• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
4  *     Author: Alex Williamson <alex.williamson@redhat.com>
5  *
6  * Derived from original vfio:
7  * Copyright 2010 Cisco Systems, Inc.  All rights reserved.
8  * Author: Tom Lyon, pugs@cisco.com
9  */
10 
11 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12 
13 #include <linux/device.h>
14 #include <linux/eventfd.h>
15 #include <linux/file.h>
16 #include <linux/interrupt.h>
17 #include <linux/iommu.h>
18 #include <linux/module.h>
19 #include <linux/mutex.h>
20 #include <linux/notifier.h>
21 #include <linux/pci.h>
22 #include <linux/pm_runtime.h>
23 #include <linux/slab.h>
24 #include <linux/types.h>
25 #include <linux/uaccess.h>
26 #include <linux/vgaarb.h>
27 #include <linux/nospec.h>
28 #include <linux/sched/mm.h>
29 
30 #include <linux/vfio_pci_core.h>
31 
32 #define DRIVER_AUTHOR   "Alex Williamson <alex.williamson@redhat.com>"
33 #define DRIVER_DESC "core driver for VFIO based PCI devices"
34 
35 static bool nointxmask;
36 static bool disable_vga;
37 static bool disable_idle_d3;
38 
39 /* List of PF's that vfio_pci_core_sriov_configure() has been called on */
40 static DEFINE_MUTEX(vfio_pci_sriov_pfs_mutex);
41 static LIST_HEAD(vfio_pci_sriov_pfs);
42 
vfio_vga_disabled(void)43 static inline bool vfio_vga_disabled(void)
44 {
45 #ifdef CONFIG_VFIO_PCI_VGA
46 	return disable_vga;
47 #else
48 	return true;
49 #endif
50 }
51 
52 /*
53  * Our VGA arbiter participation is limited since we don't know anything
54  * about the device itself.  However, if the device is the only VGA device
55  * downstream of a bridge and VFIO VGA support is disabled, then we can
56  * safely return legacy VGA IO and memory as not decoded since the user
57  * has no way to get to it and routing can be disabled externally at the
58  * bridge.
59  */
vfio_pci_set_decode(struct pci_dev * pdev,bool single_vga)60 static unsigned int vfio_pci_set_decode(struct pci_dev *pdev, bool single_vga)
61 {
62 	struct pci_dev *tmp = NULL;
63 	unsigned char max_busnr;
64 	unsigned int decodes;
65 
66 	if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
67 		return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
68 		       VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
69 
70 	max_busnr = pci_bus_max_busnr(pdev->bus);
71 	decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
72 
73 	while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
74 		if (tmp == pdev ||
75 		    pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
76 		    pci_is_root_bus(tmp->bus))
77 			continue;
78 
79 		if (tmp->bus->number >= pdev->bus->number &&
80 		    tmp->bus->number <= max_busnr) {
81 			pci_dev_put(tmp);
82 			decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
83 			break;
84 		}
85 	}
86 
87 	return decodes;
88 }
89 
vfio_pci_probe_mmaps(struct vfio_pci_core_device * vdev)90 static void vfio_pci_probe_mmaps(struct vfio_pci_core_device *vdev)
91 {
92 	struct resource *res;
93 	int i;
94 	struct vfio_pci_dummy_resource *dummy_res;
95 
96 	for (i = 0; i < PCI_STD_NUM_BARS; i++) {
97 		int bar = i + PCI_STD_RESOURCES;
98 
99 		res = &vdev->pdev->resource[bar];
100 
101 		if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
102 			goto no_mmap;
103 
104 		if (!(res->flags & IORESOURCE_MEM))
105 			goto no_mmap;
106 
107 		/*
108 		 * The PCI core shouldn't set up a resource with a
109 		 * type but zero size. But there may be bugs that
110 		 * cause us to do that.
111 		 */
112 		if (!resource_size(res))
113 			goto no_mmap;
114 
115 		if (resource_size(res) >= PAGE_SIZE) {
116 			vdev->bar_mmap_supported[bar] = true;
117 			continue;
118 		}
119 
120 		if (!(res->start & ~PAGE_MASK)) {
121 			/*
122 			 * Add a dummy resource to reserve the remainder
123 			 * of the exclusive page in case that hot-add
124 			 * device's bar is assigned into it.
125 			 */
126 			dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
127 			if (dummy_res == NULL)
128 				goto no_mmap;
129 
130 			dummy_res->resource.name = "vfio sub-page reserved";
131 			dummy_res->resource.start = res->end + 1;
132 			dummy_res->resource.end = res->start + PAGE_SIZE - 1;
133 			dummy_res->resource.flags = res->flags;
134 			if (request_resource(res->parent,
135 						&dummy_res->resource)) {
136 				kfree(dummy_res);
137 				goto no_mmap;
138 			}
139 			dummy_res->index = bar;
140 			list_add(&dummy_res->res_next,
141 					&vdev->dummy_resources_list);
142 			vdev->bar_mmap_supported[bar] = true;
143 			continue;
144 		}
145 		/*
146 		 * Here we don't handle the case when the BAR is not page
147 		 * aligned because we can't expect the BAR will be
148 		 * assigned into the same location in a page in guest
149 		 * when we passthrough the BAR. And it's hard to access
150 		 * this BAR in userspace because we have no way to get
151 		 * the BAR's location in a page.
152 		 */
153 no_mmap:
154 		vdev->bar_mmap_supported[bar] = false;
155 	}
156 }
157 
158 struct vfio_pci_group_info;
159 static bool vfio_pci_dev_set_try_reset(struct vfio_device_set *dev_set);
160 static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set,
161 				      struct vfio_pci_group_info *groups);
162 
163 /*
164  * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
165  * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
166  * If a device implements the former but not the latter we would typically
167  * expect broken_intx_masking be set and require an exclusive interrupt.
168  * However since we do have control of the device's ability to assert INTx,
169  * we can instead pretend that the device does not implement INTx, virtualizing
170  * the pin register to report zero and maintaining DisINTx set on the host.
171  */
vfio_pci_nointx(struct pci_dev * pdev)172 static bool vfio_pci_nointx(struct pci_dev *pdev)
173 {
174 	switch (pdev->vendor) {
175 	case PCI_VENDOR_ID_INTEL:
176 		switch (pdev->device) {
177 		/* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
178 		case 0x1572:
179 		case 0x1574:
180 		case 0x1580 ... 0x1581:
181 		case 0x1583 ... 0x158b:
182 		case 0x37d0 ... 0x37d2:
183 		/* X550 */
184 		case 0x1563:
185 			return true;
186 		default:
187 			return false;
188 		}
189 	}
190 
191 	return false;
192 }
193 
vfio_pci_probe_power_state(struct vfio_pci_core_device * vdev)194 static void vfio_pci_probe_power_state(struct vfio_pci_core_device *vdev)
195 {
196 	struct pci_dev *pdev = vdev->pdev;
197 	u16 pmcsr;
198 
199 	if (!pdev->pm_cap)
200 		return;
201 
202 	pci_read_config_word(pdev, pdev->pm_cap + PCI_PM_CTRL, &pmcsr);
203 
204 	vdev->needs_pm_restore = !(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
205 }
206 
207 /*
208  * pci_set_power_state() wrapper handling devices which perform a soft reset on
209  * D3->D0 transition.  Save state prior to D0/1/2->D3, stash it on the vdev,
210  * restore when returned to D0.  Saved separately from pci_saved_state for use
211  * by PM capability emulation and separately from pci_dev internal saved state
212  * to avoid it being overwritten and consumed around other resets.
213  */
vfio_pci_set_power_state(struct vfio_pci_core_device * vdev,pci_power_t state)214 int vfio_pci_set_power_state(struct vfio_pci_core_device *vdev, pci_power_t state)
215 {
216 	struct pci_dev *pdev = vdev->pdev;
217 	bool needs_restore = false, needs_save = false;
218 	int ret;
219 
220 	if (vdev->needs_pm_restore) {
221 		if (pdev->current_state < PCI_D3hot && state >= PCI_D3hot) {
222 			pci_save_state(pdev);
223 			needs_save = true;
224 		}
225 
226 		if (pdev->current_state >= PCI_D3hot && state <= PCI_D0)
227 			needs_restore = true;
228 	}
229 
230 	ret = pci_set_power_state(pdev, state);
231 
232 	if (!ret) {
233 		/* D3 might be unsupported via quirk, skip unless in D3 */
234 		if (needs_save && pdev->current_state >= PCI_D3hot) {
235 			/*
236 			 * The current PCI state will be saved locally in
237 			 * 'pm_save' during the D3hot transition. When the
238 			 * device state is changed to D0 again with the current
239 			 * function, then pci_store_saved_state() will restore
240 			 * the state and will free the memory pointed by
241 			 * 'pm_save'. There are few cases where the PCI power
242 			 * state can be changed to D0 without the involvement
243 			 * of the driver. For these cases, free the earlier
244 			 * allocated memory first before overwriting 'pm_save'
245 			 * to prevent the memory leak.
246 			 */
247 			kfree(vdev->pm_save);
248 			vdev->pm_save = pci_store_saved_state(pdev);
249 		} else if (needs_restore) {
250 			pci_load_and_free_saved_state(pdev, &vdev->pm_save);
251 			pci_restore_state(pdev);
252 		}
253 	}
254 
255 	return ret;
256 }
257 
vfio_pci_core_enable(struct vfio_pci_core_device * vdev)258 int vfio_pci_core_enable(struct vfio_pci_core_device *vdev)
259 {
260 	struct pci_dev *pdev = vdev->pdev;
261 	int ret;
262 	u16 cmd;
263 	u8 msix_pos;
264 
265 	vfio_pci_set_power_state(vdev, PCI_D0);
266 
267 	/* Don't allow our initial saved state to include busmaster */
268 	pci_clear_master(pdev);
269 
270 	ret = pci_enable_device(pdev);
271 	if (ret)
272 		return ret;
273 
274 	/* If reset fails because of the device lock, fail this path entirely */
275 	ret = pci_try_reset_function(pdev);
276 	if (ret == -EAGAIN) {
277 		pci_disable_device(pdev);
278 		return ret;
279 	}
280 
281 	vdev->reset_works = !ret;
282 	pci_save_state(pdev);
283 	vdev->pci_saved_state = pci_store_saved_state(pdev);
284 	if (!vdev->pci_saved_state)
285 		pci_dbg(pdev, "%s: Couldn't store saved state\n", __func__);
286 
287 	if (likely(!nointxmask)) {
288 		if (vfio_pci_nointx(pdev)) {
289 			pci_info(pdev, "Masking broken INTx support\n");
290 			vdev->nointx = true;
291 			pci_intx(pdev, 0);
292 		} else
293 			vdev->pci_2_3 = pci_intx_mask_supported(pdev);
294 	}
295 
296 	pci_read_config_word(pdev, PCI_COMMAND, &cmd);
297 	if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
298 		cmd &= ~PCI_COMMAND_INTX_DISABLE;
299 		pci_write_config_word(pdev, PCI_COMMAND, cmd);
300 	}
301 
302 	ret = vfio_config_init(vdev);
303 	if (ret) {
304 		kfree(vdev->pci_saved_state);
305 		vdev->pci_saved_state = NULL;
306 		pci_disable_device(pdev);
307 		return ret;
308 	}
309 
310 	msix_pos = pdev->msix_cap;
311 	if (msix_pos) {
312 		u16 flags;
313 		u32 table;
314 
315 		pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
316 		pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
317 
318 		vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
319 		vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
320 		vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
321 	} else
322 		vdev->msix_bar = 0xFF;
323 
324 	if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
325 		vdev->has_vga = true;
326 
327 
328 	return 0;
329 }
330 EXPORT_SYMBOL_GPL(vfio_pci_core_enable);
331 
vfio_pci_core_disable(struct vfio_pci_core_device * vdev)332 void vfio_pci_core_disable(struct vfio_pci_core_device *vdev)
333 {
334 	struct pci_dev *pdev = vdev->pdev;
335 	struct vfio_pci_dummy_resource *dummy_res, *tmp;
336 	struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
337 	int i, bar;
338 
339 	/* For needs_reset */
340 	lockdep_assert_held(&vdev->vdev.dev_set->lock);
341 
342 	/*
343 	 * This function can be invoked while the power state is non-D0.
344 	 * This function calls __pci_reset_function_locked() which internally
345 	 * can use pci_pm_reset() for the function reset. pci_pm_reset() will
346 	 * fail if the power state is non-D0. Also, for the devices which
347 	 * have NoSoftRst-, the reset function can cause the PCI config space
348 	 * reset without restoring the original state (saved locally in
349 	 * 'vdev->pm_save').
350 	 */
351 	vfio_pci_set_power_state(vdev, PCI_D0);
352 
353 	/* Stop the device from further DMA */
354 	pci_clear_master(pdev);
355 
356 	vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
357 				VFIO_IRQ_SET_ACTION_TRIGGER,
358 				vdev->irq_type, 0, 0, NULL);
359 
360 	/* Device closed, don't need mutex here */
361 	list_for_each_entry_safe(ioeventfd, ioeventfd_tmp,
362 				 &vdev->ioeventfds_list, next) {
363 		vfio_virqfd_disable(&ioeventfd->virqfd);
364 		list_del(&ioeventfd->next);
365 		kfree(ioeventfd);
366 	}
367 	vdev->ioeventfds_nr = 0;
368 
369 	vdev->virq_disabled = false;
370 
371 	for (i = 0; i < vdev->num_regions; i++)
372 		vdev->region[i].ops->release(vdev, &vdev->region[i]);
373 
374 	vdev->num_regions = 0;
375 	kfree(vdev->region);
376 	vdev->region = NULL; /* don't krealloc a freed pointer */
377 
378 	vfio_config_free(vdev);
379 
380 	for (i = 0; i < PCI_STD_NUM_BARS; i++) {
381 		bar = i + PCI_STD_RESOURCES;
382 		if (!vdev->barmap[bar])
383 			continue;
384 		pci_iounmap(pdev, vdev->barmap[bar]);
385 		pci_release_selected_regions(pdev, 1 << bar);
386 		vdev->barmap[bar] = NULL;
387 	}
388 
389 	list_for_each_entry_safe(dummy_res, tmp,
390 				 &vdev->dummy_resources_list, res_next) {
391 		list_del(&dummy_res->res_next);
392 		release_resource(&dummy_res->resource);
393 		kfree(dummy_res);
394 	}
395 
396 	vdev->needs_reset = true;
397 
398 	/*
399 	 * If we have saved state, restore it.  If we can reset the device,
400 	 * even better.  Resetting with current state seems better than
401 	 * nothing, but saving and restoring current state without reset
402 	 * is just busy work.
403 	 */
404 	if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
405 		pci_info(pdev, "%s: Couldn't reload saved state\n", __func__);
406 
407 		if (!vdev->reset_works)
408 			goto out;
409 
410 		pci_save_state(pdev);
411 	}
412 
413 	/*
414 	 * Disable INTx and MSI, presumably to avoid spurious interrupts
415 	 * during reset.  Stolen from pci_reset_function()
416 	 */
417 	pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
418 
419 	/*
420 	 * Try to get the locks ourselves to prevent a deadlock. The
421 	 * success of this is dependent on being able to lock the device,
422 	 * which is not always possible.
423 	 * We can not use the "try" reset interface here, which will
424 	 * overwrite the previously restored configuration information.
425 	 */
426 	if (vdev->reset_works && pci_dev_trylock(pdev)) {
427 		if (!__pci_reset_function_locked(pdev))
428 			vdev->needs_reset = false;
429 		pci_dev_unlock(pdev);
430 	}
431 
432 	pci_restore_state(pdev);
433 out:
434 	pci_disable_device(pdev);
435 
436 	if (!vfio_pci_dev_set_try_reset(vdev->vdev.dev_set) && !disable_idle_d3)
437 		vfio_pci_set_power_state(vdev, PCI_D3hot);
438 }
439 EXPORT_SYMBOL_GPL(vfio_pci_core_disable);
440 
vfio_pci_core_close_device(struct vfio_device * core_vdev)441 void vfio_pci_core_close_device(struct vfio_device *core_vdev)
442 {
443 	struct vfio_pci_core_device *vdev =
444 		container_of(core_vdev, struct vfio_pci_core_device, vdev);
445 
446 	if (vdev->sriov_pf_core_dev) {
447 		mutex_lock(&vdev->sriov_pf_core_dev->vf_token->lock);
448 		WARN_ON(!vdev->sriov_pf_core_dev->vf_token->users);
449 		vdev->sriov_pf_core_dev->vf_token->users--;
450 		mutex_unlock(&vdev->sriov_pf_core_dev->vf_token->lock);
451 	}
452 	vfio_spapr_pci_eeh_release(vdev->pdev);
453 	vfio_pci_core_disable(vdev);
454 
455 	mutex_lock(&vdev->igate);
456 	if (vdev->err_trigger) {
457 		eventfd_ctx_put(vdev->err_trigger);
458 		vdev->err_trigger = NULL;
459 	}
460 	if (vdev->req_trigger) {
461 		eventfd_ctx_put(vdev->req_trigger);
462 		vdev->req_trigger = NULL;
463 	}
464 	mutex_unlock(&vdev->igate);
465 }
466 EXPORT_SYMBOL_GPL(vfio_pci_core_close_device);
467 
vfio_pci_core_finish_enable(struct vfio_pci_core_device * vdev)468 void vfio_pci_core_finish_enable(struct vfio_pci_core_device *vdev)
469 {
470 	vfio_pci_probe_mmaps(vdev);
471 	vfio_spapr_pci_eeh_open(vdev->pdev);
472 
473 	if (vdev->sriov_pf_core_dev) {
474 		mutex_lock(&vdev->sriov_pf_core_dev->vf_token->lock);
475 		vdev->sriov_pf_core_dev->vf_token->users++;
476 		mutex_unlock(&vdev->sriov_pf_core_dev->vf_token->lock);
477 	}
478 }
479 EXPORT_SYMBOL_GPL(vfio_pci_core_finish_enable);
480 
vfio_pci_get_irq_count(struct vfio_pci_core_device * vdev,int irq_type)481 static int vfio_pci_get_irq_count(struct vfio_pci_core_device *vdev, int irq_type)
482 {
483 	if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
484 		u8 pin;
485 
486 		if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
487 		    vdev->nointx || vdev->pdev->is_virtfn)
488 			return 0;
489 
490 		pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
491 
492 		return pin ? 1 : 0;
493 	} else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
494 		u8 pos;
495 		u16 flags;
496 
497 		pos = vdev->pdev->msi_cap;
498 		if (pos) {
499 			pci_read_config_word(vdev->pdev,
500 					     pos + PCI_MSI_FLAGS, &flags);
501 			return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
502 		}
503 	} else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
504 		u8 pos;
505 		u16 flags;
506 
507 		pos = vdev->pdev->msix_cap;
508 		if (pos) {
509 			pci_read_config_word(vdev->pdev,
510 					     pos + PCI_MSIX_FLAGS, &flags);
511 
512 			return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
513 		}
514 	} else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
515 		if (pci_is_pcie(vdev->pdev))
516 			return 1;
517 	} else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
518 		return 1;
519 	}
520 
521 	return 0;
522 }
523 
vfio_pci_count_devs(struct pci_dev * pdev,void * data)524 static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
525 {
526 	(*(int *)data)++;
527 	return 0;
528 }
529 
530 struct vfio_pci_fill_info {
531 	int max;
532 	int cur;
533 	struct vfio_pci_dependent_device *devices;
534 };
535 
vfio_pci_fill_devs(struct pci_dev * pdev,void * data)536 static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
537 {
538 	struct vfio_pci_fill_info *fill = data;
539 	struct iommu_group *iommu_group;
540 
541 	if (fill->cur == fill->max)
542 		return -EAGAIN; /* Something changed, try again */
543 
544 	iommu_group = iommu_group_get(&pdev->dev);
545 	if (!iommu_group)
546 		return -EPERM; /* Cannot reset non-isolated devices */
547 
548 	fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
549 	fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
550 	fill->devices[fill->cur].bus = pdev->bus->number;
551 	fill->devices[fill->cur].devfn = pdev->devfn;
552 	fill->cur++;
553 	iommu_group_put(iommu_group);
554 	return 0;
555 }
556 
557 struct vfio_pci_group_info {
558 	int count;
559 	struct vfio_group **groups;
560 };
561 
vfio_pci_dev_below_slot(struct pci_dev * pdev,struct pci_slot * slot)562 static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
563 {
564 	for (; pdev; pdev = pdev->bus->self)
565 		if (pdev->bus == slot->bus)
566 			return (pdev->slot == slot);
567 	return false;
568 }
569 
570 struct vfio_pci_walk_info {
571 	int (*fn)(struct pci_dev *pdev, void *data);
572 	void *data;
573 	struct pci_dev *pdev;
574 	bool slot;
575 	int ret;
576 };
577 
vfio_pci_walk_wrapper(struct pci_dev * pdev,void * data)578 static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
579 {
580 	struct vfio_pci_walk_info *walk = data;
581 
582 	if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
583 		walk->ret = walk->fn(pdev, walk->data);
584 
585 	return walk->ret;
586 }
587 
vfio_pci_for_each_slot_or_bus(struct pci_dev * pdev,int (* fn)(struct pci_dev *,void * data),void * data,bool slot)588 static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
589 					 int (*fn)(struct pci_dev *,
590 						   void *data), void *data,
591 					 bool slot)
592 {
593 	struct vfio_pci_walk_info walk = {
594 		.fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
595 	};
596 
597 	pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
598 
599 	return walk.ret;
600 }
601 
msix_mmappable_cap(struct vfio_pci_core_device * vdev,struct vfio_info_cap * caps)602 static int msix_mmappable_cap(struct vfio_pci_core_device *vdev,
603 			      struct vfio_info_cap *caps)
604 {
605 	struct vfio_info_cap_header header = {
606 		.id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
607 		.version = 1
608 	};
609 
610 	return vfio_info_add_capability(caps, &header, sizeof(header));
611 }
612 
vfio_pci_register_dev_region(struct vfio_pci_core_device * vdev,unsigned int type,unsigned int subtype,const struct vfio_pci_regops * ops,size_t size,u32 flags,void * data)613 int vfio_pci_register_dev_region(struct vfio_pci_core_device *vdev,
614 				 unsigned int type, unsigned int subtype,
615 				 const struct vfio_pci_regops *ops,
616 				 size_t size, u32 flags, void *data)
617 {
618 	struct vfio_pci_region *region;
619 
620 	region = krealloc(vdev->region,
621 			  (vdev->num_regions + 1) * sizeof(*region),
622 			  GFP_KERNEL);
623 	if (!region)
624 		return -ENOMEM;
625 
626 	vdev->region = region;
627 	vdev->region[vdev->num_regions].type = type;
628 	vdev->region[vdev->num_regions].subtype = subtype;
629 	vdev->region[vdev->num_regions].ops = ops;
630 	vdev->region[vdev->num_regions].size = size;
631 	vdev->region[vdev->num_regions].flags = flags;
632 	vdev->region[vdev->num_regions].data = data;
633 
634 	vdev->num_regions++;
635 
636 	return 0;
637 }
638 EXPORT_SYMBOL_GPL(vfio_pci_register_dev_region);
639 
vfio_pci_core_ioctl(struct vfio_device * core_vdev,unsigned int cmd,unsigned long arg)640 long vfio_pci_core_ioctl(struct vfio_device *core_vdev, unsigned int cmd,
641 		unsigned long arg)
642 {
643 	struct vfio_pci_core_device *vdev =
644 		container_of(core_vdev, struct vfio_pci_core_device, vdev);
645 	unsigned long minsz;
646 
647 	if (cmd == VFIO_DEVICE_GET_INFO) {
648 		struct vfio_device_info info;
649 		struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
650 		unsigned long capsz;
651 		int ret;
652 
653 		minsz = offsetofend(struct vfio_device_info, num_irqs);
654 
655 		/* For backward compatibility, cannot require this */
656 		capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
657 
658 		if (copy_from_user(&info, (void __user *)arg, minsz))
659 			return -EFAULT;
660 
661 		if (info.argsz < minsz)
662 			return -EINVAL;
663 
664 		if (info.argsz >= capsz) {
665 			minsz = capsz;
666 			info.cap_offset = 0;
667 		}
668 
669 		info.flags = VFIO_DEVICE_FLAGS_PCI;
670 
671 		if (vdev->reset_works)
672 			info.flags |= VFIO_DEVICE_FLAGS_RESET;
673 
674 		info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
675 		info.num_irqs = VFIO_PCI_NUM_IRQS;
676 
677 		ret = vfio_pci_info_zdev_add_caps(vdev, &caps);
678 		if (ret && ret != -ENODEV) {
679 			pci_warn(vdev->pdev, "Failed to setup zPCI info capabilities\n");
680 			return ret;
681 		}
682 
683 		if (caps.size) {
684 			info.flags |= VFIO_DEVICE_FLAGS_CAPS;
685 			if (info.argsz < sizeof(info) + caps.size) {
686 				info.argsz = sizeof(info) + caps.size;
687 			} else {
688 				vfio_info_cap_shift(&caps, sizeof(info));
689 				if (copy_to_user((void __user *)arg +
690 						  sizeof(info), caps.buf,
691 						  caps.size)) {
692 					kfree(caps.buf);
693 					return -EFAULT;
694 				}
695 				info.cap_offset = sizeof(info);
696 			}
697 
698 			kfree(caps.buf);
699 		}
700 
701 		return copy_to_user((void __user *)arg, &info, minsz) ?
702 			-EFAULT : 0;
703 
704 	} else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
705 		struct pci_dev *pdev = vdev->pdev;
706 		struct vfio_region_info info;
707 		struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
708 		int i, ret;
709 
710 		minsz = offsetofend(struct vfio_region_info, offset);
711 
712 		if (copy_from_user(&info, (void __user *)arg, minsz))
713 			return -EFAULT;
714 
715 		if (info.argsz < minsz)
716 			return -EINVAL;
717 
718 		switch (info.index) {
719 		case VFIO_PCI_CONFIG_REGION_INDEX:
720 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
721 			info.size = pdev->cfg_size;
722 			info.flags = VFIO_REGION_INFO_FLAG_READ |
723 				     VFIO_REGION_INFO_FLAG_WRITE;
724 			break;
725 		case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
726 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
727 			info.size = pci_resource_len(pdev, info.index);
728 			if (!info.size) {
729 				info.flags = 0;
730 				break;
731 			}
732 
733 			info.flags = VFIO_REGION_INFO_FLAG_READ |
734 				     VFIO_REGION_INFO_FLAG_WRITE;
735 			if (vdev->bar_mmap_supported[info.index]) {
736 				info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
737 				if (info.index == vdev->msix_bar) {
738 					ret = msix_mmappable_cap(vdev, &caps);
739 					if (ret)
740 						return ret;
741 				}
742 			}
743 
744 			break;
745 		case VFIO_PCI_ROM_REGION_INDEX:
746 		{
747 			void __iomem *io;
748 			size_t size;
749 			u16 cmd;
750 
751 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
752 			info.flags = 0;
753 
754 			/* Report the BAR size, not the ROM size */
755 			info.size = pci_resource_len(pdev, info.index);
756 			if (!info.size) {
757 				/* Shadow ROMs appear as PCI option ROMs */
758 				if (pdev->resource[PCI_ROM_RESOURCE].flags &
759 							IORESOURCE_ROM_SHADOW)
760 					info.size = 0x20000;
761 				else
762 					break;
763 			}
764 
765 			/*
766 			 * Is it really there?  Enable memory decode for
767 			 * implicit access in pci_map_rom().
768 			 */
769 			cmd = vfio_pci_memory_lock_and_enable(vdev);
770 			io = pci_map_rom(pdev, &size);
771 			if (io) {
772 				info.flags = VFIO_REGION_INFO_FLAG_READ;
773 				pci_unmap_rom(pdev, io);
774 			} else {
775 				info.size = 0;
776 			}
777 			vfio_pci_memory_unlock_and_restore(vdev, cmd);
778 
779 			break;
780 		}
781 		case VFIO_PCI_VGA_REGION_INDEX:
782 			if (!vdev->has_vga)
783 				return -EINVAL;
784 
785 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
786 			info.size = 0xc0000;
787 			info.flags = VFIO_REGION_INFO_FLAG_READ |
788 				     VFIO_REGION_INFO_FLAG_WRITE;
789 
790 			break;
791 		default:
792 		{
793 			struct vfio_region_info_cap_type cap_type = {
794 					.header.id = VFIO_REGION_INFO_CAP_TYPE,
795 					.header.version = 1 };
796 
797 			if (info.index >=
798 			    VFIO_PCI_NUM_REGIONS + vdev->num_regions)
799 				return -EINVAL;
800 			info.index = array_index_nospec(info.index,
801 							VFIO_PCI_NUM_REGIONS +
802 							vdev->num_regions);
803 
804 			i = info.index - VFIO_PCI_NUM_REGIONS;
805 
806 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
807 			info.size = vdev->region[i].size;
808 			info.flags = vdev->region[i].flags;
809 
810 			cap_type.type = vdev->region[i].type;
811 			cap_type.subtype = vdev->region[i].subtype;
812 
813 			ret = vfio_info_add_capability(&caps, &cap_type.header,
814 						       sizeof(cap_type));
815 			if (ret)
816 				return ret;
817 
818 			if (vdev->region[i].ops->add_capability) {
819 				ret = vdev->region[i].ops->add_capability(vdev,
820 						&vdev->region[i], &caps);
821 				if (ret)
822 					return ret;
823 			}
824 		}
825 		}
826 
827 		if (caps.size) {
828 			info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
829 			if (info.argsz < sizeof(info) + caps.size) {
830 				info.argsz = sizeof(info) + caps.size;
831 				info.cap_offset = 0;
832 			} else {
833 				vfio_info_cap_shift(&caps, sizeof(info));
834 				if (copy_to_user((void __user *)arg +
835 						  sizeof(info), caps.buf,
836 						  caps.size)) {
837 					kfree(caps.buf);
838 					return -EFAULT;
839 				}
840 				info.cap_offset = sizeof(info);
841 			}
842 
843 			kfree(caps.buf);
844 		}
845 
846 		return copy_to_user((void __user *)arg, &info, minsz) ?
847 			-EFAULT : 0;
848 
849 	} else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
850 		struct vfio_irq_info info;
851 
852 		minsz = offsetofend(struct vfio_irq_info, count);
853 
854 		if (copy_from_user(&info, (void __user *)arg, minsz))
855 			return -EFAULT;
856 
857 		if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
858 			return -EINVAL;
859 
860 		switch (info.index) {
861 		case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
862 		case VFIO_PCI_REQ_IRQ_INDEX:
863 			break;
864 		case VFIO_PCI_ERR_IRQ_INDEX:
865 			if (pci_is_pcie(vdev->pdev))
866 				break;
867 			fallthrough;
868 		default:
869 			return -EINVAL;
870 		}
871 
872 		info.flags = VFIO_IRQ_INFO_EVENTFD;
873 
874 		info.count = vfio_pci_get_irq_count(vdev, info.index);
875 
876 		if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
877 			info.flags |= (VFIO_IRQ_INFO_MASKABLE |
878 				       VFIO_IRQ_INFO_AUTOMASKED);
879 		else
880 			info.flags |= VFIO_IRQ_INFO_NORESIZE;
881 
882 		return copy_to_user((void __user *)arg, &info, minsz) ?
883 			-EFAULT : 0;
884 
885 	} else if (cmd == VFIO_DEVICE_SET_IRQS) {
886 		struct vfio_irq_set hdr;
887 		u8 *data = NULL;
888 		int max, ret = 0;
889 		size_t data_size = 0;
890 
891 		minsz = offsetofend(struct vfio_irq_set, count);
892 
893 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
894 			return -EFAULT;
895 
896 		max = vfio_pci_get_irq_count(vdev, hdr.index);
897 
898 		ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
899 						 VFIO_PCI_NUM_IRQS, &data_size);
900 		if (ret)
901 			return ret;
902 
903 		if (data_size) {
904 			data = memdup_user((void __user *)(arg + minsz),
905 					    data_size);
906 			if (IS_ERR(data))
907 				return PTR_ERR(data);
908 		}
909 
910 		mutex_lock(&vdev->igate);
911 
912 		ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
913 					      hdr.start, hdr.count, data);
914 
915 		mutex_unlock(&vdev->igate);
916 		kfree(data);
917 
918 		return ret;
919 
920 	} else if (cmd == VFIO_DEVICE_RESET) {
921 		int ret;
922 
923 		if (!vdev->reset_works)
924 			return -EINVAL;
925 
926 		vfio_pci_zap_and_down_write_memory_lock(vdev);
927 
928 		/*
929 		 * This function can be invoked while the power state is non-D0.
930 		 * If pci_try_reset_function() has been called while the power
931 		 * state is non-D0, then pci_try_reset_function() will
932 		 * internally set the power state to D0 without vfio driver
933 		 * involvement. For the devices which have NoSoftRst-, the
934 		 * reset function can cause the PCI config space reset without
935 		 * restoring the original state (saved locally in
936 		 * 'vdev->pm_save').
937 		 */
938 		vfio_pci_set_power_state(vdev, PCI_D0);
939 
940 		ret = pci_try_reset_function(vdev->pdev);
941 		up_write(&vdev->memory_lock);
942 
943 		return ret;
944 
945 	} else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
946 		struct vfio_pci_hot_reset_info hdr;
947 		struct vfio_pci_fill_info fill = { 0 };
948 		struct vfio_pci_dependent_device *devices = NULL;
949 		bool slot = false;
950 		int ret = 0;
951 
952 		minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
953 
954 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
955 			return -EFAULT;
956 
957 		if (hdr.argsz < minsz)
958 			return -EINVAL;
959 
960 		hdr.flags = 0;
961 
962 		/* Can we do a slot or bus reset or neither? */
963 		if (!pci_probe_reset_slot(vdev->pdev->slot))
964 			slot = true;
965 		else if (pci_probe_reset_bus(vdev->pdev->bus))
966 			return -ENODEV;
967 
968 		/* How many devices are affected? */
969 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
970 						    vfio_pci_count_devs,
971 						    &fill.max, slot);
972 		if (ret)
973 			return ret;
974 
975 		WARN_ON(!fill.max); /* Should always be at least one */
976 
977 		/*
978 		 * If there's enough space, fill it now, otherwise return
979 		 * -ENOSPC and the number of devices affected.
980 		 */
981 		if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
982 			ret = -ENOSPC;
983 			hdr.count = fill.max;
984 			goto reset_info_exit;
985 		}
986 
987 		devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
988 		if (!devices)
989 			return -ENOMEM;
990 
991 		fill.devices = devices;
992 
993 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
994 						    vfio_pci_fill_devs,
995 						    &fill, slot);
996 
997 		/*
998 		 * If a device was removed between counting and filling,
999 		 * we may come up short of fill.max.  If a device was
1000 		 * added, we'll have a return of -EAGAIN above.
1001 		 */
1002 		if (!ret)
1003 			hdr.count = fill.cur;
1004 
1005 reset_info_exit:
1006 		if (copy_to_user((void __user *)arg, &hdr, minsz))
1007 			ret = -EFAULT;
1008 
1009 		if (!ret) {
1010 			if (copy_to_user((void __user *)(arg + minsz), devices,
1011 					 hdr.count * sizeof(*devices)))
1012 				ret = -EFAULT;
1013 		}
1014 
1015 		kfree(devices);
1016 		return ret;
1017 
1018 	} else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
1019 		struct vfio_pci_hot_reset hdr;
1020 		int32_t *group_fds;
1021 		struct vfio_group **groups;
1022 		struct vfio_pci_group_info info;
1023 		bool slot = false;
1024 		int group_idx, count = 0, ret = 0;
1025 
1026 		minsz = offsetofend(struct vfio_pci_hot_reset, count);
1027 
1028 		if (copy_from_user(&hdr, (void __user *)arg, minsz))
1029 			return -EFAULT;
1030 
1031 		if (hdr.argsz < minsz || hdr.flags)
1032 			return -EINVAL;
1033 
1034 		/* Can we do a slot or bus reset or neither? */
1035 		if (!pci_probe_reset_slot(vdev->pdev->slot))
1036 			slot = true;
1037 		else if (pci_probe_reset_bus(vdev->pdev->bus))
1038 			return -ENODEV;
1039 
1040 		/*
1041 		 * We can't let userspace give us an arbitrarily large
1042 		 * buffer to copy, so verify how many we think there
1043 		 * could be.  Note groups can have multiple devices so
1044 		 * one group per device is the max.
1045 		 */
1046 		ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1047 						    vfio_pci_count_devs,
1048 						    &count, slot);
1049 		if (ret)
1050 			return ret;
1051 
1052 		/* Somewhere between 1 and count is OK */
1053 		if (!hdr.count || hdr.count > count)
1054 			return -EINVAL;
1055 
1056 		group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
1057 		groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
1058 		if (!group_fds || !groups) {
1059 			kfree(group_fds);
1060 			kfree(groups);
1061 			return -ENOMEM;
1062 		}
1063 
1064 		if (copy_from_user(group_fds, (void __user *)(arg + minsz),
1065 				   hdr.count * sizeof(*group_fds))) {
1066 			kfree(group_fds);
1067 			kfree(groups);
1068 			return -EFAULT;
1069 		}
1070 
1071 		/*
1072 		 * For each group_fd, get the group through the vfio external
1073 		 * user interface and store the group and iommu ID.  This
1074 		 * ensures the group is held across the reset.
1075 		 */
1076 		for (group_idx = 0; group_idx < hdr.count; group_idx++) {
1077 			struct vfio_group *group;
1078 			struct fd f = fdget(group_fds[group_idx]);
1079 			if (!f.file) {
1080 				ret = -EBADF;
1081 				break;
1082 			}
1083 
1084 			group = vfio_group_get_external_user(f.file);
1085 			fdput(f);
1086 			if (IS_ERR(group)) {
1087 				ret = PTR_ERR(group);
1088 				break;
1089 			}
1090 
1091 			groups[group_idx] = group;
1092 		}
1093 
1094 		kfree(group_fds);
1095 
1096 		/* release reference to groups on error */
1097 		if (ret)
1098 			goto hot_reset_release;
1099 
1100 		info.count = hdr.count;
1101 		info.groups = groups;
1102 
1103 		ret = vfio_pci_dev_set_hot_reset(vdev->vdev.dev_set, &info);
1104 
1105 hot_reset_release:
1106 		for (group_idx--; group_idx >= 0; group_idx--)
1107 			vfio_group_put_external_user(groups[group_idx]);
1108 
1109 		kfree(groups);
1110 		return ret;
1111 	} else if (cmd == VFIO_DEVICE_IOEVENTFD) {
1112 		struct vfio_device_ioeventfd ioeventfd;
1113 		int count;
1114 
1115 		minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1116 
1117 		if (copy_from_user(&ioeventfd, (void __user *)arg, minsz))
1118 			return -EFAULT;
1119 
1120 		if (ioeventfd.argsz < minsz)
1121 			return -EINVAL;
1122 
1123 		if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1124 			return -EINVAL;
1125 
1126 		count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1127 
1128 		if (hweight8(count) != 1 || ioeventfd.fd < -1)
1129 			return -EINVAL;
1130 
1131 		return vfio_pci_ioeventfd(vdev, ioeventfd.offset,
1132 					  ioeventfd.data, count, ioeventfd.fd);
1133 	} else if (cmd == VFIO_DEVICE_FEATURE) {
1134 		struct vfio_device_feature feature;
1135 		uuid_t uuid;
1136 
1137 		minsz = offsetofend(struct vfio_device_feature, flags);
1138 
1139 		if (copy_from_user(&feature, (void __user *)arg, minsz))
1140 			return -EFAULT;
1141 
1142 		if (feature.argsz < minsz)
1143 			return -EINVAL;
1144 
1145 		/* Check unknown flags */
1146 		if (feature.flags & ~(VFIO_DEVICE_FEATURE_MASK |
1147 				      VFIO_DEVICE_FEATURE_SET |
1148 				      VFIO_DEVICE_FEATURE_GET |
1149 				      VFIO_DEVICE_FEATURE_PROBE))
1150 			return -EINVAL;
1151 
1152 		/* GET & SET are mutually exclusive except with PROBE */
1153 		if (!(feature.flags & VFIO_DEVICE_FEATURE_PROBE) &&
1154 		    (feature.flags & VFIO_DEVICE_FEATURE_SET) &&
1155 		    (feature.flags & VFIO_DEVICE_FEATURE_GET))
1156 			return -EINVAL;
1157 
1158 		switch (feature.flags & VFIO_DEVICE_FEATURE_MASK) {
1159 		case VFIO_DEVICE_FEATURE_PCI_VF_TOKEN:
1160 			if (!vdev->vf_token)
1161 				return -ENOTTY;
1162 
1163 			/*
1164 			 * We do not support GET of the VF Token UUID as this
1165 			 * could expose the token of the previous device user.
1166 			 */
1167 			if (feature.flags & VFIO_DEVICE_FEATURE_GET)
1168 				return -EINVAL;
1169 
1170 			if (feature.flags & VFIO_DEVICE_FEATURE_PROBE)
1171 				return 0;
1172 
1173 			/* Don't SET unless told to do so */
1174 			if (!(feature.flags & VFIO_DEVICE_FEATURE_SET))
1175 				return -EINVAL;
1176 
1177 			if (feature.argsz < minsz + sizeof(uuid))
1178 				return -EINVAL;
1179 
1180 			if (copy_from_user(&uuid, (void __user *)(arg + minsz),
1181 					   sizeof(uuid)))
1182 				return -EFAULT;
1183 
1184 			mutex_lock(&vdev->vf_token->lock);
1185 			uuid_copy(&vdev->vf_token->uuid, &uuid);
1186 			mutex_unlock(&vdev->vf_token->lock);
1187 
1188 			return 0;
1189 		default:
1190 			return -ENOTTY;
1191 		}
1192 	}
1193 
1194 	return -ENOTTY;
1195 }
1196 EXPORT_SYMBOL_GPL(vfio_pci_core_ioctl);
1197 
vfio_pci_rw(struct vfio_pci_core_device * vdev,char __user * buf,size_t count,loff_t * ppos,bool iswrite)1198 static ssize_t vfio_pci_rw(struct vfio_pci_core_device *vdev, char __user *buf,
1199 			   size_t count, loff_t *ppos, bool iswrite)
1200 {
1201 	unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1202 
1203 	if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1204 		return -EINVAL;
1205 
1206 	switch (index) {
1207 	case VFIO_PCI_CONFIG_REGION_INDEX:
1208 		return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1209 
1210 	case VFIO_PCI_ROM_REGION_INDEX:
1211 		if (iswrite)
1212 			return -EINVAL;
1213 		return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1214 
1215 	case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1216 		return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1217 
1218 	case VFIO_PCI_VGA_REGION_INDEX:
1219 		return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1220 	default:
1221 		index -= VFIO_PCI_NUM_REGIONS;
1222 		return vdev->region[index].ops->rw(vdev, buf,
1223 						   count, ppos, iswrite);
1224 	}
1225 
1226 	return -EINVAL;
1227 }
1228 
vfio_pci_core_read(struct vfio_device * core_vdev,char __user * buf,size_t count,loff_t * ppos)1229 ssize_t vfio_pci_core_read(struct vfio_device *core_vdev, char __user *buf,
1230 		size_t count, loff_t *ppos)
1231 {
1232 	struct vfio_pci_core_device *vdev =
1233 		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1234 
1235 	if (!count)
1236 		return 0;
1237 
1238 	return vfio_pci_rw(vdev, buf, count, ppos, false);
1239 }
1240 EXPORT_SYMBOL_GPL(vfio_pci_core_read);
1241 
vfio_pci_core_write(struct vfio_device * core_vdev,const char __user * buf,size_t count,loff_t * ppos)1242 ssize_t vfio_pci_core_write(struct vfio_device *core_vdev, const char __user *buf,
1243 		size_t count, loff_t *ppos)
1244 {
1245 	struct vfio_pci_core_device *vdev =
1246 		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1247 
1248 	if (!count)
1249 		return 0;
1250 
1251 	return vfio_pci_rw(vdev, (char __user *)buf, count, ppos, true);
1252 }
1253 EXPORT_SYMBOL_GPL(vfio_pci_core_write);
1254 
1255 /* Return 1 on zap and vma_lock acquired, 0 on contention (only with @try) */
vfio_pci_zap_and_vma_lock(struct vfio_pci_core_device * vdev,bool try)1256 static int vfio_pci_zap_and_vma_lock(struct vfio_pci_core_device *vdev, bool try)
1257 {
1258 	struct vfio_pci_mmap_vma *mmap_vma, *tmp;
1259 
1260 	/*
1261 	 * Lock ordering:
1262 	 * vma_lock is nested under mmap_lock for vm_ops callback paths.
1263 	 * The memory_lock semaphore is used by both code paths calling
1264 	 * into this function to zap vmas and the vm_ops.fault callback
1265 	 * to protect the memory enable state of the device.
1266 	 *
1267 	 * When zapping vmas we need to maintain the mmap_lock => vma_lock
1268 	 * ordering, which requires using vma_lock to walk vma_list to
1269 	 * acquire an mm, then dropping vma_lock to get the mmap_lock and
1270 	 * reacquiring vma_lock.  This logic is derived from similar
1271 	 * requirements in uverbs_user_mmap_disassociate().
1272 	 *
1273 	 * mmap_lock must always be the top-level lock when it is taken.
1274 	 * Therefore we can only hold the memory_lock write lock when
1275 	 * vma_list is empty, as we'd need to take mmap_lock to clear
1276 	 * entries.  vma_list can only be guaranteed empty when holding
1277 	 * vma_lock, thus memory_lock is nested under vma_lock.
1278 	 *
1279 	 * This enables the vm_ops.fault callback to acquire vma_lock,
1280 	 * followed by memory_lock read lock, while already holding
1281 	 * mmap_lock without risk of deadlock.
1282 	 */
1283 	while (1) {
1284 		struct mm_struct *mm = NULL;
1285 
1286 		if (try) {
1287 			if (!mutex_trylock(&vdev->vma_lock))
1288 				return 0;
1289 		} else {
1290 			mutex_lock(&vdev->vma_lock);
1291 		}
1292 		while (!list_empty(&vdev->vma_list)) {
1293 			mmap_vma = list_first_entry(&vdev->vma_list,
1294 						    struct vfio_pci_mmap_vma,
1295 						    vma_next);
1296 			mm = mmap_vma->vma->vm_mm;
1297 			if (mmget_not_zero(mm))
1298 				break;
1299 
1300 			list_del(&mmap_vma->vma_next);
1301 			kfree(mmap_vma);
1302 			mm = NULL;
1303 		}
1304 		if (!mm)
1305 			return 1;
1306 		mutex_unlock(&vdev->vma_lock);
1307 
1308 		if (try) {
1309 			if (!mmap_read_trylock(mm)) {
1310 				mmput(mm);
1311 				return 0;
1312 			}
1313 		} else {
1314 			mmap_read_lock(mm);
1315 		}
1316 		if (try) {
1317 			if (!mutex_trylock(&vdev->vma_lock)) {
1318 				mmap_read_unlock(mm);
1319 				mmput(mm);
1320 				return 0;
1321 			}
1322 		} else {
1323 			mutex_lock(&vdev->vma_lock);
1324 		}
1325 		list_for_each_entry_safe(mmap_vma, tmp,
1326 					 &vdev->vma_list, vma_next) {
1327 			struct vm_area_struct *vma = mmap_vma->vma;
1328 
1329 			if (vma->vm_mm != mm)
1330 				continue;
1331 
1332 			list_del(&mmap_vma->vma_next);
1333 			kfree(mmap_vma);
1334 
1335 			zap_vma_ptes(vma, vma->vm_start,
1336 				     vma->vm_end - vma->vm_start);
1337 		}
1338 		mutex_unlock(&vdev->vma_lock);
1339 		mmap_read_unlock(mm);
1340 		mmput(mm);
1341 	}
1342 }
1343 
vfio_pci_zap_and_down_write_memory_lock(struct vfio_pci_core_device * vdev)1344 void vfio_pci_zap_and_down_write_memory_lock(struct vfio_pci_core_device *vdev)
1345 {
1346 	vfio_pci_zap_and_vma_lock(vdev, false);
1347 	down_write(&vdev->memory_lock);
1348 	mutex_unlock(&vdev->vma_lock);
1349 }
1350 
vfio_pci_memory_lock_and_enable(struct vfio_pci_core_device * vdev)1351 u16 vfio_pci_memory_lock_and_enable(struct vfio_pci_core_device *vdev)
1352 {
1353 	u16 cmd;
1354 
1355 	down_write(&vdev->memory_lock);
1356 	pci_read_config_word(vdev->pdev, PCI_COMMAND, &cmd);
1357 	if (!(cmd & PCI_COMMAND_MEMORY))
1358 		pci_write_config_word(vdev->pdev, PCI_COMMAND,
1359 				      cmd | PCI_COMMAND_MEMORY);
1360 
1361 	return cmd;
1362 }
1363 
vfio_pci_memory_unlock_and_restore(struct vfio_pci_core_device * vdev,u16 cmd)1364 void vfio_pci_memory_unlock_and_restore(struct vfio_pci_core_device *vdev, u16 cmd)
1365 {
1366 	pci_write_config_word(vdev->pdev, PCI_COMMAND, cmd);
1367 	up_write(&vdev->memory_lock);
1368 }
1369 
1370 /* Caller holds vma_lock */
__vfio_pci_add_vma(struct vfio_pci_core_device * vdev,struct vm_area_struct * vma)1371 static int __vfio_pci_add_vma(struct vfio_pci_core_device *vdev,
1372 			      struct vm_area_struct *vma)
1373 {
1374 	struct vfio_pci_mmap_vma *mmap_vma;
1375 
1376 	mmap_vma = kmalloc(sizeof(*mmap_vma), GFP_KERNEL);
1377 	if (!mmap_vma)
1378 		return -ENOMEM;
1379 
1380 	mmap_vma->vma = vma;
1381 	list_add(&mmap_vma->vma_next, &vdev->vma_list);
1382 
1383 	return 0;
1384 }
1385 
1386 /*
1387  * Zap mmaps on open so that we can fault them in on access and therefore
1388  * our vma_list only tracks mappings accessed since last zap.
1389  */
vfio_pci_mmap_open(struct vm_area_struct * vma)1390 static void vfio_pci_mmap_open(struct vm_area_struct *vma)
1391 {
1392 	zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1393 }
1394 
vfio_pci_mmap_close(struct vm_area_struct * vma)1395 static void vfio_pci_mmap_close(struct vm_area_struct *vma)
1396 {
1397 	struct vfio_pci_core_device *vdev = vma->vm_private_data;
1398 	struct vfio_pci_mmap_vma *mmap_vma;
1399 
1400 	mutex_lock(&vdev->vma_lock);
1401 	list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1402 		if (mmap_vma->vma == vma) {
1403 			list_del(&mmap_vma->vma_next);
1404 			kfree(mmap_vma);
1405 			break;
1406 		}
1407 	}
1408 	mutex_unlock(&vdev->vma_lock);
1409 }
1410 
vfio_pci_mmap_fault(struct vm_fault * vmf)1411 static vm_fault_t vfio_pci_mmap_fault(struct vm_fault *vmf)
1412 {
1413 	struct vm_area_struct *vma = vmf->vma;
1414 	struct vfio_pci_core_device *vdev = vma->vm_private_data;
1415 	struct vfio_pci_mmap_vma *mmap_vma;
1416 	vm_fault_t ret = VM_FAULT_NOPAGE;
1417 
1418 	mutex_lock(&vdev->vma_lock);
1419 	down_read(&vdev->memory_lock);
1420 
1421 	if (!__vfio_pci_memory_enabled(vdev)) {
1422 		ret = VM_FAULT_SIGBUS;
1423 		goto up_out;
1424 	}
1425 
1426 	/*
1427 	 * We populate the whole vma on fault, so we need to test whether
1428 	 * the vma has already been mapped, such as for concurrent faults
1429 	 * to the same vma.  io_remap_pfn_range() will trigger a BUG_ON if
1430 	 * we ask it to fill the same range again.
1431 	 */
1432 	list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1433 		if (mmap_vma->vma == vma)
1434 			goto up_out;
1435 	}
1436 
1437 	if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1438 			       vma->vm_end - vma->vm_start,
1439 			       vma->vm_page_prot)) {
1440 		ret = VM_FAULT_SIGBUS;
1441 		zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1442 		goto up_out;
1443 	}
1444 
1445 	if (__vfio_pci_add_vma(vdev, vma)) {
1446 		ret = VM_FAULT_OOM;
1447 		zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1448 	}
1449 
1450 up_out:
1451 	up_read(&vdev->memory_lock);
1452 	mutex_unlock(&vdev->vma_lock);
1453 	return ret;
1454 }
1455 
1456 static const struct vm_operations_struct vfio_pci_mmap_ops = {
1457 	.open = vfio_pci_mmap_open,
1458 	.close = vfio_pci_mmap_close,
1459 	.fault = vfio_pci_mmap_fault,
1460 };
1461 
vfio_pci_core_mmap(struct vfio_device * core_vdev,struct vm_area_struct * vma)1462 int vfio_pci_core_mmap(struct vfio_device *core_vdev, struct vm_area_struct *vma)
1463 {
1464 	struct vfio_pci_core_device *vdev =
1465 		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1466 	struct pci_dev *pdev = vdev->pdev;
1467 	unsigned int index;
1468 	u64 phys_len, req_len, pgoff, req_start;
1469 	int ret;
1470 
1471 	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1472 
1473 	if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1474 		return -EINVAL;
1475 	if (vma->vm_end < vma->vm_start)
1476 		return -EINVAL;
1477 	if ((vma->vm_flags & VM_SHARED) == 0)
1478 		return -EINVAL;
1479 	if (index >= VFIO_PCI_NUM_REGIONS) {
1480 		int regnum = index - VFIO_PCI_NUM_REGIONS;
1481 		struct vfio_pci_region *region = vdev->region + regnum;
1482 
1483 		if (region->ops && region->ops->mmap &&
1484 		    (region->flags & VFIO_REGION_INFO_FLAG_MMAP))
1485 			return region->ops->mmap(vdev, region, vma);
1486 		return -EINVAL;
1487 	}
1488 	if (index >= VFIO_PCI_ROM_REGION_INDEX)
1489 		return -EINVAL;
1490 	if (!vdev->bar_mmap_supported[index])
1491 		return -EINVAL;
1492 
1493 	phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1494 	req_len = vma->vm_end - vma->vm_start;
1495 	pgoff = vma->vm_pgoff &
1496 		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1497 	req_start = pgoff << PAGE_SHIFT;
1498 
1499 	if (req_start + req_len > phys_len)
1500 		return -EINVAL;
1501 
1502 	/*
1503 	 * Even though we don't make use of the barmap for the mmap,
1504 	 * we need to request the region and the barmap tracks that.
1505 	 */
1506 	if (!vdev->barmap[index]) {
1507 		ret = pci_request_selected_regions(pdev,
1508 						   1 << index, "vfio-pci");
1509 		if (ret)
1510 			return ret;
1511 
1512 		vdev->barmap[index] = pci_iomap(pdev, index, 0);
1513 		if (!vdev->barmap[index]) {
1514 			pci_release_selected_regions(pdev, 1 << index);
1515 			return -ENOMEM;
1516 		}
1517 	}
1518 
1519 	vma->vm_private_data = vdev;
1520 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1521 	vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1522 
1523 	/*
1524 	 * See remap_pfn_range(), called from vfio_pci_fault() but we can't
1525 	 * change vm_flags within the fault handler.  Set them now.
1526 	 */
1527 	vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1528 	vma->vm_ops = &vfio_pci_mmap_ops;
1529 
1530 	return 0;
1531 }
1532 EXPORT_SYMBOL_GPL(vfio_pci_core_mmap);
1533 
vfio_pci_core_request(struct vfio_device * core_vdev,unsigned int count)1534 void vfio_pci_core_request(struct vfio_device *core_vdev, unsigned int count)
1535 {
1536 	struct vfio_pci_core_device *vdev =
1537 		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1538 	struct pci_dev *pdev = vdev->pdev;
1539 
1540 	mutex_lock(&vdev->igate);
1541 
1542 	if (vdev->req_trigger) {
1543 		if (!(count % 10))
1544 			pci_notice_ratelimited(pdev,
1545 				"Relaying device request to user (#%u)\n",
1546 				count);
1547 		eventfd_signal(vdev->req_trigger, 1);
1548 	} else if (count == 0) {
1549 		pci_warn(pdev,
1550 			"No device request channel registered, blocked until released by user\n");
1551 	}
1552 
1553 	mutex_unlock(&vdev->igate);
1554 }
1555 EXPORT_SYMBOL_GPL(vfio_pci_core_request);
1556 
vfio_pci_validate_vf_token(struct vfio_pci_core_device * vdev,bool vf_token,uuid_t * uuid)1557 static int vfio_pci_validate_vf_token(struct vfio_pci_core_device *vdev,
1558 				      bool vf_token, uuid_t *uuid)
1559 {
1560 	/*
1561 	 * There's always some degree of trust or collaboration between SR-IOV
1562 	 * PF and VFs, even if just that the PF hosts the SR-IOV capability and
1563 	 * can disrupt VFs with a reset, but often the PF has more explicit
1564 	 * access to deny service to the VF or access data passed through the
1565 	 * VF.  We therefore require an opt-in via a shared VF token (UUID) to
1566 	 * represent this trust.  This both prevents that a VF driver might
1567 	 * assume the PF driver is a trusted, in-kernel driver, and also that
1568 	 * a PF driver might be replaced with a rogue driver, unknown to in-use
1569 	 * VF drivers.
1570 	 *
1571 	 * Therefore when presented with a VF, if the PF is a vfio device and
1572 	 * it is bound to the vfio-pci driver, the user needs to provide a VF
1573 	 * token to access the device, in the form of appending a vf_token to
1574 	 * the device name, for example:
1575 	 *
1576 	 * "0000:04:10.0 vf_token=bd8d9d2b-5a5f-4f5a-a211-f591514ba1f3"
1577 	 *
1578 	 * When presented with a PF which has VFs in use, the user must also
1579 	 * provide the current VF token to prove collaboration with existing
1580 	 * VF users.  If VFs are not in use, the VF token provided for the PF
1581 	 * device will act to set the VF token.
1582 	 *
1583 	 * If the VF token is provided but unused, an error is generated.
1584 	 */
1585 	if (vdev->pdev->is_virtfn) {
1586 		struct vfio_pci_core_device *pf_vdev = vdev->sriov_pf_core_dev;
1587 		bool match;
1588 
1589 		if (!pf_vdev) {
1590 			if (!vf_token)
1591 				return 0; /* PF is not vfio-pci, no VF token */
1592 
1593 			pci_info_ratelimited(vdev->pdev,
1594 				"VF token incorrectly provided, PF not bound to vfio-pci\n");
1595 			return -EINVAL;
1596 		}
1597 
1598 		if (!vf_token) {
1599 			pci_info_ratelimited(vdev->pdev,
1600 				"VF token required to access device\n");
1601 			return -EACCES;
1602 		}
1603 
1604 		mutex_lock(&pf_vdev->vf_token->lock);
1605 		match = uuid_equal(uuid, &pf_vdev->vf_token->uuid);
1606 		mutex_unlock(&pf_vdev->vf_token->lock);
1607 
1608 		if (!match) {
1609 			pci_info_ratelimited(vdev->pdev,
1610 				"Incorrect VF token provided for device\n");
1611 			return -EACCES;
1612 		}
1613 	} else if (vdev->vf_token) {
1614 		mutex_lock(&vdev->vf_token->lock);
1615 		if (vdev->vf_token->users) {
1616 			if (!vf_token) {
1617 				mutex_unlock(&vdev->vf_token->lock);
1618 				pci_info_ratelimited(vdev->pdev,
1619 					"VF token required to access device\n");
1620 				return -EACCES;
1621 			}
1622 
1623 			if (!uuid_equal(uuid, &vdev->vf_token->uuid)) {
1624 				mutex_unlock(&vdev->vf_token->lock);
1625 				pci_info_ratelimited(vdev->pdev,
1626 					"Incorrect VF token provided for device\n");
1627 				return -EACCES;
1628 			}
1629 		} else if (vf_token) {
1630 			uuid_copy(&vdev->vf_token->uuid, uuid);
1631 		}
1632 
1633 		mutex_unlock(&vdev->vf_token->lock);
1634 	} else if (vf_token) {
1635 		pci_info_ratelimited(vdev->pdev,
1636 			"VF token incorrectly provided, not a PF or VF\n");
1637 		return -EINVAL;
1638 	}
1639 
1640 	return 0;
1641 }
1642 
1643 #define VF_TOKEN_ARG "vf_token="
1644 
vfio_pci_core_match(struct vfio_device * core_vdev,char * buf)1645 int vfio_pci_core_match(struct vfio_device *core_vdev, char *buf)
1646 {
1647 	struct vfio_pci_core_device *vdev =
1648 		container_of(core_vdev, struct vfio_pci_core_device, vdev);
1649 	bool vf_token = false;
1650 	uuid_t uuid;
1651 	int ret;
1652 
1653 	if (strncmp(pci_name(vdev->pdev), buf, strlen(pci_name(vdev->pdev))))
1654 		return 0; /* No match */
1655 
1656 	if (strlen(buf) > strlen(pci_name(vdev->pdev))) {
1657 		buf += strlen(pci_name(vdev->pdev));
1658 
1659 		if (*buf != ' ')
1660 			return 0; /* No match: non-whitespace after name */
1661 
1662 		while (*buf) {
1663 			if (*buf == ' ') {
1664 				buf++;
1665 				continue;
1666 			}
1667 
1668 			if (!vf_token && !strncmp(buf, VF_TOKEN_ARG,
1669 						  strlen(VF_TOKEN_ARG))) {
1670 				buf += strlen(VF_TOKEN_ARG);
1671 
1672 				if (strlen(buf) < UUID_STRING_LEN)
1673 					return -EINVAL;
1674 
1675 				ret = uuid_parse(buf, &uuid);
1676 				if (ret)
1677 					return ret;
1678 
1679 				vf_token = true;
1680 				buf += UUID_STRING_LEN;
1681 			} else {
1682 				/* Unknown/duplicate option */
1683 				return -EINVAL;
1684 			}
1685 		}
1686 	}
1687 
1688 	ret = vfio_pci_validate_vf_token(vdev, vf_token, &uuid);
1689 	if (ret)
1690 		return ret;
1691 
1692 	return 1; /* Match */
1693 }
1694 EXPORT_SYMBOL_GPL(vfio_pci_core_match);
1695 
vfio_pci_bus_notifier(struct notifier_block * nb,unsigned long action,void * data)1696 static int vfio_pci_bus_notifier(struct notifier_block *nb,
1697 				 unsigned long action, void *data)
1698 {
1699 	struct vfio_pci_core_device *vdev = container_of(nb,
1700 						    struct vfio_pci_core_device, nb);
1701 	struct device *dev = data;
1702 	struct pci_dev *pdev = to_pci_dev(dev);
1703 	struct pci_dev *physfn = pci_physfn(pdev);
1704 
1705 	if (action == BUS_NOTIFY_ADD_DEVICE &&
1706 	    pdev->is_virtfn && physfn == vdev->pdev) {
1707 		pci_info(vdev->pdev, "Captured SR-IOV VF %s driver_override\n",
1708 			 pci_name(pdev));
1709 		pdev->driver_override = kasprintf(GFP_KERNEL, "%s",
1710 						  vdev->vdev.ops->name);
1711 	} else if (action == BUS_NOTIFY_BOUND_DRIVER &&
1712 		   pdev->is_virtfn && physfn == vdev->pdev) {
1713 		struct pci_driver *drv = pci_dev_driver(pdev);
1714 
1715 		if (drv && drv != pci_dev_driver(vdev->pdev))
1716 			pci_warn(vdev->pdev,
1717 				 "VF %s bound to driver %s while PF bound to driver %s\n",
1718 				 pci_name(pdev), drv->name,
1719 				 pci_dev_driver(vdev->pdev)->name);
1720 	}
1721 
1722 	return 0;
1723 }
1724 
vfio_pci_vf_init(struct vfio_pci_core_device * vdev)1725 static int vfio_pci_vf_init(struct vfio_pci_core_device *vdev)
1726 {
1727 	struct pci_dev *pdev = vdev->pdev;
1728 	struct vfio_pci_core_device *cur;
1729 	struct pci_dev *physfn;
1730 	int ret;
1731 
1732 	if (pdev->is_virtfn) {
1733 		/*
1734 		 * If this VF was created by our vfio_pci_core_sriov_configure()
1735 		 * then we can find the PF vfio_pci_core_device now, and due to
1736 		 * the locking in pci_disable_sriov() it cannot change until
1737 		 * this VF device driver is removed.
1738 		 */
1739 		physfn = pci_physfn(vdev->pdev);
1740 		mutex_lock(&vfio_pci_sriov_pfs_mutex);
1741 		list_for_each_entry(cur, &vfio_pci_sriov_pfs, sriov_pfs_item) {
1742 			if (cur->pdev == physfn) {
1743 				vdev->sriov_pf_core_dev = cur;
1744 				break;
1745 			}
1746 		}
1747 		mutex_unlock(&vfio_pci_sriov_pfs_mutex);
1748 		return 0;
1749 	}
1750 
1751 	/* Not a SRIOV PF */
1752 	if (!pdev->is_physfn)
1753 		return 0;
1754 
1755 	vdev->vf_token = kzalloc(sizeof(*vdev->vf_token), GFP_KERNEL);
1756 	if (!vdev->vf_token)
1757 		return -ENOMEM;
1758 
1759 	mutex_init(&vdev->vf_token->lock);
1760 	uuid_gen(&vdev->vf_token->uuid);
1761 
1762 	vdev->nb.notifier_call = vfio_pci_bus_notifier;
1763 	ret = bus_register_notifier(&pci_bus_type, &vdev->nb);
1764 	if (ret) {
1765 		kfree(vdev->vf_token);
1766 		return ret;
1767 	}
1768 	return 0;
1769 }
1770 
vfio_pci_vf_uninit(struct vfio_pci_core_device * vdev)1771 static void vfio_pci_vf_uninit(struct vfio_pci_core_device *vdev)
1772 {
1773 	if (!vdev->vf_token)
1774 		return;
1775 
1776 	bus_unregister_notifier(&pci_bus_type, &vdev->nb);
1777 	WARN_ON(vdev->vf_token->users);
1778 	mutex_destroy(&vdev->vf_token->lock);
1779 	kfree(vdev->vf_token);
1780 }
1781 
vfio_pci_vga_init(struct vfio_pci_core_device * vdev)1782 static int vfio_pci_vga_init(struct vfio_pci_core_device *vdev)
1783 {
1784 	struct pci_dev *pdev = vdev->pdev;
1785 	int ret;
1786 
1787 	if (!vfio_pci_is_vga(pdev))
1788 		return 0;
1789 
1790 	ret = vga_client_register(pdev, vfio_pci_set_decode);
1791 	if (ret)
1792 		return ret;
1793 	vga_set_legacy_decoding(pdev, vfio_pci_set_decode(pdev, false));
1794 	return 0;
1795 }
1796 
vfio_pci_vga_uninit(struct vfio_pci_core_device * vdev)1797 static void vfio_pci_vga_uninit(struct vfio_pci_core_device *vdev)
1798 {
1799 	struct pci_dev *pdev = vdev->pdev;
1800 
1801 	if (!vfio_pci_is_vga(pdev))
1802 		return;
1803 	vga_client_unregister(pdev);
1804 	vga_set_legacy_decoding(pdev, VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
1805 					      VGA_RSRC_LEGACY_IO |
1806 					      VGA_RSRC_LEGACY_MEM);
1807 }
1808 
vfio_pci_core_init_device(struct vfio_pci_core_device * vdev,struct pci_dev * pdev,const struct vfio_device_ops * vfio_pci_ops)1809 void vfio_pci_core_init_device(struct vfio_pci_core_device *vdev,
1810 			       struct pci_dev *pdev,
1811 			       const struct vfio_device_ops *vfio_pci_ops)
1812 {
1813 	vfio_init_group_dev(&vdev->vdev, &pdev->dev, vfio_pci_ops);
1814 	vdev->pdev = pdev;
1815 	vdev->irq_type = VFIO_PCI_NUM_IRQS;
1816 	mutex_init(&vdev->igate);
1817 	spin_lock_init(&vdev->irqlock);
1818 	mutex_init(&vdev->ioeventfds_lock);
1819 	INIT_LIST_HEAD(&vdev->dummy_resources_list);
1820 	INIT_LIST_HEAD(&vdev->ioeventfds_list);
1821 	mutex_init(&vdev->vma_lock);
1822 	INIT_LIST_HEAD(&vdev->vma_list);
1823 	INIT_LIST_HEAD(&vdev->sriov_pfs_item);
1824 	init_rwsem(&vdev->memory_lock);
1825 }
1826 EXPORT_SYMBOL_GPL(vfio_pci_core_init_device);
1827 
vfio_pci_core_uninit_device(struct vfio_pci_core_device * vdev)1828 void vfio_pci_core_uninit_device(struct vfio_pci_core_device *vdev)
1829 {
1830 	mutex_destroy(&vdev->igate);
1831 	mutex_destroy(&vdev->ioeventfds_lock);
1832 	mutex_destroy(&vdev->vma_lock);
1833 	vfio_uninit_group_dev(&vdev->vdev);
1834 	kfree(vdev->region);
1835 	kfree(vdev->pm_save);
1836 }
1837 EXPORT_SYMBOL_GPL(vfio_pci_core_uninit_device);
1838 
vfio_pci_core_register_device(struct vfio_pci_core_device * vdev)1839 int vfio_pci_core_register_device(struct vfio_pci_core_device *vdev)
1840 {
1841 	struct pci_dev *pdev = vdev->pdev;
1842 	struct iommu_group *group;
1843 	int ret;
1844 
1845 	if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1846 		return -EINVAL;
1847 
1848 	/*
1849 	 * Prevent binding to PFs with VFs enabled, the VFs might be in use
1850 	 * by the host or other users.  We cannot capture the VFs if they
1851 	 * already exist, nor can we track VF users.  Disabling SR-IOV here
1852 	 * would initiate removing the VFs, which would unbind the driver,
1853 	 * which is prone to blocking if that VF is also in use by vfio-pci.
1854 	 * Just reject these PFs and let the user sort it out.
1855 	 */
1856 	if (pci_num_vf(pdev)) {
1857 		pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
1858 		return -EBUSY;
1859 	}
1860 
1861 	group = vfio_iommu_group_get(&pdev->dev);
1862 	if (!group)
1863 		return -EINVAL;
1864 
1865 	if (pci_is_root_bus(pdev->bus)) {
1866 		ret = vfio_assign_device_set(&vdev->vdev, vdev);
1867 	} else if (!pci_probe_reset_slot(pdev->slot)) {
1868 		ret = vfio_assign_device_set(&vdev->vdev, pdev->slot);
1869 	} else {
1870 		/*
1871 		 * If there is no slot reset support for this device, the whole
1872 		 * bus needs to be grouped together to support bus-wide resets.
1873 		 */
1874 		ret = vfio_assign_device_set(&vdev->vdev, pdev->bus);
1875 	}
1876 
1877 	if (ret)
1878 		goto out_group_put;
1879 	ret = vfio_pci_vf_init(vdev);
1880 	if (ret)
1881 		goto out_group_put;
1882 	ret = vfio_pci_vga_init(vdev);
1883 	if (ret)
1884 		goto out_vf;
1885 
1886 	vfio_pci_probe_power_state(vdev);
1887 
1888 	if (!disable_idle_d3) {
1889 		/*
1890 		 * pci-core sets the device power state to an unknown value at
1891 		 * bootup and after being removed from a driver.  The only
1892 		 * transition it allows from this unknown state is to D0, which
1893 		 * typically happens when a driver calls pci_enable_device().
1894 		 * We're not ready to enable the device yet, but we do want to
1895 		 * be able to get to D3.  Therefore first do a D0 transition
1896 		 * before going to D3.
1897 		 */
1898 		vfio_pci_set_power_state(vdev, PCI_D0);
1899 		vfio_pci_set_power_state(vdev, PCI_D3hot);
1900 	}
1901 
1902 	ret = vfio_register_group_dev(&vdev->vdev);
1903 	if (ret)
1904 		goto out_power;
1905 	return 0;
1906 
1907 out_power:
1908 	if (!disable_idle_d3)
1909 		vfio_pci_set_power_state(vdev, PCI_D0);
1910 out_vf:
1911 	vfio_pci_vf_uninit(vdev);
1912 out_group_put:
1913 	vfio_iommu_group_put(group, &pdev->dev);
1914 	return ret;
1915 }
1916 EXPORT_SYMBOL_GPL(vfio_pci_core_register_device);
1917 
vfio_pci_core_unregister_device(struct vfio_pci_core_device * vdev)1918 void vfio_pci_core_unregister_device(struct vfio_pci_core_device *vdev)
1919 {
1920 	struct pci_dev *pdev = vdev->pdev;
1921 
1922 	vfio_pci_core_sriov_configure(pdev, 0);
1923 
1924 	vfio_unregister_group_dev(&vdev->vdev);
1925 
1926 	vfio_pci_vf_uninit(vdev);
1927 	vfio_pci_vga_uninit(vdev);
1928 
1929 	vfio_iommu_group_put(pdev->dev.iommu_group, &pdev->dev);
1930 
1931 	if (!disable_idle_d3)
1932 		vfio_pci_set_power_state(vdev, PCI_D0);
1933 }
1934 EXPORT_SYMBOL_GPL(vfio_pci_core_unregister_device);
1935 
vfio_pci_aer_err_detected(struct pci_dev * pdev,pci_channel_state_t state)1936 static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
1937 						  pci_channel_state_t state)
1938 {
1939 	struct vfio_pci_core_device *vdev;
1940 	struct vfio_device *device;
1941 
1942 	device = vfio_device_get_from_dev(&pdev->dev);
1943 	if (device == NULL)
1944 		return PCI_ERS_RESULT_DISCONNECT;
1945 
1946 	vdev = container_of(device, struct vfio_pci_core_device, vdev);
1947 
1948 	mutex_lock(&vdev->igate);
1949 
1950 	if (vdev->err_trigger)
1951 		eventfd_signal(vdev->err_trigger, 1);
1952 
1953 	mutex_unlock(&vdev->igate);
1954 
1955 	vfio_device_put(device);
1956 
1957 	return PCI_ERS_RESULT_CAN_RECOVER;
1958 }
1959 
vfio_pci_core_sriov_configure(struct pci_dev * pdev,int nr_virtfn)1960 int vfio_pci_core_sriov_configure(struct pci_dev *pdev, int nr_virtfn)
1961 {
1962 	struct vfio_pci_core_device *vdev;
1963 	struct vfio_device *device;
1964 	int ret = 0;
1965 
1966 	device_lock_assert(&pdev->dev);
1967 
1968 	device = vfio_device_get_from_dev(&pdev->dev);
1969 	if (!device)
1970 		return -ENODEV;
1971 
1972 	vdev = container_of(device, struct vfio_pci_core_device, vdev);
1973 
1974 	if (nr_virtfn) {
1975 		mutex_lock(&vfio_pci_sriov_pfs_mutex);
1976 		/*
1977 		 * The thread that adds the vdev to the list is the only thread
1978 		 * that gets to call pci_enable_sriov() and we will only allow
1979 		 * it to be called once without going through
1980 		 * pci_disable_sriov()
1981 		 */
1982 		if (!list_empty(&vdev->sriov_pfs_item)) {
1983 			ret = -EINVAL;
1984 			goto out_unlock;
1985 		}
1986 		list_add_tail(&vdev->sriov_pfs_item, &vfio_pci_sriov_pfs);
1987 		mutex_unlock(&vfio_pci_sriov_pfs_mutex);
1988 		ret = pci_enable_sriov(pdev, nr_virtfn);
1989 		if (ret)
1990 			goto out_del;
1991 		ret = nr_virtfn;
1992 		goto out_put;
1993 	}
1994 
1995 	pci_disable_sriov(pdev);
1996 
1997 out_del:
1998 	mutex_lock(&vfio_pci_sriov_pfs_mutex);
1999 	list_del_init(&vdev->sriov_pfs_item);
2000 out_unlock:
2001 	mutex_unlock(&vfio_pci_sriov_pfs_mutex);
2002 out_put:
2003 	vfio_device_put(device);
2004 	return ret;
2005 }
2006 EXPORT_SYMBOL_GPL(vfio_pci_core_sriov_configure);
2007 
2008 const struct pci_error_handlers vfio_pci_core_err_handlers = {
2009 	.error_detected = vfio_pci_aer_err_detected,
2010 };
2011 EXPORT_SYMBOL_GPL(vfio_pci_core_err_handlers);
2012 
vfio_dev_in_groups(struct vfio_pci_core_device * vdev,struct vfio_pci_group_info * groups)2013 static bool vfio_dev_in_groups(struct vfio_pci_core_device *vdev,
2014 			       struct vfio_pci_group_info *groups)
2015 {
2016 	unsigned int i;
2017 
2018 	for (i = 0; i < groups->count; i++)
2019 		if (groups->groups[i] == vdev->vdev.group)
2020 			return true;
2021 	return false;
2022 }
2023 
vfio_pci_is_device_in_set(struct pci_dev * pdev,void * data)2024 static int vfio_pci_is_device_in_set(struct pci_dev *pdev, void *data)
2025 {
2026 	struct vfio_device_set *dev_set = data;
2027 	struct vfio_device *cur;
2028 
2029 	list_for_each_entry(cur, &dev_set->device_list, dev_set_list)
2030 		if (cur->dev == &pdev->dev)
2031 			return 0;
2032 	return -EBUSY;
2033 }
2034 
2035 /*
2036  * vfio-core considers a group to be viable and will create a vfio_device even
2037  * if some devices are bound to drivers like pci-stub or pcieport. Here we
2038  * require all PCI devices to be inside our dev_set since that ensures they stay
2039  * put and that every driver controlling the device can co-ordinate with the
2040  * device reset.
2041  *
2042  * Returns the pci_dev to pass to pci_reset_bus() if every PCI device to be
2043  * reset is inside the dev_set, and pci_reset_bus() can succeed. NULL otherwise.
2044  */
2045 static struct pci_dev *
vfio_pci_dev_set_resettable(struct vfio_device_set * dev_set)2046 vfio_pci_dev_set_resettable(struct vfio_device_set *dev_set)
2047 {
2048 	struct pci_dev *pdev;
2049 
2050 	lockdep_assert_held(&dev_set->lock);
2051 
2052 	/*
2053 	 * By definition all PCI devices in the dev_set share the same PCI
2054 	 * reset, so any pci_dev will have the same outcomes for
2055 	 * pci_probe_reset_*() and pci_reset_bus().
2056 	 */
2057 	pdev = list_first_entry(&dev_set->device_list,
2058 				struct vfio_pci_core_device,
2059 				vdev.dev_set_list)->pdev;
2060 
2061 	/* pci_reset_bus() is supported */
2062 	if (pci_probe_reset_slot(pdev->slot) && pci_probe_reset_bus(pdev->bus))
2063 		return NULL;
2064 
2065 	if (vfio_pci_for_each_slot_or_bus(pdev, vfio_pci_is_device_in_set,
2066 					  dev_set,
2067 					  !pci_probe_reset_slot(pdev->slot)))
2068 		return NULL;
2069 	return pdev;
2070 }
2071 
2072 /*
2073  * We need to get memory_lock for each device, but devices can share mmap_lock,
2074  * therefore we need to zap and hold the vma_lock for each device, and only then
2075  * get each memory_lock.
2076  */
vfio_pci_dev_set_hot_reset(struct vfio_device_set * dev_set,struct vfio_pci_group_info * groups)2077 static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set,
2078 				      struct vfio_pci_group_info *groups)
2079 {
2080 	struct vfio_pci_core_device *cur_mem;
2081 	struct vfio_pci_core_device *cur_vma;
2082 	struct vfio_pci_core_device *cur;
2083 	struct pci_dev *pdev;
2084 	bool is_mem = true;
2085 	int ret;
2086 
2087 	mutex_lock(&dev_set->lock);
2088 	cur_mem = list_first_entry(&dev_set->device_list,
2089 				   struct vfio_pci_core_device,
2090 				   vdev.dev_set_list);
2091 
2092 	pdev = vfio_pci_dev_set_resettable(dev_set);
2093 	if (!pdev) {
2094 		ret = -EINVAL;
2095 		goto err_unlock;
2096 	}
2097 
2098 	list_for_each_entry(cur_vma, &dev_set->device_list, vdev.dev_set_list) {
2099 		/*
2100 		 * Test whether all the affected devices are contained by the
2101 		 * set of groups provided by the user.
2102 		 */
2103 		if (!vfio_dev_in_groups(cur_vma, groups)) {
2104 			ret = -EINVAL;
2105 			goto err_undo;
2106 		}
2107 
2108 		/*
2109 		 * Locking multiple devices is prone to deadlock, runaway and
2110 		 * unwind if we hit contention.
2111 		 */
2112 		if (!vfio_pci_zap_and_vma_lock(cur_vma, true)) {
2113 			ret = -EBUSY;
2114 			goto err_undo;
2115 		}
2116 	}
2117 	cur_vma = NULL;
2118 
2119 	list_for_each_entry(cur_mem, &dev_set->device_list, vdev.dev_set_list) {
2120 		if (!down_write_trylock(&cur_mem->memory_lock)) {
2121 			ret = -EBUSY;
2122 			goto err_undo;
2123 		}
2124 		mutex_unlock(&cur_mem->vma_lock);
2125 	}
2126 	cur_mem = NULL;
2127 
2128 	/*
2129 	 * The pci_reset_bus() will reset all the devices in the bus.
2130 	 * The power state can be non-D0 for some of the devices in the bus.
2131 	 * For these devices, the pci_reset_bus() will internally set
2132 	 * the power state to D0 without vfio driver involvement.
2133 	 * For the devices which have NoSoftRst-, the reset function can
2134 	 * cause the PCI config space reset without restoring the original
2135 	 * state (saved locally in 'vdev->pm_save').
2136 	 */
2137 	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list)
2138 		vfio_pci_set_power_state(cur, PCI_D0);
2139 
2140 	ret = pci_reset_bus(pdev);
2141 
2142 err_undo:
2143 	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2144 		if (cur == cur_mem)
2145 			is_mem = false;
2146 		if (cur == cur_vma)
2147 			break;
2148 		if (is_mem)
2149 			up_write(&cur->memory_lock);
2150 		else
2151 			mutex_unlock(&cur->vma_lock);
2152 	}
2153 err_unlock:
2154 	mutex_unlock(&dev_set->lock);
2155 	return ret;
2156 }
2157 
vfio_pci_dev_set_needs_reset(struct vfio_device_set * dev_set)2158 static bool vfio_pci_dev_set_needs_reset(struct vfio_device_set *dev_set)
2159 {
2160 	struct vfio_pci_core_device *cur;
2161 	bool needs_reset = false;
2162 
2163 	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2164 		/* No VFIO device in the set can have an open device FD */
2165 		if (cur->vdev.open_count)
2166 			return false;
2167 		needs_reset |= cur->needs_reset;
2168 	}
2169 	return needs_reset;
2170 }
2171 
2172 /*
2173  * If a bus or slot reset is available for the provided dev_set and:
2174  *  - All of the devices affected by that bus or slot reset are unused
2175  *  - At least one of the affected devices is marked dirty via
2176  *    needs_reset (such as by lack of FLR support)
2177  * Then attempt to perform that bus or slot reset.
2178  * Returns true if the dev_set was reset.
2179  */
vfio_pci_dev_set_try_reset(struct vfio_device_set * dev_set)2180 static bool vfio_pci_dev_set_try_reset(struct vfio_device_set *dev_set)
2181 {
2182 	struct vfio_pci_core_device *cur;
2183 	struct pci_dev *pdev;
2184 	int ret;
2185 
2186 	if (!vfio_pci_dev_set_needs_reset(dev_set))
2187 		return false;
2188 
2189 	pdev = vfio_pci_dev_set_resettable(dev_set);
2190 	if (!pdev)
2191 		return false;
2192 
2193 	/*
2194 	 * The pci_reset_bus() will reset all the devices in the bus.
2195 	 * The power state can be non-D0 for some of the devices in the bus.
2196 	 * For these devices, the pci_reset_bus() will internally set
2197 	 * the power state to D0 without vfio driver involvement.
2198 	 * For the devices which have NoSoftRst-, the reset function can
2199 	 * cause the PCI config space reset without restoring the original
2200 	 * state (saved locally in 'vdev->pm_save').
2201 	 */
2202 	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list)
2203 		vfio_pci_set_power_state(cur, PCI_D0);
2204 
2205 	ret = pci_reset_bus(pdev);
2206 	if (ret)
2207 		return false;
2208 
2209 	list_for_each_entry(cur, &dev_set->device_list, vdev.dev_set_list) {
2210 		cur->needs_reset = false;
2211 		if (!disable_idle_d3)
2212 			vfio_pci_set_power_state(cur, PCI_D3hot);
2213 	}
2214 	return true;
2215 }
2216 
vfio_pci_core_set_params(bool is_nointxmask,bool is_disable_vga,bool is_disable_idle_d3)2217 void vfio_pci_core_set_params(bool is_nointxmask, bool is_disable_vga,
2218 			      bool is_disable_idle_d3)
2219 {
2220 	nointxmask = is_nointxmask;
2221 	disable_vga = is_disable_vga;
2222 	disable_idle_d3 = is_disable_idle_d3;
2223 }
2224 EXPORT_SYMBOL_GPL(vfio_pci_core_set_params);
2225 
vfio_pci_core_cleanup(void)2226 static void vfio_pci_core_cleanup(void)
2227 {
2228 	vfio_pci_uninit_perm_bits();
2229 }
2230 
vfio_pci_core_init(void)2231 static int __init vfio_pci_core_init(void)
2232 {
2233 	/* Allocate shared config space permission data used by all devices */
2234 	return vfio_pci_init_perm_bits();
2235 }
2236 
2237 module_init(vfio_pci_core_init);
2238 module_exit(vfio_pci_core_cleanup);
2239 
2240 MODULE_LICENSE("GPL v2");
2241 MODULE_AUTHOR(DRIVER_AUTHOR);
2242 MODULE_DESCRIPTION(DRIVER_DESC);
2243