xref: /openbmc/linux/drivers/vfio/pci/vfio_pci_core.c (revision 2326f3cd)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
4  *     Author: Alex Williamson <alex.williamson@redhat.com>
5  *
6  * Derived from original vfio:
7  * Copyright 2010 Cisco Systems, Inc.  All rights reserved.
8  * Author: Tom Lyon, pugs@cisco.com
9  */
10 
11 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12 
13 #include <linux/device.h>
14 #include <linux/eventfd.h>
15 #include <linux/file.h>
16 #include <linux/interrupt.h>
17 #include <linux/iommu.h>
18 #include <linux/module.h>
19 #include <linux/mutex.h>
20 #include <linux/notifier.h>
21 #include <linux/pci.h>
22 #include <linux/pm_runtime.h>
23 #include <linux/slab.h>
24 #include <linux/types.h>
25 #include <linux/uaccess.h>
26 #include <linux/vgaarb.h>
27 #include <linux/nospec.h>
28 #include <linux/sched/mm.h>
29 
30 #include <linux/vfio_pci_core.h>
31 
32 #define DRIVER_AUTHOR   "Alex Williamson <alex.williamson@redhat.com>"
33 #define DRIVER_DESC "core driver for VFIO based PCI devices"
34 
35 static bool nointxmask;
36 static bool disable_vga;
37 static bool disable_idle_d3;
38 
39 static inline bool vfio_vga_disabled(void)
40 {
41 #ifdef CONFIG_VFIO_PCI_VGA
42 	return disable_vga;
43 #else
44 	return true;
45 #endif
46 }
47 
48 /*
49  * Our VGA arbiter participation is limited since we don't know anything
50  * about the device itself.  However, if the device is the only VGA device
51  * downstream of a bridge and VFIO VGA support is disabled, then we can
52  * safely return legacy VGA IO and memory as not decoded since the user
53  * has no way to get to it and routing can be disabled externally at the
54  * bridge.
55  */
56 static unsigned int vfio_pci_set_decode(struct pci_dev *pdev, bool single_vga)
57 {
58 	struct pci_dev *tmp = NULL;
59 	unsigned char max_busnr;
60 	unsigned int decodes;
61 
62 	if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
63 		return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
64 		       VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
65 
66 	max_busnr = pci_bus_max_busnr(pdev->bus);
67 	decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
68 
69 	while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
70 		if (tmp == pdev ||
71 		    pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
72 		    pci_is_root_bus(tmp->bus))
73 			continue;
74 
75 		if (tmp->bus->number >= pdev->bus->number &&
76 		    tmp->bus->number <= max_busnr) {
77 			pci_dev_put(tmp);
78 			decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
79 			break;
80 		}
81 	}
82 
83 	return decodes;
84 }
85 
86 static void vfio_pci_probe_mmaps(struct vfio_pci_core_device *vdev)
87 {
88 	struct resource *res;
89 	int i;
90 	struct vfio_pci_dummy_resource *dummy_res;
91 
92 	for (i = 0; i < PCI_STD_NUM_BARS; i++) {
93 		int bar = i + PCI_STD_RESOURCES;
94 
95 		res = &vdev->pdev->resource[bar];
96 
97 		if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
98 			goto no_mmap;
99 
100 		if (!(res->flags & IORESOURCE_MEM))
101 			goto no_mmap;
102 
103 		/*
104 		 * The PCI core shouldn't set up a resource with a
105 		 * type but zero size. But there may be bugs that
106 		 * cause us to do that.
107 		 */
108 		if (!resource_size(res))
109 			goto no_mmap;
110 
111 		if (resource_size(res) >= PAGE_SIZE) {
112 			vdev->bar_mmap_supported[bar] = true;
113 			continue;
114 		}
115 
116 		if (!(res->start & ~PAGE_MASK)) {
117 			/*
118 			 * Add a dummy resource to reserve the remainder
119 			 * of the exclusive page in case that hot-add
120 			 * device's bar is assigned into it.
121 			 */
122 			dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
123 			if (dummy_res == NULL)
124 				goto no_mmap;
125 
126 			dummy_res->resource.name = "vfio sub-page reserved";
127 			dummy_res->resource.start = res->end + 1;
128 			dummy_res->resource.end = res->start + PAGE_SIZE - 1;
129 			dummy_res->resource.flags = res->flags;
130 			if (request_resource(res->parent,
131 						&dummy_res->resource)) {
132 				kfree(dummy_res);
133 				goto no_mmap;
134 			}
135 			dummy_res->index = bar;
136 			list_add(&dummy_res->res_next,
137 					&vdev->dummy_resources_list);
138 			vdev->bar_mmap_supported[bar] = true;
139 			continue;
140 		}
141 		/*
142 		 * Here we don't handle the case when the BAR is not page
143 		 * aligned because we can't expect the BAR will be
144 		 * assigned into the same location in a page in guest
145 		 * when we passthrough the BAR. And it's hard to access
146 		 * this BAR in userspace because we have no way to get
147 		 * the BAR's location in a page.
148 		 */
149 no_mmap:
150 		vdev->bar_mmap_supported[bar] = false;
151 	}
152 }
153 
154 struct vfio_pci_group_info;
155 static bool vfio_pci_dev_set_try_reset(struct vfio_device_set *dev_set);
156 static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set,
157 				      struct vfio_pci_group_info *groups);
158 
159 /*
160  * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
161  * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
162  * If a device implements the former but not the latter we would typically
163  * expect broken_intx_masking be set and require an exclusive interrupt.
164  * However since we do have control of the device's ability to assert INTx,
165  * we can instead pretend that the device does not implement INTx, virtualizing
166  * the pin register to report zero and maintaining DisINTx set on the host.
167  */
168 static bool vfio_pci_nointx(struct pci_dev *pdev)
169 {
170 	switch (pdev->vendor) {
171 	case PCI_VENDOR_ID_INTEL:
172 		switch (pdev->device) {
173 		/* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
174 		case 0x1572:
175 		case 0x1574:
176 		case 0x1580 ... 0x1581:
177 		case 0x1583 ... 0x158b:
178 		case 0x37d0 ... 0x37d2:
179 		/* X550 */
180 		case 0x1563:
181 			return true;
182 		default:
183 			return false;
184 		}
185 	}
186 
187 	return false;
188 }
189 
190 static void vfio_pci_probe_power_state(struct vfio_pci_core_device *vdev)
191 {
192 	struct pci_dev *pdev = vdev->pdev;
193 	u16 pmcsr;
194 
195 	if (!pdev->pm_cap)
196 		return;
197 
198 	pci_read_config_word(pdev, pdev->pm_cap + PCI_PM_CTRL, &pmcsr);
199 
200 	vdev->needs_pm_restore = !(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
201 }
202 
203 /*
204  * pci_set_power_state() wrapper handling devices which perform a soft reset on
205  * D3->D0 transition.  Save state prior to D0/1/2->D3, stash it on the vdev,
206  * restore when returned to D0.  Saved separately from pci_saved_state for use
207  * by PM capability emulation and separately from pci_dev internal saved state
208  * to avoid it being overwritten and consumed around other resets.
209  */
210 int vfio_pci_set_power_state(struct vfio_pci_core_device *vdev, pci_power_t state)
211 {
212 	struct pci_dev *pdev = vdev->pdev;
213 	bool needs_restore = false, needs_save = false;
214 	int ret;
215 
216 	if (vdev->needs_pm_restore) {
217 		if (pdev->current_state < PCI_D3hot && state >= PCI_D3hot) {
218 			pci_save_state(pdev);
219 			needs_save = true;
220 		}
221 
222 		if (pdev->current_state >= PCI_D3hot && state <= PCI_D0)
223 			needs_restore = true;
224 	}
225 
226 	ret = pci_set_power_state(pdev, state);
227 
228 	if (!ret) {
229 		/* D3 might be unsupported via quirk, skip unless in D3 */
230 		if (needs_save && pdev->current_state >= PCI_D3hot) {
231 			vdev->pm_save = pci_store_saved_state(pdev);
232 		} else if (needs_restore) {
233 			pci_load_and_free_saved_state(pdev, &vdev->pm_save);
234 			pci_restore_state(pdev);
235 		}
236 	}
237 
238 	return ret;
239 }
240 
241 int vfio_pci_core_enable(struct vfio_pci_core_device *vdev)
242 {
243 	struct pci_dev *pdev = vdev->pdev;
244 	int ret;
245 	u16 cmd;
246 	u8 msix_pos;
247 
248 	vfio_pci_set_power_state(vdev, PCI_D0);
249 
250 	/* Don't allow our initial saved state to include busmaster */
251 	pci_clear_master(pdev);
252 
253 	ret = pci_enable_device(pdev);
254 	if (ret)
255 		return ret;
256 
257 	/* If reset fails because of the device lock, fail this path entirely */
258 	ret = pci_try_reset_function(pdev);
259 	if (ret == -EAGAIN) {
260 		pci_disable_device(pdev);
261 		return ret;
262 	}
263 
264 	vdev->reset_works = !ret;
265 	pci_save_state(pdev);
266 	vdev->pci_saved_state = pci_store_saved_state(pdev);
267 	if (!vdev->pci_saved_state)
268 		pci_dbg(pdev, "%s: Couldn't store saved state\n", __func__);
269 
270 	if (likely(!nointxmask)) {
271 		if (vfio_pci_nointx(pdev)) {
272 			pci_info(pdev, "Masking broken INTx support\n");
273 			vdev->nointx = true;
274 			pci_intx(pdev, 0);
275 		} else
276 			vdev->pci_2_3 = pci_intx_mask_supported(pdev);
277 	}
278 
279 	pci_read_config_word(pdev, PCI_COMMAND, &cmd);
280 	if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
281 		cmd &= ~PCI_COMMAND_INTX_DISABLE;
282 		pci_write_config_word(pdev, PCI_COMMAND, cmd);
283 	}
284 
285 	ret = vfio_config_init(vdev);
286 	if (ret) {
287 		kfree(vdev->pci_saved_state);
288 		vdev->pci_saved_state = NULL;
289 		pci_disable_device(pdev);
290 		return ret;
291 	}
292 
293 	msix_pos = pdev->msix_cap;
294 	if (msix_pos) {
295 		u16 flags;
296 		u32 table;
297 
298 		pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
299 		pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
300 
301 		vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
302 		vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
303 		vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
304 	} else
305 		vdev->msix_bar = 0xFF;
306 
307 	if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
308 		vdev->has_vga = true;
309 
310 
311 	return 0;
312 }
313 EXPORT_SYMBOL_GPL(vfio_pci_core_enable);
314 
315 void vfio_pci_core_disable(struct vfio_pci_core_device *vdev)
316 {
317 	struct pci_dev *pdev = vdev->pdev;
318 	struct vfio_pci_dummy_resource *dummy_res, *tmp;
319 	struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
320 	int i, bar;
321 
322 	/* For needs_reset */
323 	lockdep_assert_held(&vdev->vdev.dev_set->lock);
324 
325 	/* Stop the device from further DMA */
326 	pci_clear_master(pdev);
327 
328 	vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
329 				VFIO_IRQ_SET_ACTION_TRIGGER,
330 				vdev->irq_type, 0, 0, NULL);
331 
332 	/* Device closed, don't need mutex here */
333 	list_for_each_entry_safe(ioeventfd, ioeventfd_tmp,
334 				 &vdev->ioeventfds_list, next) {
335 		vfio_virqfd_disable(&ioeventfd->virqfd);
336 		list_del(&ioeventfd->next);
337 		kfree(ioeventfd);
338 	}
339 	vdev->ioeventfds_nr = 0;
340 
341 	vdev->virq_disabled = false;
342 
343 	for (i = 0; i < vdev->num_regions; i++)
344 		vdev->region[i].ops->release(vdev, &vdev->region[i]);
345 
346 	vdev->num_regions = 0;
347 	kfree(vdev->region);
348 	vdev->region = NULL; /* don't krealloc a freed pointer */
349 
350 	vfio_config_free(vdev);
351 
352 	for (i = 0; i < PCI_STD_NUM_BARS; i++) {
353 		bar = i + PCI_STD_RESOURCES;
354 		if (!vdev->barmap[bar])
355 			continue;
356 		pci_iounmap(pdev, vdev->barmap[bar]);
357 		pci_release_selected_regions(pdev, 1 << bar);
358 		vdev->barmap[bar] = NULL;
359 	}
360 
361 	list_for_each_entry_safe(dummy_res, tmp,
362 				 &vdev->dummy_resources_list, res_next) {
363 		list_del(&dummy_res->res_next);
364 		release_resource(&dummy_res->resource);
365 		kfree(dummy_res);
366 	}
367 
368 	vdev->needs_reset = true;
369 
370 	/*
371 	 * If we have saved state, restore it.  If we can reset the device,
372 	 * even better.  Resetting with current state seems better than
373 	 * nothing, but saving and restoring current state without reset
374 	 * is just busy work.
375 	 */
376 	if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
377 		pci_info(pdev, "%s: Couldn't reload saved state\n", __func__);
378 
379 		if (!vdev->reset_works)
380 			goto out;
381 
382 		pci_save_state(pdev);
383 	}
384 
385 	/*
386 	 * Disable INTx and MSI, presumably to avoid spurious interrupts
387 	 * during reset.  Stolen from pci_reset_function()
388 	 */
389 	pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
390 
391 	/*
392 	 * Try to get the locks ourselves to prevent a deadlock. The
393 	 * success of this is dependent on being able to lock the device,
394 	 * which is not always possible.
395 	 * We can not use the "try" reset interface here, which will
396 	 * overwrite the previously restored configuration information.
397 	 */
398 	if (vdev->reset_works && pci_dev_trylock(pdev)) {
399 		if (!__pci_reset_function_locked(pdev))
400 			vdev->needs_reset = false;
401 		pci_dev_unlock(pdev);
402 	}
403 
404 	pci_restore_state(pdev);
405 out:
406 	pci_disable_device(pdev);
407 
408 	if (!vfio_pci_dev_set_try_reset(vdev->vdev.dev_set) && !disable_idle_d3)
409 		vfio_pci_set_power_state(vdev, PCI_D3hot);
410 }
411 EXPORT_SYMBOL_GPL(vfio_pci_core_disable);
412 
413 static struct vfio_pci_core_device *get_pf_vdev(struct vfio_pci_core_device *vdev)
414 {
415 	struct pci_dev *physfn = pci_physfn(vdev->pdev);
416 	struct vfio_device *pf_dev;
417 
418 	if (!vdev->pdev->is_virtfn)
419 		return NULL;
420 
421 	pf_dev = vfio_device_get_from_dev(&physfn->dev);
422 	if (!pf_dev)
423 		return NULL;
424 
425 	if (pci_dev_driver(physfn) != pci_dev_driver(vdev->pdev)) {
426 		vfio_device_put(pf_dev);
427 		return NULL;
428 	}
429 
430 	return container_of(pf_dev, struct vfio_pci_core_device, vdev);
431 }
432 
433 static void vfio_pci_vf_token_user_add(struct vfio_pci_core_device *vdev, int val)
434 {
435 	struct vfio_pci_core_device *pf_vdev = get_pf_vdev(vdev);
436 
437 	if (!pf_vdev)
438 		return;
439 
440 	mutex_lock(&pf_vdev->vf_token->lock);
441 	pf_vdev->vf_token->users += val;
442 	WARN_ON(pf_vdev->vf_token->users < 0);
443 	mutex_unlock(&pf_vdev->vf_token->lock);
444 
445 	vfio_device_put(&pf_vdev->vdev);
446 }
447 
448 void vfio_pci_core_close_device(struct vfio_device *core_vdev)
449 {
450 	struct vfio_pci_core_device *vdev =
451 		container_of(core_vdev, struct vfio_pci_core_device, vdev);
452 
453 	vfio_pci_vf_token_user_add(vdev, -1);
454 	vfio_spapr_pci_eeh_release(vdev->pdev);
455 	vfio_pci_core_disable(vdev);
456 
457 	mutex_lock(&vdev->igate);
458 	if (vdev->err_trigger) {
459 		eventfd_ctx_put(vdev->err_trigger);
460 		vdev->err_trigger = NULL;
461 	}
462 	if (vdev->req_trigger) {
463 		eventfd_ctx_put(vdev->req_trigger);
464 		vdev->req_trigger = NULL;
465 	}
466 	mutex_unlock(&vdev->igate);
467 }
468 EXPORT_SYMBOL_GPL(vfio_pci_core_close_device);
469 
470 void vfio_pci_core_finish_enable(struct vfio_pci_core_device *vdev)
471 {
472 	vfio_pci_probe_mmaps(vdev);
473 	vfio_spapr_pci_eeh_open(vdev->pdev);
474 	vfio_pci_vf_token_user_add(vdev, 1);
475 }
476 EXPORT_SYMBOL_GPL(vfio_pci_core_finish_enable);
477 
478 static int vfio_pci_get_irq_count(struct vfio_pci_core_device *vdev, int irq_type)
479 {
480 	if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
481 		u8 pin;
482 
483 		if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
484 		    vdev->nointx || vdev->pdev->is_virtfn)
485 			return 0;
486 
487 		pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
488 
489 		return pin ? 1 : 0;
490 	} else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
491 		u8 pos;
492 		u16 flags;
493 
494 		pos = vdev->pdev->msi_cap;
495 		if (pos) {
496 			pci_read_config_word(vdev->pdev,
497 					     pos + PCI_MSI_FLAGS, &flags);
498 			return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
499 		}
500 	} else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
501 		u8 pos;
502 		u16 flags;
503 
504 		pos = vdev->pdev->msix_cap;
505 		if (pos) {
506 			pci_read_config_word(vdev->pdev,
507 					     pos + PCI_MSIX_FLAGS, &flags);
508 
509 			return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
510 		}
511 	} else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
512 		if (pci_is_pcie(vdev->pdev))
513 			return 1;
514 	} else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
515 		return 1;
516 	}
517 
518 	return 0;
519 }
520 
521 static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
522 {
523 	(*(int *)data)++;
524 	return 0;
525 }
526 
527 struct vfio_pci_fill_info {
528 	int max;
529 	int cur;
530 	struct vfio_pci_dependent_device *devices;
531 };
532 
533 static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
534 {
535 	struct vfio_pci_fill_info *fill = data;
536 	struct iommu_group *iommu_group;
537 
538 	if (fill->cur == fill->max)
539 		return -EAGAIN; /* Something changed, try again */
540 
541 	iommu_group = iommu_group_get(&pdev->dev);
542 	if (!iommu_group)
543 		return -EPERM; /* Cannot reset non-isolated devices */
544 
545 	fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
546 	fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
547 	fill->devices[fill->cur].bus = pdev->bus->number;
548 	fill->devices[fill->cur].devfn = pdev->devfn;
549 	fill->cur++;
550 	iommu_group_put(iommu_group);
551 	return 0;
552 }
553 
554 struct vfio_pci_group_info {
555 	int count;
556 	struct vfio_group **groups;
557 };
558 
559 static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
560 {
561 	for (; pdev; pdev = pdev->bus->self)
562 		if (pdev->bus == slot->bus)
563 			return (pdev->slot == slot);
564 	return false;
565 }
566 
567 struct vfio_pci_walk_info {
568 	int (*fn)(struct pci_dev *, void *data);
569 	void *data;
570 	struct pci_dev *pdev;
571 	bool slot;
572 	int ret;
573 };
574 
575 static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
576 {
577 	struct vfio_pci_walk_info *walk = data;
578 
579 	if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
580 		walk->ret = walk->fn(pdev, walk->data);
581 
582 	return walk->ret;
583 }
584 
585 static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
586 					 int (*fn)(struct pci_dev *,
587 						   void *data), void *data,
588 					 bool slot)
589 {
590 	struct vfio_pci_walk_info walk = {
591 		.fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
592 	};
593 
594 	pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
595 
596 	return walk.ret;
597 }
598 
599 static int msix_mmappable_cap(struct vfio_pci_core_device *vdev,
600 			      struct vfio_info_cap *caps)
601 {
602 	struct vfio_info_cap_header header = {
603 		.id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
604 		.version = 1
605 	};
606 
607 	return vfio_info_add_capability(caps, &header, sizeof(header));
608 }
609 
610 int vfio_pci_register_dev_region(struct vfio_pci_core_device *vdev,
611 				 unsigned int type, unsigned int subtype,
612 				 const struct vfio_pci_regops *ops,
613 				 size_t size, u32 flags, void *data)
614 {
615 	struct vfio_pci_region *region;
616 
617 	region = krealloc(vdev->region,
618 			  (vdev->num_regions + 1) * sizeof(*region),
619 			  GFP_KERNEL);
620 	if (!region)
621 		return -ENOMEM;
622 
623 	vdev->region = region;
624 	vdev->region[vdev->num_regions].type = type;
625 	vdev->region[vdev->num_regions].subtype = subtype;
626 	vdev->region[vdev->num_regions].ops = ops;
627 	vdev->region[vdev->num_regions].size = size;
628 	vdev->region[vdev->num_regions].flags = flags;
629 	vdev->region[vdev->num_regions].data = data;
630 
631 	vdev->num_regions++;
632 
633 	return 0;
634 }
635 EXPORT_SYMBOL_GPL(vfio_pci_register_dev_region);
636 
637 long vfio_pci_core_ioctl(struct vfio_device *core_vdev, unsigned int cmd,
638 		unsigned long arg)
639 {
640 	struct vfio_pci_core_device *vdev =
641 		container_of(core_vdev, struct vfio_pci_core_device, vdev);
642 	unsigned long minsz;
643 
644 	if (cmd == VFIO_DEVICE_GET_INFO) {
645 		struct vfio_device_info info;
646 		struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
647 		unsigned long capsz;
648 		int ret;
649 
650 		minsz = offsetofend(struct vfio_device_info, num_irqs);
651 
652 		/* For backward compatibility, cannot require this */
653 		capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
654 
655 		if (copy_from_user(&info, (void __user *)arg, minsz))
656 			return -EFAULT;
657 
658 		if (info.argsz < minsz)
659 			return -EINVAL;
660 
661 		if (info.argsz >= capsz) {
662 			minsz = capsz;
663 			info.cap_offset = 0;
664 		}
665 
666 		info.flags = VFIO_DEVICE_FLAGS_PCI;
667 
668 		if (vdev->reset_works)
669 			info.flags |= VFIO_DEVICE_FLAGS_RESET;
670 
671 		info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
672 		info.num_irqs = VFIO_PCI_NUM_IRQS;
673 
674 		ret = vfio_pci_info_zdev_add_caps(vdev, &caps);
675 		if (ret && ret != -ENODEV) {
676 			pci_warn(vdev->pdev, "Failed to setup zPCI info capabilities\n");
677 			return ret;
678 		}
679 
680 		if (caps.size) {
681 			info.flags |= VFIO_DEVICE_FLAGS_CAPS;
682 			if (info.argsz < sizeof(info) + caps.size) {
683 				info.argsz = sizeof(info) + caps.size;
684 			} else {
685 				vfio_info_cap_shift(&caps, sizeof(info));
686 				if (copy_to_user((void __user *)arg +
687 						  sizeof(info), caps.buf,
688 						  caps.size)) {
689 					kfree(caps.buf);
690 					return -EFAULT;
691 				}
692 				info.cap_offset = sizeof(info);
693 			}
694 
695 			kfree(caps.buf);
696 		}
697 
698 		return copy_to_user((void __user *)arg, &info, minsz) ?
699 			-EFAULT : 0;
700 
701 	} else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
702 		struct pci_dev *pdev = vdev->pdev;
703 		struct vfio_region_info info;
704 		struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
705 		int i, ret;
706 
707 		minsz = offsetofend(struct vfio_region_info, offset);
708 
709 		if (copy_from_user(&info, (void __user *)arg, minsz))
710 			return -EFAULT;
711 
712 		if (info.argsz < minsz)
713 			return -EINVAL;
714 
715 		switch (info.index) {
716 		case VFIO_PCI_CONFIG_REGION_INDEX:
717 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
718 			info.size = pdev->cfg_size;
719 			info.flags = VFIO_REGION_INFO_FLAG_READ |
720 				     VFIO_REGION_INFO_FLAG_WRITE;
721 			break;
722 		case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
723 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
724 			info.size = pci_resource_len(pdev, info.index);
725 			if (!info.size) {
726 				info.flags = 0;
727 				break;
728 			}
729 
730 			info.flags = VFIO_REGION_INFO_FLAG_READ |
731 				     VFIO_REGION_INFO_FLAG_WRITE;
732 			if (vdev->bar_mmap_supported[info.index]) {
733 				info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
734 				if (info.index == vdev->msix_bar) {
735 					ret = msix_mmappable_cap(vdev, &caps);
736 					if (ret)
737 						return ret;
738 				}
739 			}
740 
741 			break;
742 		case VFIO_PCI_ROM_REGION_INDEX:
743 		{
744 			void __iomem *io;
745 			size_t size;
746 			u16 cmd;
747 
748 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
749 			info.flags = 0;
750 
751 			/* Report the BAR size, not the ROM size */
752 			info.size = pci_resource_len(pdev, info.index);
753 			if (!info.size) {
754 				/* Shadow ROMs appear as PCI option ROMs */
755 				if (pdev->resource[PCI_ROM_RESOURCE].flags &
756 							IORESOURCE_ROM_SHADOW)
757 					info.size = 0x20000;
758 				else
759 					break;
760 			}
761 
762 			/*
763 			 * Is it really there?  Enable memory decode for
764 			 * implicit access in pci_map_rom().
765 			 */
766 			cmd = vfio_pci_memory_lock_and_enable(vdev);
767 			io = pci_map_rom(pdev, &size);
768 			if (io) {
769 				info.flags = VFIO_REGION_INFO_FLAG_READ;
770 				pci_unmap_rom(pdev, io);
771 			} else {
772 				info.size = 0;
773 			}
774 			vfio_pci_memory_unlock_and_restore(vdev, cmd);
775 
776 			break;
777 		}
778 		case VFIO_PCI_VGA_REGION_INDEX:
779 			if (!vdev->has_vga)
780 				return -EINVAL;
781 
782 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
783 			info.size = 0xc0000;
784 			info.flags = VFIO_REGION_INFO_FLAG_READ |
785 				     VFIO_REGION_INFO_FLAG_WRITE;
786 
787 			break;
788 		default:
789 		{
790 			struct vfio_region_info_cap_type cap_type = {
791 					.header.id = VFIO_REGION_INFO_CAP_TYPE,
792 					.header.version = 1 };
793 
794 			if (info.index >=
795 			    VFIO_PCI_NUM_REGIONS + vdev->num_regions)
796 				return -EINVAL;
797 			info.index = array_index_nospec(info.index,
798 							VFIO_PCI_NUM_REGIONS +
799 							vdev->num_regions);
800 
801 			i = info.index - VFIO_PCI_NUM_REGIONS;
802 
803 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
804 			info.size = vdev->region[i].size;
805 			info.flags = vdev->region[i].flags;
806 
807 			cap_type.type = vdev->region[i].type;
808 			cap_type.subtype = vdev->region[i].subtype;
809 
810 			ret = vfio_info_add_capability(&caps, &cap_type.header,
811 						       sizeof(cap_type));
812 			if (ret)
813 				return ret;
814 
815 			if (vdev->region[i].ops->add_capability) {
816 				ret = vdev->region[i].ops->add_capability(vdev,
817 						&vdev->region[i], &caps);
818 				if (ret)
819 					return ret;
820 			}
821 		}
822 		}
823 
824 		if (caps.size) {
825 			info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
826 			if (info.argsz < sizeof(info) + caps.size) {
827 				info.argsz = sizeof(info) + caps.size;
828 				info.cap_offset = 0;
829 			} else {
830 				vfio_info_cap_shift(&caps, sizeof(info));
831 				if (copy_to_user((void __user *)arg +
832 						  sizeof(info), caps.buf,
833 						  caps.size)) {
834 					kfree(caps.buf);
835 					return -EFAULT;
836 				}
837 				info.cap_offset = sizeof(info);
838 			}
839 
840 			kfree(caps.buf);
841 		}
842 
843 		return copy_to_user((void __user *)arg, &info, minsz) ?
844 			-EFAULT : 0;
845 
846 	} else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
847 		struct vfio_irq_info info;
848 
849 		minsz = offsetofend(struct vfio_irq_info, count);
850 
851 		if (copy_from_user(&info, (void __user *)arg, minsz))
852 			return -EFAULT;
853 
854 		if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
855 			return -EINVAL;
856 
857 		switch (info.index) {
858 		case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
859 		case VFIO_PCI_REQ_IRQ_INDEX:
860 			break;
861 		case VFIO_PCI_ERR_IRQ_INDEX:
862 			if (pci_is_pcie(vdev->pdev))
863 				break;
864 			fallthrough;
865 		default:
866 			return -EINVAL;
867 		}
868 
869 		info.flags = VFIO_IRQ_INFO_EVENTFD;
870 
871 		info.count = vfio_pci_get_irq_count(vdev, info.index);
872 
873 		if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
874 			info.flags |= (VFIO_IRQ_INFO_MASKABLE |
875 				       VFIO_IRQ_INFO_AUTOMASKED);
876 		else
877 			info.flags |= VFIO_IRQ_INFO_NORESIZE;
878 
879 		return copy_to_user((void __user *)arg, &info, minsz) ?
880 			-EFAULT : 0;
881 
882 	} else if (cmd == VFIO_DEVICE_SET_IRQS) {
883 		struct vfio_irq_set hdr;
884 		u8 *data = NULL;
885 		int max, ret = 0;
886 		size_t data_size = 0;
887 
888 		minsz = offsetofend(struct vfio_irq_set, count);
889 
890 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
891 			return -EFAULT;
892 
893 		max = vfio_pci_get_irq_count(vdev, hdr.index);
894 
895 		ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
896 						 VFIO_PCI_NUM_IRQS, &data_size);
897 		if (ret)
898 			return ret;
899 
900 		if (data_size) {
901 			data = memdup_user((void __user *)(arg + minsz),
902 					    data_size);
903 			if (IS_ERR(data))
904 				return PTR_ERR(data);
905 		}
906 
907 		mutex_lock(&vdev->igate);
908 
909 		ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
910 					      hdr.start, hdr.count, data);
911 
912 		mutex_unlock(&vdev->igate);
913 		kfree(data);
914 
915 		return ret;
916 
917 	} else if (cmd == VFIO_DEVICE_RESET) {
918 		int ret;
919 
920 		if (!vdev->reset_works)
921 			return -EINVAL;
922 
923 		vfio_pci_zap_and_down_write_memory_lock(vdev);
924 		ret = pci_try_reset_function(vdev->pdev);
925 		up_write(&vdev->memory_lock);
926 
927 		return ret;
928 
929 	} else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
930 		struct vfio_pci_hot_reset_info hdr;
931 		struct vfio_pci_fill_info fill = { 0 };
932 		struct vfio_pci_dependent_device *devices = NULL;
933 		bool slot = false;
934 		int ret = 0;
935 
936 		minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
937 
938 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
939 			return -EFAULT;
940 
941 		if (hdr.argsz < minsz)
942 			return -EINVAL;
943 
944 		hdr.flags = 0;
945 
946 		/* Can we do a slot or bus reset or neither? */
947 		if (!pci_probe_reset_slot(vdev->pdev->slot))
948 			slot = true;
949 		else if (pci_probe_reset_bus(vdev->pdev->bus))
950 			return -ENODEV;
951 
952 		/* How many devices are affected? */
953 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
954 						    vfio_pci_count_devs,
955 						    &fill.max, slot);
956 		if (ret)
957 			return ret;
958 
959 		WARN_ON(!fill.max); /* Should always be at least one */
960 
961 		/*
962 		 * If there's enough space, fill it now, otherwise return
963 		 * -ENOSPC and the number of devices affected.
964 		 */
965 		if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
966 			ret = -ENOSPC;
967 			hdr.count = fill.max;
968 			goto reset_info_exit;
969 		}
970 
971 		devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
972 		if (!devices)
973 			return -ENOMEM;
974 
975 		fill.devices = devices;
976 
977 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
978 						    vfio_pci_fill_devs,
979 						    &fill, slot);
980 
981 		/*
982 		 * If a device was removed between counting and filling,
983 		 * we may come up short of fill.max.  If a device was
984 		 * added, we'll have a return of -EAGAIN above.
985 		 */
986 		if (!ret)
987 			hdr.count = fill.cur;
988 
989 reset_info_exit:
990 		if (copy_to_user((void __user *)arg, &hdr, minsz))
991 			ret = -EFAULT;
992 
993 		if (!ret) {
994 			if (copy_to_user((void __user *)(arg + minsz), devices,
995 					 hdr.count * sizeof(*devices)))
996 				ret = -EFAULT;
997 		}
998 
999 		kfree(devices);
1000 		return ret;
1001 
1002 	} else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
1003 		struct vfio_pci_hot_reset hdr;
1004 		int32_t *group_fds;
1005 		struct vfio_group **groups;
1006 		struct vfio_pci_group_info info;
1007 		bool slot = false;
1008 		int group_idx, count = 0, ret = 0;
1009 
1010 		minsz = offsetofend(struct vfio_pci_hot_reset, count);
1011 
1012 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
1013 			return -EFAULT;
1014 
1015 		if (hdr.argsz < minsz || hdr.flags)
1016 			return -EINVAL;
1017 
1018 		/* Can we do a slot or bus reset or neither? */
1019 		if (!pci_probe_reset_slot(vdev->pdev->slot))
1020 			slot = true;
1021 		else if (pci_probe_reset_bus(vdev->pdev->bus))
1022 			return -ENODEV;
1023 
1024 		/*
1025 		 * We can't let userspace give us an arbitrarily large
1026 		 * buffer to copy, so verify how many we think there
1027 		 * could be.  Note groups can have multiple devices so
1028 		 * one group per device is the max.
1029 		 */
1030 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1031 						    vfio_pci_count_devs,
1032 						    &count, slot);
1033 		if (ret)
1034 			return ret;
1035 
1036 		/* Somewhere between 1 and count is OK */
1037 		if (!hdr.count || hdr.count > count)
1038 			return -EINVAL;
1039 
1040 		group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
1041 		groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
1042 		if (!group_fds || !groups) {
1043 			kfree(group_fds);
1044 			kfree(groups);
1045 			return -ENOMEM;
1046 		}
1047 
1048 		if (copy_from_user(group_fds, (void __user *)(arg + minsz),
1049 				   hdr.count * sizeof(*group_fds))) {
1050 			kfree(group_fds);
1051 			kfree(groups);
1052 			return -EFAULT;
1053 		}
1054 
1055 		/*
1056 		 * For each group_fd, get the group through the vfio external
1057 		 * user interface and store the group and iommu ID.  This
1058 		 * ensures the group is held across the reset.
1059 		 */
1060 		for (group_idx = 0; group_idx < hdr.count; group_idx++) {
1061 			struct vfio_group *group;
1062 			struct fd f = fdget(group_fds[group_idx]);
1063 			if (!f.file) {
1064 				ret = -EBADF;
1065 				break;
1066 			}
1067 
1068 			group = vfio_group_get_external_user(f.file);
1069 			fdput(f);
1070 			if (IS_ERR(group)) {
1071 				ret = PTR_ERR(group);
1072 				break;
1073 			}
1074 
1075 			groups[group_idx] = group;
1076 		}
1077 
1078 		kfree(group_fds);
1079 
1080 		/* release reference to groups on error */
1081 		if (ret)
1082 			goto hot_reset_release;
1083 
1084 		info.count = hdr.count;
1085 		info.groups = groups;
1086 
1087 		ret = vfio_pci_dev_set_hot_reset(vdev->vdev.dev_set, &info);
1088 
1089 hot_reset_release:
1090 		for (group_idx--; group_idx >= 0; group_idx--)
1091 			vfio_group_put_external_user(groups[group_idx]);
1092 
1093 		kfree(groups);
1094 		return ret;
1095 	} else if (cmd == VFIO_DEVICE_IOEVENTFD) {
1096 		struct vfio_device_ioeventfd ioeventfd;
1097 		int count;
1098 
1099 		minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1100 
1101 		if (copy_from_user(&ioeventfd, (void __user *)arg, minsz))
1102 			return -EFAULT;
1103 
1104 		if (ioeventfd.argsz < minsz)
1105 			return -EINVAL;
1106 
1107 		if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1108 			return -EINVAL;
1109 
1110 		count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1111 
1112 		if (hweight8(count) != 1 || ioeventfd.fd < -1)
1113 			return -EINVAL;
1114 
1115 		return vfio_pci_ioeventfd(vdev, ioeventfd.offset,
1116 					  ioeventfd.data, count, ioeventfd.fd);
1117 	} else if (cmd == VFIO_DEVICE_FEATURE) {
1118 		struct vfio_device_feature feature;
1119 		uuid_t uuid;
1120 
1121 		minsz = offsetofend(struct vfio_device_feature, flags);
1122 
1123 		if (copy_from_user(&feature, (void __user *)arg, minsz))
1124 			return -EFAULT;
1125 
1126 		if (feature.argsz < minsz)
1127 			return -EINVAL;
1128 
1129 		/* Check unknown flags */
1130 		if (feature.flags & ~(VFIO_DEVICE_FEATURE_MASK |
1131 				      VFIO_DEVICE_FEATURE_SET |
1132 				      VFIO_DEVICE_FEATURE_GET |
1133 				      VFIO_DEVICE_FEATURE_PROBE))
1134 			return -EINVAL;
1135 
1136 		/* GET & SET are mutually exclusive except with PROBE */
1137 		if (!(feature.flags & VFIO_DEVICE_FEATURE_PROBE) &&
1138 		    (feature.flags & VFIO_DEVICE_FEATURE_SET) &&
1139 		    (feature.flags & VFIO_DEVICE_FEATURE_GET))
1140 			return -EINVAL;
1141 
1142 		switch (feature.flags & VFIO_DEVICE_FEATURE_MASK) {
1143 		case VFIO_DEVICE_FEATURE_PCI_VF_TOKEN:
1144 			if (!vdev->vf_token)
1145 				return -ENOTTY;
1146 
1147 			/*
1148 			 * We do not support GET of the VF Token UUID as this
1149 			 * could expose the token of the previous device user.
1150 			 */
1151 			if (feature.flags & VFIO_DEVICE_FEATURE_GET)
1152 				return -EINVAL;
1153 
1154 			if (feature.flags & VFIO_DEVICE_FEATURE_PROBE)
1155 				return 0;
1156 
1157 			/* Don't SET unless told to do so */
1158 			if (!(feature.flags & VFIO_DEVICE_FEATURE_SET))
1159 				return -EINVAL;
1160 
1161 			if (feature.argsz < minsz + sizeof(uuid))
1162 				return -EINVAL;
1163 
1164 			if (copy_from_user(&uuid, (void __user *)(arg + minsz),
1165 					   sizeof(uuid)))
1166 				return -EFAULT;
1167 
1168 			mutex_lock(&vdev->vf_token->lock);
1169 			uuid_copy(&vdev->vf_token->uuid, &uuid);
1170 			mutex_unlock(&vdev->vf_token->lock);
1171 
1172 			return 0;
1173 		default:
1174 			return -ENOTTY;
1175 		}
1176 	}
1177 
1178 	return -ENOTTY;
1179 }
1180 EXPORT_SYMBOL_GPL(vfio_pci_core_ioctl);
1181 
1182 static ssize_t vfio_pci_rw(struct vfio_pci_core_device *vdev, char __user *buf,
1183 			   size_t count, loff_t *ppos, bool iswrite)
1184 {
1185 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1186 
1187 	if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1188 		return -EINVAL;
1189 
1190 	switch (index) {
1191 	case VFIO_PCI_CONFIG_REGION_INDEX:
1192 		return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1193 
1194 	case VFIO_PCI_ROM_REGION_INDEX:
1195 		if (iswrite)
1196 			return -EINVAL;
1197 		return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1198 
1199 	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1200 		return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1201 
1202 	case VFIO_PCI_VGA_REGION_INDEX:
1203 		return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1204 	default:
1205 		index -= VFIO_PCI_NUM_REGIONS;
1206 		return vdev->region[index].ops->rw(vdev, buf,
1207 						   count, ppos, iswrite);
1208 	}
1209 
1210 	return -EINVAL;
1211 }
1212 
1213 ssize_t vfio_pci_core_read(struct vfio_device *core_vdev, char __user *buf,
1214 		size_t count, loff_t *ppos)
1215 {
1216 	struct vfio_pci_core_device *vdev =
1217 		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1218 
1219 	if (!count)
1220 		return 0;
1221 
1222 	return vfio_pci_rw(vdev, buf, count, ppos, false);
1223 }
1224 EXPORT_SYMBOL_GPL(vfio_pci_core_read);
1225 
1226 ssize_t vfio_pci_core_write(struct vfio_device *core_vdev, const char __user *buf,
1227 		size_t count, loff_t *ppos)
1228 {
1229 	struct vfio_pci_core_device *vdev =
1230 		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1231 
1232 	if (!count)
1233 		return 0;
1234 
1235 	return vfio_pci_rw(vdev, (char __user *)buf, count, ppos, true);
1236 }
1237 EXPORT_SYMBOL_GPL(vfio_pci_core_write);
1238 
1239 /* Return 1 on zap and vma_lock acquired, 0 on contention (only with @try) */
1240 static int vfio_pci_zap_and_vma_lock(struct vfio_pci_core_device *vdev, bool try)
1241 {
1242 	struct vfio_pci_mmap_vma *mmap_vma, *tmp;
1243 
1244 	/*
1245 	 * Lock ordering:
1246 	 * vma_lock is nested under mmap_lock for vm_ops callback paths.
1247 	 * The memory_lock semaphore is used by both code paths calling
1248 	 * into this function to zap vmas and the vm_ops.fault callback
1249 	 * to protect the memory enable state of the device.
1250 	 *
1251 	 * When zapping vmas we need to maintain the mmap_lock => vma_lock
1252 	 * ordering, which requires using vma_lock to walk vma_list to
1253 	 * acquire an mm, then dropping vma_lock to get the mmap_lock and
1254 	 * reacquiring vma_lock.  This logic is derived from similar
1255 	 * requirements in uverbs_user_mmap_disassociate().
1256 	 *
1257 	 * mmap_lock must always be the top-level lock when it is taken.
1258 	 * Therefore we can only hold the memory_lock write lock when
1259 	 * vma_list is empty, as we'd need to take mmap_lock to clear
1260 	 * entries.  vma_list can only be guaranteed empty when holding
1261 	 * vma_lock, thus memory_lock is nested under vma_lock.
1262 	 *
1263 	 * This enables the vm_ops.fault callback to acquire vma_lock,
1264 	 * followed by memory_lock read lock, while already holding
1265 	 * mmap_lock without risk of deadlock.
1266 	 */
1267 	while (1) {
1268 		struct mm_struct *mm = NULL;
1269 
1270 		if (try) {
1271 			if (!mutex_trylock(&vdev->vma_lock))
1272 				return 0;
1273 		} else {
1274 			mutex_lock(&vdev->vma_lock);
1275 		}
1276 		while (!list_empty(&vdev->vma_list)) {
1277 			mmap_vma = list_first_entry(&vdev->vma_list,
1278 						    struct vfio_pci_mmap_vma,
1279 						    vma_next);
1280 			mm = mmap_vma->vma->vm_mm;
1281 			if (mmget_not_zero(mm))
1282 				break;
1283 
1284 			list_del(&mmap_vma->vma_next);
1285 			kfree(mmap_vma);
1286 			mm = NULL;
1287 		}
1288 		if (!mm)
1289 			return 1;
1290 		mutex_unlock(&vdev->vma_lock);
1291 
1292 		if (try) {
1293 			if (!mmap_read_trylock(mm)) {
1294 				mmput(mm);
1295 				return 0;
1296 			}
1297 		} else {
1298 			mmap_read_lock(mm);
1299 		}
1300 		if (try) {
1301 			if (!mutex_trylock(&vdev->vma_lock)) {
1302 				mmap_read_unlock(mm);
1303 				mmput(mm);
1304 				return 0;
1305 			}
1306 		} else {
1307 			mutex_lock(&vdev->vma_lock);
1308 		}
1309 		list_for_each_entry_safe(mmap_vma, tmp,
1310 					 &vdev->vma_list, vma_next) {
1311 			struct vm_area_struct *vma = mmap_vma->vma;
1312 
1313 			if (vma->vm_mm != mm)
1314 				continue;
1315 
1316 			list_del(&mmap_vma->vma_next);
1317 			kfree(mmap_vma);
1318 
1319 			zap_vma_ptes(vma, vma->vm_start,
1320 				     vma->vm_end - vma->vm_start);
1321 		}
1322 		mutex_unlock(&vdev->vma_lock);
1323 		mmap_read_unlock(mm);
1324 		mmput(mm);
1325 	}
1326 }
1327 
1328 void vfio_pci_zap_and_down_write_memory_lock(struct vfio_pci_core_device *vdev)
1329 {
1330 	vfio_pci_zap_and_vma_lock(vdev, false);
1331 	down_write(&vdev->memory_lock);
1332 	mutex_unlock(&vdev->vma_lock);
1333 }
1334 
1335 u16 vfio_pci_memory_lock_and_enable(struct vfio_pci_core_device *vdev)
1336 {
1337 	u16 cmd;
1338 
1339 	down_write(&vdev->memory_lock);
1340 	pci_read_config_word(vdev->pdev, PCI_COMMAND, &cmd);
1341 	if (!(cmd & PCI_COMMAND_MEMORY))
1342 		pci_write_config_word(vdev->pdev, PCI_COMMAND,
1343 				      cmd | PCI_COMMAND_MEMORY);
1344 
1345 	return cmd;
1346 }
1347 
1348 void vfio_pci_memory_unlock_and_restore(struct vfio_pci_core_device *vdev, u16 cmd)
1349 {
1350 	pci_write_config_word(vdev->pdev, PCI_COMMAND, cmd);
1351 	up_write(&vdev->memory_lock);
1352 }
1353 
1354 /* Caller holds vma_lock */
1355 static int __vfio_pci_add_vma(struct vfio_pci_core_device *vdev,
1356 			      struct vm_area_struct *vma)
1357 {
1358 	struct vfio_pci_mmap_vma *mmap_vma;
1359 
1360 	mmap_vma = kmalloc(sizeof(*mmap_vma), GFP_KERNEL);
1361 	if (!mmap_vma)
1362 		return -ENOMEM;
1363 
1364 	mmap_vma->vma = vma;
1365 	list_add(&mmap_vma->vma_next, &vdev->vma_list);
1366 
1367 	return 0;
1368 }
1369 
1370 /*
1371  * Zap mmaps on open so that we can fault them in on access and therefore
1372  * our vma_list only tracks mappings accessed since last zap.
1373  */
1374 static void vfio_pci_mmap_open(struct vm_area_struct *vma)
1375 {
1376 	zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1377 }
1378 
1379 static void vfio_pci_mmap_close(struct vm_area_struct *vma)
1380 {
1381 	struct vfio_pci_core_device *vdev = vma->vm_private_data;
1382 	struct vfio_pci_mmap_vma *mmap_vma;
1383 
1384 	mutex_lock(&vdev->vma_lock);
1385 	list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1386 		if (mmap_vma->vma == vma) {
1387 			list_del(&mmap_vma->vma_next);
1388 			kfree(mmap_vma);
1389 			break;
1390 		}
1391 	}
1392 	mutex_unlock(&vdev->vma_lock);
1393 }
1394 
1395 static vm_fault_t vfio_pci_mmap_fault(struct vm_fault *vmf)
1396 {
1397 	struct vm_area_struct *vma = vmf->vma;
1398 	struct vfio_pci_core_device *vdev = vma->vm_private_data;
1399 	struct vfio_pci_mmap_vma *mmap_vma;
1400 	vm_fault_t ret = VM_FAULT_NOPAGE;
1401 
1402 	mutex_lock(&vdev->vma_lock);
1403 	down_read(&vdev->memory_lock);
1404 
1405 	if (!__vfio_pci_memory_enabled(vdev)) {
1406 		ret = VM_FAULT_SIGBUS;
1407 		goto up_out;
1408 	}
1409 
1410 	/*
1411 	 * We populate the whole vma on fault, so we need to test whether
1412 	 * the vma has already been mapped, such as for concurrent faults
1413 	 * to the same vma.  io_remap_pfn_range() will trigger a BUG_ON if
1414 	 * we ask it to fill the same range again.
1415 	 */
1416 	list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1417 		if (mmap_vma->vma == vma)
1418 			goto up_out;
1419 	}
1420 
1421 	if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1422 			       vma->vm_end - vma->vm_start,
1423 			       vma->vm_page_prot)) {
1424 		ret = VM_FAULT_SIGBUS;
1425 		zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1426 		goto up_out;
1427 	}
1428 
1429 	if (__vfio_pci_add_vma(vdev, vma)) {
1430 		ret = VM_FAULT_OOM;
1431 		zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1432 	}
1433 
1434 up_out:
1435 	up_read(&vdev->memory_lock);
1436 	mutex_unlock(&vdev->vma_lock);
1437 	return ret;
1438 }
1439 
1440 static const struct vm_operations_struct vfio_pci_mmap_ops = {
1441 	.open = vfio_pci_mmap_open,
1442 	.close = vfio_pci_mmap_close,
1443 	.fault = vfio_pci_mmap_fault,
1444 };
1445 
1446 int vfio_pci_core_mmap(struct vfio_device *core_vdev, struct vm_area_struct *vma)
1447 {
1448 	struct vfio_pci_core_device *vdev =
1449 		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1450 	struct pci_dev *pdev = vdev->pdev;
1451 	unsigned int index;
1452 	u64 phys_len, req_len, pgoff, req_start;
1453 	int ret;
1454 
1455 	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1456 
1457 	if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1458 		return -EINVAL;
1459 	if (vma->vm_end < vma->vm_start)
1460 		return -EINVAL;
1461 	if ((vma->vm_flags & VM_SHARED) == 0)
1462 		return -EINVAL;
1463 	if (index >= VFIO_PCI_NUM_REGIONS) {
1464 		int regnum = index - VFIO_PCI_NUM_REGIONS;
1465 		struct vfio_pci_region *region = vdev->region + regnum;
1466 
1467 		if (region->ops && region->ops->mmap &&
1468 		    (region->flags & VFIO_REGION_INFO_FLAG_MMAP))
1469 			return region->ops->mmap(vdev, region, vma);
1470 		return -EINVAL;
1471 	}
1472 	if (index >= VFIO_PCI_ROM_REGION_INDEX)
1473 		return -EINVAL;
1474 	if (!vdev->bar_mmap_supported[index])
1475 		return -EINVAL;
1476 
1477 	phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1478 	req_len = vma->vm_end - vma->vm_start;
1479 	pgoff = vma->vm_pgoff &
1480 		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1481 	req_start = pgoff << PAGE_SHIFT;
1482 
1483 	if (req_start + req_len > phys_len)
1484 		return -EINVAL;
1485 
1486 	/*
1487 	 * Even though we don't make use of the barmap for the mmap,
1488 	 * we need to request the region and the barmap tracks that.
1489 	 */
1490 	if (!vdev->barmap[index]) {
1491 		ret = pci_request_selected_regions(pdev,
1492 						   1 << index, "vfio-pci");
1493 		if (ret)
1494 			return ret;
1495 
1496 		vdev->barmap[index] = pci_iomap(pdev, index, 0);
1497 		if (!vdev->barmap[index]) {
1498 			pci_release_selected_regions(pdev, 1 << index);
1499 			return -ENOMEM;
1500 		}
1501 	}
1502 
1503 	vma->vm_private_data = vdev;
1504 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1505 	vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1506 
1507 	/*
1508 	 * See remap_pfn_range(), called from vfio_pci_fault() but we can't
1509 	 * change vm_flags within the fault handler.  Set them now.
1510 	 */
1511 	vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1512 	vma->vm_ops = &vfio_pci_mmap_ops;
1513 
1514 	return 0;
1515 }
1516 EXPORT_SYMBOL_GPL(vfio_pci_core_mmap);
1517 
1518 void vfio_pci_core_request(struct vfio_device *core_vdev, unsigned int count)
1519 {
1520 	struct vfio_pci_core_device *vdev =
1521 		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1522 	struct pci_dev *pdev = vdev->pdev;
1523 
1524 	mutex_lock(&vdev->igate);
1525 
1526 	if (vdev->req_trigger) {
1527 		if (!(count % 10))
1528 			pci_notice_ratelimited(pdev,
1529 				"Relaying device request to user (#%u)\n",
1530 				count);
1531 		eventfd_signal(vdev->req_trigger, 1);
1532 	} else if (count == 0) {
1533 		pci_warn(pdev,
1534 			"No device request channel registered, blocked until released by user\n");
1535 	}
1536 
1537 	mutex_unlock(&vdev->igate);
1538 }
1539 EXPORT_SYMBOL_GPL(vfio_pci_core_request);
1540 
1541 static int vfio_pci_validate_vf_token(struct vfio_pci_core_device *vdev,
1542 				      bool vf_token, uuid_t *uuid)
1543 {
1544 	/*
1545 	 * There's always some degree of trust or collaboration between SR-IOV
1546 	 * PF and VFs, even if just that the PF hosts the SR-IOV capability and
1547 	 * can disrupt VFs with a reset, but often the PF has more explicit
1548 	 * access to deny service to the VF or access data passed through the
1549 	 * VF.  We therefore require an opt-in via a shared VF token (UUID) to
1550 	 * represent this trust.  This both prevents that a VF driver might
1551 	 * assume the PF driver is a trusted, in-kernel driver, and also that
1552 	 * a PF driver might be replaced with a rogue driver, unknown to in-use
1553 	 * VF drivers.
1554 	 *
1555 	 * Therefore when presented with a VF, if the PF is a vfio device and
1556 	 * it is bound to the vfio-pci driver, the user needs to provide a VF
1557 	 * token to access the device, in the form of appending a vf_token to
1558 	 * the device name, for example:
1559 	 *
1560 	 * "0000:04:10.0 vf_token=bd8d9d2b-5a5f-4f5a-a211-f591514ba1f3"
1561 	 *
1562 	 * When presented with a PF which has VFs in use, the user must also
1563 	 * provide the current VF token to prove collaboration with existing
1564 	 * VF users.  If VFs are not in use, the VF token provided for the PF
1565 	 * device will act to set the VF token.
1566 	 *
1567 	 * If the VF token is provided but unused, an error is generated.
1568 	 */
1569 	if (!vdev->pdev->is_virtfn && !vdev->vf_token && !vf_token)
1570 		return 0; /* No VF token provided or required */
1571 
1572 	if (vdev->pdev->is_virtfn) {
1573 		struct vfio_pci_core_device *pf_vdev = get_pf_vdev(vdev);
1574 		bool match;
1575 
1576 		if (!pf_vdev) {
1577 			if (!vf_token)
1578 				return 0; /* PF is not vfio-pci, no VF token */
1579 
1580 			pci_info_ratelimited(vdev->pdev,
1581 				"VF token incorrectly provided, PF not bound to vfio-pci\n");
1582 			return -EINVAL;
1583 		}
1584 
1585 		if (!vf_token) {
1586 			vfio_device_put(&pf_vdev->vdev);
1587 			pci_info_ratelimited(vdev->pdev,
1588 				"VF token required to access device\n");
1589 			return -EACCES;
1590 		}
1591 
1592 		mutex_lock(&pf_vdev->vf_token->lock);
1593 		match = uuid_equal(uuid, &pf_vdev->vf_token->uuid);
1594 		mutex_unlock(&pf_vdev->vf_token->lock);
1595 
1596 		vfio_device_put(&pf_vdev->vdev);
1597 
1598 		if (!match) {
1599 			pci_info_ratelimited(vdev->pdev,
1600 				"Incorrect VF token provided for device\n");
1601 			return -EACCES;
1602 		}
1603 	} else if (vdev->vf_token) {
1604 		mutex_lock(&vdev->vf_token->lock);
1605 		if (vdev->vf_token->users) {
1606 			if (!vf_token) {
1607 				mutex_unlock(&vdev->vf_token->lock);
1608 				pci_info_ratelimited(vdev->pdev,
1609 					"VF token required to access device\n");
1610 				return -EACCES;
1611 			}
1612 
1613 			if (!uuid_equal(uuid, &vdev->vf_token->uuid)) {
1614 				mutex_unlock(&vdev->vf_token->lock);
1615 				pci_info_ratelimited(vdev->pdev,
1616 					"Incorrect VF token provided for device\n");
1617 				return -EACCES;
1618 			}
1619 		} else if (vf_token) {
1620 			uuid_copy(&vdev->vf_token->uuid, uuid);
1621 		}
1622 
1623 		mutex_unlock(&vdev->vf_token->lock);
1624 	} else if (vf_token) {
1625 		pci_info_ratelimited(vdev->pdev,
1626 			"VF token incorrectly provided, not a PF or VF\n");
1627 		return -EINVAL;
1628 	}
1629 
1630 	return 0;
1631 }
1632 
1633 #define VF_TOKEN_ARG "vf_token="
1634 
1635 int vfio_pci_core_match(struct vfio_device *core_vdev, char *buf)
1636 {
1637 	struct vfio_pci_core_device *vdev =
1638 		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1639 	bool vf_token = false;
1640 	uuid_t uuid;
1641 	int ret;
1642 
1643 	if (strncmp(pci_name(vdev->pdev), buf, strlen(pci_name(vdev->pdev))))
1644 		return 0; /* No match */
1645 
1646 	if (strlen(buf) > strlen(pci_name(vdev->pdev))) {
1647 		buf += strlen(pci_name(vdev->pdev));
1648 
1649 		if (*buf != ' ')
1650 			return 0; /* No match: non-whitespace after name */
1651 
1652 		while (*buf) {
1653 			if (*buf == ' ') {
1654 				buf++;
1655 				continue;
1656 			}
1657 
1658 			if (!vf_token && !strncmp(buf, VF_TOKEN_ARG,
1659 						  strlen(VF_TOKEN_ARG))) {
1660 				buf += strlen(VF_TOKEN_ARG);
1661 
1662 				if (strlen(buf) < UUID_STRING_LEN)
1663 					return -EINVAL;
1664 
1665 				ret = uuid_parse(buf, &uuid);
1666 				if (ret)
1667 					return ret;
1668 
1669 				vf_token = true;
1670 				buf += UUID_STRING_LEN;
1671 			} else {
1672 				/* Unknown/duplicate option */
1673 				return -EINVAL;
1674 			}
1675 		}
1676 	}
1677 
1678 	ret = vfio_pci_validate_vf_token(vdev, vf_token, &uuid);
1679 	if (ret)
1680 		return ret;
1681 
1682 	return 1; /* Match */
1683 }
1684 EXPORT_SYMBOL_GPL(vfio_pci_core_match);
1685 
1686 static int vfio_pci_bus_notifier(struct notifier_block *nb,
1687 				 unsigned long action, void *data)
1688 {
1689 	struct vfio_pci_core_device *vdev = container_of(nb,
1690 						    struct vfio_pci_core_device, nb);
1691 	struct device *dev = data;
1692 	struct pci_dev *pdev = to_pci_dev(dev);
1693 	struct pci_dev *physfn = pci_physfn(pdev);
1694 
1695 	if (action == BUS_NOTIFY_ADD_DEVICE &&
1696 	    pdev->is_virtfn && physfn == vdev->pdev) {
1697 		pci_info(vdev->pdev, "Captured SR-IOV VF %s driver_override\n",
1698 			 pci_name(pdev));
1699 		pdev->driver_override = kasprintf(GFP_KERNEL, "%s",
1700 						  vdev->vdev.ops->name);
1701 	} else if (action == BUS_NOTIFY_BOUND_DRIVER &&
1702 		   pdev->is_virtfn && physfn == vdev->pdev) {
1703 		struct pci_driver *drv = pci_dev_driver(pdev);
1704 
1705 		if (drv && drv != pci_dev_driver(vdev->pdev))
1706 			pci_warn(vdev->pdev,
1707 				 "VF %s bound to driver %s while PF bound to driver %s\n",
1708 				 pci_name(pdev), drv->name,
1709 				 pci_dev_driver(vdev->pdev)->name);
1710 	}
1711 
1712 	return 0;
1713 }
1714 
1715 static int vfio_pci_vf_init(struct vfio_pci_core_device *vdev)
1716 {
1717 	struct pci_dev *pdev = vdev->pdev;
1718 	int ret;
1719 
1720 	if (!pdev->is_physfn)
1721 		return 0;
1722 
1723 	vdev->vf_token = kzalloc(sizeof(*vdev->vf_token), GFP_KERNEL);
1724 	if (!vdev->vf_token)
1725 		return -ENOMEM;
1726 
1727 	mutex_init(&vdev->vf_token->lock);
1728 	uuid_gen(&vdev->vf_token->uuid);
1729 
1730 	vdev->nb.notifier_call = vfio_pci_bus_notifier;
1731 	ret = bus_register_notifier(&pci_bus_type, &vdev->nb);
1732 	if (ret) {
1733 		kfree(vdev->vf_token);
1734 		return ret;
1735 	}
1736 	return 0;
1737 }
1738 
1739 static void vfio_pci_vf_uninit(struct vfio_pci_core_device *vdev)
1740 {
1741 	if (!vdev->vf_token)
1742 		return;
1743 
1744 	bus_unregister_notifier(&pci_bus_type, &vdev->nb);
1745 	WARN_ON(vdev->vf_token->users);
1746 	mutex_destroy(&vdev->vf_token->lock);
1747 	kfree(vdev->vf_token);
1748 }
1749 
1750 static int vfio_pci_vga_init(struct vfio_pci_core_device *vdev)
1751 {
1752 	struct pci_dev *pdev = vdev->pdev;
1753 	int ret;
1754 
1755 	if (!vfio_pci_is_vga(pdev))
1756 		return 0;
1757 
1758 	ret = vga_client_register(pdev, vfio_pci_set_decode);
1759 	if (ret)
1760 		return ret;
1761 	vga_set_legacy_decoding(pdev, vfio_pci_set_decode(pdev, false));
1762 	return 0;
1763 }
1764 
1765 static void vfio_pci_vga_uninit(struct vfio_pci_core_device *vdev)
1766 {
1767 	struct pci_dev *pdev = vdev->pdev;
1768 
1769 	if (!vfio_pci_is_vga(pdev))
1770 		return;
1771 	vga_client_unregister(pdev);
1772 	vga_set_legacy_decoding(pdev, VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
1773 					      VGA_RSRC_LEGACY_IO |
1774 					      VGA_RSRC_LEGACY_MEM);
1775 }
1776 
1777 void vfio_pci_core_init_device(struct vfio_pci_core_device *vdev,
1778 			       struct pci_dev *pdev,
1779 			       const struct vfio_device_ops *vfio_pci_ops)
1780 {
1781 	vfio_init_group_dev(&vdev->vdev, &pdev->dev, vfio_pci_ops);
1782 	vdev->pdev = pdev;
1783 	vdev->irq_type = VFIO_PCI_NUM_IRQS;
1784 	mutex_init(&vdev->igate);
1785 	spin_lock_init(&vdev->irqlock);
1786 	mutex_init(&vdev->ioeventfds_lock);
1787 	INIT_LIST_HEAD(&vdev->dummy_resources_list);
1788 	INIT_LIST_HEAD(&vdev->ioeventfds_list);
1789 	mutex_init(&vdev->vma_lock);
1790 	INIT_LIST_HEAD(&vdev->vma_list);
1791 	init_rwsem(&vdev->memory_lock);
1792 }
1793 EXPORT_SYMBOL_GPL(vfio_pci_core_init_device);
1794 
1795 void vfio_pci_core_uninit_device(struct vfio_pci_core_device *vdev)
1796 {
1797 	mutex_destroy(&vdev->igate);
1798 	mutex_destroy(&vdev->ioeventfds_lock);
1799 	mutex_destroy(&vdev->vma_lock);
1800 	vfio_uninit_group_dev(&vdev->vdev);
1801 	kfree(vdev->region);
1802 	kfree(vdev->pm_save);
1803 }
1804 EXPORT_SYMBOL_GPL(vfio_pci_core_uninit_device);
1805 
1806 int vfio_pci_core_register_device(struct vfio_pci_core_device *vdev)
1807 {
1808 	struct pci_dev *pdev = vdev->pdev;
1809 	struct iommu_group *group;
1810 	int ret;
1811 
1812 	if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1813 		return -EINVAL;
1814 
1815 	/*
1816 	 * Prevent binding to PFs with VFs enabled, the VFs might be in use
1817 	 * by the host or other users.  We cannot capture the VFs if they
1818 	 * already exist, nor can we track VF users.  Disabling SR-IOV here
1819 	 * would initiate removing the VFs, which would unbind the driver,
1820 	 * which is prone to blocking if that VF is also in use by vfio-pci.
1821 	 * Just reject these PFs and let the user sort it out.
1822 	 */
1823 	if (pci_num_vf(pdev)) {
1824 		pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
1825 		return -EBUSY;
1826 	}
1827 
1828 	group = vfio_iommu_group_get(&pdev->dev);
1829 	if (!group)
1830 		return -EINVAL;
1831 
1832 	if (pci_is_root_bus(pdev->bus)) {
1833 		ret = vfio_assign_device_set(&vdev->vdev, vdev);
1834 	} else if (!pci_probe_reset_slot(pdev->slot)) {
1835 		ret = vfio_assign_device_set(&vdev->vdev, pdev->slot);
1836 	} else {
1837 		/*
1838 		 * If there is no slot reset support for this device, the whole
1839 		 * bus needs to be grouped together to support bus-wide resets.
1840 		 */
1841 		ret = vfio_assign_device_set(&vdev->vdev, pdev->bus);
1842 	}
1843 
1844 	if (ret)
1845 		goto out_group_put;
1846 	ret = vfio_pci_vf_init(vdev);
1847 	if (ret)
1848 		goto out_group_put;
1849 	ret = vfio_pci_vga_init(vdev);
1850 	if (ret)
1851 		goto out_vf;
1852 
1853 	vfio_pci_probe_power_state(vdev);
1854 
1855 	if (!disable_idle_d3) {
1856 		/*
1857 		 * pci-core sets the device power state to an unknown value at
1858 		 * bootup and after being removed from a driver.  The only
1859 		 * transition it allows from this unknown state is to D0, which
1860 		 * typically happens when a driver calls pci_enable_device().
1861 		 * We're not ready to enable the device yet, but we do want to
1862 		 * be able to get to D3.  Therefore first do a D0 transition
1863 		 * before going to D3.
1864 		 */
1865 		vfio_pci_set_power_state(vdev, PCI_D0);
1866 		vfio_pci_set_power_state(vdev, PCI_D3hot);
1867 	}
1868 
1869 	ret = vfio_register_group_dev(&vdev->vdev);
1870 	if (ret)
1871 		goto out_power;
1872 	return 0;
1873 
1874 out_power:
1875 	if (!disable_idle_d3)
1876 		vfio_pci_set_power_state(vdev, PCI_D0);
1877 out_vf:
1878 	vfio_pci_vf_uninit(vdev);
1879 out_group_put:
1880 	vfio_iommu_group_put(group, &pdev->dev);
1881 	return ret;
1882 }
1883 EXPORT_SYMBOL_GPL(vfio_pci_core_register_device);
1884 
1885 void vfio_pci_core_unregister_device(struct vfio_pci_core_device *vdev)
1886 {
1887 	struct pci_dev *pdev = vdev->pdev;
1888 
1889 	pci_disable_sriov(pdev);
1890 
1891 	vfio_unregister_group_dev(&vdev->vdev);
1892 
1893 	vfio_pci_vf_uninit(vdev);
1894 	vfio_pci_vga_uninit(vdev);
1895 
1896 	vfio_iommu_group_put(pdev->dev.iommu_group, &pdev->dev);
1897 
1898 	if (!disable_idle_d3)
1899 		vfio_pci_set_power_state(vdev, PCI_D0);
1900 }
1901 EXPORT_SYMBOL_GPL(vfio_pci_core_unregister_device);
1902 
1903 static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
1904 						  pci_channel_state_t state)
1905 {
1906 	struct vfio_pci_core_device *vdev;
1907 	struct vfio_device *device;
1908 
1909 	device = vfio_device_get_from_dev(&pdev->dev);
1910 	if (device == NULL)
1911 		return PCI_ERS_RESULT_DISCONNECT;
1912 
1913 	vdev = container_of(device, struct vfio_pci_core_device, vdev);
1914 
1915 	mutex_lock(&vdev->igate);
1916 
1917 	if (vdev->err_trigger)
1918 		eventfd_signal(vdev->err_trigger, 1);
1919 
1920 	mutex_unlock(&vdev->igate);
1921 
1922 	vfio_device_put(device);
1923 
1924 	return PCI_ERS_RESULT_CAN_RECOVER;
1925 }
1926 
1927 int vfio_pci_core_sriov_configure(struct pci_dev *pdev, int nr_virtfn)
1928 {
1929 	struct vfio_device *device;
1930 	int ret = 0;
1931 
1932 	device = vfio_device_get_from_dev(&pdev->dev);
1933 	if (!device)
1934 		return -ENODEV;
1935 
1936 	if (nr_virtfn == 0)
1937 		pci_disable_sriov(pdev);
1938 	else
1939 		ret = pci_enable_sriov(pdev, nr_virtfn);
1940 
1941 	vfio_device_put(device);
1942 
1943 	return ret < 0 ? ret : nr_virtfn;
1944 }
1945 EXPORT_SYMBOL_GPL(vfio_pci_core_sriov_configure);
1946 
1947 const struct pci_error_handlers vfio_pci_core_err_handlers = {
1948 	.error_detected = vfio_pci_aer_err_detected,
1949 };
1950 EXPORT_SYMBOL_GPL(vfio_pci_core_err_handlers);
1951 
1952 static bool vfio_dev_in_groups(struct vfio_pci_core_device *vdev,
1953 			       struct vfio_pci_group_info *groups)
1954 {
1955 	unsigned int i;
1956 
1957 	for (i = 0; i < groups->count; i++)
1958 		if (groups->groups[i] == vdev->vdev.group)
1959 			return true;
1960 	return false;
1961 }
1962 
1963 static int vfio_pci_is_device_in_set(struct pci_dev *pdev, void *data)
1964 {
1965 	struct vfio_device_set *dev_set = data;
1966 	struct vfio_device *cur;
1967 
1968 	list_for_each_entry(cur, &dev_set->device_list, dev_set_list)
1969 		if (cur->dev == &pdev->dev)
1970 			return 0;
1971 	return -EBUSY;
1972 }
1973 
1974 /*
1975  * vfio-core considers a group to be viable and will create a vfio_device even
1976  * if some devices are bound to drivers like pci-stub or pcieport. Here we
1977  * require all PCI devices to be inside our dev_set since that ensures they stay
1978  * put and that every driver controlling the device can co-ordinate with the
1979  * device reset.
1980  *
1981  * Returns the pci_dev to pass to pci_reset_bus() if every PCI device to be
1982  * reset is inside the dev_set, and pci_reset_bus() can succeed. NULL otherwise.
1983  */
1984 static struct pci_dev *
1985 vfio_pci_dev_set_resettable(struct vfio_device_set *dev_set)
1986 {
1987 	struct pci_dev *pdev;
1988 
1989 	lockdep_assert_held(&dev_set->lock);
1990 
1991 	/*
1992 	 * By definition all PCI devices in the dev_set share the same PCI
1993 	 * reset, so any pci_dev will have the same outcomes for
1994 	 * pci_probe_reset_*() and pci_reset_bus().
1995 	 */
1996 	pdev = list_first_entry(&dev_set->device_list,
1997 				struct vfio_pci_core_device,
1998 				vdev.dev_set_list)->pdev;
1999 
2000 	/* pci_reset_bus() is supported */
2001 	if (pci_probe_reset_slot(pdev->slot) && pci_probe_reset_bus(pdev->bus))
2002 		return NULL;
2003 
2004 	if (vfio_pci_for_each_slot_or_bus(pdev, vfio_pci_is_device_in_set,
2005 					  dev_set,
2006 					  !pci_probe_reset_slot(pdev->slot)))
2007 		return NULL;
2008 	return pdev;
2009 }
2010 
2011 /*
2012  * We need to get memory_lock for each device, but devices can share mmap_lock,
2013  * therefore we need to zap and hold the vma_lock for each device, and only then
2014  * get each memory_lock.
2015  */
2016 static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set,
2017 				      struct vfio_pci_group_info *groups)
2018 {
2019 	struct vfio_pci_core_device *cur_mem;
2020 	struct vfio_pci_core_device *cur_vma;
2021 	struct vfio_pci_core_device *cur;
2022 	struct pci_dev *pdev;
2023 	bool is_mem = true;
2024 	int ret;
2025 
2026 	mutex_lock(&dev_set->lock);
2027 	cur_mem = list_first_entry(&dev_set->device_list,
2028 				   struct vfio_pci_core_device,
2029 				   vdev.dev_set_list);
2030 
2031 	pdev = vfio_pci_dev_set_resettable(dev_set);
2032 	if (!pdev) {
2033 		ret = -EINVAL;
2034 		goto err_unlock;
2035 	}
2036 
2037 	list_for_each_entry(cur_vma, &dev_set->device_list, vdev.dev_set_list) {
2038 		/*
2039 		 * Test whether all the affected devices are contained by the
2040 		 * set of groups provided by the user.
2041 		 */
2042 		if (!vfio_dev_in_groups(cur_vma, groups)) {
2043 			ret = -EINVAL;
2044 			goto err_undo;
2045 		}
2046 
2047 		/*
2048 		 * Locking multiple devices is prone to deadlock, runaway and
2049 		 * unwind if we hit contention.
2050 		 */
2051 		if (!vfio_pci_zap_and_vma_lock(cur_vma, true)) {
2052 			ret = -EBUSY;
2053 			goto err_undo;
2054 		}
2055 	}
2056 	cur_vma = NULL;
2057 
2058 	list_for_each_entry(cur_mem, &dev_set->device_list, vdev.dev_set_list) {
2059 		if (!down_write_trylock(&cur_mem->memory_lock)) {
2060 			ret = -EBUSY;
2061 			goto err_undo;
2062 		}
2063 		mutex_unlock(&cur_mem->vma_lock);
2064 	}
2065 	cur_mem = NULL;
2066 
2067 	ret = pci_reset_bus(pdev);
2068 
2069 err_undo:
2070 	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2071 		if (cur == cur_mem)
2072 			is_mem = false;
2073 		if (cur == cur_vma)
2074 			break;
2075 		if (is_mem)
2076 			up_write(&cur->memory_lock);
2077 		else
2078 			mutex_unlock(&cur->vma_lock);
2079 	}
2080 err_unlock:
2081 	mutex_unlock(&dev_set->lock);
2082 	return ret;
2083 }
2084 
2085 static bool vfio_pci_dev_set_needs_reset(struct vfio_device_set *dev_set)
2086 {
2087 	struct vfio_pci_core_device *cur;
2088 	bool needs_reset = false;
2089 
2090 	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2091 		/* No VFIO device in the set can have an open device FD */
2092 		if (cur->vdev.open_count)
2093 			return false;
2094 		needs_reset |= cur->needs_reset;
2095 	}
2096 	return needs_reset;
2097 }
2098 
2099 /*
2100  * If a bus or slot reset is available for the provided dev_set and:
2101  *  - All of the devices affected by that bus or slot reset are unused
2102  *  - At least one of the affected devices is marked dirty via
2103  *    needs_reset (such as by lack of FLR support)
2104  * Then attempt to perform that bus or slot reset.
2105  * Returns true if the dev_set was reset.
2106  */
2107 static bool vfio_pci_dev_set_try_reset(struct vfio_device_set *dev_set)
2108 {
2109 	struct vfio_pci_core_device *cur;
2110 	struct pci_dev *pdev;
2111 	int ret;
2112 
2113 	if (!vfio_pci_dev_set_needs_reset(dev_set))
2114 		return false;
2115 
2116 	pdev = vfio_pci_dev_set_resettable(dev_set);
2117 	if (!pdev)
2118 		return false;
2119 
2120 	ret = pci_reset_bus(pdev);
2121 	if (ret)
2122 		return false;
2123 
2124 	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2125 		cur->needs_reset = false;
2126 		if (!disable_idle_d3)
2127 			vfio_pci_set_power_state(cur, PCI_D3hot);
2128 	}
2129 	return true;
2130 }
2131 
2132 void vfio_pci_core_set_params(bool is_nointxmask, bool is_disable_vga,
2133 			      bool is_disable_idle_d3)
2134 {
2135 	nointxmask = is_nointxmask;
2136 	disable_vga = is_disable_vga;
2137 	disable_idle_d3 = is_disable_idle_d3;
2138 }
2139 EXPORT_SYMBOL_GPL(vfio_pci_core_set_params);
2140 
2141 static void vfio_pci_core_cleanup(void)
2142 {
2143 	vfio_pci_uninit_perm_bits();
2144 }
2145 
2146 static int __init vfio_pci_core_init(void)
2147 {
2148 	/* Allocate shared config space permission data used by all devices */
2149 	return vfio_pci_init_perm_bits();
2150 }
2151 
2152 module_init(vfio_pci_core_init);
2153 module_exit(vfio_pci_core_cleanup);
2154 
2155 MODULE_LICENSE("GPL v2");
2156 MODULE_AUTHOR(DRIVER_AUTHOR);
2157 MODULE_DESCRIPTION(DRIVER_DESC);
2158