xref: /openbmc/linux/drivers/vfio/pci/vfio_pci.c (revision 82df5b73)
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/vfio.h>
27 #include <linux/vgaarb.h>
28 #include <linux/nospec.h>
29 #include <linux/sched/mm.h>
30 
31 #include "vfio_pci_private.h"
32 
33 #define DRIVER_VERSION  "0.2"
34 #define DRIVER_AUTHOR   "Alex Williamson <alex.williamson@redhat.com>"
35 #define DRIVER_DESC     "VFIO PCI - User Level meta-driver"
36 
37 static char ids[1024] __initdata;
38 module_param_string(ids, ids, sizeof(ids), 0);
39 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");
40 
41 static bool nointxmask;
42 module_param_named(nointxmask, nointxmask, bool, S_IRUGO | S_IWUSR);
43 MODULE_PARM_DESC(nointxmask,
44 		  "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.");
45 
46 #ifdef CONFIG_VFIO_PCI_VGA
47 static bool disable_vga;
48 module_param(disable_vga, bool, S_IRUGO);
49 MODULE_PARM_DESC(disable_vga, "Disable VGA resource access through vfio-pci");
50 #endif
51 
52 static bool disable_idle_d3;
53 module_param(disable_idle_d3, bool, S_IRUGO | S_IWUSR);
54 MODULE_PARM_DESC(disable_idle_d3,
55 		 "Disable using the PCI D3 low power state for idle, unused devices");
56 
57 static bool enable_sriov;
58 #ifdef CONFIG_PCI_IOV
59 module_param(enable_sriov, bool, 0644);
60 MODULE_PARM_DESC(enable_sriov, "Enable support for SR-IOV configuration.  Enabling SR-IOV on a PF typically requires support of the userspace PF driver, enabling VFs without such support may result in non-functional VFs or PF.");
61 #endif
62 
63 static inline bool vfio_vga_disabled(void)
64 {
65 #ifdef CONFIG_VFIO_PCI_VGA
66 	return disable_vga;
67 #else
68 	return true;
69 #endif
70 }
71 
72 /*
73  * Our VGA arbiter participation is limited since we don't know anything
74  * about the device itself.  However, if the device is the only VGA device
75  * downstream of a bridge and VFIO VGA support is disabled, then we can
76  * safely return legacy VGA IO and memory as not decoded since the user
77  * has no way to get to it and routing can be disabled externally at the
78  * bridge.
79  */
80 static unsigned int vfio_pci_set_vga_decode(void *opaque, bool single_vga)
81 {
82 	struct vfio_pci_device *vdev = opaque;
83 	struct pci_dev *tmp = NULL, *pdev = vdev->pdev;
84 	unsigned char max_busnr;
85 	unsigned int decodes;
86 
87 	if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
88 		return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
89 		       VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
90 
91 	max_busnr = pci_bus_max_busnr(pdev->bus);
92 	decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
93 
94 	while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
95 		if (tmp == pdev ||
96 		    pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
97 		    pci_is_root_bus(tmp->bus))
98 			continue;
99 
100 		if (tmp->bus->number >= pdev->bus->number &&
101 		    tmp->bus->number <= max_busnr) {
102 			pci_dev_put(tmp);
103 			decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
104 			break;
105 		}
106 	}
107 
108 	return decodes;
109 }
110 
111 static inline bool vfio_pci_is_vga(struct pci_dev *pdev)
112 {
113 	return (pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA;
114 }
115 
116 static void vfio_pci_probe_mmaps(struct vfio_pci_device *vdev)
117 {
118 	struct resource *res;
119 	int i;
120 	struct vfio_pci_dummy_resource *dummy_res;
121 
122 	INIT_LIST_HEAD(&vdev->dummy_resources_list);
123 
124 	for (i = 0; i < PCI_STD_NUM_BARS; i++) {
125 		int bar = i + PCI_STD_RESOURCES;
126 
127 		res = &vdev->pdev->resource[bar];
128 
129 		if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
130 			goto no_mmap;
131 
132 		if (!(res->flags & IORESOURCE_MEM))
133 			goto no_mmap;
134 
135 		/*
136 		 * The PCI core shouldn't set up a resource with a
137 		 * type but zero size. But there may be bugs that
138 		 * cause us to do that.
139 		 */
140 		if (!resource_size(res))
141 			goto no_mmap;
142 
143 		if (resource_size(res) >= PAGE_SIZE) {
144 			vdev->bar_mmap_supported[bar] = true;
145 			continue;
146 		}
147 
148 		if (!(res->start & ~PAGE_MASK)) {
149 			/*
150 			 * Add a dummy resource to reserve the remainder
151 			 * of the exclusive page in case that hot-add
152 			 * device's bar is assigned into it.
153 			 */
154 			dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
155 			if (dummy_res == NULL)
156 				goto no_mmap;
157 
158 			dummy_res->resource.name = "vfio sub-page reserved";
159 			dummy_res->resource.start = res->end + 1;
160 			dummy_res->resource.end = res->start + PAGE_SIZE - 1;
161 			dummy_res->resource.flags = res->flags;
162 			if (request_resource(res->parent,
163 						&dummy_res->resource)) {
164 				kfree(dummy_res);
165 				goto no_mmap;
166 			}
167 			dummy_res->index = bar;
168 			list_add(&dummy_res->res_next,
169 					&vdev->dummy_resources_list);
170 			vdev->bar_mmap_supported[bar] = true;
171 			continue;
172 		}
173 		/*
174 		 * Here we don't handle the case when the BAR is not page
175 		 * aligned because we can't expect the BAR will be
176 		 * assigned into the same location in a page in guest
177 		 * when we passthrough the BAR. And it's hard to access
178 		 * this BAR in userspace because we have no way to get
179 		 * the BAR's location in a page.
180 		 */
181 no_mmap:
182 		vdev->bar_mmap_supported[bar] = false;
183 	}
184 }
185 
186 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev);
187 static void vfio_pci_disable(struct vfio_pci_device *vdev);
188 static int vfio_pci_try_zap_and_vma_lock_cb(struct pci_dev *pdev, void *data);
189 
190 /*
191  * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
192  * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
193  * If a device implements the former but not the latter we would typically
194  * expect broken_intx_masking be set and require an exclusive interrupt.
195  * However since we do have control of the device's ability to assert INTx,
196  * we can instead pretend that the device does not implement INTx, virtualizing
197  * the pin register to report zero and maintaining DisINTx set on the host.
198  */
199 static bool vfio_pci_nointx(struct pci_dev *pdev)
200 {
201 	switch (pdev->vendor) {
202 	case PCI_VENDOR_ID_INTEL:
203 		switch (pdev->device) {
204 		/* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
205 		case 0x1572:
206 		case 0x1574:
207 		case 0x1580 ... 0x1581:
208 		case 0x1583 ... 0x158b:
209 		case 0x37d0 ... 0x37d2:
210 			return true;
211 		default:
212 			return false;
213 		}
214 	}
215 
216 	return false;
217 }
218 
219 static void vfio_pci_probe_power_state(struct vfio_pci_device *vdev)
220 {
221 	struct pci_dev *pdev = vdev->pdev;
222 	u16 pmcsr;
223 
224 	if (!pdev->pm_cap)
225 		return;
226 
227 	pci_read_config_word(pdev, pdev->pm_cap + PCI_PM_CTRL, &pmcsr);
228 
229 	vdev->needs_pm_restore = !(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
230 }
231 
232 /*
233  * pci_set_power_state() wrapper handling devices which perform a soft reset on
234  * D3->D0 transition.  Save state prior to D0/1/2->D3, stash it on the vdev,
235  * restore when returned to D0.  Saved separately from pci_saved_state for use
236  * by PM capability emulation and separately from pci_dev internal saved state
237  * to avoid it being overwritten and consumed around other resets.
238  */
239 int vfio_pci_set_power_state(struct vfio_pci_device *vdev, pci_power_t state)
240 {
241 	struct pci_dev *pdev = vdev->pdev;
242 	bool needs_restore = false, needs_save = false;
243 	int ret;
244 
245 	if (vdev->needs_pm_restore) {
246 		if (pdev->current_state < PCI_D3hot && state >= PCI_D3hot) {
247 			pci_save_state(pdev);
248 			needs_save = true;
249 		}
250 
251 		if (pdev->current_state >= PCI_D3hot && state <= PCI_D0)
252 			needs_restore = true;
253 	}
254 
255 	ret = pci_set_power_state(pdev, state);
256 
257 	if (!ret) {
258 		/* D3 might be unsupported via quirk, skip unless in D3 */
259 		if (needs_save && pdev->current_state >= PCI_D3hot) {
260 			vdev->pm_save = pci_store_saved_state(pdev);
261 		} else if (needs_restore) {
262 			pci_load_and_free_saved_state(pdev, &vdev->pm_save);
263 			pci_restore_state(pdev);
264 		}
265 	}
266 
267 	return ret;
268 }
269 
270 static int vfio_pci_enable(struct vfio_pci_device *vdev)
271 {
272 	struct pci_dev *pdev = vdev->pdev;
273 	int ret;
274 	u16 cmd;
275 	u8 msix_pos;
276 
277 	vfio_pci_set_power_state(vdev, PCI_D0);
278 
279 	/* Don't allow our initial saved state to include busmaster */
280 	pci_clear_master(pdev);
281 
282 	ret = pci_enable_device(pdev);
283 	if (ret)
284 		return ret;
285 
286 	/* If reset fails because of the device lock, fail this path entirely */
287 	ret = pci_try_reset_function(pdev);
288 	if (ret == -EAGAIN) {
289 		pci_disable_device(pdev);
290 		return ret;
291 	}
292 
293 	vdev->reset_works = !ret;
294 	pci_save_state(pdev);
295 	vdev->pci_saved_state = pci_store_saved_state(pdev);
296 	if (!vdev->pci_saved_state)
297 		pci_dbg(pdev, "%s: Couldn't store saved state\n", __func__);
298 
299 	if (likely(!nointxmask)) {
300 		if (vfio_pci_nointx(pdev)) {
301 			pci_info(pdev, "Masking broken INTx support\n");
302 			vdev->nointx = true;
303 			pci_intx(pdev, 0);
304 		} else
305 			vdev->pci_2_3 = pci_intx_mask_supported(pdev);
306 	}
307 
308 	pci_read_config_word(pdev, PCI_COMMAND, &cmd);
309 	if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
310 		cmd &= ~PCI_COMMAND_INTX_DISABLE;
311 		pci_write_config_word(pdev, PCI_COMMAND, cmd);
312 	}
313 
314 	ret = vfio_config_init(vdev);
315 	if (ret) {
316 		kfree(vdev->pci_saved_state);
317 		vdev->pci_saved_state = NULL;
318 		pci_disable_device(pdev);
319 		return ret;
320 	}
321 
322 	msix_pos = pdev->msix_cap;
323 	if (msix_pos) {
324 		u16 flags;
325 		u32 table;
326 
327 		pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
328 		pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
329 
330 		vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
331 		vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
332 		vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
333 	} else
334 		vdev->msix_bar = 0xFF;
335 
336 	if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
337 		vdev->has_vga = true;
338 
339 
340 	if (vfio_pci_is_vga(pdev) &&
341 	    pdev->vendor == PCI_VENDOR_ID_INTEL &&
342 	    IS_ENABLED(CONFIG_VFIO_PCI_IGD)) {
343 		ret = vfio_pci_igd_init(vdev);
344 		if (ret) {
345 			pci_warn(pdev, "Failed to setup Intel IGD regions\n");
346 			goto disable_exit;
347 		}
348 	}
349 
350 	if (pdev->vendor == PCI_VENDOR_ID_NVIDIA &&
351 	    IS_ENABLED(CONFIG_VFIO_PCI_NVLINK2)) {
352 		ret = vfio_pci_nvdia_v100_nvlink2_init(vdev);
353 		if (ret && ret != -ENODEV) {
354 			pci_warn(pdev, "Failed to setup NVIDIA NV2 RAM region\n");
355 			goto disable_exit;
356 		}
357 	}
358 
359 	if (pdev->vendor == PCI_VENDOR_ID_IBM &&
360 	    IS_ENABLED(CONFIG_VFIO_PCI_NVLINK2)) {
361 		ret = vfio_pci_ibm_npu2_init(vdev);
362 		if (ret && ret != -ENODEV) {
363 			pci_warn(pdev, "Failed to setup NVIDIA NV2 ATSD region\n");
364 			goto disable_exit;
365 		}
366 	}
367 
368 	vfio_pci_probe_mmaps(vdev);
369 
370 	return 0;
371 
372 disable_exit:
373 	vfio_pci_disable(vdev);
374 	return ret;
375 }
376 
377 static void vfio_pci_disable(struct vfio_pci_device *vdev)
378 {
379 	struct pci_dev *pdev = vdev->pdev;
380 	struct vfio_pci_dummy_resource *dummy_res, *tmp;
381 	struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
382 	int i, bar;
383 
384 	/* Stop the device from further DMA */
385 	pci_clear_master(pdev);
386 
387 	vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
388 				VFIO_IRQ_SET_ACTION_TRIGGER,
389 				vdev->irq_type, 0, 0, NULL);
390 
391 	/* Device closed, don't need mutex here */
392 	list_for_each_entry_safe(ioeventfd, ioeventfd_tmp,
393 				 &vdev->ioeventfds_list, next) {
394 		vfio_virqfd_disable(&ioeventfd->virqfd);
395 		list_del(&ioeventfd->next);
396 		kfree(ioeventfd);
397 	}
398 	vdev->ioeventfds_nr = 0;
399 
400 	vdev->virq_disabled = false;
401 
402 	for (i = 0; i < vdev->num_regions; i++)
403 		vdev->region[i].ops->release(vdev, &vdev->region[i]);
404 
405 	vdev->num_regions = 0;
406 	kfree(vdev->region);
407 	vdev->region = NULL; /* don't krealloc a freed pointer */
408 
409 	vfio_config_free(vdev);
410 
411 	for (i = 0; i < PCI_STD_NUM_BARS; i++) {
412 		bar = i + PCI_STD_RESOURCES;
413 		if (!vdev->barmap[bar])
414 			continue;
415 		pci_iounmap(pdev, vdev->barmap[bar]);
416 		pci_release_selected_regions(pdev, 1 << bar);
417 		vdev->barmap[bar] = NULL;
418 	}
419 
420 	list_for_each_entry_safe(dummy_res, tmp,
421 				 &vdev->dummy_resources_list, res_next) {
422 		list_del(&dummy_res->res_next);
423 		release_resource(&dummy_res->resource);
424 		kfree(dummy_res);
425 	}
426 
427 	vdev->needs_reset = true;
428 
429 	/*
430 	 * If we have saved state, restore it.  If we can reset the device,
431 	 * even better.  Resetting with current state seems better than
432 	 * nothing, but saving and restoring current state without reset
433 	 * is just busy work.
434 	 */
435 	if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
436 		pci_info(pdev, "%s: Couldn't reload saved state\n", __func__);
437 
438 		if (!vdev->reset_works)
439 			goto out;
440 
441 		pci_save_state(pdev);
442 	}
443 
444 	/*
445 	 * Disable INTx and MSI, presumably to avoid spurious interrupts
446 	 * during reset.  Stolen from pci_reset_function()
447 	 */
448 	pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
449 
450 	/*
451 	 * Try to get the locks ourselves to prevent a deadlock. The
452 	 * success of this is dependent on being able to lock the device,
453 	 * which is not always possible.
454 	 * We can not use the "try" reset interface here, which will
455 	 * overwrite the previously restored configuration information.
456 	 */
457 	if (vdev->reset_works && pci_cfg_access_trylock(pdev)) {
458 		if (device_trylock(&pdev->dev)) {
459 			if (!__pci_reset_function_locked(pdev))
460 				vdev->needs_reset = false;
461 			device_unlock(&pdev->dev);
462 		}
463 		pci_cfg_access_unlock(pdev);
464 	}
465 
466 	pci_restore_state(pdev);
467 out:
468 	pci_disable_device(pdev);
469 
470 	vfio_pci_try_bus_reset(vdev);
471 
472 	if (!disable_idle_d3)
473 		vfio_pci_set_power_state(vdev, PCI_D3hot);
474 }
475 
476 static struct pci_driver vfio_pci_driver;
477 
478 static struct vfio_pci_device *get_pf_vdev(struct vfio_pci_device *vdev,
479 					   struct vfio_device **pf_dev)
480 {
481 	struct pci_dev *physfn = pci_physfn(vdev->pdev);
482 
483 	if (!vdev->pdev->is_virtfn)
484 		return NULL;
485 
486 	*pf_dev = vfio_device_get_from_dev(&physfn->dev);
487 	if (!*pf_dev)
488 		return NULL;
489 
490 	if (pci_dev_driver(physfn) != &vfio_pci_driver) {
491 		vfio_device_put(*pf_dev);
492 		return NULL;
493 	}
494 
495 	return vfio_device_data(*pf_dev);
496 }
497 
498 static void vfio_pci_vf_token_user_add(struct vfio_pci_device *vdev, int val)
499 {
500 	struct vfio_device *pf_dev;
501 	struct vfio_pci_device *pf_vdev = get_pf_vdev(vdev, &pf_dev);
502 
503 	if (!pf_vdev)
504 		return;
505 
506 	mutex_lock(&pf_vdev->vf_token->lock);
507 	pf_vdev->vf_token->users += val;
508 	WARN_ON(pf_vdev->vf_token->users < 0);
509 	mutex_unlock(&pf_vdev->vf_token->lock);
510 
511 	vfio_device_put(pf_dev);
512 }
513 
514 static void vfio_pci_release(void *device_data)
515 {
516 	struct vfio_pci_device *vdev = device_data;
517 
518 	mutex_lock(&vdev->reflck->lock);
519 
520 	if (!(--vdev->refcnt)) {
521 		vfio_pci_vf_token_user_add(vdev, -1);
522 		vfio_spapr_pci_eeh_release(vdev->pdev);
523 		vfio_pci_disable(vdev);
524 		if (vdev->err_trigger)
525 			eventfd_ctx_put(vdev->err_trigger);
526 		if (vdev->req_trigger)
527 			eventfd_ctx_put(vdev->req_trigger);
528 	}
529 
530 	mutex_unlock(&vdev->reflck->lock);
531 
532 	module_put(THIS_MODULE);
533 }
534 
535 static int vfio_pci_open(void *device_data)
536 {
537 	struct vfio_pci_device *vdev = device_data;
538 	int ret = 0;
539 
540 	if (!try_module_get(THIS_MODULE))
541 		return -ENODEV;
542 
543 	mutex_lock(&vdev->reflck->lock);
544 
545 	if (!vdev->refcnt) {
546 		ret = vfio_pci_enable(vdev);
547 		if (ret)
548 			goto error;
549 
550 		vfio_spapr_pci_eeh_open(vdev->pdev);
551 		vfio_pci_vf_token_user_add(vdev, 1);
552 	}
553 	vdev->refcnt++;
554 error:
555 	mutex_unlock(&vdev->reflck->lock);
556 	if (ret)
557 		module_put(THIS_MODULE);
558 	return ret;
559 }
560 
561 static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
562 {
563 	if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
564 		u8 pin;
565 
566 		if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
567 		    vdev->nointx || vdev->pdev->is_virtfn)
568 			return 0;
569 
570 		pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
571 
572 		return pin ? 1 : 0;
573 	} else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
574 		u8 pos;
575 		u16 flags;
576 
577 		pos = vdev->pdev->msi_cap;
578 		if (pos) {
579 			pci_read_config_word(vdev->pdev,
580 					     pos + PCI_MSI_FLAGS, &flags);
581 			return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
582 		}
583 	} else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
584 		u8 pos;
585 		u16 flags;
586 
587 		pos = vdev->pdev->msix_cap;
588 		if (pos) {
589 			pci_read_config_word(vdev->pdev,
590 					     pos + PCI_MSIX_FLAGS, &flags);
591 
592 			return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
593 		}
594 	} else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
595 		if (pci_is_pcie(vdev->pdev))
596 			return 1;
597 	} else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
598 		return 1;
599 	}
600 
601 	return 0;
602 }
603 
604 static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
605 {
606 	(*(int *)data)++;
607 	return 0;
608 }
609 
610 struct vfio_pci_fill_info {
611 	int max;
612 	int cur;
613 	struct vfio_pci_dependent_device *devices;
614 };
615 
616 static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
617 {
618 	struct vfio_pci_fill_info *fill = data;
619 	struct iommu_group *iommu_group;
620 
621 	if (fill->cur == fill->max)
622 		return -EAGAIN; /* Something changed, try again */
623 
624 	iommu_group = iommu_group_get(&pdev->dev);
625 	if (!iommu_group)
626 		return -EPERM; /* Cannot reset non-isolated devices */
627 
628 	fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
629 	fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
630 	fill->devices[fill->cur].bus = pdev->bus->number;
631 	fill->devices[fill->cur].devfn = pdev->devfn;
632 	fill->cur++;
633 	iommu_group_put(iommu_group);
634 	return 0;
635 }
636 
637 struct vfio_pci_group_entry {
638 	struct vfio_group *group;
639 	int id;
640 };
641 
642 struct vfio_pci_group_info {
643 	int count;
644 	struct vfio_pci_group_entry *groups;
645 };
646 
647 static int vfio_pci_validate_devs(struct pci_dev *pdev, void *data)
648 {
649 	struct vfio_pci_group_info *info = data;
650 	struct iommu_group *group;
651 	int id, i;
652 
653 	group = iommu_group_get(&pdev->dev);
654 	if (!group)
655 		return -EPERM;
656 
657 	id = iommu_group_id(group);
658 
659 	for (i = 0; i < info->count; i++)
660 		if (info->groups[i].id == id)
661 			break;
662 
663 	iommu_group_put(group);
664 
665 	return (i == info->count) ? -EINVAL : 0;
666 }
667 
668 static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
669 {
670 	for (; pdev; pdev = pdev->bus->self)
671 		if (pdev->bus == slot->bus)
672 			return (pdev->slot == slot);
673 	return false;
674 }
675 
676 struct vfio_pci_walk_info {
677 	int (*fn)(struct pci_dev *, void *data);
678 	void *data;
679 	struct pci_dev *pdev;
680 	bool slot;
681 	int ret;
682 };
683 
684 static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
685 {
686 	struct vfio_pci_walk_info *walk = data;
687 
688 	if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
689 		walk->ret = walk->fn(pdev, walk->data);
690 
691 	return walk->ret;
692 }
693 
694 static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
695 					 int (*fn)(struct pci_dev *,
696 						   void *data), void *data,
697 					 bool slot)
698 {
699 	struct vfio_pci_walk_info walk = {
700 		.fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
701 	};
702 
703 	pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
704 
705 	return walk.ret;
706 }
707 
708 static int msix_mmappable_cap(struct vfio_pci_device *vdev,
709 			      struct vfio_info_cap *caps)
710 {
711 	struct vfio_info_cap_header header = {
712 		.id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
713 		.version = 1
714 	};
715 
716 	return vfio_info_add_capability(caps, &header, sizeof(header));
717 }
718 
719 int vfio_pci_register_dev_region(struct vfio_pci_device *vdev,
720 				 unsigned int type, unsigned int subtype,
721 				 const struct vfio_pci_regops *ops,
722 				 size_t size, u32 flags, void *data)
723 {
724 	struct vfio_pci_region *region;
725 
726 	region = krealloc(vdev->region,
727 			  (vdev->num_regions + 1) * sizeof(*region),
728 			  GFP_KERNEL);
729 	if (!region)
730 		return -ENOMEM;
731 
732 	vdev->region = region;
733 	vdev->region[vdev->num_regions].type = type;
734 	vdev->region[vdev->num_regions].subtype = subtype;
735 	vdev->region[vdev->num_regions].ops = ops;
736 	vdev->region[vdev->num_regions].size = size;
737 	vdev->region[vdev->num_regions].flags = flags;
738 	vdev->region[vdev->num_regions].data = data;
739 
740 	vdev->num_regions++;
741 
742 	return 0;
743 }
744 
745 struct vfio_devices {
746 	struct vfio_device **devices;
747 	int cur_index;
748 	int max_index;
749 };
750 
751 static long vfio_pci_ioctl(void *device_data,
752 			   unsigned int cmd, unsigned long arg)
753 {
754 	struct vfio_pci_device *vdev = device_data;
755 	unsigned long minsz;
756 
757 	if (cmd == VFIO_DEVICE_GET_INFO) {
758 		struct vfio_device_info info;
759 
760 		minsz = offsetofend(struct vfio_device_info, num_irqs);
761 
762 		if (copy_from_user(&info, (void __user *)arg, minsz))
763 			return -EFAULT;
764 
765 		if (info.argsz < minsz)
766 			return -EINVAL;
767 
768 		info.flags = VFIO_DEVICE_FLAGS_PCI;
769 
770 		if (vdev->reset_works)
771 			info.flags |= VFIO_DEVICE_FLAGS_RESET;
772 
773 		info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
774 		info.num_irqs = VFIO_PCI_NUM_IRQS;
775 
776 		return copy_to_user((void __user *)arg, &info, minsz) ?
777 			-EFAULT : 0;
778 
779 	} else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
780 		struct pci_dev *pdev = vdev->pdev;
781 		struct vfio_region_info info;
782 		struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
783 		int i, ret;
784 
785 		minsz = offsetofend(struct vfio_region_info, offset);
786 
787 		if (copy_from_user(&info, (void __user *)arg, minsz))
788 			return -EFAULT;
789 
790 		if (info.argsz < minsz)
791 			return -EINVAL;
792 
793 		switch (info.index) {
794 		case VFIO_PCI_CONFIG_REGION_INDEX:
795 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
796 			info.size = pdev->cfg_size;
797 			info.flags = VFIO_REGION_INFO_FLAG_READ |
798 				     VFIO_REGION_INFO_FLAG_WRITE;
799 			break;
800 		case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
801 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
802 			info.size = pci_resource_len(pdev, info.index);
803 			if (!info.size) {
804 				info.flags = 0;
805 				break;
806 			}
807 
808 			info.flags = VFIO_REGION_INFO_FLAG_READ |
809 				     VFIO_REGION_INFO_FLAG_WRITE;
810 			if (vdev->bar_mmap_supported[info.index]) {
811 				info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
812 				if (info.index == vdev->msix_bar) {
813 					ret = msix_mmappable_cap(vdev, &caps);
814 					if (ret)
815 						return ret;
816 				}
817 			}
818 
819 			break;
820 		case VFIO_PCI_ROM_REGION_INDEX:
821 		{
822 			void __iomem *io;
823 			size_t size;
824 			u16 cmd;
825 
826 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
827 			info.flags = 0;
828 
829 			/* Report the BAR size, not the ROM size */
830 			info.size = pci_resource_len(pdev, info.index);
831 			if (!info.size) {
832 				/* Shadow ROMs appear as PCI option ROMs */
833 				if (pdev->resource[PCI_ROM_RESOURCE].flags &
834 							IORESOURCE_ROM_SHADOW)
835 					info.size = 0x20000;
836 				else
837 					break;
838 			}
839 
840 			/*
841 			 * Is it really there?  Enable memory decode for
842 			 * implicit access in pci_map_rom().
843 			 */
844 			cmd = vfio_pci_memory_lock_and_enable(vdev);
845 			io = pci_map_rom(pdev, &size);
846 			if (io) {
847 				info.flags = VFIO_REGION_INFO_FLAG_READ;
848 				pci_unmap_rom(pdev, io);
849 			} else {
850 				info.size = 0;
851 			}
852 			vfio_pci_memory_unlock_and_restore(vdev, cmd);
853 
854 			break;
855 		}
856 		case VFIO_PCI_VGA_REGION_INDEX:
857 			if (!vdev->has_vga)
858 				return -EINVAL;
859 
860 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
861 			info.size = 0xc0000;
862 			info.flags = VFIO_REGION_INFO_FLAG_READ |
863 				     VFIO_REGION_INFO_FLAG_WRITE;
864 
865 			break;
866 		default:
867 		{
868 			struct vfio_region_info_cap_type cap_type = {
869 					.header.id = VFIO_REGION_INFO_CAP_TYPE,
870 					.header.version = 1 };
871 
872 			if (info.index >=
873 			    VFIO_PCI_NUM_REGIONS + vdev->num_regions)
874 				return -EINVAL;
875 			info.index = array_index_nospec(info.index,
876 							VFIO_PCI_NUM_REGIONS +
877 							vdev->num_regions);
878 
879 			i = info.index - VFIO_PCI_NUM_REGIONS;
880 
881 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
882 			info.size = vdev->region[i].size;
883 			info.flags = vdev->region[i].flags;
884 
885 			cap_type.type = vdev->region[i].type;
886 			cap_type.subtype = vdev->region[i].subtype;
887 
888 			ret = vfio_info_add_capability(&caps, &cap_type.header,
889 						       sizeof(cap_type));
890 			if (ret)
891 				return ret;
892 
893 			if (vdev->region[i].ops->add_capability) {
894 				ret = vdev->region[i].ops->add_capability(vdev,
895 						&vdev->region[i], &caps);
896 				if (ret)
897 					return ret;
898 			}
899 		}
900 		}
901 
902 		if (caps.size) {
903 			info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
904 			if (info.argsz < sizeof(info) + caps.size) {
905 				info.argsz = sizeof(info) + caps.size;
906 				info.cap_offset = 0;
907 			} else {
908 				vfio_info_cap_shift(&caps, sizeof(info));
909 				if (copy_to_user((void __user *)arg +
910 						  sizeof(info), caps.buf,
911 						  caps.size)) {
912 					kfree(caps.buf);
913 					return -EFAULT;
914 				}
915 				info.cap_offset = sizeof(info);
916 			}
917 
918 			kfree(caps.buf);
919 		}
920 
921 		return copy_to_user((void __user *)arg, &info, minsz) ?
922 			-EFAULT : 0;
923 
924 	} else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
925 		struct vfio_irq_info info;
926 
927 		minsz = offsetofend(struct vfio_irq_info, count);
928 
929 		if (copy_from_user(&info, (void __user *)arg, minsz))
930 			return -EFAULT;
931 
932 		if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
933 			return -EINVAL;
934 
935 		switch (info.index) {
936 		case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
937 		case VFIO_PCI_REQ_IRQ_INDEX:
938 			break;
939 		case VFIO_PCI_ERR_IRQ_INDEX:
940 			if (pci_is_pcie(vdev->pdev))
941 				break;
942 		/* fall through */
943 		default:
944 			return -EINVAL;
945 		}
946 
947 		info.flags = VFIO_IRQ_INFO_EVENTFD;
948 
949 		info.count = vfio_pci_get_irq_count(vdev, info.index);
950 
951 		if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
952 			info.flags |= (VFIO_IRQ_INFO_MASKABLE |
953 				       VFIO_IRQ_INFO_AUTOMASKED);
954 		else
955 			info.flags |= VFIO_IRQ_INFO_NORESIZE;
956 
957 		return copy_to_user((void __user *)arg, &info, minsz) ?
958 			-EFAULT : 0;
959 
960 	} else if (cmd == VFIO_DEVICE_SET_IRQS) {
961 		struct vfio_irq_set hdr;
962 		u8 *data = NULL;
963 		int max, ret = 0;
964 		size_t data_size = 0;
965 
966 		minsz = offsetofend(struct vfio_irq_set, count);
967 
968 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
969 			return -EFAULT;
970 
971 		max = vfio_pci_get_irq_count(vdev, hdr.index);
972 
973 		ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
974 						 VFIO_PCI_NUM_IRQS, &data_size);
975 		if (ret)
976 			return ret;
977 
978 		if (data_size) {
979 			data = memdup_user((void __user *)(arg + minsz),
980 					    data_size);
981 			if (IS_ERR(data))
982 				return PTR_ERR(data);
983 		}
984 
985 		mutex_lock(&vdev->igate);
986 
987 		ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
988 					      hdr.start, hdr.count, data);
989 
990 		mutex_unlock(&vdev->igate);
991 		kfree(data);
992 
993 		return ret;
994 
995 	} else if (cmd == VFIO_DEVICE_RESET) {
996 		int ret;
997 
998 		if (!vdev->reset_works)
999 			return -EINVAL;
1000 
1001 		vfio_pci_zap_and_down_write_memory_lock(vdev);
1002 		ret = pci_try_reset_function(vdev->pdev);
1003 		up_write(&vdev->memory_lock);
1004 
1005 		return ret;
1006 
1007 	} else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
1008 		struct vfio_pci_hot_reset_info hdr;
1009 		struct vfio_pci_fill_info fill = { 0 };
1010 		struct vfio_pci_dependent_device *devices = NULL;
1011 		bool slot = false;
1012 		int ret = 0;
1013 
1014 		minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
1015 
1016 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
1017 			return -EFAULT;
1018 
1019 		if (hdr.argsz < minsz)
1020 			return -EINVAL;
1021 
1022 		hdr.flags = 0;
1023 
1024 		/* Can we do a slot or bus reset or neither? */
1025 		if (!pci_probe_reset_slot(vdev->pdev->slot))
1026 			slot = true;
1027 		else if (pci_probe_reset_bus(vdev->pdev->bus))
1028 			return -ENODEV;
1029 
1030 		/* How many devices are affected? */
1031 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1032 						    vfio_pci_count_devs,
1033 						    &fill.max, slot);
1034 		if (ret)
1035 			return ret;
1036 
1037 		WARN_ON(!fill.max); /* Should always be at least one */
1038 
1039 		/*
1040 		 * If there's enough space, fill it now, otherwise return
1041 		 * -ENOSPC and the number of devices affected.
1042 		 */
1043 		if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
1044 			ret = -ENOSPC;
1045 			hdr.count = fill.max;
1046 			goto reset_info_exit;
1047 		}
1048 
1049 		devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
1050 		if (!devices)
1051 			return -ENOMEM;
1052 
1053 		fill.devices = devices;
1054 
1055 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1056 						    vfio_pci_fill_devs,
1057 						    &fill, slot);
1058 
1059 		/*
1060 		 * If a device was removed between counting and filling,
1061 		 * we may come up short of fill.max.  If a device was
1062 		 * added, we'll have a return of -EAGAIN above.
1063 		 */
1064 		if (!ret)
1065 			hdr.count = fill.cur;
1066 
1067 reset_info_exit:
1068 		if (copy_to_user((void __user *)arg, &hdr, minsz))
1069 			ret = -EFAULT;
1070 
1071 		if (!ret) {
1072 			if (copy_to_user((void __user *)(arg + minsz), devices,
1073 					 hdr.count * sizeof(*devices)))
1074 				ret = -EFAULT;
1075 		}
1076 
1077 		kfree(devices);
1078 		return ret;
1079 
1080 	} else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
1081 		struct vfio_pci_hot_reset hdr;
1082 		int32_t *group_fds;
1083 		struct vfio_pci_group_entry *groups;
1084 		struct vfio_pci_group_info info;
1085 		struct vfio_devices devs = { .cur_index = 0 };
1086 		bool slot = false;
1087 		int i, group_idx, mem_idx = 0, count = 0, ret = 0;
1088 
1089 		minsz = offsetofend(struct vfio_pci_hot_reset, count);
1090 
1091 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
1092 			return -EFAULT;
1093 
1094 		if (hdr.argsz < minsz || hdr.flags)
1095 			return -EINVAL;
1096 
1097 		/* Can we do a slot or bus reset or neither? */
1098 		if (!pci_probe_reset_slot(vdev->pdev->slot))
1099 			slot = true;
1100 		else if (pci_probe_reset_bus(vdev->pdev->bus))
1101 			return -ENODEV;
1102 
1103 		/*
1104 		 * We can't let userspace give us an arbitrarily large
1105 		 * buffer to copy, so verify how many we think there
1106 		 * could be.  Note groups can have multiple devices so
1107 		 * one group per device is the max.
1108 		 */
1109 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1110 						    vfio_pci_count_devs,
1111 						    &count, slot);
1112 		if (ret)
1113 			return ret;
1114 
1115 		/* Somewhere between 1 and count is OK */
1116 		if (!hdr.count || hdr.count > count)
1117 			return -EINVAL;
1118 
1119 		group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
1120 		groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
1121 		if (!group_fds || !groups) {
1122 			kfree(group_fds);
1123 			kfree(groups);
1124 			return -ENOMEM;
1125 		}
1126 
1127 		if (copy_from_user(group_fds, (void __user *)(arg + minsz),
1128 				   hdr.count * sizeof(*group_fds))) {
1129 			kfree(group_fds);
1130 			kfree(groups);
1131 			return -EFAULT;
1132 		}
1133 
1134 		/*
1135 		 * For each group_fd, get the group through the vfio external
1136 		 * user interface and store the group and iommu ID.  This
1137 		 * ensures the group is held across the reset.
1138 		 */
1139 		for (group_idx = 0; group_idx < hdr.count; group_idx++) {
1140 			struct vfio_group *group;
1141 			struct fd f = fdget(group_fds[group_idx]);
1142 			if (!f.file) {
1143 				ret = -EBADF;
1144 				break;
1145 			}
1146 
1147 			group = vfio_group_get_external_user(f.file);
1148 			fdput(f);
1149 			if (IS_ERR(group)) {
1150 				ret = PTR_ERR(group);
1151 				break;
1152 			}
1153 
1154 			groups[group_idx].group = group;
1155 			groups[group_idx].id =
1156 					vfio_external_user_iommu_id(group);
1157 		}
1158 
1159 		kfree(group_fds);
1160 
1161 		/* release reference to groups on error */
1162 		if (ret)
1163 			goto hot_reset_release;
1164 
1165 		info.count = hdr.count;
1166 		info.groups = groups;
1167 
1168 		/*
1169 		 * Test whether all the affected devices are contained
1170 		 * by the set of groups provided by the user.
1171 		 */
1172 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1173 						    vfio_pci_validate_devs,
1174 						    &info, slot);
1175 		if (ret)
1176 			goto hot_reset_release;
1177 
1178 		devs.max_index = count;
1179 		devs.devices = kcalloc(count, sizeof(struct vfio_device *),
1180 				       GFP_KERNEL);
1181 		if (!devs.devices) {
1182 			ret = -ENOMEM;
1183 			goto hot_reset_release;
1184 		}
1185 
1186 		/*
1187 		 * We need to get memory_lock for each device, but devices
1188 		 * can share mmap_lock, therefore we need to zap and hold
1189 		 * the vma_lock for each device, and only then get each
1190 		 * memory_lock.
1191 		 */
1192 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1193 					    vfio_pci_try_zap_and_vma_lock_cb,
1194 					    &devs, slot);
1195 		if (ret)
1196 			goto hot_reset_release;
1197 
1198 		for (; mem_idx < devs.cur_index; mem_idx++) {
1199 			struct vfio_pci_device *tmp;
1200 
1201 			tmp = vfio_device_data(devs.devices[mem_idx]);
1202 
1203 			ret = down_write_trylock(&tmp->memory_lock);
1204 			if (!ret) {
1205 				ret = -EBUSY;
1206 				goto hot_reset_release;
1207 			}
1208 			mutex_unlock(&tmp->vma_lock);
1209 		}
1210 
1211 		/* User has access, do the reset */
1212 		ret = pci_reset_bus(vdev->pdev);
1213 
1214 hot_reset_release:
1215 		for (i = 0; i < devs.cur_index; i++) {
1216 			struct vfio_device *device;
1217 			struct vfio_pci_device *tmp;
1218 
1219 			device = devs.devices[i];
1220 			tmp = vfio_device_data(device);
1221 
1222 			if (i < mem_idx)
1223 				up_write(&tmp->memory_lock);
1224 			else
1225 				mutex_unlock(&tmp->vma_lock);
1226 			vfio_device_put(device);
1227 		}
1228 		kfree(devs.devices);
1229 
1230 		for (group_idx--; group_idx >= 0; group_idx--)
1231 			vfio_group_put_external_user(groups[group_idx].group);
1232 
1233 		kfree(groups);
1234 		return ret;
1235 	} else if (cmd == VFIO_DEVICE_IOEVENTFD) {
1236 		struct vfio_device_ioeventfd ioeventfd;
1237 		int count;
1238 
1239 		minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1240 
1241 		if (copy_from_user(&ioeventfd, (void __user *)arg, minsz))
1242 			return -EFAULT;
1243 
1244 		if (ioeventfd.argsz < minsz)
1245 			return -EINVAL;
1246 
1247 		if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1248 			return -EINVAL;
1249 
1250 		count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1251 
1252 		if (hweight8(count) != 1 || ioeventfd.fd < -1)
1253 			return -EINVAL;
1254 
1255 		return vfio_pci_ioeventfd(vdev, ioeventfd.offset,
1256 					  ioeventfd.data, count, ioeventfd.fd);
1257 	} else if (cmd == VFIO_DEVICE_FEATURE) {
1258 		struct vfio_device_feature feature;
1259 		uuid_t uuid;
1260 
1261 		minsz = offsetofend(struct vfio_device_feature, flags);
1262 
1263 		if (copy_from_user(&feature, (void __user *)arg, minsz))
1264 			return -EFAULT;
1265 
1266 		if (feature.argsz < minsz)
1267 			return -EINVAL;
1268 
1269 		/* Check unknown flags */
1270 		if (feature.flags & ~(VFIO_DEVICE_FEATURE_MASK |
1271 				      VFIO_DEVICE_FEATURE_SET |
1272 				      VFIO_DEVICE_FEATURE_GET |
1273 				      VFIO_DEVICE_FEATURE_PROBE))
1274 			return -EINVAL;
1275 
1276 		/* GET & SET are mutually exclusive except with PROBE */
1277 		if (!(feature.flags & VFIO_DEVICE_FEATURE_PROBE) &&
1278 		    (feature.flags & VFIO_DEVICE_FEATURE_SET) &&
1279 		    (feature.flags & VFIO_DEVICE_FEATURE_GET))
1280 			return -EINVAL;
1281 
1282 		switch (feature.flags & VFIO_DEVICE_FEATURE_MASK) {
1283 		case VFIO_DEVICE_FEATURE_PCI_VF_TOKEN:
1284 			if (!vdev->vf_token)
1285 				return -ENOTTY;
1286 
1287 			/*
1288 			 * We do not support GET of the VF Token UUID as this
1289 			 * could expose the token of the previous device user.
1290 			 */
1291 			if (feature.flags & VFIO_DEVICE_FEATURE_GET)
1292 				return -EINVAL;
1293 
1294 			if (feature.flags & VFIO_DEVICE_FEATURE_PROBE)
1295 				return 0;
1296 
1297 			/* Don't SET unless told to do so */
1298 			if (!(feature.flags & VFIO_DEVICE_FEATURE_SET))
1299 				return -EINVAL;
1300 
1301 			if (feature.argsz < minsz + sizeof(uuid))
1302 				return -EINVAL;
1303 
1304 			if (copy_from_user(&uuid, (void __user *)(arg + minsz),
1305 					   sizeof(uuid)))
1306 				return -EFAULT;
1307 
1308 			mutex_lock(&vdev->vf_token->lock);
1309 			uuid_copy(&vdev->vf_token->uuid, &uuid);
1310 			mutex_unlock(&vdev->vf_token->lock);
1311 
1312 			return 0;
1313 		default:
1314 			return -ENOTTY;
1315 		}
1316 	}
1317 
1318 	return -ENOTTY;
1319 }
1320 
1321 static ssize_t vfio_pci_rw(void *device_data, char __user *buf,
1322 			   size_t count, loff_t *ppos, bool iswrite)
1323 {
1324 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1325 	struct vfio_pci_device *vdev = device_data;
1326 
1327 	if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1328 		return -EINVAL;
1329 
1330 	switch (index) {
1331 	case VFIO_PCI_CONFIG_REGION_INDEX:
1332 		return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1333 
1334 	case VFIO_PCI_ROM_REGION_INDEX:
1335 		if (iswrite)
1336 			return -EINVAL;
1337 		return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1338 
1339 	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1340 		return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1341 
1342 	case VFIO_PCI_VGA_REGION_INDEX:
1343 		return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1344 	default:
1345 		index -= VFIO_PCI_NUM_REGIONS;
1346 		return vdev->region[index].ops->rw(vdev, buf,
1347 						   count, ppos, iswrite);
1348 	}
1349 
1350 	return -EINVAL;
1351 }
1352 
1353 static ssize_t vfio_pci_read(void *device_data, char __user *buf,
1354 			     size_t count, loff_t *ppos)
1355 {
1356 	if (!count)
1357 		return 0;
1358 
1359 	return vfio_pci_rw(device_data, buf, count, ppos, false);
1360 }
1361 
1362 static ssize_t vfio_pci_write(void *device_data, const char __user *buf,
1363 			      size_t count, loff_t *ppos)
1364 {
1365 	if (!count)
1366 		return 0;
1367 
1368 	return vfio_pci_rw(device_data, (char __user *)buf, count, ppos, true);
1369 }
1370 
1371 /* Return 1 on zap and vma_lock acquired, 0 on contention (only with @try) */
1372 static int vfio_pci_zap_and_vma_lock(struct vfio_pci_device *vdev, bool try)
1373 {
1374 	struct vfio_pci_mmap_vma *mmap_vma, *tmp;
1375 
1376 	/*
1377 	 * Lock ordering:
1378 	 * vma_lock is nested under mmap_lock for vm_ops callback paths.
1379 	 * The memory_lock semaphore is used by both code paths calling
1380 	 * into this function to zap vmas and the vm_ops.fault callback
1381 	 * to protect the memory enable state of the device.
1382 	 *
1383 	 * When zapping vmas we need to maintain the mmap_lock => vma_lock
1384 	 * ordering, which requires using vma_lock to walk vma_list to
1385 	 * acquire an mm, then dropping vma_lock to get the mmap_lock and
1386 	 * reacquiring vma_lock.  This logic is derived from similar
1387 	 * requirements in uverbs_user_mmap_disassociate().
1388 	 *
1389 	 * mmap_lock must always be the top-level lock when it is taken.
1390 	 * Therefore we can only hold the memory_lock write lock when
1391 	 * vma_list is empty, as we'd need to take mmap_lock to clear
1392 	 * entries.  vma_list can only be guaranteed empty when holding
1393 	 * vma_lock, thus memory_lock is nested under vma_lock.
1394 	 *
1395 	 * This enables the vm_ops.fault callback to acquire vma_lock,
1396 	 * followed by memory_lock read lock, while already holding
1397 	 * mmap_lock without risk of deadlock.
1398 	 */
1399 	while (1) {
1400 		struct mm_struct *mm = NULL;
1401 
1402 		if (try) {
1403 			if (!mutex_trylock(&vdev->vma_lock))
1404 				return 0;
1405 		} else {
1406 			mutex_lock(&vdev->vma_lock);
1407 		}
1408 		while (!list_empty(&vdev->vma_list)) {
1409 			mmap_vma = list_first_entry(&vdev->vma_list,
1410 						    struct vfio_pci_mmap_vma,
1411 						    vma_next);
1412 			mm = mmap_vma->vma->vm_mm;
1413 			if (mmget_not_zero(mm))
1414 				break;
1415 
1416 			list_del(&mmap_vma->vma_next);
1417 			kfree(mmap_vma);
1418 			mm = NULL;
1419 		}
1420 		if (!mm)
1421 			return 1;
1422 		mutex_unlock(&vdev->vma_lock);
1423 
1424 		if (try) {
1425 			if (!mmap_read_trylock(mm)) {
1426 				mmput(mm);
1427 				return 0;
1428 			}
1429 		} else {
1430 			mmap_read_lock(mm);
1431 		}
1432 		if (mmget_still_valid(mm)) {
1433 			if (try) {
1434 				if (!mutex_trylock(&vdev->vma_lock)) {
1435 					mmap_read_unlock(mm);
1436 					mmput(mm);
1437 					return 0;
1438 				}
1439 			} else {
1440 				mutex_lock(&vdev->vma_lock);
1441 			}
1442 			list_for_each_entry_safe(mmap_vma, tmp,
1443 						 &vdev->vma_list, vma_next) {
1444 				struct vm_area_struct *vma = mmap_vma->vma;
1445 
1446 				if (vma->vm_mm != mm)
1447 					continue;
1448 
1449 				list_del(&mmap_vma->vma_next);
1450 				kfree(mmap_vma);
1451 
1452 				zap_vma_ptes(vma, vma->vm_start,
1453 					     vma->vm_end - vma->vm_start);
1454 			}
1455 			mutex_unlock(&vdev->vma_lock);
1456 		}
1457 		mmap_read_unlock(mm);
1458 		mmput(mm);
1459 	}
1460 }
1461 
1462 void vfio_pci_zap_and_down_write_memory_lock(struct vfio_pci_device *vdev)
1463 {
1464 	vfio_pci_zap_and_vma_lock(vdev, false);
1465 	down_write(&vdev->memory_lock);
1466 	mutex_unlock(&vdev->vma_lock);
1467 }
1468 
1469 u16 vfio_pci_memory_lock_and_enable(struct vfio_pci_device *vdev)
1470 {
1471 	u16 cmd;
1472 
1473 	down_write(&vdev->memory_lock);
1474 	pci_read_config_word(vdev->pdev, PCI_COMMAND, &cmd);
1475 	if (!(cmd & PCI_COMMAND_MEMORY))
1476 		pci_write_config_word(vdev->pdev, PCI_COMMAND,
1477 				      cmd | PCI_COMMAND_MEMORY);
1478 
1479 	return cmd;
1480 }
1481 
1482 void vfio_pci_memory_unlock_and_restore(struct vfio_pci_device *vdev, u16 cmd)
1483 {
1484 	pci_write_config_word(vdev->pdev, PCI_COMMAND, cmd);
1485 	up_write(&vdev->memory_lock);
1486 }
1487 
1488 /* Caller holds vma_lock */
1489 static int __vfio_pci_add_vma(struct vfio_pci_device *vdev,
1490 			      struct vm_area_struct *vma)
1491 {
1492 	struct vfio_pci_mmap_vma *mmap_vma;
1493 
1494 	mmap_vma = kmalloc(sizeof(*mmap_vma), GFP_KERNEL);
1495 	if (!mmap_vma)
1496 		return -ENOMEM;
1497 
1498 	mmap_vma->vma = vma;
1499 	list_add(&mmap_vma->vma_next, &vdev->vma_list);
1500 
1501 	return 0;
1502 }
1503 
1504 /*
1505  * Zap mmaps on open so that we can fault them in on access and therefore
1506  * our vma_list only tracks mappings accessed since last zap.
1507  */
1508 static void vfio_pci_mmap_open(struct vm_area_struct *vma)
1509 {
1510 	zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1511 }
1512 
1513 static void vfio_pci_mmap_close(struct vm_area_struct *vma)
1514 {
1515 	struct vfio_pci_device *vdev = vma->vm_private_data;
1516 	struct vfio_pci_mmap_vma *mmap_vma;
1517 
1518 	mutex_lock(&vdev->vma_lock);
1519 	list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1520 		if (mmap_vma->vma == vma) {
1521 			list_del(&mmap_vma->vma_next);
1522 			kfree(mmap_vma);
1523 			break;
1524 		}
1525 	}
1526 	mutex_unlock(&vdev->vma_lock);
1527 }
1528 
1529 static vm_fault_t vfio_pci_mmap_fault(struct vm_fault *vmf)
1530 {
1531 	struct vm_area_struct *vma = vmf->vma;
1532 	struct vfio_pci_device *vdev = vma->vm_private_data;
1533 	vm_fault_t ret = VM_FAULT_NOPAGE;
1534 
1535 	mutex_lock(&vdev->vma_lock);
1536 	down_read(&vdev->memory_lock);
1537 
1538 	if (!__vfio_pci_memory_enabled(vdev)) {
1539 		ret = VM_FAULT_SIGBUS;
1540 		mutex_unlock(&vdev->vma_lock);
1541 		goto up_out;
1542 	}
1543 
1544 	if (__vfio_pci_add_vma(vdev, vma)) {
1545 		ret = VM_FAULT_OOM;
1546 		mutex_unlock(&vdev->vma_lock);
1547 		goto up_out;
1548 	}
1549 
1550 	mutex_unlock(&vdev->vma_lock);
1551 
1552 	if (remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1553 			    vma->vm_end - vma->vm_start, vma->vm_page_prot))
1554 		ret = VM_FAULT_SIGBUS;
1555 
1556 up_out:
1557 	up_read(&vdev->memory_lock);
1558 	return ret;
1559 }
1560 
1561 static const struct vm_operations_struct vfio_pci_mmap_ops = {
1562 	.open = vfio_pci_mmap_open,
1563 	.close = vfio_pci_mmap_close,
1564 	.fault = vfio_pci_mmap_fault,
1565 };
1566 
1567 static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma)
1568 {
1569 	struct vfio_pci_device *vdev = device_data;
1570 	struct pci_dev *pdev = vdev->pdev;
1571 	unsigned int index;
1572 	u64 phys_len, req_len, pgoff, req_start;
1573 	int ret;
1574 
1575 	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1576 
1577 	if (vma->vm_end < vma->vm_start)
1578 		return -EINVAL;
1579 	if ((vma->vm_flags & VM_SHARED) == 0)
1580 		return -EINVAL;
1581 	if (index >= VFIO_PCI_NUM_REGIONS) {
1582 		int regnum = index - VFIO_PCI_NUM_REGIONS;
1583 		struct vfio_pci_region *region = vdev->region + regnum;
1584 
1585 		if (region && region->ops && region->ops->mmap &&
1586 		    (region->flags & VFIO_REGION_INFO_FLAG_MMAP))
1587 			return region->ops->mmap(vdev, region, vma);
1588 		return -EINVAL;
1589 	}
1590 	if (index >= VFIO_PCI_ROM_REGION_INDEX)
1591 		return -EINVAL;
1592 	if (!vdev->bar_mmap_supported[index])
1593 		return -EINVAL;
1594 
1595 	phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1596 	req_len = vma->vm_end - vma->vm_start;
1597 	pgoff = vma->vm_pgoff &
1598 		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1599 	req_start = pgoff << PAGE_SHIFT;
1600 
1601 	if (req_start + req_len > phys_len)
1602 		return -EINVAL;
1603 
1604 	/*
1605 	 * Even though we don't make use of the barmap for the mmap,
1606 	 * we need to request the region and the barmap tracks that.
1607 	 */
1608 	if (!vdev->barmap[index]) {
1609 		ret = pci_request_selected_regions(pdev,
1610 						   1 << index, "vfio-pci");
1611 		if (ret)
1612 			return ret;
1613 
1614 		vdev->barmap[index] = pci_iomap(pdev, index, 0);
1615 		if (!vdev->barmap[index]) {
1616 			pci_release_selected_regions(pdev, 1 << index);
1617 			return -ENOMEM;
1618 		}
1619 	}
1620 
1621 	vma->vm_private_data = vdev;
1622 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1623 	vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1624 
1625 	/*
1626 	 * See remap_pfn_range(), called from vfio_pci_fault() but we can't
1627 	 * change vm_flags within the fault handler.  Set them now.
1628 	 */
1629 	vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1630 	vma->vm_ops = &vfio_pci_mmap_ops;
1631 
1632 	return 0;
1633 }
1634 
1635 static void vfio_pci_request(void *device_data, unsigned int count)
1636 {
1637 	struct vfio_pci_device *vdev = device_data;
1638 	struct pci_dev *pdev = vdev->pdev;
1639 
1640 	mutex_lock(&vdev->igate);
1641 
1642 	if (vdev->req_trigger) {
1643 		if (!(count % 10))
1644 			pci_notice_ratelimited(pdev,
1645 				"Relaying device request to user (#%u)\n",
1646 				count);
1647 		eventfd_signal(vdev->req_trigger, 1);
1648 	} else if (count == 0) {
1649 		pci_warn(pdev,
1650 			"No device request channel registered, blocked until released by user\n");
1651 	}
1652 
1653 	mutex_unlock(&vdev->igate);
1654 }
1655 
1656 static int vfio_pci_validate_vf_token(struct vfio_pci_device *vdev,
1657 				      bool vf_token, uuid_t *uuid)
1658 {
1659 	/*
1660 	 * There's always some degree of trust or collaboration between SR-IOV
1661 	 * PF and VFs, even if just that the PF hosts the SR-IOV capability and
1662 	 * can disrupt VFs with a reset, but often the PF has more explicit
1663 	 * access to deny service to the VF or access data passed through the
1664 	 * VF.  We therefore require an opt-in via a shared VF token (UUID) to
1665 	 * represent this trust.  This both prevents that a VF driver might
1666 	 * assume the PF driver is a trusted, in-kernel driver, and also that
1667 	 * a PF driver might be replaced with a rogue driver, unknown to in-use
1668 	 * VF drivers.
1669 	 *
1670 	 * Therefore when presented with a VF, if the PF is a vfio device and
1671 	 * it is bound to the vfio-pci driver, the user needs to provide a VF
1672 	 * token to access the device, in the form of appending a vf_token to
1673 	 * the device name, for example:
1674 	 *
1675 	 * "0000:04:10.0 vf_token=bd8d9d2b-5a5f-4f5a-a211-f591514ba1f3"
1676 	 *
1677 	 * When presented with a PF which has VFs in use, the user must also
1678 	 * provide the current VF token to prove collaboration with existing
1679 	 * VF users.  If VFs are not in use, the VF token provided for the PF
1680 	 * device will act to set the VF token.
1681 	 *
1682 	 * If the VF token is provided but unused, an error is generated.
1683 	 */
1684 	if (!vdev->pdev->is_virtfn && !vdev->vf_token && !vf_token)
1685 		return 0; /* No VF token provided or required */
1686 
1687 	if (vdev->pdev->is_virtfn) {
1688 		struct vfio_device *pf_dev;
1689 		struct vfio_pci_device *pf_vdev = get_pf_vdev(vdev, &pf_dev);
1690 		bool match;
1691 
1692 		if (!pf_vdev) {
1693 			if (!vf_token)
1694 				return 0; /* PF is not vfio-pci, no VF token */
1695 
1696 			pci_info_ratelimited(vdev->pdev,
1697 				"VF token incorrectly provided, PF not bound to vfio-pci\n");
1698 			return -EINVAL;
1699 		}
1700 
1701 		if (!vf_token) {
1702 			vfio_device_put(pf_dev);
1703 			pci_info_ratelimited(vdev->pdev,
1704 				"VF token required to access device\n");
1705 			return -EACCES;
1706 		}
1707 
1708 		mutex_lock(&pf_vdev->vf_token->lock);
1709 		match = uuid_equal(uuid, &pf_vdev->vf_token->uuid);
1710 		mutex_unlock(&pf_vdev->vf_token->lock);
1711 
1712 		vfio_device_put(pf_dev);
1713 
1714 		if (!match) {
1715 			pci_info_ratelimited(vdev->pdev,
1716 				"Incorrect VF token provided for device\n");
1717 			return -EACCES;
1718 		}
1719 	} else if (vdev->vf_token) {
1720 		mutex_lock(&vdev->vf_token->lock);
1721 		if (vdev->vf_token->users) {
1722 			if (!vf_token) {
1723 				mutex_unlock(&vdev->vf_token->lock);
1724 				pci_info_ratelimited(vdev->pdev,
1725 					"VF token required to access device\n");
1726 				return -EACCES;
1727 			}
1728 
1729 			if (!uuid_equal(uuid, &vdev->vf_token->uuid)) {
1730 				mutex_unlock(&vdev->vf_token->lock);
1731 				pci_info_ratelimited(vdev->pdev,
1732 					"Incorrect VF token provided for device\n");
1733 				return -EACCES;
1734 			}
1735 		} else if (vf_token) {
1736 			uuid_copy(&vdev->vf_token->uuid, uuid);
1737 		}
1738 
1739 		mutex_unlock(&vdev->vf_token->lock);
1740 	} else if (vf_token) {
1741 		pci_info_ratelimited(vdev->pdev,
1742 			"VF token incorrectly provided, not a PF or VF\n");
1743 		return -EINVAL;
1744 	}
1745 
1746 	return 0;
1747 }
1748 
1749 #define VF_TOKEN_ARG "vf_token="
1750 
1751 static int vfio_pci_match(void *device_data, char *buf)
1752 {
1753 	struct vfio_pci_device *vdev = device_data;
1754 	bool vf_token = false;
1755 	uuid_t uuid;
1756 	int ret;
1757 
1758 	if (strncmp(pci_name(vdev->pdev), buf, strlen(pci_name(vdev->pdev))))
1759 		return 0; /* No match */
1760 
1761 	if (strlen(buf) > strlen(pci_name(vdev->pdev))) {
1762 		buf += strlen(pci_name(vdev->pdev));
1763 
1764 		if (*buf != ' ')
1765 			return 0; /* No match: non-whitespace after name */
1766 
1767 		while (*buf) {
1768 			if (*buf == ' ') {
1769 				buf++;
1770 				continue;
1771 			}
1772 
1773 			if (!vf_token && !strncmp(buf, VF_TOKEN_ARG,
1774 						  strlen(VF_TOKEN_ARG))) {
1775 				buf += strlen(VF_TOKEN_ARG);
1776 
1777 				if (strlen(buf) < UUID_STRING_LEN)
1778 					return -EINVAL;
1779 
1780 				ret = uuid_parse(buf, &uuid);
1781 				if (ret)
1782 					return ret;
1783 
1784 				vf_token = true;
1785 				buf += UUID_STRING_LEN;
1786 			} else {
1787 				/* Unknown/duplicate option */
1788 				return -EINVAL;
1789 			}
1790 		}
1791 	}
1792 
1793 	ret = vfio_pci_validate_vf_token(vdev, vf_token, &uuid);
1794 	if (ret)
1795 		return ret;
1796 
1797 	return 1; /* Match */
1798 }
1799 
1800 static const struct vfio_device_ops vfio_pci_ops = {
1801 	.name		= "vfio-pci",
1802 	.open		= vfio_pci_open,
1803 	.release	= vfio_pci_release,
1804 	.ioctl		= vfio_pci_ioctl,
1805 	.read		= vfio_pci_read,
1806 	.write		= vfio_pci_write,
1807 	.mmap		= vfio_pci_mmap,
1808 	.request	= vfio_pci_request,
1809 	.match		= vfio_pci_match,
1810 };
1811 
1812 static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev);
1813 static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck);
1814 static struct pci_driver vfio_pci_driver;
1815 
1816 static int vfio_pci_bus_notifier(struct notifier_block *nb,
1817 				 unsigned long action, void *data)
1818 {
1819 	struct vfio_pci_device *vdev = container_of(nb,
1820 						    struct vfio_pci_device, nb);
1821 	struct device *dev = data;
1822 	struct pci_dev *pdev = to_pci_dev(dev);
1823 	struct pci_dev *physfn = pci_physfn(pdev);
1824 
1825 	if (action == BUS_NOTIFY_ADD_DEVICE &&
1826 	    pdev->is_virtfn && physfn == vdev->pdev) {
1827 		pci_info(vdev->pdev, "Captured SR-IOV VF %s driver_override\n",
1828 			 pci_name(pdev));
1829 		pdev->driver_override = kasprintf(GFP_KERNEL, "%s",
1830 						  vfio_pci_ops.name);
1831 	} else if (action == BUS_NOTIFY_BOUND_DRIVER &&
1832 		   pdev->is_virtfn && physfn == vdev->pdev) {
1833 		struct pci_driver *drv = pci_dev_driver(pdev);
1834 
1835 		if (drv && drv != &vfio_pci_driver)
1836 			pci_warn(vdev->pdev,
1837 				 "VF %s bound to driver %s while PF bound to vfio-pci\n",
1838 				 pci_name(pdev), drv->name);
1839 	}
1840 
1841 	return 0;
1842 }
1843 
1844 static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1845 {
1846 	struct vfio_pci_device *vdev;
1847 	struct iommu_group *group;
1848 	int ret;
1849 
1850 	if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1851 		return -EINVAL;
1852 
1853 	/*
1854 	 * Prevent binding to PFs with VFs enabled, the VFs might be in use
1855 	 * by the host or other users.  We cannot capture the VFs if they
1856 	 * already exist, nor can we track VF users.  Disabling SR-IOV here
1857 	 * would initiate removing the VFs, which would unbind the driver,
1858 	 * which is prone to blocking if that VF is also in use by vfio-pci.
1859 	 * Just reject these PFs and let the user sort it out.
1860 	 */
1861 	if (pci_num_vf(pdev)) {
1862 		pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
1863 		return -EBUSY;
1864 	}
1865 
1866 	group = vfio_iommu_group_get(&pdev->dev);
1867 	if (!group)
1868 		return -EINVAL;
1869 
1870 	vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
1871 	if (!vdev) {
1872 		ret = -ENOMEM;
1873 		goto out_group_put;
1874 	}
1875 
1876 	vdev->pdev = pdev;
1877 	vdev->irq_type = VFIO_PCI_NUM_IRQS;
1878 	mutex_init(&vdev->igate);
1879 	spin_lock_init(&vdev->irqlock);
1880 	mutex_init(&vdev->ioeventfds_lock);
1881 	INIT_LIST_HEAD(&vdev->ioeventfds_list);
1882 	mutex_init(&vdev->vma_lock);
1883 	INIT_LIST_HEAD(&vdev->vma_list);
1884 	init_rwsem(&vdev->memory_lock);
1885 
1886 	ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
1887 	if (ret)
1888 		goto out_free;
1889 
1890 	ret = vfio_pci_reflck_attach(vdev);
1891 	if (ret)
1892 		goto out_del_group_dev;
1893 
1894 	if (pdev->is_physfn) {
1895 		vdev->vf_token = kzalloc(sizeof(*vdev->vf_token), GFP_KERNEL);
1896 		if (!vdev->vf_token) {
1897 			ret = -ENOMEM;
1898 			goto out_reflck;
1899 		}
1900 
1901 		mutex_init(&vdev->vf_token->lock);
1902 		uuid_gen(&vdev->vf_token->uuid);
1903 
1904 		vdev->nb.notifier_call = vfio_pci_bus_notifier;
1905 		ret = bus_register_notifier(&pci_bus_type, &vdev->nb);
1906 		if (ret)
1907 			goto out_vf_token;
1908 	}
1909 
1910 	if (vfio_pci_is_vga(pdev)) {
1911 		vga_client_register(pdev, vdev, NULL, vfio_pci_set_vga_decode);
1912 		vga_set_legacy_decoding(pdev,
1913 					vfio_pci_set_vga_decode(vdev, false));
1914 	}
1915 
1916 	vfio_pci_probe_power_state(vdev);
1917 
1918 	if (!disable_idle_d3) {
1919 		/*
1920 		 * pci-core sets the device power state to an unknown value at
1921 		 * bootup and after being removed from a driver.  The only
1922 		 * transition it allows from this unknown state is to D0, which
1923 		 * typically happens when a driver calls pci_enable_device().
1924 		 * We're not ready to enable the device yet, but we do want to
1925 		 * be able to get to D3.  Therefore first do a D0 transition
1926 		 * before going to D3.
1927 		 */
1928 		vfio_pci_set_power_state(vdev, PCI_D0);
1929 		vfio_pci_set_power_state(vdev, PCI_D3hot);
1930 	}
1931 
1932 	return ret;
1933 
1934 out_vf_token:
1935 	kfree(vdev->vf_token);
1936 out_reflck:
1937 	vfio_pci_reflck_put(vdev->reflck);
1938 out_del_group_dev:
1939 	vfio_del_group_dev(&pdev->dev);
1940 out_free:
1941 	kfree(vdev);
1942 out_group_put:
1943 	vfio_iommu_group_put(group, &pdev->dev);
1944 	return ret;
1945 }
1946 
1947 static void vfio_pci_remove(struct pci_dev *pdev)
1948 {
1949 	struct vfio_pci_device *vdev;
1950 
1951 	pci_disable_sriov(pdev);
1952 
1953 	vdev = vfio_del_group_dev(&pdev->dev);
1954 	if (!vdev)
1955 		return;
1956 
1957 	if (vdev->vf_token) {
1958 		WARN_ON(vdev->vf_token->users);
1959 		mutex_destroy(&vdev->vf_token->lock);
1960 		kfree(vdev->vf_token);
1961 	}
1962 
1963 	if (vdev->nb.notifier_call)
1964 		bus_unregister_notifier(&pci_bus_type, &vdev->nb);
1965 
1966 	vfio_pci_reflck_put(vdev->reflck);
1967 
1968 	vfio_iommu_group_put(pdev->dev.iommu_group, &pdev->dev);
1969 	kfree(vdev->region);
1970 	mutex_destroy(&vdev->ioeventfds_lock);
1971 
1972 	if (!disable_idle_d3)
1973 		vfio_pci_set_power_state(vdev, PCI_D0);
1974 
1975 	kfree(vdev->pm_save);
1976 	kfree(vdev);
1977 
1978 	if (vfio_pci_is_vga(pdev)) {
1979 		vga_client_register(pdev, NULL, NULL, NULL);
1980 		vga_set_legacy_decoding(pdev,
1981 				VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
1982 				VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM);
1983 	}
1984 }
1985 
1986 static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
1987 						  pci_channel_state_t state)
1988 {
1989 	struct vfio_pci_device *vdev;
1990 	struct vfio_device *device;
1991 
1992 	device = vfio_device_get_from_dev(&pdev->dev);
1993 	if (device == NULL)
1994 		return PCI_ERS_RESULT_DISCONNECT;
1995 
1996 	vdev = vfio_device_data(device);
1997 	if (vdev == NULL) {
1998 		vfio_device_put(device);
1999 		return PCI_ERS_RESULT_DISCONNECT;
2000 	}
2001 
2002 	mutex_lock(&vdev->igate);
2003 
2004 	if (vdev->err_trigger)
2005 		eventfd_signal(vdev->err_trigger, 1);
2006 
2007 	mutex_unlock(&vdev->igate);
2008 
2009 	vfio_device_put(device);
2010 
2011 	return PCI_ERS_RESULT_CAN_RECOVER;
2012 }
2013 
2014 static int vfio_pci_sriov_configure(struct pci_dev *pdev, int nr_virtfn)
2015 {
2016 	struct vfio_pci_device *vdev;
2017 	struct vfio_device *device;
2018 	int ret = 0;
2019 
2020 	might_sleep();
2021 
2022 	if (!enable_sriov)
2023 		return -ENOENT;
2024 
2025 	device = vfio_device_get_from_dev(&pdev->dev);
2026 	if (!device)
2027 		return -ENODEV;
2028 
2029 	vdev = vfio_device_data(device);
2030 	if (!vdev) {
2031 		vfio_device_put(device);
2032 		return -ENODEV;
2033 	}
2034 
2035 	if (nr_virtfn == 0)
2036 		pci_disable_sriov(pdev);
2037 	else
2038 		ret = pci_enable_sriov(pdev, nr_virtfn);
2039 
2040 	vfio_device_put(device);
2041 
2042 	return ret < 0 ? ret : nr_virtfn;
2043 }
2044 
2045 static const struct pci_error_handlers vfio_err_handlers = {
2046 	.error_detected = vfio_pci_aer_err_detected,
2047 };
2048 
2049 static struct pci_driver vfio_pci_driver = {
2050 	.name			= "vfio-pci",
2051 	.id_table		= NULL, /* only dynamic ids */
2052 	.probe			= vfio_pci_probe,
2053 	.remove			= vfio_pci_remove,
2054 	.sriov_configure	= vfio_pci_sriov_configure,
2055 	.err_handler		= &vfio_err_handlers,
2056 };
2057 
2058 static DEFINE_MUTEX(reflck_lock);
2059 
2060 static struct vfio_pci_reflck *vfio_pci_reflck_alloc(void)
2061 {
2062 	struct vfio_pci_reflck *reflck;
2063 
2064 	reflck = kzalloc(sizeof(*reflck), GFP_KERNEL);
2065 	if (!reflck)
2066 		return ERR_PTR(-ENOMEM);
2067 
2068 	kref_init(&reflck->kref);
2069 	mutex_init(&reflck->lock);
2070 
2071 	return reflck;
2072 }
2073 
2074 static void vfio_pci_reflck_get(struct vfio_pci_reflck *reflck)
2075 {
2076 	kref_get(&reflck->kref);
2077 }
2078 
2079 static int vfio_pci_reflck_find(struct pci_dev *pdev, void *data)
2080 {
2081 	struct vfio_pci_reflck **preflck = data;
2082 	struct vfio_device *device;
2083 	struct vfio_pci_device *vdev;
2084 
2085 	device = vfio_device_get_from_dev(&pdev->dev);
2086 	if (!device)
2087 		return 0;
2088 
2089 	if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2090 		vfio_device_put(device);
2091 		return 0;
2092 	}
2093 
2094 	vdev = vfio_device_data(device);
2095 
2096 	if (vdev->reflck) {
2097 		vfio_pci_reflck_get(vdev->reflck);
2098 		*preflck = vdev->reflck;
2099 		vfio_device_put(device);
2100 		return 1;
2101 	}
2102 
2103 	vfio_device_put(device);
2104 	return 0;
2105 }
2106 
2107 static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev)
2108 {
2109 	bool slot = !pci_probe_reset_slot(vdev->pdev->slot);
2110 
2111 	mutex_lock(&reflck_lock);
2112 
2113 	if (pci_is_root_bus(vdev->pdev->bus) ||
2114 	    vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_reflck_find,
2115 					  &vdev->reflck, slot) <= 0)
2116 		vdev->reflck = vfio_pci_reflck_alloc();
2117 
2118 	mutex_unlock(&reflck_lock);
2119 
2120 	return PTR_ERR_OR_ZERO(vdev->reflck);
2121 }
2122 
2123 static void vfio_pci_reflck_release(struct kref *kref)
2124 {
2125 	struct vfio_pci_reflck *reflck = container_of(kref,
2126 						      struct vfio_pci_reflck,
2127 						      kref);
2128 
2129 	kfree(reflck);
2130 	mutex_unlock(&reflck_lock);
2131 }
2132 
2133 static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck)
2134 {
2135 	kref_put_mutex(&reflck->kref, vfio_pci_reflck_release, &reflck_lock);
2136 }
2137 
2138 static int vfio_pci_get_unused_devs(struct pci_dev *pdev, void *data)
2139 {
2140 	struct vfio_devices *devs = data;
2141 	struct vfio_device *device;
2142 	struct vfio_pci_device *vdev;
2143 
2144 	if (devs->cur_index == devs->max_index)
2145 		return -ENOSPC;
2146 
2147 	device = vfio_device_get_from_dev(&pdev->dev);
2148 	if (!device)
2149 		return -EINVAL;
2150 
2151 	if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2152 		vfio_device_put(device);
2153 		return -EBUSY;
2154 	}
2155 
2156 	vdev = vfio_device_data(device);
2157 
2158 	/* Fault if the device is not unused */
2159 	if (vdev->refcnt) {
2160 		vfio_device_put(device);
2161 		return -EBUSY;
2162 	}
2163 
2164 	devs->devices[devs->cur_index++] = device;
2165 	return 0;
2166 }
2167 
2168 static int vfio_pci_try_zap_and_vma_lock_cb(struct pci_dev *pdev, void *data)
2169 {
2170 	struct vfio_devices *devs = data;
2171 	struct vfio_device *device;
2172 	struct vfio_pci_device *vdev;
2173 
2174 	if (devs->cur_index == devs->max_index)
2175 		return -ENOSPC;
2176 
2177 	device = vfio_device_get_from_dev(&pdev->dev);
2178 	if (!device)
2179 		return -EINVAL;
2180 
2181 	if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2182 		vfio_device_put(device);
2183 		return -EBUSY;
2184 	}
2185 
2186 	vdev = vfio_device_data(device);
2187 
2188 	/*
2189 	 * Locking multiple devices is prone to deadlock, runaway and
2190 	 * unwind if we hit contention.
2191 	 */
2192 	if (!vfio_pci_zap_and_vma_lock(vdev, true)) {
2193 		vfio_device_put(device);
2194 		return -EBUSY;
2195 	}
2196 
2197 	devs->devices[devs->cur_index++] = device;
2198 	return 0;
2199 }
2200 
2201 /*
2202  * If a bus or slot reset is available for the provided device and:
2203  *  - All of the devices affected by that bus or slot reset are unused
2204  *    (!refcnt)
2205  *  - At least one of the affected devices is marked dirty via
2206  *    needs_reset (such as by lack of FLR support)
2207  * Then attempt to perform that bus or slot reset.  Callers are required
2208  * to hold vdev->reflck->lock, protecting the bus/slot reset group from
2209  * concurrent opens.  A vfio_device reference is acquired for each device
2210  * to prevent unbinds during the reset operation.
2211  *
2212  * NB: vfio-core considers a group to be viable even if some devices are
2213  * bound to drivers like pci-stub or pcieport.  Here we require all devices
2214  * to be bound to vfio_pci since that's the only way we can be sure they
2215  * stay put.
2216  */
2217 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev)
2218 {
2219 	struct vfio_devices devs = { .cur_index = 0 };
2220 	int i = 0, ret = -EINVAL;
2221 	bool slot = false;
2222 	struct vfio_pci_device *tmp;
2223 
2224 	if (!pci_probe_reset_slot(vdev->pdev->slot))
2225 		slot = true;
2226 	else if (pci_probe_reset_bus(vdev->pdev->bus))
2227 		return;
2228 
2229 	if (vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
2230 					  &i, slot) || !i)
2231 		return;
2232 
2233 	devs.max_index = i;
2234 	devs.devices = kcalloc(i, sizeof(struct vfio_device *), GFP_KERNEL);
2235 	if (!devs.devices)
2236 		return;
2237 
2238 	if (vfio_pci_for_each_slot_or_bus(vdev->pdev,
2239 					  vfio_pci_get_unused_devs,
2240 					  &devs, slot))
2241 		goto put_devs;
2242 
2243 	/* Does at least one need a reset? */
2244 	for (i = 0; i < devs.cur_index; i++) {
2245 		tmp = vfio_device_data(devs.devices[i]);
2246 		if (tmp->needs_reset) {
2247 			ret = pci_reset_bus(vdev->pdev);
2248 			break;
2249 		}
2250 	}
2251 
2252 put_devs:
2253 	for (i = 0; i < devs.cur_index; i++) {
2254 		tmp = vfio_device_data(devs.devices[i]);
2255 
2256 		/*
2257 		 * If reset was successful, affected devices no longer need
2258 		 * a reset and we should return all the collateral devices
2259 		 * to low power.  If not successful, we either didn't reset
2260 		 * the bus or timed out waiting for it, so let's not touch
2261 		 * the power state.
2262 		 */
2263 		if (!ret) {
2264 			tmp->needs_reset = false;
2265 
2266 			if (tmp != vdev && !disable_idle_d3)
2267 				vfio_pci_set_power_state(tmp, PCI_D3hot);
2268 		}
2269 
2270 		vfio_device_put(devs.devices[i]);
2271 	}
2272 
2273 	kfree(devs.devices);
2274 }
2275 
2276 static void __exit vfio_pci_cleanup(void)
2277 {
2278 	pci_unregister_driver(&vfio_pci_driver);
2279 	vfio_pci_uninit_perm_bits();
2280 }
2281 
2282 static void __init vfio_pci_fill_ids(void)
2283 {
2284 	char *p, *id;
2285 	int rc;
2286 
2287 	/* no ids passed actually */
2288 	if (ids[0] == '\0')
2289 		return;
2290 
2291 	/* add ids specified in the module parameter */
2292 	p = ids;
2293 	while ((id = strsep(&p, ","))) {
2294 		unsigned int vendor, device, subvendor = PCI_ANY_ID,
2295 			subdevice = PCI_ANY_ID, class = 0, class_mask = 0;
2296 		int fields;
2297 
2298 		if (!strlen(id))
2299 			continue;
2300 
2301 		fields = sscanf(id, "%x:%x:%x:%x:%x:%x",
2302 				&vendor, &device, &subvendor, &subdevice,
2303 				&class, &class_mask);
2304 
2305 		if (fields < 2) {
2306 			pr_warn("invalid id string \"%s\"\n", id);
2307 			continue;
2308 		}
2309 
2310 		rc = pci_add_dynid(&vfio_pci_driver, vendor, device,
2311 				   subvendor, subdevice, class, class_mask, 0);
2312 		if (rc)
2313 			pr_warn("failed to add dynamic id [%04x:%04x[%04x:%04x]] class %#08x/%08x (%d)\n",
2314 				vendor, device, subvendor, subdevice,
2315 				class, class_mask, rc);
2316 		else
2317 			pr_info("add [%04x:%04x[%04x:%04x]] class %#08x/%08x\n",
2318 				vendor, device, subvendor, subdevice,
2319 				class, class_mask);
2320 	}
2321 }
2322 
2323 static int __init vfio_pci_init(void)
2324 {
2325 	int ret;
2326 
2327 	/* Allocate shared config space permision data used by all devices */
2328 	ret = vfio_pci_init_perm_bits();
2329 	if (ret)
2330 		return ret;
2331 
2332 	/* Register and scan for devices */
2333 	ret = pci_register_driver(&vfio_pci_driver);
2334 	if (ret)
2335 		goto out_driver;
2336 
2337 	vfio_pci_fill_ids();
2338 
2339 	return 0;
2340 
2341 out_driver:
2342 	vfio_pci_uninit_perm_bits();
2343 	return ret;
2344 }
2345 
2346 module_init(vfio_pci_init);
2347 module_exit(vfio_pci_cleanup);
2348 
2349 MODULE_VERSION(DRIVER_VERSION);
2350 MODULE_LICENSE("GPL v2");
2351 MODULE_AUTHOR(DRIVER_AUTHOR);
2352 MODULE_DESCRIPTION(DRIVER_DESC);
2353