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
37static char ids[1024] __initdata;
38module_param_string(ids, ids, sizeof(ids), 0);
39MODULE_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
41static bool nointxmask;
42module_param_named(nointxmask, nointxmask, bool, S_IRUGO | S_IWUSR);
43MODULE_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
47static bool disable_vga;
48module_param(disable_vga, bool, S_IRUGO);
49MODULE_PARM_DESC(disable_vga, "Disable VGA resource access through vfio-pci");
50#endif
51
52static bool disable_idle_d3;
53module_param(disable_idle_d3, bool, S_IRUGO | S_IWUSR);
54MODULE_PARM_DESC(disable_idle_d3,
55		 "Disable using the PCI D3 low power state for idle, unused devices");
56
57static bool enable_sriov;
58#ifdef CONFIG_PCI_IOV
59module_param(enable_sriov, bool, 0644);
60MODULE_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
63static bool disable_denylist;
64module_param(disable_denylist, bool, 0444);
65MODULE_PARM_DESC(disable_denylist, "Disable use of device denylist. Disabling the denylist allows binding to devices with known errata that may lead to exploitable stability or security issues when accessed by untrusted users.");
66
67static inline bool vfio_vga_disabled(void)
68{
69#ifdef CONFIG_VFIO_PCI_VGA
70	return disable_vga;
71#else
72	return true;
73#endif
74}
75
76static bool vfio_pci_dev_in_denylist(struct pci_dev *pdev)
77{
78	switch (pdev->vendor) {
79	case PCI_VENDOR_ID_INTEL:
80		switch (pdev->device) {
81		case PCI_DEVICE_ID_INTEL_QAT_C3XXX:
82		case PCI_DEVICE_ID_INTEL_QAT_C3XXX_VF:
83		case PCI_DEVICE_ID_INTEL_QAT_C62X:
84		case PCI_DEVICE_ID_INTEL_QAT_C62X_VF:
85		case PCI_DEVICE_ID_INTEL_QAT_DH895XCC:
86		case PCI_DEVICE_ID_INTEL_QAT_DH895XCC_VF:
87			return true;
88		default:
89			return false;
90		}
91	}
92
93	return false;
94}
95
96static bool vfio_pci_is_denylisted(struct pci_dev *pdev)
97{
98	if (!vfio_pci_dev_in_denylist(pdev))
99		return false;
100
101	if (disable_denylist) {
102		pci_warn(pdev,
103			 "device denylist disabled - allowing device %04x:%04x.\n",
104			 pdev->vendor, pdev->device);
105		return false;
106	}
107
108	pci_warn(pdev, "%04x:%04x exists in vfio-pci device denylist, driver probing disallowed.\n",
109		 pdev->vendor, pdev->device);
110
111	return true;
112}
113
114/*
115 * Our VGA arbiter participation is limited since we don't know anything
116 * about the device itself.  However, if the device is the only VGA device
117 * downstream of a bridge and VFIO VGA support is disabled, then we can
118 * safely return legacy VGA IO and memory as not decoded since the user
119 * has no way to get to it and routing can be disabled externally at the
120 * bridge.
121 */
122static unsigned int vfio_pci_set_vga_decode(void *opaque, bool single_vga)
123{
124	struct vfio_pci_device *vdev = opaque;
125	struct pci_dev *tmp = NULL, *pdev = vdev->pdev;
126	unsigned char max_busnr;
127	unsigned int decodes;
128
129	if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
130		return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
131		       VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
132
133	max_busnr = pci_bus_max_busnr(pdev->bus);
134	decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
135
136	while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
137		if (tmp == pdev ||
138		    pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
139		    pci_is_root_bus(tmp->bus))
140			continue;
141
142		if (tmp->bus->number >= pdev->bus->number &&
143		    tmp->bus->number <= max_busnr) {
144			pci_dev_put(tmp);
145			decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
146			break;
147		}
148	}
149
150	return decodes;
151}
152
153static inline bool vfio_pci_is_vga(struct pci_dev *pdev)
154{
155	return (pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA;
156}
157
158static void vfio_pci_probe_mmaps(struct vfio_pci_device *vdev)
159{
160	struct resource *res;
161	int i;
162	struct vfio_pci_dummy_resource *dummy_res;
163
164	for (i = 0; i < PCI_STD_NUM_BARS; i++) {
165		int bar = i + PCI_STD_RESOURCES;
166
167		res = &vdev->pdev->resource[bar];
168
169		if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
170			goto no_mmap;
171
172		if (!(res->flags & IORESOURCE_MEM))
173			goto no_mmap;
174
175		/*
176		 * The PCI core shouldn't set up a resource with a
177		 * type but zero size. But there may be bugs that
178		 * cause us to do that.
179		 */
180		if (!resource_size(res))
181			goto no_mmap;
182
183		if (resource_size(res) >= PAGE_SIZE) {
184			vdev->bar_mmap_supported[bar] = true;
185			continue;
186		}
187
188		if (!(res->start & ~PAGE_MASK)) {
189			/*
190			 * Add a dummy resource to reserve the remainder
191			 * of the exclusive page in case that hot-add
192			 * device's bar is assigned into it.
193			 */
194			dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
195			if (dummy_res == NULL)
196				goto no_mmap;
197
198			dummy_res->resource.name = "vfio sub-page reserved";
199			dummy_res->resource.start = res->end + 1;
200			dummy_res->resource.end = res->start + PAGE_SIZE - 1;
201			dummy_res->resource.flags = res->flags;
202			if (request_resource(res->parent,
203						&dummy_res->resource)) {
204				kfree(dummy_res);
205				goto no_mmap;
206			}
207			dummy_res->index = bar;
208			list_add(&dummy_res->res_next,
209					&vdev->dummy_resources_list);
210			vdev->bar_mmap_supported[bar] = true;
211			continue;
212		}
213		/*
214		 * Here we don't handle the case when the BAR is not page
215		 * aligned because we can't expect the BAR will be
216		 * assigned into the same location in a page in guest
217		 * when we passthrough the BAR. And it's hard to access
218		 * this BAR in userspace because we have no way to get
219		 * the BAR's location in a page.
220		 */
221no_mmap:
222		vdev->bar_mmap_supported[bar] = false;
223	}
224}
225
226static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev);
227static void vfio_pci_disable(struct vfio_pci_device *vdev);
228static int vfio_pci_try_zap_and_vma_lock_cb(struct pci_dev *pdev, void *data);
229
230/*
231 * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
232 * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
233 * If a device implements the former but not the latter we would typically
234 * expect broken_intx_masking be set and require an exclusive interrupt.
235 * However since we do have control of the device's ability to assert INTx,
236 * we can instead pretend that the device does not implement INTx, virtualizing
237 * the pin register to report zero and maintaining DisINTx set on the host.
238 */
239static bool vfio_pci_nointx(struct pci_dev *pdev)
240{
241	switch (pdev->vendor) {
242	case PCI_VENDOR_ID_INTEL:
243		switch (pdev->device) {
244		/* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
245		case 0x1572:
246		case 0x1574:
247		case 0x1580 ... 0x1581:
248		case 0x1583 ... 0x158b:
249		case 0x37d0 ... 0x37d2:
250		/* X550 */
251		case 0x1563:
252			return true;
253		default:
254			return false;
255		}
256	}
257
258	return false;
259}
260
261static void vfio_pci_probe_power_state(struct vfio_pci_device *vdev)
262{
263	struct pci_dev *pdev = vdev->pdev;
264	u16 pmcsr;
265
266	if (!pdev->pm_cap)
267		return;
268
269	pci_read_config_word(pdev, pdev->pm_cap + PCI_PM_CTRL, &pmcsr);
270
271	vdev->needs_pm_restore = !(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
272}
273
274/*
275 * pci_set_power_state() wrapper handling devices which perform a soft reset on
276 * D3->D0 transition.  Save state prior to D0/1/2->D3, stash it on the vdev,
277 * restore when returned to D0.  Saved separately from pci_saved_state for use
278 * by PM capability emulation and separately from pci_dev internal saved state
279 * to avoid it being overwritten and consumed around other resets.
280 */
281int vfio_pci_set_power_state(struct vfio_pci_device *vdev, pci_power_t state)
282{
283	struct pci_dev *pdev = vdev->pdev;
284	bool needs_restore = false, needs_save = false;
285	int ret;
286
287	if (vdev->needs_pm_restore) {
288		if (pdev->current_state < PCI_D3hot && state >= PCI_D3hot) {
289			pci_save_state(pdev);
290			needs_save = true;
291		}
292
293		if (pdev->current_state >= PCI_D3hot && state <= PCI_D0)
294			needs_restore = true;
295	}
296
297	ret = pci_set_power_state(pdev, state);
298
299	if (!ret) {
300		/* D3 might be unsupported via quirk, skip unless in D3 */
301		if (needs_save && pdev->current_state >= PCI_D3hot) {
302			vdev->pm_save = pci_store_saved_state(pdev);
303		} else if (needs_restore) {
304			pci_load_and_free_saved_state(pdev, &vdev->pm_save);
305			pci_restore_state(pdev);
306		}
307	}
308
309	return ret;
310}
311
312static int vfio_pci_enable(struct vfio_pci_device *vdev)
313{
314	struct pci_dev *pdev = vdev->pdev;
315	int ret;
316	u16 cmd;
317	u8 msix_pos;
318
319	vfio_pci_set_power_state(vdev, PCI_D0);
320
321	/* Don't allow our initial saved state to include busmaster */
322	pci_clear_master(pdev);
323
324	ret = pci_enable_device(pdev);
325	if (ret)
326		return ret;
327
328	/* If reset fails because of the device lock, fail this path entirely */
329	ret = pci_try_reset_function(pdev);
330	if (ret == -EAGAIN) {
331		pci_disable_device(pdev);
332		return ret;
333	}
334
335	vdev->reset_works = !ret;
336	pci_save_state(pdev);
337	vdev->pci_saved_state = pci_store_saved_state(pdev);
338	if (!vdev->pci_saved_state)
339		pci_dbg(pdev, "%s: Couldn't store saved state\n", __func__);
340
341	if (likely(!nointxmask)) {
342		if (vfio_pci_nointx(pdev)) {
343			pci_info(pdev, "Masking broken INTx support\n");
344			vdev->nointx = true;
345			pci_intx(pdev, 0);
346		} else
347			vdev->pci_2_3 = pci_intx_mask_supported(pdev);
348	}
349
350	pci_read_config_word(pdev, PCI_COMMAND, &cmd);
351	if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
352		cmd &= ~PCI_COMMAND_INTX_DISABLE;
353		pci_write_config_word(pdev, PCI_COMMAND, cmd);
354	}
355
356	ret = vfio_config_init(vdev);
357	if (ret) {
358		kfree(vdev->pci_saved_state);
359		vdev->pci_saved_state = NULL;
360		pci_disable_device(pdev);
361		return ret;
362	}
363
364	msix_pos = pdev->msix_cap;
365	if (msix_pos) {
366		u16 flags;
367		u32 table;
368
369		pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
370		pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
371
372		vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
373		vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
374		vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
375	} else
376		vdev->msix_bar = 0xFF;
377
378	if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
379		vdev->has_vga = true;
380
381
382	if (vfio_pci_is_vga(pdev) &&
383	    pdev->vendor == PCI_VENDOR_ID_INTEL &&
384	    IS_ENABLED(CONFIG_VFIO_PCI_IGD)) {
385		ret = vfio_pci_igd_init(vdev);
386		if (ret && ret != -ENODEV) {
387			pci_warn(pdev, "Failed to setup Intel IGD regions\n");
388			goto disable_exit;
389		}
390	}
391
392	if (pdev->vendor == PCI_VENDOR_ID_NVIDIA &&
393	    IS_ENABLED(CONFIG_VFIO_PCI_NVLINK2)) {
394		ret = vfio_pci_nvdia_v100_nvlink2_init(vdev);
395		if (ret && ret != -ENODEV) {
396			pci_warn(pdev, "Failed to setup NVIDIA NV2 RAM region\n");
397			goto disable_exit;
398		}
399	}
400
401	if (pdev->vendor == PCI_VENDOR_ID_IBM &&
402	    IS_ENABLED(CONFIG_VFIO_PCI_NVLINK2)) {
403		ret = vfio_pci_ibm_npu2_init(vdev);
404		if (ret && ret != -ENODEV) {
405			pci_warn(pdev, "Failed to setup NVIDIA NV2 ATSD region\n");
406			goto disable_exit;
407		}
408	}
409
410	vfio_pci_probe_mmaps(vdev);
411
412	return 0;
413
414disable_exit:
415	vfio_pci_disable(vdev);
416	return ret;
417}
418
419static void vfio_pci_disable(struct vfio_pci_device *vdev)
420{
421	struct pci_dev *pdev = vdev->pdev;
422	struct vfio_pci_dummy_resource *dummy_res, *tmp;
423	struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
424	int i, bar;
425
426	/* Stop the device from further DMA */
427	pci_clear_master(pdev);
428
429	vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
430				VFIO_IRQ_SET_ACTION_TRIGGER,
431				vdev->irq_type, 0, 0, NULL);
432
433	/* Device closed, don't need mutex here */
434	list_for_each_entry_safe(ioeventfd, ioeventfd_tmp,
435				 &vdev->ioeventfds_list, next) {
436		vfio_virqfd_disable(&ioeventfd->virqfd);
437		list_del(&ioeventfd->next);
438		kfree(ioeventfd);
439	}
440	vdev->ioeventfds_nr = 0;
441
442	vdev->virq_disabled = false;
443
444	for (i = 0; i < vdev->num_regions; i++)
445		vdev->region[i].ops->release(vdev, &vdev->region[i]);
446
447	vdev->num_regions = 0;
448	kfree(vdev->region);
449	vdev->region = NULL; /* don't krealloc a freed pointer */
450
451	vfio_config_free(vdev);
452
453	for (i = 0; i < PCI_STD_NUM_BARS; i++) {
454		bar = i + PCI_STD_RESOURCES;
455		if (!vdev->barmap[bar])
456			continue;
457		pci_iounmap(pdev, vdev->barmap[bar]);
458		pci_release_selected_regions(pdev, 1 << bar);
459		vdev->barmap[bar] = NULL;
460	}
461
462	list_for_each_entry_safe(dummy_res, tmp,
463				 &vdev->dummy_resources_list, res_next) {
464		list_del(&dummy_res->res_next);
465		release_resource(&dummy_res->resource);
466		kfree(dummy_res);
467	}
468
469	vdev->needs_reset = true;
470
471	/*
472	 * If we have saved state, restore it.  If we can reset the device,
473	 * even better.  Resetting with current state seems better than
474	 * nothing, but saving and restoring current state without reset
475	 * is just busy work.
476	 */
477	if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
478		pci_info(pdev, "%s: Couldn't reload saved state\n", __func__);
479
480		if (!vdev->reset_works)
481			goto out;
482
483		pci_save_state(pdev);
484	}
485
486	/*
487	 * Disable INTx and MSI, presumably to avoid spurious interrupts
488	 * during reset.  Stolen from pci_reset_function()
489	 */
490	pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
491
492	/*
493	 * Try to get the locks ourselves to prevent a deadlock. The
494	 * success of this is dependent on being able to lock the device,
495	 * which is not always possible.
496	 * We can not use the "try" reset interface here, which will
497	 * overwrite the previously restored configuration information.
498	 */
499	if (vdev->reset_works && pci_cfg_access_trylock(pdev)) {
500		if (device_trylock(&pdev->dev)) {
501			if (!__pci_reset_function_locked(pdev))
502				vdev->needs_reset = false;
503			device_unlock(&pdev->dev);
504		}
505		pci_cfg_access_unlock(pdev);
506	}
507
508	pci_restore_state(pdev);
509out:
510	pci_disable_device(pdev);
511
512	vfio_pci_try_bus_reset(vdev);
513
514	if (!disable_idle_d3)
515		vfio_pci_set_power_state(vdev, PCI_D3hot);
516}
517
518static struct pci_driver vfio_pci_driver;
519
520static struct vfio_pci_device *get_pf_vdev(struct vfio_pci_device *vdev,
521					   struct vfio_device **pf_dev)
522{
523	struct pci_dev *physfn = pci_physfn(vdev->pdev);
524
525	if (!vdev->pdev->is_virtfn)
526		return NULL;
527
528	*pf_dev = vfio_device_get_from_dev(&physfn->dev);
529	if (!*pf_dev)
530		return NULL;
531
532	if (pci_dev_driver(physfn) != &vfio_pci_driver) {
533		vfio_device_put(*pf_dev);
534		return NULL;
535	}
536
537	return vfio_device_data(*pf_dev);
538}
539
540static void vfio_pci_vf_token_user_add(struct vfio_pci_device *vdev, int val)
541{
542	struct vfio_device *pf_dev;
543	struct vfio_pci_device *pf_vdev = get_pf_vdev(vdev, &pf_dev);
544
545	if (!pf_vdev)
546		return;
547
548	mutex_lock(&pf_vdev->vf_token->lock);
549	pf_vdev->vf_token->users += val;
550	WARN_ON(pf_vdev->vf_token->users < 0);
551	mutex_unlock(&pf_vdev->vf_token->lock);
552
553	vfio_device_put(pf_dev);
554}
555
556static void vfio_pci_release(void *device_data)
557{
558	struct vfio_pci_device *vdev = device_data;
559
560	mutex_lock(&vdev->reflck->lock);
561
562	if (!(--vdev->refcnt)) {
563		vfio_pci_vf_token_user_add(vdev, -1);
564		vfio_spapr_pci_eeh_release(vdev->pdev);
565		vfio_pci_disable(vdev);
566
567		mutex_lock(&vdev->igate);
568		if (vdev->err_trigger) {
569			eventfd_ctx_put(vdev->err_trigger);
570			vdev->err_trigger = NULL;
571		}
572		if (vdev->req_trigger) {
573			eventfd_ctx_put(vdev->req_trigger);
574			vdev->req_trigger = NULL;
575		}
576		mutex_unlock(&vdev->igate);
577	}
578
579	mutex_unlock(&vdev->reflck->lock);
580
581	module_put(THIS_MODULE);
582}
583
584static int vfio_pci_open(void *device_data)
585{
586	struct vfio_pci_device *vdev = device_data;
587	int ret = 0;
588
589	if (!try_module_get(THIS_MODULE))
590		return -ENODEV;
591
592	mutex_lock(&vdev->reflck->lock);
593
594	if (!vdev->refcnt) {
595		ret = vfio_pci_enable(vdev);
596		if (ret)
597			goto error;
598
599		vfio_spapr_pci_eeh_open(vdev->pdev);
600		vfio_pci_vf_token_user_add(vdev, 1);
601	}
602	vdev->refcnt++;
603error:
604	mutex_unlock(&vdev->reflck->lock);
605	if (ret)
606		module_put(THIS_MODULE);
607	return ret;
608}
609
610static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
611{
612	if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
613		u8 pin;
614
615		if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
616		    vdev->nointx || vdev->pdev->is_virtfn)
617			return 0;
618
619		pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
620
621		return pin ? 1 : 0;
622	} else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
623		u8 pos;
624		u16 flags;
625
626		pos = vdev->pdev->msi_cap;
627		if (pos) {
628			pci_read_config_word(vdev->pdev,
629					     pos + PCI_MSI_FLAGS, &flags);
630			return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
631		}
632	} else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
633		u8 pos;
634		u16 flags;
635
636		pos = vdev->pdev->msix_cap;
637		if (pos) {
638			pci_read_config_word(vdev->pdev,
639					     pos + PCI_MSIX_FLAGS, &flags);
640
641			return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
642		}
643	} else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
644		if (pci_is_pcie(vdev->pdev))
645			return 1;
646	} else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
647		return 1;
648	}
649
650	return 0;
651}
652
653static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
654{
655	(*(int *)data)++;
656	return 0;
657}
658
659struct vfio_pci_fill_info {
660	int max;
661	int cur;
662	struct vfio_pci_dependent_device *devices;
663};
664
665static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
666{
667	struct vfio_pci_fill_info *fill = data;
668	struct iommu_group *iommu_group;
669
670	if (fill->cur == fill->max)
671		return -EAGAIN; /* Something changed, try again */
672
673	iommu_group = iommu_group_get(&pdev->dev);
674	if (!iommu_group)
675		return -EPERM; /* Cannot reset non-isolated devices */
676
677	fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
678	fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
679	fill->devices[fill->cur].bus = pdev->bus->number;
680	fill->devices[fill->cur].devfn = pdev->devfn;
681	fill->cur++;
682	iommu_group_put(iommu_group);
683	return 0;
684}
685
686struct vfio_pci_group_entry {
687	struct vfio_group *group;
688	int id;
689};
690
691struct vfio_pci_group_info {
692	int count;
693	struct vfio_pci_group_entry *groups;
694};
695
696static int vfio_pci_validate_devs(struct pci_dev *pdev, void *data)
697{
698	struct vfio_pci_group_info *info = data;
699	struct iommu_group *group;
700	int id, i;
701
702	group = iommu_group_get(&pdev->dev);
703	if (!group)
704		return -EPERM;
705
706	id = iommu_group_id(group);
707
708	for (i = 0; i < info->count; i++)
709		if (info->groups[i].id == id)
710			break;
711
712	iommu_group_put(group);
713
714	return (i == info->count) ? -EINVAL : 0;
715}
716
717static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
718{
719	for (; pdev; pdev = pdev->bus->self)
720		if (pdev->bus == slot->bus)
721			return (pdev->slot == slot);
722	return false;
723}
724
725struct vfio_pci_walk_info {
726	int (*fn)(struct pci_dev *, void *data);
727	void *data;
728	struct pci_dev *pdev;
729	bool slot;
730	int ret;
731};
732
733static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
734{
735	struct vfio_pci_walk_info *walk = data;
736
737	if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
738		walk->ret = walk->fn(pdev, walk->data);
739
740	return walk->ret;
741}
742
743static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
744					 int (*fn)(struct pci_dev *,
745						   void *data), void *data,
746					 bool slot)
747{
748	struct vfio_pci_walk_info walk = {
749		.fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
750	};
751
752	pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
753
754	return walk.ret;
755}
756
757static int msix_mmappable_cap(struct vfio_pci_device *vdev,
758			      struct vfio_info_cap *caps)
759{
760	struct vfio_info_cap_header header = {
761		.id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
762		.version = 1
763	};
764
765	return vfio_info_add_capability(caps, &header, sizeof(header));
766}
767
768int vfio_pci_register_dev_region(struct vfio_pci_device *vdev,
769				 unsigned int type, unsigned int subtype,
770				 const struct vfio_pci_regops *ops,
771				 size_t size, u32 flags, void *data)
772{
773	struct vfio_pci_region *region;
774
775	region = krealloc(vdev->region,
776			  (vdev->num_regions + 1) * sizeof(*region),
777			  GFP_KERNEL);
778	if (!region)
779		return -ENOMEM;
780
781	vdev->region = region;
782	vdev->region[vdev->num_regions].type = type;
783	vdev->region[vdev->num_regions].subtype = subtype;
784	vdev->region[vdev->num_regions].ops = ops;
785	vdev->region[vdev->num_regions].size = size;
786	vdev->region[vdev->num_regions].flags = flags;
787	vdev->region[vdev->num_regions].data = data;
788
789	vdev->num_regions++;
790
791	return 0;
792}
793
794struct vfio_devices {
795	struct vfio_device **devices;
796	int cur_index;
797	int max_index;
798};
799
800static long vfio_pci_ioctl(void *device_data,
801			   unsigned int cmd, unsigned long arg)
802{
803	struct vfio_pci_device *vdev = device_data;
804	unsigned long minsz;
805
806	if (cmd == VFIO_DEVICE_GET_INFO) {
807		struct vfio_device_info info;
808		struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
809		unsigned long capsz;
810
811		minsz = offsetofend(struct vfio_device_info, num_irqs);
812
813		/* For backward compatibility, cannot require this */
814		capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
815
816		if (copy_from_user(&info, (void __user *)arg, minsz))
817			return -EFAULT;
818
819		if (info.argsz < minsz)
820			return -EINVAL;
821
822		if (info.argsz >= capsz) {
823			minsz = capsz;
824			info.cap_offset = 0;
825		}
826
827		info.flags = VFIO_DEVICE_FLAGS_PCI;
828
829		if (vdev->reset_works)
830			info.flags |= VFIO_DEVICE_FLAGS_RESET;
831
832		info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
833		info.num_irqs = VFIO_PCI_NUM_IRQS;
834
835		if (IS_ENABLED(CONFIG_VFIO_PCI_ZDEV)) {
836			int ret = vfio_pci_info_zdev_add_caps(vdev, &caps);
837
838			if (ret && ret != -ENODEV) {
839				pci_warn(vdev->pdev, "Failed to setup zPCI info capabilities\n");
840				return ret;
841			}
842		}
843
844		if (caps.size) {
845			info.flags |= VFIO_DEVICE_FLAGS_CAPS;
846			if (info.argsz < sizeof(info) + caps.size) {
847				info.argsz = sizeof(info) + caps.size;
848			} else {
849				vfio_info_cap_shift(&caps, sizeof(info));
850				if (copy_to_user((void __user *)arg +
851						  sizeof(info), caps.buf,
852						  caps.size)) {
853					kfree(caps.buf);
854					return -EFAULT;
855				}
856				info.cap_offset = sizeof(info);
857			}
858
859			kfree(caps.buf);
860		}
861
862		return copy_to_user((void __user *)arg, &info, minsz) ?
863			-EFAULT : 0;
864
865	} else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
866		struct pci_dev *pdev = vdev->pdev;
867		struct vfio_region_info info;
868		struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
869		int i, ret;
870
871		minsz = offsetofend(struct vfio_region_info, offset);
872
873		if (copy_from_user(&info, (void __user *)arg, minsz))
874			return -EFAULT;
875
876		if (info.argsz < minsz)
877			return -EINVAL;
878
879		switch (info.index) {
880		case VFIO_PCI_CONFIG_REGION_INDEX:
881			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
882			info.size = pdev->cfg_size;
883			info.flags = VFIO_REGION_INFO_FLAG_READ |
884				     VFIO_REGION_INFO_FLAG_WRITE;
885			break;
886		case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
887			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
888			info.size = pci_resource_len(pdev, info.index);
889			if (!info.size) {
890				info.flags = 0;
891				break;
892			}
893
894			info.flags = VFIO_REGION_INFO_FLAG_READ |
895				     VFIO_REGION_INFO_FLAG_WRITE;
896			if (vdev->bar_mmap_supported[info.index]) {
897				info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
898				if (info.index == vdev->msix_bar) {
899					ret = msix_mmappable_cap(vdev, &caps);
900					if (ret)
901						return ret;
902				}
903			}
904
905			break;
906		case VFIO_PCI_ROM_REGION_INDEX:
907		{
908			void __iomem *io;
909			size_t size;
910			u16 cmd;
911
912			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
913			info.flags = 0;
914
915			/* Report the BAR size, not the ROM size */
916			info.size = pci_resource_len(pdev, info.index);
917			if (!info.size) {
918				/* Shadow ROMs appear as PCI option ROMs */
919				if (pdev->resource[PCI_ROM_RESOURCE].flags &
920							IORESOURCE_ROM_SHADOW)
921					info.size = 0x20000;
922				else
923					break;
924			}
925
926			/*
927			 * Is it really there?  Enable memory decode for
928			 * implicit access in pci_map_rom().
929			 */
930			cmd = vfio_pci_memory_lock_and_enable(vdev);
931			io = pci_map_rom(pdev, &size);
932			if (io) {
933				info.flags = VFIO_REGION_INFO_FLAG_READ;
934				pci_unmap_rom(pdev, io);
935			} else {
936				info.size = 0;
937			}
938			vfio_pci_memory_unlock_and_restore(vdev, cmd);
939
940			break;
941		}
942		case VFIO_PCI_VGA_REGION_INDEX:
943			if (!vdev->has_vga)
944				return -EINVAL;
945
946			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
947			info.size = 0xc0000;
948			info.flags = VFIO_REGION_INFO_FLAG_READ |
949				     VFIO_REGION_INFO_FLAG_WRITE;
950
951			break;
952		default:
953		{
954			struct vfio_region_info_cap_type cap_type = {
955					.header.id = VFIO_REGION_INFO_CAP_TYPE,
956					.header.version = 1 };
957
958			if (info.index >=
959			    VFIO_PCI_NUM_REGIONS + vdev->num_regions)
960				return -EINVAL;
961			info.index = array_index_nospec(info.index,
962							VFIO_PCI_NUM_REGIONS +
963							vdev->num_regions);
964
965			i = info.index - VFIO_PCI_NUM_REGIONS;
966
967			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
968			info.size = vdev->region[i].size;
969			info.flags = vdev->region[i].flags;
970
971			cap_type.type = vdev->region[i].type;
972			cap_type.subtype = vdev->region[i].subtype;
973
974			ret = vfio_info_add_capability(&caps, &cap_type.header,
975						       sizeof(cap_type));
976			if (ret)
977				return ret;
978
979			if (vdev->region[i].ops->add_capability) {
980				ret = vdev->region[i].ops->add_capability(vdev,
981						&vdev->region[i], &caps);
982				if (ret)
983					return ret;
984			}
985		}
986		}
987
988		if (caps.size) {
989			info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
990			if (info.argsz < sizeof(info) + caps.size) {
991				info.argsz = sizeof(info) + caps.size;
992				info.cap_offset = 0;
993			} else {
994				vfio_info_cap_shift(&caps, sizeof(info));
995				if (copy_to_user((void __user *)arg +
996						  sizeof(info), caps.buf,
997						  caps.size)) {
998					kfree(caps.buf);
999					return -EFAULT;
1000				}
1001				info.cap_offset = sizeof(info);
1002			}
1003
1004			kfree(caps.buf);
1005		}
1006
1007		return copy_to_user((void __user *)arg, &info, minsz) ?
1008			-EFAULT : 0;
1009
1010	} else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
1011		struct vfio_irq_info info;
1012
1013		minsz = offsetofend(struct vfio_irq_info, count);
1014
1015		if (copy_from_user(&info, (void __user *)arg, minsz))
1016			return -EFAULT;
1017
1018		if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
1019			return -EINVAL;
1020
1021		switch (info.index) {
1022		case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
1023		case VFIO_PCI_REQ_IRQ_INDEX:
1024			break;
1025		case VFIO_PCI_ERR_IRQ_INDEX:
1026			if (pci_is_pcie(vdev->pdev))
1027				break;
1028			fallthrough;
1029		default:
1030			return -EINVAL;
1031		}
1032
1033		info.flags = VFIO_IRQ_INFO_EVENTFD;
1034
1035		info.count = vfio_pci_get_irq_count(vdev, info.index);
1036
1037		if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
1038			info.flags |= (VFIO_IRQ_INFO_MASKABLE |
1039				       VFIO_IRQ_INFO_AUTOMASKED);
1040		else
1041			info.flags |= VFIO_IRQ_INFO_NORESIZE;
1042
1043		return copy_to_user((void __user *)arg, &info, minsz) ?
1044			-EFAULT : 0;
1045
1046	} else if (cmd == VFIO_DEVICE_SET_IRQS) {
1047		struct vfio_irq_set hdr;
1048		u8 *data = NULL;
1049		int max, ret = 0;
1050		size_t data_size = 0;
1051
1052		minsz = offsetofend(struct vfio_irq_set, count);
1053
1054		if (copy_from_user(&hdr, (void __user *)arg, minsz))
1055			return -EFAULT;
1056
1057		max = vfio_pci_get_irq_count(vdev, hdr.index);
1058
1059		ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
1060						 VFIO_PCI_NUM_IRQS, &data_size);
1061		if (ret)
1062			return ret;
1063
1064		if (data_size) {
1065			data = memdup_user((void __user *)(arg + minsz),
1066					    data_size);
1067			if (IS_ERR(data))
1068				return PTR_ERR(data);
1069		}
1070
1071		mutex_lock(&vdev->igate);
1072
1073		ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
1074					      hdr.start, hdr.count, data);
1075
1076		mutex_unlock(&vdev->igate);
1077		kfree(data);
1078
1079		return ret;
1080
1081	} else if (cmd == VFIO_DEVICE_RESET) {
1082		int ret;
1083
1084		if (!vdev->reset_works)
1085			return -EINVAL;
1086
1087		vfio_pci_zap_and_down_write_memory_lock(vdev);
1088		ret = pci_try_reset_function(vdev->pdev);
1089		up_write(&vdev->memory_lock);
1090
1091		return ret;
1092
1093	} else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
1094		struct vfio_pci_hot_reset_info hdr;
1095		struct vfio_pci_fill_info fill = { 0 };
1096		struct vfio_pci_dependent_device *devices = NULL;
1097		bool slot = false;
1098		int ret = 0;
1099
1100		minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
1101
1102		if (copy_from_user(&hdr, (void __user *)arg, minsz))
1103			return -EFAULT;
1104
1105		if (hdr.argsz < minsz)
1106			return -EINVAL;
1107
1108		hdr.flags = 0;
1109
1110		/* Can we do a slot or bus reset or neither? */
1111		if (!pci_probe_reset_slot(vdev->pdev->slot))
1112			slot = true;
1113		else if (pci_probe_reset_bus(vdev->pdev->bus))
1114			return -ENODEV;
1115
1116		/* How many devices are affected? */
1117		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1118						    vfio_pci_count_devs,
1119						    &fill.max, slot);
1120		if (ret)
1121			return ret;
1122
1123		WARN_ON(!fill.max); /* Should always be at least one */
1124
1125		/*
1126		 * If there's enough space, fill it now, otherwise return
1127		 * -ENOSPC and the number of devices affected.
1128		 */
1129		if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
1130			ret = -ENOSPC;
1131			hdr.count = fill.max;
1132			goto reset_info_exit;
1133		}
1134
1135		devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
1136		if (!devices)
1137			return -ENOMEM;
1138
1139		fill.devices = devices;
1140
1141		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1142						    vfio_pci_fill_devs,
1143						    &fill, slot);
1144
1145		/*
1146		 * If a device was removed between counting and filling,
1147		 * we may come up short of fill.max.  If a device was
1148		 * added, we'll have a return of -EAGAIN above.
1149		 */
1150		if (!ret)
1151			hdr.count = fill.cur;
1152
1153reset_info_exit:
1154		if (copy_to_user((void __user *)arg, &hdr, minsz))
1155			ret = -EFAULT;
1156
1157		if (!ret) {
1158			if (copy_to_user((void __user *)(arg + minsz), devices,
1159					 hdr.count * sizeof(*devices)))
1160				ret = -EFAULT;
1161		}
1162
1163		kfree(devices);
1164		return ret;
1165
1166	} else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
1167		struct vfio_pci_hot_reset hdr;
1168		int32_t *group_fds;
1169		struct vfio_pci_group_entry *groups;
1170		struct vfio_pci_group_info info;
1171		struct vfio_devices devs = { .cur_index = 0 };
1172		bool slot = false;
1173		int i, group_idx, mem_idx = 0, count = 0, ret = 0;
1174
1175		minsz = offsetofend(struct vfio_pci_hot_reset, count);
1176
1177		if (copy_from_user(&hdr, (void __user *)arg, minsz))
1178			return -EFAULT;
1179
1180		if (hdr.argsz < minsz || hdr.flags)
1181			return -EINVAL;
1182
1183		/* Can we do a slot or bus reset or neither? */
1184		if (!pci_probe_reset_slot(vdev->pdev->slot))
1185			slot = true;
1186		else if (pci_probe_reset_bus(vdev->pdev->bus))
1187			return -ENODEV;
1188
1189		/*
1190		 * We can't let userspace give us an arbitrarily large
1191		 * buffer to copy, so verify how many we think there
1192		 * could be.  Note groups can have multiple devices so
1193		 * one group per device is the max.
1194		 */
1195		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1196						    vfio_pci_count_devs,
1197						    &count, slot);
1198		if (ret)
1199			return ret;
1200
1201		/* Somewhere between 1 and count is OK */
1202		if (!hdr.count || hdr.count > count)
1203			return -EINVAL;
1204
1205		group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
1206		groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
1207		if (!group_fds || !groups) {
1208			kfree(group_fds);
1209			kfree(groups);
1210			return -ENOMEM;
1211		}
1212
1213		if (copy_from_user(group_fds, (void __user *)(arg + minsz),
1214				   hdr.count * sizeof(*group_fds))) {
1215			kfree(group_fds);
1216			kfree(groups);
1217			return -EFAULT;
1218		}
1219
1220		/*
1221		 * For each group_fd, get the group through the vfio external
1222		 * user interface and store the group and iommu ID.  This
1223		 * ensures the group is held across the reset.
1224		 */
1225		for (group_idx = 0; group_idx < hdr.count; group_idx++) {
1226			struct vfio_group *group;
1227			struct fd f = fdget(group_fds[group_idx]);
1228			if (!f.file) {
1229				ret = -EBADF;
1230				break;
1231			}
1232
1233			group = vfio_group_get_external_user(f.file);
1234			fdput(f);
1235			if (IS_ERR(group)) {
1236				ret = PTR_ERR(group);
1237				break;
1238			}
1239
1240			groups[group_idx].group = group;
1241			groups[group_idx].id =
1242					vfio_external_user_iommu_id(group);
1243		}
1244
1245		kfree(group_fds);
1246
1247		/* release reference to groups on error */
1248		if (ret)
1249			goto hot_reset_release;
1250
1251		info.count = hdr.count;
1252		info.groups = groups;
1253
1254		/*
1255		 * Test whether all the affected devices are contained
1256		 * by the set of groups provided by the user.
1257		 */
1258		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1259						    vfio_pci_validate_devs,
1260						    &info, slot);
1261		if (ret)
1262			goto hot_reset_release;
1263
1264		devs.max_index = count;
1265		devs.devices = kcalloc(count, sizeof(struct vfio_device *),
1266				       GFP_KERNEL);
1267		if (!devs.devices) {
1268			ret = -ENOMEM;
1269			goto hot_reset_release;
1270		}
1271
1272		/*
1273		 * We need to get memory_lock for each device, but devices
1274		 * can share mmap_lock, therefore we need to zap and hold
1275		 * the vma_lock for each device, and only then get each
1276		 * memory_lock.
1277		 */
1278		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1279					    vfio_pci_try_zap_and_vma_lock_cb,
1280					    &devs, slot);
1281		if (ret)
1282			goto hot_reset_release;
1283
1284		for (; mem_idx < devs.cur_index; mem_idx++) {
1285			struct vfio_pci_device *tmp;
1286
1287			tmp = vfio_device_data(devs.devices[mem_idx]);
1288
1289			ret = down_write_trylock(&tmp->memory_lock);
1290			if (!ret) {
1291				ret = -EBUSY;
1292				goto hot_reset_release;
1293			}
1294			mutex_unlock(&tmp->vma_lock);
1295		}
1296
1297		/* User has access, do the reset */
1298		ret = pci_reset_bus(vdev->pdev);
1299
1300hot_reset_release:
1301		for (i = 0; i < devs.cur_index; i++) {
1302			struct vfio_device *device;
1303			struct vfio_pci_device *tmp;
1304
1305			device = devs.devices[i];
1306			tmp = vfio_device_data(device);
1307
1308			if (i < mem_idx)
1309				up_write(&tmp->memory_lock);
1310			else
1311				mutex_unlock(&tmp->vma_lock);
1312			vfio_device_put(device);
1313		}
1314		kfree(devs.devices);
1315
1316		for (group_idx--; group_idx >= 0; group_idx--)
1317			vfio_group_put_external_user(groups[group_idx].group);
1318
1319		kfree(groups);
1320		return ret;
1321	} else if (cmd == VFIO_DEVICE_IOEVENTFD) {
1322		struct vfio_device_ioeventfd ioeventfd;
1323		int count;
1324
1325		minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1326
1327		if (copy_from_user(&ioeventfd, (void __user *)arg, minsz))
1328			return -EFAULT;
1329
1330		if (ioeventfd.argsz < minsz)
1331			return -EINVAL;
1332
1333		if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1334			return -EINVAL;
1335
1336		count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1337
1338		if (hweight8(count) != 1 || ioeventfd.fd < -1)
1339			return -EINVAL;
1340
1341		return vfio_pci_ioeventfd(vdev, ioeventfd.offset,
1342					  ioeventfd.data, count, ioeventfd.fd);
1343	} else if (cmd == VFIO_DEVICE_FEATURE) {
1344		struct vfio_device_feature feature;
1345		uuid_t uuid;
1346
1347		minsz = offsetofend(struct vfio_device_feature, flags);
1348
1349		if (copy_from_user(&feature, (void __user *)arg, minsz))
1350			return -EFAULT;
1351
1352		if (feature.argsz < minsz)
1353			return -EINVAL;
1354
1355		/* Check unknown flags */
1356		if (feature.flags & ~(VFIO_DEVICE_FEATURE_MASK |
1357				      VFIO_DEVICE_FEATURE_SET |
1358				      VFIO_DEVICE_FEATURE_GET |
1359				      VFIO_DEVICE_FEATURE_PROBE))
1360			return -EINVAL;
1361
1362		/* GET & SET are mutually exclusive except with PROBE */
1363		if (!(feature.flags & VFIO_DEVICE_FEATURE_PROBE) &&
1364		    (feature.flags & VFIO_DEVICE_FEATURE_SET) &&
1365		    (feature.flags & VFIO_DEVICE_FEATURE_GET))
1366			return -EINVAL;
1367
1368		switch (feature.flags & VFIO_DEVICE_FEATURE_MASK) {
1369		case VFIO_DEVICE_FEATURE_PCI_VF_TOKEN:
1370			if (!vdev->vf_token)
1371				return -ENOTTY;
1372
1373			/*
1374			 * We do not support GET of the VF Token UUID as this
1375			 * could expose the token of the previous device user.
1376			 */
1377			if (feature.flags & VFIO_DEVICE_FEATURE_GET)
1378				return -EINVAL;
1379
1380			if (feature.flags & VFIO_DEVICE_FEATURE_PROBE)
1381				return 0;
1382
1383			/* Don't SET unless told to do so */
1384			if (!(feature.flags & VFIO_DEVICE_FEATURE_SET))
1385				return -EINVAL;
1386
1387			if (feature.argsz < minsz + sizeof(uuid))
1388				return -EINVAL;
1389
1390			if (copy_from_user(&uuid, (void __user *)(arg + minsz),
1391					   sizeof(uuid)))
1392				return -EFAULT;
1393
1394			mutex_lock(&vdev->vf_token->lock);
1395			uuid_copy(&vdev->vf_token->uuid, &uuid);
1396			mutex_unlock(&vdev->vf_token->lock);
1397
1398			return 0;
1399		default:
1400			return -ENOTTY;
1401		}
1402	}
1403
1404	return -ENOTTY;
1405}
1406
1407static ssize_t vfio_pci_rw(void *device_data, char __user *buf,
1408			   size_t count, loff_t *ppos, bool iswrite)
1409{
1410	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1411	struct vfio_pci_device *vdev = device_data;
1412
1413	if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1414		return -EINVAL;
1415
1416	switch (index) {
1417	case VFIO_PCI_CONFIG_REGION_INDEX:
1418		return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1419
1420	case VFIO_PCI_ROM_REGION_INDEX:
1421		if (iswrite)
1422			return -EINVAL;
1423		return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1424
1425	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1426		return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1427
1428	case VFIO_PCI_VGA_REGION_INDEX:
1429		return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1430	default:
1431		index -= VFIO_PCI_NUM_REGIONS;
1432		return vdev->region[index].ops->rw(vdev, buf,
1433						   count, ppos, iswrite);
1434	}
1435
1436	return -EINVAL;
1437}
1438
1439static ssize_t vfio_pci_read(void *device_data, char __user *buf,
1440			     size_t count, loff_t *ppos)
1441{
1442	if (!count)
1443		return 0;
1444
1445	return vfio_pci_rw(device_data, buf, count, ppos, false);
1446}
1447
1448static ssize_t vfio_pci_write(void *device_data, const char __user *buf,
1449			      size_t count, loff_t *ppos)
1450{
1451	if (!count)
1452		return 0;
1453
1454	return vfio_pci_rw(device_data, (char __user *)buf, count, ppos, true);
1455}
1456
1457/* Return 1 on zap and vma_lock acquired, 0 on contention (only with @try) */
1458static int vfio_pci_zap_and_vma_lock(struct vfio_pci_device *vdev, bool try)
1459{
1460	struct vfio_pci_mmap_vma *mmap_vma, *tmp;
1461
1462	/*
1463	 * Lock ordering:
1464	 * vma_lock is nested under mmap_lock for vm_ops callback paths.
1465	 * The memory_lock semaphore is used by both code paths calling
1466	 * into this function to zap vmas and the vm_ops.fault callback
1467	 * to protect the memory enable state of the device.
1468	 *
1469	 * When zapping vmas we need to maintain the mmap_lock => vma_lock
1470	 * ordering, which requires using vma_lock to walk vma_list to
1471	 * acquire an mm, then dropping vma_lock to get the mmap_lock and
1472	 * reacquiring vma_lock.  This logic is derived from similar
1473	 * requirements in uverbs_user_mmap_disassociate().
1474	 *
1475	 * mmap_lock must always be the top-level lock when it is taken.
1476	 * Therefore we can only hold the memory_lock write lock when
1477	 * vma_list is empty, as we'd need to take mmap_lock to clear
1478	 * entries.  vma_list can only be guaranteed empty when holding
1479	 * vma_lock, thus memory_lock is nested under vma_lock.
1480	 *
1481	 * This enables the vm_ops.fault callback to acquire vma_lock,
1482	 * followed by memory_lock read lock, while already holding
1483	 * mmap_lock without risk of deadlock.
1484	 */
1485	while (1) {
1486		struct mm_struct *mm = NULL;
1487
1488		if (try) {
1489			if (!mutex_trylock(&vdev->vma_lock))
1490				return 0;
1491		} else {
1492			mutex_lock(&vdev->vma_lock);
1493		}
1494		while (!list_empty(&vdev->vma_list)) {
1495			mmap_vma = list_first_entry(&vdev->vma_list,
1496						    struct vfio_pci_mmap_vma,
1497						    vma_next);
1498			mm = mmap_vma->vma->vm_mm;
1499			if (mmget_not_zero(mm))
1500				break;
1501
1502			list_del(&mmap_vma->vma_next);
1503			kfree(mmap_vma);
1504			mm = NULL;
1505		}
1506		if (!mm)
1507			return 1;
1508		mutex_unlock(&vdev->vma_lock);
1509
1510		if (try) {
1511			if (!mmap_read_trylock(mm)) {
1512				mmput(mm);
1513				return 0;
1514			}
1515		} else {
1516			mmap_read_lock(mm);
1517		}
1518		if (try) {
1519			if (!mutex_trylock(&vdev->vma_lock)) {
1520				mmap_read_unlock(mm);
1521				mmput(mm);
1522				return 0;
1523			}
1524		} else {
1525			mutex_lock(&vdev->vma_lock);
1526		}
1527		list_for_each_entry_safe(mmap_vma, tmp,
1528					 &vdev->vma_list, vma_next) {
1529			struct vm_area_struct *vma = mmap_vma->vma;
1530
1531			if (vma->vm_mm != mm)
1532				continue;
1533
1534			list_del(&mmap_vma->vma_next);
1535			kfree(mmap_vma);
1536
1537			zap_vma_ptes(vma, vma->vm_start,
1538				     vma->vm_end - vma->vm_start);
1539		}
1540		mutex_unlock(&vdev->vma_lock);
1541		mmap_read_unlock(mm);
1542		mmput(mm);
1543	}
1544}
1545
1546void vfio_pci_zap_and_down_write_memory_lock(struct vfio_pci_device *vdev)
1547{
1548	vfio_pci_zap_and_vma_lock(vdev, false);
1549	down_write(&vdev->memory_lock);
1550	mutex_unlock(&vdev->vma_lock);
1551}
1552
1553u16 vfio_pci_memory_lock_and_enable(struct vfio_pci_device *vdev)
1554{
1555	u16 cmd;
1556
1557	down_write(&vdev->memory_lock);
1558	pci_read_config_word(vdev->pdev, PCI_COMMAND, &cmd);
1559	if (!(cmd & PCI_COMMAND_MEMORY))
1560		pci_write_config_word(vdev->pdev, PCI_COMMAND,
1561				      cmd | PCI_COMMAND_MEMORY);
1562
1563	return cmd;
1564}
1565
1566void vfio_pci_memory_unlock_and_restore(struct vfio_pci_device *vdev, u16 cmd)
1567{
1568	pci_write_config_word(vdev->pdev, PCI_COMMAND, cmd);
1569	up_write(&vdev->memory_lock);
1570}
1571
1572/* Caller holds vma_lock */
1573static int __vfio_pci_add_vma(struct vfio_pci_device *vdev,
1574			      struct vm_area_struct *vma)
1575{
1576	struct vfio_pci_mmap_vma *mmap_vma;
1577
1578	mmap_vma = kmalloc(sizeof(*mmap_vma), GFP_KERNEL);
1579	if (!mmap_vma)
1580		return -ENOMEM;
1581
1582	mmap_vma->vma = vma;
1583	list_add(&mmap_vma->vma_next, &vdev->vma_list);
1584
1585	return 0;
1586}
1587
1588/*
1589 * Zap mmaps on open so that we can fault them in on access and therefore
1590 * our vma_list only tracks mappings accessed since last zap.
1591 */
1592static void vfio_pci_mmap_open(struct vm_area_struct *vma)
1593{
1594	zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1595}
1596
1597static void vfio_pci_mmap_close(struct vm_area_struct *vma)
1598{
1599	struct vfio_pci_device *vdev = vma->vm_private_data;
1600	struct vfio_pci_mmap_vma *mmap_vma;
1601
1602	mutex_lock(&vdev->vma_lock);
1603	list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1604		if (mmap_vma->vma == vma) {
1605			list_del(&mmap_vma->vma_next);
1606			kfree(mmap_vma);
1607			break;
1608		}
1609	}
1610	mutex_unlock(&vdev->vma_lock);
1611}
1612
1613static vm_fault_t vfio_pci_mmap_fault(struct vm_fault *vmf)
1614{
1615	struct vm_area_struct *vma = vmf->vma;
1616	struct vfio_pci_device *vdev = vma->vm_private_data;
1617	struct vfio_pci_mmap_vma *mmap_vma;
1618	vm_fault_t ret = VM_FAULT_NOPAGE;
1619
1620	mutex_lock(&vdev->vma_lock);
1621	down_read(&vdev->memory_lock);
1622
1623	if (!__vfio_pci_memory_enabled(vdev)) {
1624		ret = VM_FAULT_SIGBUS;
1625		goto up_out;
1626	}
1627
1628	/*
1629	 * We populate the whole vma on fault, so we need to test whether
1630	 * the vma has already been mapped, such as for concurrent faults
1631	 * to the same vma.  io_remap_pfn_range() will trigger a BUG_ON if
1632	 * we ask it to fill the same range again.
1633	 */
1634	list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1635		if (mmap_vma->vma == vma)
1636			goto up_out;
1637	}
1638
1639	if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1640			       vma->vm_end - vma->vm_start,
1641			       vma->vm_page_prot)) {
1642		ret = VM_FAULT_SIGBUS;
1643		zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1644		goto up_out;
1645	}
1646
1647	if (__vfio_pci_add_vma(vdev, vma)) {
1648		ret = VM_FAULT_OOM;
1649		zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1650	}
1651
1652up_out:
1653	up_read(&vdev->memory_lock);
1654	mutex_unlock(&vdev->vma_lock);
1655	return ret;
1656}
1657
1658static const struct vm_operations_struct vfio_pci_mmap_ops = {
1659	.open = vfio_pci_mmap_open,
1660	.close = vfio_pci_mmap_close,
1661	.fault = vfio_pci_mmap_fault,
1662};
1663
1664static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma)
1665{
1666	struct vfio_pci_device *vdev = device_data;
1667	struct pci_dev *pdev = vdev->pdev;
1668	unsigned int index;
1669	u64 phys_len, req_len, pgoff, req_start;
1670	int ret;
1671
1672	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1673
1674	if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1675		return -EINVAL;
1676	if (vma->vm_end < vma->vm_start)
1677		return -EINVAL;
1678	if ((vma->vm_flags & VM_SHARED) == 0)
1679		return -EINVAL;
1680	if (index >= VFIO_PCI_NUM_REGIONS) {
1681		int regnum = index - VFIO_PCI_NUM_REGIONS;
1682		struct vfio_pci_region *region = vdev->region + regnum;
1683
1684		if (region->ops && region->ops->mmap &&
1685		    (region->flags & VFIO_REGION_INFO_FLAG_MMAP))
1686			return region->ops->mmap(vdev, region, vma);
1687		return -EINVAL;
1688	}
1689	if (index >= VFIO_PCI_ROM_REGION_INDEX)
1690		return -EINVAL;
1691	if (!vdev->bar_mmap_supported[index])
1692		return -EINVAL;
1693
1694	phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1695	req_len = vma->vm_end - vma->vm_start;
1696	pgoff = vma->vm_pgoff &
1697		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1698	req_start = pgoff << PAGE_SHIFT;
1699
1700	if (req_start + req_len > phys_len)
1701		return -EINVAL;
1702
1703	/*
1704	 * Even though we don't make use of the barmap for the mmap,
1705	 * we need to request the region and the barmap tracks that.
1706	 */
1707	if (!vdev->barmap[index]) {
1708		ret = pci_request_selected_regions(pdev,
1709						   1 << index, "vfio-pci");
1710		if (ret)
1711			return ret;
1712
1713		vdev->barmap[index] = pci_iomap(pdev, index, 0);
1714		if (!vdev->barmap[index]) {
1715			pci_release_selected_regions(pdev, 1 << index);
1716			return -ENOMEM;
1717		}
1718	}
1719
1720	vma->vm_private_data = vdev;
1721	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1722	vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1723
1724	/*
1725	 * See remap_pfn_range(), called from vfio_pci_fault() but we can't
1726	 * change vm_flags within the fault handler.  Set them now.
1727	 */
1728	vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1729	vma->vm_ops = &vfio_pci_mmap_ops;
1730
1731	return 0;
1732}
1733
1734static void vfio_pci_request(void *device_data, unsigned int count)
1735{
1736	struct vfio_pci_device *vdev = device_data;
1737	struct pci_dev *pdev = vdev->pdev;
1738
1739	mutex_lock(&vdev->igate);
1740
1741	if (vdev->req_trigger) {
1742		if (!(count % 10))
1743			pci_notice_ratelimited(pdev,
1744				"Relaying device request to user (#%u)\n",
1745				count);
1746		eventfd_signal(vdev->req_trigger, 1);
1747	} else if (count == 0) {
1748		pci_warn(pdev,
1749			"No device request channel registered, blocked until released by user\n");
1750	}
1751
1752	mutex_unlock(&vdev->igate);
1753}
1754
1755static int vfio_pci_validate_vf_token(struct vfio_pci_device *vdev,
1756				      bool vf_token, uuid_t *uuid)
1757{
1758	/*
1759	 * There's always some degree of trust or collaboration between SR-IOV
1760	 * PF and VFs, even if just that the PF hosts the SR-IOV capability and
1761	 * can disrupt VFs with a reset, but often the PF has more explicit
1762	 * access to deny service to the VF or access data passed through the
1763	 * VF.  We therefore require an opt-in via a shared VF token (UUID) to
1764	 * represent this trust.  This both prevents that a VF driver might
1765	 * assume the PF driver is a trusted, in-kernel driver, and also that
1766	 * a PF driver might be replaced with a rogue driver, unknown to in-use
1767	 * VF drivers.
1768	 *
1769	 * Therefore when presented with a VF, if the PF is a vfio device and
1770	 * it is bound to the vfio-pci driver, the user needs to provide a VF
1771	 * token to access the device, in the form of appending a vf_token to
1772	 * the device name, for example:
1773	 *
1774	 * "0000:04:10.0 vf_token=bd8d9d2b-5a5f-4f5a-a211-f591514ba1f3"
1775	 *
1776	 * When presented with a PF which has VFs in use, the user must also
1777	 * provide the current VF token to prove collaboration with existing
1778	 * VF users.  If VFs are not in use, the VF token provided for the PF
1779	 * device will act to set the VF token.
1780	 *
1781	 * If the VF token is provided but unused, an error is generated.
1782	 */
1783	if (!vdev->pdev->is_virtfn && !vdev->vf_token && !vf_token)
1784		return 0; /* No VF token provided or required */
1785
1786	if (vdev->pdev->is_virtfn) {
1787		struct vfio_device *pf_dev;
1788		struct vfio_pci_device *pf_vdev = get_pf_vdev(vdev, &pf_dev);
1789		bool match;
1790
1791		if (!pf_vdev) {
1792			if (!vf_token)
1793				return 0; /* PF is not vfio-pci, no VF token */
1794
1795			pci_info_ratelimited(vdev->pdev,
1796				"VF token incorrectly provided, PF not bound to vfio-pci\n");
1797			return -EINVAL;
1798		}
1799
1800		if (!vf_token) {
1801			vfio_device_put(pf_dev);
1802			pci_info_ratelimited(vdev->pdev,
1803				"VF token required to access device\n");
1804			return -EACCES;
1805		}
1806
1807		mutex_lock(&pf_vdev->vf_token->lock);
1808		match = uuid_equal(uuid, &pf_vdev->vf_token->uuid);
1809		mutex_unlock(&pf_vdev->vf_token->lock);
1810
1811		vfio_device_put(pf_dev);
1812
1813		if (!match) {
1814			pci_info_ratelimited(vdev->pdev,
1815				"Incorrect VF token provided for device\n");
1816			return -EACCES;
1817		}
1818	} else if (vdev->vf_token) {
1819		mutex_lock(&vdev->vf_token->lock);
1820		if (vdev->vf_token->users) {
1821			if (!vf_token) {
1822				mutex_unlock(&vdev->vf_token->lock);
1823				pci_info_ratelimited(vdev->pdev,
1824					"VF token required to access device\n");
1825				return -EACCES;
1826			}
1827
1828			if (!uuid_equal(uuid, &vdev->vf_token->uuid)) {
1829				mutex_unlock(&vdev->vf_token->lock);
1830				pci_info_ratelimited(vdev->pdev,
1831					"Incorrect VF token provided for device\n");
1832				return -EACCES;
1833			}
1834		} else if (vf_token) {
1835			uuid_copy(&vdev->vf_token->uuid, uuid);
1836		}
1837
1838		mutex_unlock(&vdev->vf_token->lock);
1839	} else if (vf_token) {
1840		pci_info_ratelimited(vdev->pdev,
1841			"VF token incorrectly provided, not a PF or VF\n");
1842		return -EINVAL;
1843	}
1844
1845	return 0;
1846}
1847
1848#define VF_TOKEN_ARG "vf_token="
1849
1850static int vfio_pci_match(void *device_data, char *buf)
1851{
1852	struct vfio_pci_device *vdev = device_data;
1853	bool vf_token = false;
1854	uuid_t uuid;
1855	int ret;
1856
1857	if (strncmp(pci_name(vdev->pdev), buf, strlen(pci_name(vdev->pdev))))
1858		return 0; /* No match */
1859
1860	if (strlen(buf) > strlen(pci_name(vdev->pdev))) {
1861		buf += strlen(pci_name(vdev->pdev));
1862
1863		if (*buf != ' ')
1864			return 0; /* No match: non-whitespace after name */
1865
1866		while (*buf) {
1867			if (*buf == ' ') {
1868				buf++;
1869				continue;
1870			}
1871
1872			if (!vf_token && !strncmp(buf, VF_TOKEN_ARG,
1873						  strlen(VF_TOKEN_ARG))) {
1874				buf += strlen(VF_TOKEN_ARG);
1875
1876				if (strlen(buf) < UUID_STRING_LEN)
1877					return -EINVAL;
1878
1879				ret = uuid_parse(buf, &uuid);
1880				if (ret)
1881					return ret;
1882
1883				vf_token = true;
1884				buf += UUID_STRING_LEN;
1885			} else {
1886				/* Unknown/duplicate option */
1887				return -EINVAL;
1888			}
1889		}
1890	}
1891
1892	ret = vfio_pci_validate_vf_token(vdev, vf_token, &uuid);
1893	if (ret)
1894		return ret;
1895
1896	return 1; /* Match */
1897}
1898
1899static const struct vfio_device_ops vfio_pci_ops = {
1900	.name		= "vfio-pci",
1901	.open		= vfio_pci_open,
1902	.release	= vfio_pci_release,
1903	.ioctl		= vfio_pci_ioctl,
1904	.read		= vfio_pci_read,
1905	.write		= vfio_pci_write,
1906	.mmap		= vfio_pci_mmap,
1907	.request	= vfio_pci_request,
1908	.match		= vfio_pci_match,
1909};
1910
1911static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev);
1912static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck);
1913
1914static int vfio_pci_bus_notifier(struct notifier_block *nb,
1915				 unsigned long action, void *data)
1916{
1917	struct vfio_pci_device *vdev = container_of(nb,
1918						    struct vfio_pci_device, nb);
1919	struct device *dev = data;
1920	struct pci_dev *pdev = to_pci_dev(dev);
1921	struct pci_dev *physfn = pci_physfn(pdev);
1922
1923	if (action == BUS_NOTIFY_ADD_DEVICE &&
1924	    pdev->is_virtfn && physfn == vdev->pdev) {
1925		pci_info(vdev->pdev, "Captured SR-IOV VF %s driver_override\n",
1926			 pci_name(pdev));
1927		pdev->driver_override = kasprintf(GFP_KERNEL, "%s",
1928						  vfio_pci_ops.name);
1929	} else if (action == BUS_NOTIFY_BOUND_DRIVER &&
1930		   pdev->is_virtfn && physfn == vdev->pdev) {
1931		struct pci_driver *drv = pci_dev_driver(pdev);
1932
1933		if (drv && drv != &vfio_pci_driver)
1934			pci_warn(vdev->pdev,
1935				 "VF %s bound to driver %s while PF bound to vfio-pci\n",
1936				 pci_name(pdev), drv->name);
1937	}
1938
1939	return 0;
1940}
1941
1942static int vfio_pci_vf_init(struct vfio_pci_device *vdev)
1943{
1944	struct pci_dev *pdev = vdev->pdev;
1945	int ret;
1946
1947	if (!pdev->is_physfn)
1948		return 0;
1949
1950	vdev->vf_token = kzalloc(sizeof(*vdev->vf_token), GFP_KERNEL);
1951	if (!vdev->vf_token)
1952		return -ENOMEM;
1953
1954	mutex_init(&vdev->vf_token->lock);
1955	uuid_gen(&vdev->vf_token->uuid);
1956
1957	vdev->nb.notifier_call = vfio_pci_bus_notifier;
1958	ret = bus_register_notifier(&pci_bus_type, &vdev->nb);
1959	if (ret) {
1960		kfree(vdev->vf_token);
1961		return ret;
1962	}
1963	return 0;
1964}
1965
1966static void vfio_pci_vf_uninit(struct vfio_pci_device *vdev)
1967{
1968	if (!vdev->vf_token)
1969		return;
1970
1971	bus_unregister_notifier(&pci_bus_type, &vdev->nb);
1972	WARN_ON(vdev->vf_token->users);
1973	mutex_destroy(&vdev->vf_token->lock);
1974	kfree(vdev->vf_token);
1975}
1976
1977static int vfio_pci_vga_init(struct vfio_pci_device *vdev)
1978{
1979	struct pci_dev *pdev = vdev->pdev;
1980	int ret;
1981
1982	if (!vfio_pci_is_vga(pdev))
1983		return 0;
1984
1985	ret = vga_client_register(pdev, vdev, NULL, vfio_pci_set_vga_decode);
1986	if (ret)
1987		return ret;
1988	vga_set_legacy_decoding(pdev, vfio_pci_set_vga_decode(vdev, false));
1989	return 0;
1990}
1991
1992static void vfio_pci_vga_uninit(struct vfio_pci_device *vdev)
1993{
1994	struct pci_dev *pdev = vdev->pdev;
1995
1996	if (!vfio_pci_is_vga(pdev))
1997		return;
1998	vga_client_register(pdev, NULL, NULL, NULL);
1999	vga_set_legacy_decoding(pdev, VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
2000					      VGA_RSRC_LEGACY_IO |
2001					      VGA_RSRC_LEGACY_MEM);
2002}
2003
2004static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2005{
2006	struct vfio_pci_device *vdev;
2007	struct iommu_group *group;
2008	int ret;
2009
2010	if (vfio_pci_is_denylisted(pdev))
2011		return -EINVAL;
2012
2013	if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
2014		return -EINVAL;
2015
2016	/*
2017	 * Prevent binding to PFs with VFs enabled, the VFs might be in use
2018	 * by the host or other users.  We cannot capture the VFs if they
2019	 * already exist, nor can we track VF users.  Disabling SR-IOV here
2020	 * would initiate removing the VFs, which would unbind the driver,
2021	 * which is prone to blocking if that VF is also in use by vfio-pci.
2022	 * Just reject these PFs and let the user sort it out.
2023	 */
2024	if (pci_num_vf(pdev)) {
2025		pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
2026		return -EBUSY;
2027	}
2028
2029	group = vfio_iommu_group_get(&pdev->dev);
2030	if (!group)
2031		return -EINVAL;
2032
2033	vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
2034	if (!vdev) {
2035		ret = -ENOMEM;
2036		goto out_group_put;
2037	}
2038
2039	vdev->pdev = pdev;
2040	vdev->irq_type = VFIO_PCI_NUM_IRQS;
2041	mutex_init(&vdev->igate);
2042	spin_lock_init(&vdev->irqlock);
2043	mutex_init(&vdev->ioeventfds_lock);
2044	INIT_LIST_HEAD(&vdev->dummy_resources_list);
2045	INIT_LIST_HEAD(&vdev->ioeventfds_list);
2046	mutex_init(&vdev->vma_lock);
2047	INIT_LIST_HEAD(&vdev->vma_list);
2048	init_rwsem(&vdev->memory_lock);
2049
2050	ret = vfio_pci_reflck_attach(vdev);
2051	if (ret)
2052		goto out_free;
2053	ret = vfio_pci_vf_init(vdev);
2054	if (ret)
2055		goto out_reflck;
2056	ret = vfio_pci_vga_init(vdev);
2057	if (ret)
2058		goto out_vf;
2059
2060	vfio_pci_probe_power_state(vdev);
2061
2062	if (!disable_idle_d3) {
2063		/*
2064		 * pci-core sets the device power state to an unknown value at
2065		 * bootup and after being removed from a driver.  The only
2066		 * transition it allows from this unknown state is to D0, which
2067		 * typically happens when a driver calls pci_enable_device().
2068		 * We're not ready to enable the device yet, but we do want to
2069		 * be able to get to D3.  Therefore first do a D0 transition
2070		 * before going to D3.
2071		 */
2072		vfio_pci_set_power_state(vdev, PCI_D0);
2073		vfio_pci_set_power_state(vdev, PCI_D3hot);
2074	}
2075
2076	ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
2077	if (ret)
2078		goto out_power;
2079	return 0;
2080
2081out_power:
2082	if (!disable_idle_d3)
2083		vfio_pci_set_power_state(vdev, PCI_D0);
2084out_vf:
2085	vfio_pci_vf_uninit(vdev);
2086out_reflck:
2087	vfio_pci_reflck_put(vdev->reflck);
2088out_free:
2089	kfree(vdev->pm_save);
2090	kfree(vdev);
2091out_group_put:
2092	vfio_iommu_group_put(group, &pdev->dev);
2093	return ret;
2094}
2095
2096static void vfio_pci_remove(struct pci_dev *pdev)
2097{
2098	struct vfio_pci_device *vdev;
2099
2100	pci_disable_sriov(pdev);
2101
2102	vdev = vfio_del_group_dev(&pdev->dev);
2103	if (!vdev)
2104		return;
2105
2106	vfio_pci_vf_uninit(vdev);
2107	vfio_pci_reflck_put(vdev->reflck);
2108	vfio_pci_vga_uninit(vdev);
2109
2110	vfio_iommu_group_put(pdev->dev.iommu_group, &pdev->dev);
2111
2112	if (!disable_idle_d3)
2113		vfio_pci_set_power_state(vdev, PCI_D0);
2114
2115	mutex_destroy(&vdev->ioeventfds_lock);
2116	kfree(vdev->region);
2117	kfree(vdev->pm_save);
2118	kfree(vdev);
2119}
2120
2121static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
2122						  pci_channel_state_t state)
2123{
2124	struct vfio_pci_device *vdev;
2125	struct vfio_device *device;
2126
2127	device = vfio_device_get_from_dev(&pdev->dev);
2128	if (device == NULL)
2129		return PCI_ERS_RESULT_DISCONNECT;
2130
2131	vdev = vfio_device_data(device);
2132	if (vdev == NULL) {
2133		vfio_device_put(device);
2134		return PCI_ERS_RESULT_DISCONNECT;
2135	}
2136
2137	mutex_lock(&vdev->igate);
2138
2139	if (vdev->err_trigger)
2140		eventfd_signal(vdev->err_trigger, 1);
2141
2142	mutex_unlock(&vdev->igate);
2143
2144	vfio_device_put(device);
2145
2146	return PCI_ERS_RESULT_CAN_RECOVER;
2147}
2148
2149static int vfio_pci_sriov_configure(struct pci_dev *pdev, int nr_virtfn)
2150{
2151	struct vfio_pci_device *vdev;
2152	struct vfio_device *device;
2153	int ret = 0;
2154
2155	might_sleep();
2156
2157	if (!enable_sriov)
2158		return -ENOENT;
2159
2160	device = vfio_device_get_from_dev(&pdev->dev);
2161	if (!device)
2162		return -ENODEV;
2163
2164	vdev = vfio_device_data(device);
2165	if (!vdev) {
2166		vfio_device_put(device);
2167		return -ENODEV;
2168	}
2169
2170	if (nr_virtfn == 0)
2171		pci_disable_sriov(pdev);
2172	else
2173		ret = pci_enable_sriov(pdev, nr_virtfn);
2174
2175	vfio_device_put(device);
2176
2177	return ret < 0 ? ret : nr_virtfn;
2178}
2179
2180static const struct pci_error_handlers vfio_err_handlers = {
2181	.error_detected = vfio_pci_aer_err_detected,
2182};
2183
2184static struct pci_driver vfio_pci_driver = {
2185	.name			= "vfio-pci",
2186	.id_table		= NULL, /* only dynamic ids */
2187	.probe			= vfio_pci_probe,
2188	.remove			= vfio_pci_remove,
2189	.sriov_configure	= vfio_pci_sriov_configure,
2190	.err_handler		= &vfio_err_handlers,
2191};
2192
2193static DEFINE_MUTEX(reflck_lock);
2194
2195static struct vfio_pci_reflck *vfio_pci_reflck_alloc(void)
2196{
2197	struct vfio_pci_reflck *reflck;
2198
2199	reflck = kzalloc(sizeof(*reflck), GFP_KERNEL);
2200	if (!reflck)
2201		return ERR_PTR(-ENOMEM);
2202
2203	kref_init(&reflck->kref);
2204	mutex_init(&reflck->lock);
2205
2206	return reflck;
2207}
2208
2209static void vfio_pci_reflck_get(struct vfio_pci_reflck *reflck)
2210{
2211	kref_get(&reflck->kref);
2212}
2213
2214static int vfio_pci_reflck_find(struct pci_dev *pdev, void *data)
2215{
2216	struct vfio_pci_reflck **preflck = data;
2217	struct vfio_device *device;
2218	struct vfio_pci_device *vdev;
2219
2220	device = vfio_device_get_from_dev(&pdev->dev);
2221	if (!device)
2222		return 0;
2223
2224	if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2225		vfio_device_put(device);
2226		return 0;
2227	}
2228
2229	vdev = vfio_device_data(device);
2230
2231	if (vdev->reflck) {
2232		vfio_pci_reflck_get(vdev->reflck);
2233		*preflck = vdev->reflck;
2234		vfio_device_put(device);
2235		return 1;
2236	}
2237
2238	vfio_device_put(device);
2239	return 0;
2240}
2241
2242static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev)
2243{
2244	bool slot = !pci_probe_reset_slot(vdev->pdev->slot);
2245
2246	mutex_lock(&reflck_lock);
2247
2248	if (pci_is_root_bus(vdev->pdev->bus) ||
2249	    vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_reflck_find,
2250					  &vdev->reflck, slot) <= 0)
2251		vdev->reflck = vfio_pci_reflck_alloc();
2252
2253	mutex_unlock(&reflck_lock);
2254
2255	return PTR_ERR_OR_ZERO(vdev->reflck);
2256}
2257
2258static void vfio_pci_reflck_release(struct kref *kref)
2259{
2260	struct vfio_pci_reflck *reflck = container_of(kref,
2261						      struct vfio_pci_reflck,
2262						      kref);
2263
2264	kfree(reflck);
2265	mutex_unlock(&reflck_lock);
2266}
2267
2268static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck)
2269{
2270	kref_put_mutex(&reflck->kref, vfio_pci_reflck_release, &reflck_lock);
2271}
2272
2273static int vfio_pci_get_unused_devs(struct pci_dev *pdev, void *data)
2274{
2275	struct vfio_devices *devs = data;
2276	struct vfio_device *device;
2277	struct vfio_pci_device *vdev;
2278
2279	if (devs->cur_index == devs->max_index)
2280		return -ENOSPC;
2281
2282	device = vfio_device_get_from_dev(&pdev->dev);
2283	if (!device)
2284		return -EINVAL;
2285
2286	if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2287		vfio_device_put(device);
2288		return -EBUSY;
2289	}
2290
2291	vdev = vfio_device_data(device);
2292
2293	/* Fault if the device is not unused */
2294	if (vdev->refcnt) {
2295		vfio_device_put(device);
2296		return -EBUSY;
2297	}
2298
2299	devs->devices[devs->cur_index++] = device;
2300	return 0;
2301}
2302
2303static int vfio_pci_try_zap_and_vma_lock_cb(struct pci_dev *pdev, void *data)
2304{
2305	struct vfio_devices *devs = data;
2306	struct vfio_device *device;
2307	struct vfio_pci_device *vdev;
2308
2309	if (devs->cur_index == devs->max_index)
2310		return -ENOSPC;
2311
2312	device = vfio_device_get_from_dev(&pdev->dev);
2313	if (!device)
2314		return -EINVAL;
2315
2316	if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2317		vfio_device_put(device);
2318		return -EBUSY;
2319	}
2320
2321	vdev = vfio_device_data(device);
2322
2323	/*
2324	 * Locking multiple devices is prone to deadlock, runaway and
2325	 * unwind if we hit contention.
2326	 */
2327	if (!vfio_pci_zap_and_vma_lock(vdev, true)) {
2328		vfio_device_put(device);
2329		return -EBUSY;
2330	}
2331
2332	devs->devices[devs->cur_index++] = device;
2333	return 0;
2334}
2335
2336/*
2337 * If a bus or slot reset is available for the provided device and:
2338 *  - All of the devices affected by that bus or slot reset are unused
2339 *    (!refcnt)
2340 *  - At least one of the affected devices is marked dirty via
2341 *    needs_reset (such as by lack of FLR support)
2342 * Then attempt to perform that bus or slot reset.  Callers are required
2343 * to hold vdev->reflck->lock, protecting the bus/slot reset group from
2344 * concurrent opens.  A vfio_device reference is acquired for each device
2345 * to prevent unbinds during the reset operation.
2346 *
2347 * NB: vfio-core considers a group to be viable even if some devices are
2348 * bound to drivers like pci-stub or pcieport.  Here we require all devices
2349 * to be bound to vfio_pci since that's the only way we can be sure they
2350 * stay put.
2351 */
2352static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev)
2353{
2354	struct vfio_devices devs = { .cur_index = 0 };
2355	int i = 0, ret = -EINVAL;
2356	bool slot = false;
2357	struct vfio_pci_device *tmp;
2358
2359	if (!pci_probe_reset_slot(vdev->pdev->slot))
2360		slot = true;
2361	else if (pci_probe_reset_bus(vdev->pdev->bus))
2362		return;
2363
2364	if (vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
2365					  &i, slot) || !i)
2366		return;
2367
2368	devs.max_index = i;
2369	devs.devices = kcalloc(i, sizeof(struct vfio_device *), GFP_KERNEL);
2370	if (!devs.devices)
2371		return;
2372
2373	if (vfio_pci_for_each_slot_or_bus(vdev->pdev,
2374					  vfio_pci_get_unused_devs,
2375					  &devs, slot))
2376		goto put_devs;
2377
2378	/* Does at least one need a reset? */
2379	for (i = 0; i < devs.cur_index; i++) {
2380		tmp = vfio_device_data(devs.devices[i]);
2381		if (tmp->needs_reset) {
2382			ret = pci_reset_bus(vdev->pdev);
2383			break;
2384		}
2385	}
2386
2387put_devs:
2388	for (i = 0; i < devs.cur_index; i++) {
2389		tmp = vfio_device_data(devs.devices[i]);
2390
2391		/*
2392		 * If reset was successful, affected devices no longer need
2393		 * a reset and we should return all the collateral devices
2394		 * to low power.  If not successful, we either didn't reset
2395		 * the bus or timed out waiting for it, so let's not touch
2396		 * the power state.
2397		 */
2398		if (!ret) {
2399			tmp->needs_reset = false;
2400
2401			if (tmp != vdev && !disable_idle_d3)
2402				vfio_pci_set_power_state(tmp, PCI_D3hot);
2403		}
2404
2405		vfio_device_put(devs.devices[i]);
2406	}
2407
2408	kfree(devs.devices);
2409}
2410
2411static void __exit vfio_pci_cleanup(void)
2412{
2413	pci_unregister_driver(&vfio_pci_driver);
2414	vfio_pci_uninit_perm_bits();
2415}
2416
2417static void __init vfio_pci_fill_ids(void)
2418{
2419	char *p, *id;
2420	int rc;
2421
2422	/* no ids passed actually */
2423	if (ids[0] == '\0')
2424		return;
2425
2426	/* add ids specified in the module parameter */
2427	p = ids;
2428	while ((id = strsep(&p, ","))) {
2429		unsigned int vendor, device, subvendor = PCI_ANY_ID,
2430			subdevice = PCI_ANY_ID, class = 0, class_mask = 0;
2431		int fields;
2432
2433		if (!strlen(id))
2434			continue;
2435
2436		fields = sscanf(id, "%x:%x:%x:%x:%x:%x",
2437				&vendor, &device, &subvendor, &subdevice,
2438				&class, &class_mask);
2439
2440		if (fields < 2) {
2441			pr_warn("invalid id string \"%s\"\n", id);
2442			continue;
2443		}
2444
2445		rc = pci_add_dynid(&vfio_pci_driver, vendor, device,
2446				   subvendor, subdevice, class, class_mask, 0);
2447		if (rc)
2448			pr_warn("failed to add dynamic id [%04x:%04x[%04x:%04x]] class %#08x/%08x (%d)\n",
2449				vendor, device, subvendor, subdevice,
2450				class, class_mask, rc);
2451		else
2452			pr_info("add [%04x:%04x[%04x:%04x]] class %#08x/%08x\n",
2453				vendor, device, subvendor, subdevice,
2454				class, class_mask);
2455	}
2456}
2457
2458static int __init vfio_pci_init(void)
2459{
2460	int ret;
2461
2462	/* Allocate shared config space permision data used by all devices */
2463	ret = vfio_pci_init_perm_bits();
2464	if (ret)
2465		return ret;
2466
2467	/* Register and scan for devices */
2468	ret = pci_register_driver(&vfio_pci_driver);
2469	if (ret)
2470		goto out_driver;
2471
2472	vfio_pci_fill_ids();
2473
2474	if (disable_denylist)
2475		pr_warn("device denylist disabled.\n");
2476
2477	return 0;
2478
2479out_driver:
2480	vfio_pci_uninit_perm_bits();
2481	return ret;
2482}
2483
2484module_init(vfio_pci_init);
2485module_exit(vfio_pci_cleanup);
2486
2487MODULE_VERSION(DRIVER_VERSION);
2488MODULE_LICENSE("GPL v2");
2489MODULE_AUTHOR(DRIVER_AUTHOR);
2490MODULE_DESCRIPTION(DRIVER_DESC);
2491