1263 	mutex_init(&vdev->igate);
1264 	spin_lock_init(&vdev->irqlock);
1265 	mutex_init(&vdev->ioeventfds_lock);
1266 	INIT_LIST_HEAD(&vdev->ioeventfds_list);
1267 
1268 	ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
1269 	if (ret) {
1270 		vfio_iommu_group_put(group, &pdev->dev);
1271 		kfree(vdev);
1272 		return ret;
1273 	}
1274 
1275 	ret = vfio_pci_reflck_attach(vdev);
1276 	if (ret) {
1277 		vfio_del_group_dev(&pdev->dev);
1278 		vfio_iommu_group_put(group, &pdev->dev);
1279 		kfree(vdev);
1280 		return ret;
1281 	}
1282 
1283 	if (vfio_pci_is_vga(pdev)) {
1284 		vga_client_register(pdev, vdev, NULL, vfio_pci_set_vga_decode);
1285 		vga_set_legacy_decoding(pdev,
1286 					vfio_pci_set_vga_decode(vdev, false));
1287 	}
1288 
1289 	if (!disable_idle_d3) {
1290 		/*
1291 		 * pci-core sets the device power state to an unknown value at
1292 		 * bootup and after being removed from a driver.  The only
1293 		 * transition it allows from this unknown state is to D0, which
1294 		 * typically happens when a driver calls pci_enable_device().
1295 		 * We're not ready to enable the device yet, but we do want to
1296 		 * be able to get to D3.  Therefore first do a D0 transition
1297 		 * before going to D3.
1298 		 */
1299 		pci_set_power_state(pdev, PCI_D0);
1300 		pci_set_power_state(pdev, PCI_D3hot);
1301 	}
1302 
1303 	return ret;
1304 }
1305 
1306 static void vfio_pci_remove(struct pci_dev *pdev)
1307 {
1308 	struct vfio_pci_device *vdev;
1309 
1310 	vdev = vfio_del_group_dev(&pdev->dev);
1311 	if (!vdev)
1312 		return;
1313 
1314 	vfio_pci_reflck_put(vdev->reflck);
1315 
1316 	vfio_iommu_group_put(pdev->dev.iommu_group, &pdev->dev);
1317 	kfree(vdev->region);
1318 	mutex_destroy(&vdev->ioeventfds_lock);
1319 	kfree(vdev);
1320 
1321 	if (vfio_pci_is_vga(pdev)) {
1322 		vga_client_register(pdev, NULL, NULL, NULL);
1323 		vga_set_legacy_decoding(pdev,
1324 				VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
1325 				VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM);
1326 	}
1327 
1328 	if (!disable_idle_d3)
1329 		pci_set_power_state(pdev, PCI_D0);
1330 }
1331 
1332 static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
1333 						  pci_channel_state_t state)
1334 {
1335 	struct vfio_pci_device *vdev;
1336 	struct vfio_device *device;
1337 
1338 	device = vfio_device_get_from_dev(&pdev->dev);
1339 	if (device == NULL)
1340 		return PCI_ERS_RESULT_DISCONNECT;
1341 
1342 	vdev = vfio_device_data(device);
1343 	if (vdev == NULL) {
1344 		vfio_device_put(device);
1345 		return PCI_ERS_RESULT_DISCONNECT;
1346 	}
1347 
1348 	mutex_lock(&vdev->igate);
1349 
1350 	if (vdev->err_trigger)
1351 		eventfd_signal(vdev->err_trigger, 1);
1352 
1353 	mutex_unlock(&vdev->igate);
1354 
1355 	vfio_device_put(device);
1356 
1357 	return PCI_ERS_RESULT_CAN_RECOVER;
1358 }
1359 
1360 static const struct pci_error_handlers vfio_err_handlers = {
1361 	.error_detected = vfio_pci_aer_err_detected,
1362 };
1363 
1364 static struct pci_driver vfio_pci_driver = {
1365 	.name		= "vfio-pci",
1366 	.id_table	= NULL, /* only dynamic ids */
1367 	.probe		= vfio_pci_probe,
1368 	.remove		= vfio_pci_remove,
1369 	.err_handler	= &vfio_err_handlers,
1370 };
1371 
1372 static DEFINE_MUTEX(reflck_lock);
1373 
1374 static struct vfio_pci_reflck *vfio_pci_reflck_alloc(void)
1375 {
1376 	struct vfio_pci_reflck *reflck;
1377 
1378 	reflck = kzalloc(sizeof(*reflck), GFP_KERNEL);
1379 	if (!reflck)
1380 		return ERR_PTR(-ENOMEM);
1381 
1382 	kref_init(&reflck->kref);
1383 	mutex_init(&reflck->lock);
1384 
1385 	return reflck;
1386 }
1387 
1388 static void vfio_pci_reflck_get(struct vfio_pci_reflck *reflck)
1389 {
1390 	kref_get(&reflck->kref);
1391 }
1392 
1393 static int vfio_pci_reflck_find(struct pci_dev *pdev, void *data)
1394 {
1395 	struct vfio_pci_reflck **preflck = data;
1396 	struct vfio_device *device;
1397 	struct vfio_pci_device *vdev;
1398 
1399 	device = vfio_device_get_from_dev(&pdev->dev);
1400 	if (!device)
1401 		return 0;
1402 
1403 	if (pci_dev_driver(pdev) != &vfio_pci_driver) {
1404 		vfio_device_put(device);
1405 		return 0;
1406 	}
1407 
1408 	vdev = vfio_device_data(device);
1409 
1410 	if (vdev->reflck) {
1411 		vfio_pci_reflck_get(vdev->reflck);
1412 		*preflck = vdev->reflck;
1413 		vfio_device_put(device);
1414 		return 1;
1415 	}
1416 
1417 	vfio_device_put(device);
1418 	return 0;
1419 }
1420 
1421 static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev)
1422 {
1423 	bool slot = !pci_probe_reset_slot(vdev->pdev->slot);
1424 
1425 	mutex_lock(&reflck_lock);
1426 
1427 	if (pci_is_root_bus(vdev->pdev->bus) ||
1428 	    vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_reflck_find,
1429 					  &vdev->reflck, slot) <= 0)
1430 		vdev->reflck = vfio_pci_reflck_alloc();
1431 
1432 	mutex_unlock(&reflck_lock);
1433 
1434 	return PTR_ERR_OR_ZERO(vdev->reflck);
1435 }
1436 
1437 static void vfio_pci_reflck_release(struct kref *kref)
1438 {
1439 	struct vfio_pci_reflck *reflck = container_of(kref,
1440 						      struct vfio_pci_reflck,
1441 						      kref);
1442 
1443 	kfree(reflck);
1444 	mutex_unlock(&reflck_lock);
1445 }
1446 
1447 static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck)
1448 {
1449 	kref_put_mutex(&reflck->kref, vfio_pci_reflck_release, &reflck_lock);
1450 }
1451 
1452 struct vfio_devices {
1453 	struct vfio_device **devices;
1454 	int cur_index;
1455 	int max_index;
1456 };
1457 
1458 static int vfio_pci_get_unused_devs(struct pci_dev *pdev, void *data)
1459 {
1460 	struct vfio_devices *devs = data;
1461 	struct vfio_device *device;
1462 	struct vfio_pci_device *vdev;
1463 
1464 	if (devs->cur_index == devs->max_index)
1465 		return -ENOSPC;
1466 
1467 	device = vfio_device_get_from_dev(&pdev->dev);
1468 	if (!device)
1469 		return -EINVAL;
1470 
1471 	if (pci_dev_driver(pdev) != &vfio_pci_driver) {
1472 		vfio_device_put(device);
1473 		return -EBUSY;
1474 	}
1475 
1476 	vdev = vfio_device_data(device);
1477 
1478 	/* Fault if the device is not unused */
1479 	if (vdev->refcnt) {
1480 		vfio_device_put(device);
1481 		return -EBUSY;
1482 	}
1483 
1484 	devs->devices[devs->cur_index++] = device;
1485 	return 0;
1486 }
1487 
1488 /*
1489  * If a bus or slot reset is available for the provided device and:
1490  *  - All of the devices affected by that bus or slot reset are unused
1491  *    (!refcnt)
1492  *  - At least one of the affected devices is marked dirty via
1493  *    needs_reset (such as by lack of FLR support)
1494  * Then attempt to perform that bus or slot reset.  Callers are required
1495  * to hold vdev->reflck->lock, protecting the bus/slot reset group from
1496  * concurrent opens.  A vfio_device reference is acquired for each device
1497  * to prevent unbinds during the reset operation.
1498  *
1499  * NB: vfio-core considers a group to be viable even if some devices are
1500  * bound to drivers like pci-stub or pcieport.  Here we require all devices
1501  * to be bound to vfio_pci since that's the only way we can be sure they
1502  * stay put.
1503  */
1504 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev)
1505 {
1506 	struct vfio_devices devs = { .cur_index = 0 };
1507 	int i = 0, ret = -EINVAL;
1508 	bool slot = false;
1509 	struct vfio_pci_device *tmp;
1510 
1511 	if (!pci_probe_reset_slot(vdev->pdev->slot))
1512 		slot = true;
1513 	else if (pci_probe_reset_bus(vdev->pdev->bus))
1514 		return;
1515 
1516 	if (vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
1517 					  &i, slot) || !i)
1518 		return;
1519 
1520 	devs.max_index = i;
1521 	devs.devices = kcalloc(i, sizeof(struct vfio_device *), GFP_KERNEL);
1522 	if (!devs.devices)
1523 		return;
1524 
1525 	if (vfio_pci_for_each_slot_or_bus(vdev->pdev,
1526 					  vfio_pci_get_unused_devs,
1527 					  &devs, slot))
1528 		goto put_devs;
1529 
1530 	/* Does at least one need a reset? */
1531 	for (i = 0; i < devs.cur_index; i++) {
1532 		tmp = vfio_device_data(devs.devices[i]);
1533 		if (tmp->needs_reset) {
1534 			ret = pci_reset_bus(vdev->pdev);
1535 			break;
1536 		}
1537 	}
1538 
1539 put_devs:
1540 	for (i = 0; i < devs.cur_index; i++) {
1541 		tmp = vfio_device_data(devs.devices[i]);
1542 
1543 		/*
1544 		 * If reset was successful, affected devices no longer need
1545 		 * a reset and we should return all the collateral devices
1546 		 * to low power.  If not successful, we either didn't reset
1547 		 * the bus or timed out waiting for it, so let's not touch
1548 		 * the power state.
1549 		 */
1550 		if (!ret) {
1551 			tmp->needs_reset = false;
1552 
1553 			if (tmp != vdev && !disable_idle_d3)
1554 				pci_set_power_state(tmp->pdev, PCI_D3hot);
1555 		}
1556 
1557 		vfio_device_put(devs.devices[i]);
1558 	}
1559 
1560 	kfree(devs.devices);
1561 }
1562 
1563 static void __exit vfio_pci_cleanup(void)
1564 {
1565 	pci_unregister_driver(&vfio_pci_driver);
1566 	vfio_pci_uninit_perm_bits();
1567 }
1568 
1569 static void __init vfio_pci_fill_ids(void)
1570 {
1571 	char *p, *id;
1572 	int rc;
1573 
1574 	/* no ids passed actually */
1575 	if (ids[0] == '\0')
1576 		return;
1577 
1578 	/* add ids specified in the module parameter */
1579 	p = ids;
1580 	while ((id = strsep(&p, ","))) {
1581 		unsigned int vendor, device, subvendor = PCI_ANY_ID,
1582 			subdevice = PCI_ANY_ID, class = 0, class_mask = 0;
1583 		int fields;
1584 
1585 		if (!strlen(id))
1586 			continue;
1587 
1588 		fields = sscanf(id, "%x:%x:%x:%x:%x:%x",
1589 				&vendor, &device, &subvendor, &subdevice,
1590 				&class, &class_mask);
1591 
1592 		if (fields < 2) {
1593 			pr_warn("invalid id string \"%s\"\n", id);
1594 			continue;
1595 		}
1596 
1597 		rc = pci_add_dynid(&vfio_pci_driver, vendor, device,
1598 				   subvendor, subdevice, class, class_mask, 0);
1599 		if (rc)
1600 			pr_warn("failed to add dynamic id [%04hx:%04hx[%04hx:%04hx]] class %#08x/%08x (%d)\n",
1601 				vendor, device, subvendor, subdevice,
1602 				class, class_mask, rc);
1603 		else
1604 			pr_info("add [%04hx:%04hx[%04hx:%04hx]] class %#08x/%08x\n",
1605 				vendor, device, subvendor, subdevice,
1606 				class, class_mask);
1607 	}
1608 }
1609 
1610 static int __init vfio_pci_init(void)
1611 {
1612 	int ret;
1613 
1614 	/* Allocate shared config space permision data used by all devices */
1615 	ret = vfio_pci_init_perm_bits();
1616 	if (ret)
1617 		return ret;
1618 
1619 	/* Register and scan for devices */
1620 	ret = pci_register_driver(&vfio_pci_driver);
1621 	if (ret)
1622 		goto out_driver;
1623 
1624 	vfio_pci_fill_ids();
1625 
1626 	return 0;
1627 
1628 out_driver:
1629 	vfio_pci_uninit_perm_bits();
1630 	return ret;
1631 }
1632 
1633 module_init(vfio_pci_init);
1634 module_exit(vfio_pci_cleanup);
1635 
1636 MODULE_VERSION(DRIVER_VERSION);
1637 MODULE_LICENSE("GPL v2");
1638 MODULE_AUTHOR(DRIVER_AUTHOR);
1639 MODULE_DESCRIPTION(DRIVER_DESC);
1640