?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
treewide: remove redundant IS_ERR() before error code check
[linux/fpc-iii.git] / drivers / vfio / pci / vfio_pci.c
blob379a02c36e3770b653594d61d8e781253d09595d
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 #define dev_fmt pr_fmt
13
14 #include <linux/device.h>
15 #include <linux/eventfd.h>
16 #include <linux/file.h>
17 #include <linux/interrupt.h>
18 #include <linux/iommu.h>
19 #include <linux/module.h>
20 #include <linux/mutex.h>
21 #include <linux/notifier.h>
22 #include <linux/pci.h>
23 #include <linux/pm_runtime.h>
24 #include <linux/slab.h>
25 #include <linux/types.h>
26 #include <linux/uaccess.h>
27 #include <linux/vfio.h>
28 #include <linux/vgaarb.h>
29 #include <linux/nospec.h>
30
31 #include "vfio_pci_private.h"
32
33 #define DRIVER_VERSION "0.2"
34 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
35 #define DRIVER_DESC "VFIO PCI - User Level meta-driver"
36
37 static char ids[1024] __initdata;
38 module_param_string(ids, ids, sizeof(ids), 0);
39 MODULE_PARM_DESC(ids, "Initial PCI IDs to add to the vfio driver, format is \"vendor:device[:subvendor[:subdevice[:class[:class_mask]]]]\" and multiple comma separated entries can be specified");
40
41 static bool nointxmask;
42 module_param_named(nointxmask, nointxmask, bool, S_IRUGO | S_IWUSR);
43 MODULE_PARM_DESC(nointxmask,
44 "Disable support for PCI 2.3 style INTx masking. If this resolves problems for specific devices, report lspci -vvvxxx to linux-pci@vger.kernel.org so the device can be fixed automatically via the broken_intx_masking flag.");
45
46 #ifdef CONFIG_VFIO_PCI_VGA
47 static bool disable_vga;
48 module_param(disable_vga, bool, S_IRUGO);
49 MODULE_PARM_DESC(disable_vga, "Disable VGA resource access through vfio-pci");
50 #endif
51
52 static bool disable_idle_d3;
53 module_param(disable_idle_d3, bool, S_IRUGO | S_IWUSR);
54 MODULE_PARM_DESC(disable_idle_d3,
55 "Disable using the PCI D3 low power state for idle, unused devices");
56
57 static inline bool vfio_vga_disabled(void)
58 {
59 #ifdef CONFIG_VFIO_PCI_VGA
60 return disable_vga;
61 #else
62 return true;
63 #endif
64 }
65
66 /*
67 * Our VGA arbiter participation is limited since we don't know anything
68 * about the device itself. However, if the device is the only VGA device
69 * downstream of a bridge and VFIO VGA support is disabled, then we can
70 * safely return legacy VGA IO and memory as not decoded since the user
71 * has no way to get to it and routing can be disabled externally at the
72 * bridge.
73 */
74 static unsigned int vfio_pci_set_vga_decode(void *opaque, bool single_vga)
75 {
76 struct vfio_pci_device *vdev = opaque;
77 struct pci_dev *tmp = NULL, *pdev = vdev->pdev;
78 unsigned char max_busnr;
79 unsigned int decodes;
80
81 if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
82 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
83 VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
84
85 max_busnr = pci_bus_max_busnr(pdev->bus);
86 decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
87
88 while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
89 if (tmp == pdev ||
90 pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
91 pci_is_root_bus(tmp->bus))
92 continue;
93
94 if (tmp->bus->number >= pdev->bus->number &&
95 tmp->bus->number <= max_busnr) {
96 pci_dev_put(tmp);
97 decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
98 break;
99 }
100 }
101
102 return decodes;
103 }
104
105 static inline bool vfio_pci_is_vga(struct pci_dev *pdev)
106 {
107 return (pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA;
108 }
109
110 static void vfio_pci_probe_mmaps(struct vfio_pci_device *vdev)
111 {
112 struct resource *res;
113 int i;
114 struct vfio_pci_dummy_resource *dummy_res;
115
116 INIT_LIST_HEAD(&vdev->dummy_resources_list);
117
118 for (i = 0; i < PCI_STD_NUM_BARS; i++) {
119 int bar = i + PCI_STD_RESOURCES;
120
121 res = &vdev->pdev->resource[bar];
122
123 if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
124 goto no_mmap;
125
126 if (!(res->flags & IORESOURCE_MEM))
127 goto no_mmap;
128
129 /*
130 * The PCI core shouldn't set up a resource with a
131 * type but zero size. But there may be bugs that
132 * cause us to do that.
133 */
134 if (!resource_size(res))
135 goto no_mmap;
136
137 if (resource_size(res) >= PAGE_SIZE) {
138 vdev->bar_mmap_supported[bar] = true;
139 continue;
140 }
141
142 if (!(res->start & ~PAGE_MASK)) {
143 /*
144 * Add a dummy resource to reserve the remainder
145 * of the exclusive page in case that hot-add
146 * device's bar is assigned into it.
147 */
148 dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
149 if (dummy_res == NULL)
150 goto no_mmap;
151
152 dummy_res->resource.name = "vfio sub-page reserved";
153 dummy_res->resource.start = res->end + 1;
154 dummy_res->resource.end = res->start + PAGE_SIZE - 1;
155 dummy_res->resource.flags = res->flags;
156 if (request_resource(res->parent,
157 &dummy_res->resource)) {
158 kfree(dummy_res);
159 goto no_mmap;
160 }
161 dummy_res->index = bar;
162 list_add(&dummy_res->res_next,
163 &vdev->dummy_resources_list);
164 vdev->bar_mmap_supported[bar] = true;
165 continue;
166 }
167 /*
168 * Here we don't handle the case when the BAR is not page
169 * aligned because we can't expect the BAR will be
170 * assigned into the same location in a page in guest
171 * when we passthrough the BAR. And it's hard to access
172 * this BAR in userspace because we have no way to get
173 * the BAR's location in a page.
174 */
175 no_mmap:
176 vdev->bar_mmap_supported[bar] = false;
177 }
178 }
179
180 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev);
181 static void vfio_pci_disable(struct vfio_pci_device *vdev);
182
183 /*
184 * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
185 * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
186 * If a device implements the former but not the latter we would typically
187 * expect broken_intx_masking be set and require an exclusive interrupt.
188 * However since we do have control of the device's ability to assert INTx,
189 * we can instead pretend that the device does not implement INTx, virtualizing
190 * the pin register to report zero and maintaining DisINTx set on the host.
191 */
192 static bool vfio_pci_nointx(struct pci_dev *pdev)
193 {
194 switch (pdev->vendor) {
195 case PCI_VENDOR_ID_INTEL:
196 switch (pdev->device) {
197 /* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
198 case 0x1572:
199 case 0x1574:
200 case 0x1580 ... 0x1581:
201 case 0x1583 ... 0x158b:
202 case 0x37d0 ... 0x37d2:
203 return true;
204 default:
205 return false;
206 }
207 }
208
209 return false;
210 }
211
212 static void vfio_pci_probe_power_state(struct vfio_pci_device *vdev)
213 {
214 struct pci_dev *pdev = vdev->pdev;
215 u16 pmcsr;
216
217 if (!pdev->pm_cap)
218 return;
219
220 pci_read_config_word(pdev, pdev->pm_cap + PCI_PM_CTRL, &pmcsr);
221
222 vdev->needs_pm_restore = !(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
223 }
224
225 /*
226 * pci_set_power_state() wrapper handling devices which perform a soft reset on
227 * D3->D0 transition. Save state prior to D0/1/2->D3, stash it on the vdev,
228 * restore when returned to D0. Saved separately from pci_saved_state for use
229 * by PM capability emulation and separately from pci_dev internal saved state
230 * to avoid it being overwritten and consumed around other resets.
231 */
232 int vfio_pci_set_power_state(struct vfio_pci_device *vdev, pci_power_t state)
233 {
234 struct pci_dev *pdev = vdev->pdev;
235 bool needs_restore = false, needs_save = false;
236 int ret;
237
238 if (vdev->needs_pm_restore) {
239 if (pdev->current_state < PCI_D3hot && state >= PCI_D3hot) {
240 pci_save_state(pdev);
241 needs_save = true;
242 }
243
244 if (pdev->current_state >= PCI_D3hot && state <= PCI_D0)
245 needs_restore = true;
246 }
247
248 ret = pci_set_power_state(pdev, state);
249
250 if (!ret) {
251 /* D3 might be unsupported via quirk, skip unless in D3 */
252 if (needs_save && pdev->current_state >= PCI_D3hot) {
253 vdev->pm_save = pci_store_saved_state(pdev);
254 } else if (needs_restore) {
255 pci_load_and_free_saved_state(pdev, &vdev->pm_save);
256 pci_restore_state(pdev);
257 }
258 }
259
260 return ret;
261 }
262
263 static int vfio_pci_enable(struct vfio_pci_device *vdev)
264 {
265 struct pci_dev *pdev = vdev->pdev;
266 int ret;
267 u16 cmd;
268 u8 msix_pos;
269
270 vfio_pci_set_power_state(vdev, PCI_D0);
271
272 /* Don't allow our initial saved state to include busmaster */
273 pci_clear_master(pdev);
274
275 ret = pci_enable_device(pdev);
276 if (ret)
277 return ret;
278
279 /* If reset fails because of the device lock, fail this path entirely */
280 ret = pci_try_reset_function(pdev);
281 if (ret == -EAGAIN) {
282 pci_disable_device(pdev);
283 return ret;
284 }
285
286 vdev->reset_works = !ret;
287 pci_save_state(pdev);
288 vdev->pci_saved_state = pci_store_saved_state(pdev);
289 if (!vdev->pci_saved_state)
290 pci_dbg(pdev, "%s: Couldn't store saved state\n", __func__);
291
292 if (likely(!nointxmask)) {
293 if (vfio_pci_nointx(pdev)) {
294 pci_info(pdev, "Masking broken INTx support\n");
295 vdev->nointx = true;
296 pci_intx(pdev, 0);
297 } else
298 vdev->pci_2_3 = pci_intx_mask_supported(pdev);
299 }
300
301 pci_read_config_word(pdev, PCI_COMMAND, &cmd);
302 if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
303 cmd &= ~PCI_COMMAND_INTX_DISABLE;
304 pci_write_config_word(pdev, PCI_COMMAND, cmd);
305 }
306
307 ret = vfio_config_init(vdev);
308 if (ret) {
309 kfree(vdev->pci_saved_state);
310 vdev->pci_saved_state = NULL;
311 pci_disable_device(pdev);
312 return ret;
313 }
314
315 msix_pos = pdev->msix_cap;
316 if (msix_pos) {
317 u16 flags;
318 u32 table;
319
320 pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
321 pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
322
323 vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
324 vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
325 vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
326 } else
327 vdev->msix_bar = 0xFF;
328
329 if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
330 vdev->has_vga = true;
331
332
333 if (vfio_pci_is_vga(pdev) &&
334 pdev->vendor == PCI_VENDOR_ID_INTEL &&
335 IS_ENABLED(CONFIG_VFIO_PCI_IGD)) {
336 ret = vfio_pci_igd_init(vdev);
337 if (ret) {
338 pci_warn(pdev, "Failed to setup Intel IGD regions\n");
339 goto disable_exit;
340 }
341 }
342
343 if (pdev->vendor == PCI_VENDOR_ID_NVIDIA &&
344 IS_ENABLED(CONFIG_VFIO_PCI_NVLINK2)) {
345 ret = vfio_pci_nvdia_v100_nvlink2_init(vdev);
346 if (ret && ret != -ENODEV) {
347 pci_warn(pdev, "Failed to setup NVIDIA NV2 RAM region\n");
348 goto disable_exit;
349 }
350 }
351
352 if (pdev->vendor == PCI_VENDOR_ID_IBM &&
353 IS_ENABLED(CONFIG_VFIO_PCI_NVLINK2)) {
354 ret = vfio_pci_ibm_npu2_init(vdev);
355 if (ret && ret != -ENODEV) {
356 pci_warn(pdev, "Failed to setup NVIDIA NV2 ATSD region\n");
357 goto disable_exit;
358 }
359 }
360
361 vfio_pci_probe_mmaps(vdev);
362
363 return 0;
364
365 disable_exit:
366 vfio_pci_disable(vdev);
367 return ret;
368 }
369
370 static void vfio_pci_disable(struct vfio_pci_device *vdev)
371 {
372 struct pci_dev *pdev = vdev->pdev;
373 struct vfio_pci_dummy_resource *dummy_res, *tmp;
374 struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
375 int i, bar;
376
377 /* Stop the device from further DMA */
378 pci_clear_master(pdev);
379
380 vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
381 VFIO_IRQ_SET_ACTION_TRIGGER,
382 vdev->irq_type, 0, 0, NULL);
383
384 /* Device closed, don't need mutex here */
385 list_for_each_entry_safe(ioeventfd, ioeventfd_tmp,
386 &vdev->ioeventfds_list, next) {
387 vfio_virqfd_disable(&ioeventfd->virqfd);
388 list_del(&ioeventfd->next);
389 kfree(ioeventfd);
390 }
391 vdev->ioeventfds_nr = 0;
392
393 vdev->virq_disabled = false;
394
395 for (i = 0; i < vdev->num_regions; i++)
396 vdev->region[i].ops->release(vdev, &vdev->region[i]);
397
398 vdev->num_regions = 0;
399 kfree(vdev->region);
400 vdev->region = NULL; /* don't krealloc a freed pointer */
401
402 vfio_config_free(vdev);
403
404 for (i = 0; i < PCI_STD_NUM_BARS; i++) {
405 bar = i + PCI_STD_RESOURCES;
406 if (!vdev->barmap[bar])
407 continue;
408 pci_iounmap(pdev, vdev->barmap[bar]);
409 pci_release_selected_regions(pdev, 1 << bar);
410 vdev->barmap[bar] = NULL;
411 }
412
413 list_for_each_entry_safe(dummy_res, tmp,
414 &vdev->dummy_resources_list, res_next) {
415 list_del(&dummy_res->res_next);
416 release_resource(&dummy_res->resource);
417 kfree(dummy_res);
418 }
419
420 vdev->needs_reset = true;
421
422 /*
423 * If we have saved state, restore it. If we can reset the device,
424 * even better. Resetting with current state seems better than
425 * nothing, but saving and restoring current state without reset
426 * is just busy work.
427 */
428 if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
429 pci_info(pdev, "%s: Couldn't reload saved state\n", __func__);
430
431 if (!vdev->reset_works)
432 goto out;
433
434 pci_save_state(pdev);
435 }
436
437 /*
438 * Disable INTx and MSI, presumably to avoid spurious interrupts
439 * during reset. Stolen from pci_reset_function()
440 */
441 pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
442
443 /*
444 * Try to get the locks ourselves to prevent a deadlock. The
445 * success of this is dependent on being able to lock the device,
446 * which is not always possible.
447 * We can not use the "try" reset interface here, which will
448 * overwrite the previously restored configuration information.
449 */
450 if (vdev->reset_works && pci_cfg_access_trylock(pdev)) {
451 if (device_trylock(&pdev->dev)) {
452 if (!__pci_reset_function_locked(pdev))
453 vdev->needs_reset = false;
454 device_unlock(&pdev->dev);
455 }
456 pci_cfg_access_unlock(pdev);
457 }
458
459 pci_restore_state(pdev);
460 out:
461 pci_disable_device(pdev);
462
463 vfio_pci_try_bus_reset(vdev);
464
465 if (!disable_idle_d3)
466 vfio_pci_set_power_state(vdev, PCI_D3hot);
467 }
468
469 static void vfio_pci_release(void *device_data)
470 {
471 struct vfio_pci_device *vdev = device_data;
472
473 mutex_lock(&vdev->reflck->lock);
474
475 if (!(--vdev->refcnt)) {
476 vfio_spapr_pci_eeh_release(vdev->pdev);
477 vfio_pci_disable(vdev);
478 }
479
480 mutex_unlock(&vdev->reflck->lock);
481
482 module_put(THIS_MODULE);
483 }
484
485 static int vfio_pci_open(void *device_data)
486 {
487 struct vfio_pci_device *vdev = device_data;
488 int ret = 0;
489
490 if (!try_module_get(THIS_MODULE))
491 return -ENODEV;
492
493 mutex_lock(&vdev->reflck->lock);
494
495 if (!vdev->refcnt) {
496 ret = vfio_pci_enable(vdev);
497 if (ret)
498 goto error;
499
500 vfio_spapr_pci_eeh_open(vdev->pdev);
501 }
502 vdev->refcnt++;
503 error:
504 mutex_unlock(&vdev->reflck->lock);
505 if (ret)
506 module_put(THIS_MODULE);
507 return ret;
508 }
509
510 static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
511 {
512 if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
513 u8 pin;
514
515 if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
516 vdev->nointx || vdev->pdev->is_virtfn)
517 return 0;
518
519 pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
520
521 return pin ? 1 : 0;
522 } else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
523 u8 pos;
524 u16 flags;
525
526 pos = vdev->pdev->msi_cap;
527 if (pos) {
528 pci_read_config_word(vdev->pdev,
529 pos + PCI_MSI_FLAGS, &flags);
530 return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
531 }
532 } else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
533 u8 pos;
534 u16 flags;
535
536 pos = vdev->pdev->msix_cap;
537 if (pos) {
538 pci_read_config_word(vdev->pdev,
539 pos + PCI_MSIX_FLAGS, &flags);
540
541 return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
542 }
543 } else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
544 if (pci_is_pcie(vdev->pdev))
545 return 1;
546 } else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
547 return 1;
548 }
549
550 return 0;
551 }
552
553 static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
554 {
555 (*(int *)data)++;
556 return 0;
557 }
558
559 struct vfio_pci_fill_info {
560 int max;
561 int cur;
562 struct vfio_pci_dependent_device *devices;
563 };
564
565 static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
566 {
567 struct vfio_pci_fill_info *fill = data;
568 struct iommu_group *iommu_group;
569
570 if (fill->cur == fill->max)
571 return -EAGAIN; /* Something changed, try again */
572
573 iommu_group = iommu_group_get(&pdev->dev);
574 if (!iommu_group)
575 return -EPERM; /* Cannot reset non-isolated devices */
576
577 fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
578 fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
579 fill->devices[fill->cur].bus = pdev->bus->number;
580 fill->devices[fill->cur].devfn = pdev->devfn;
581 fill->cur++;
582 iommu_group_put(iommu_group);
583 return 0;
584 }
585
586 struct vfio_pci_group_entry {
587 struct vfio_group *group;
588 int id;
589 };
590
591 struct vfio_pci_group_info {
592 int count;
593 struct vfio_pci_group_entry *groups;
594 };
595
596 static int vfio_pci_validate_devs(struct pci_dev *pdev, void *data)
597 {
598 struct vfio_pci_group_info *info = data;
599 struct iommu_group *group;
600 int id, i;
601
602 group = iommu_group_get(&pdev->dev);
603 if (!group)
604 return -EPERM;
605
606 id = iommu_group_id(group);
607
608 for (i = 0; i < info->count; i++)
609 if (info->groups[i].id == id)
610 break;
611
612 iommu_group_put(group);
613
614 return (i == info->count) ? -EINVAL : 0;
615 }
616
617 static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
618 {
619 for (; pdev; pdev = pdev->bus->self)
620 if (pdev->bus == slot->bus)
621 return (pdev->slot == slot);
622 return false;
623 }
624
625 struct vfio_pci_walk_info {
626 int (*fn)(struct pci_dev *, void *data);
627 void *data;
628 struct pci_dev *pdev;
629 bool slot;
630 int ret;
631 };
632
633 static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
634 {
635 struct vfio_pci_walk_info *walk = data;
636
637 if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
638 walk->ret = walk->fn(pdev, walk->data);
639
640 return walk->ret;
641 }
642
643 static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
644 int (*fn)(struct pci_dev *,
645 void *data), void *data,
646 bool slot)
647 {
648 struct vfio_pci_walk_info walk = {
649 .fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
650 };
651
652 pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
653
654 return walk.ret;
655 }
656
657 static int msix_mmappable_cap(struct vfio_pci_device *vdev,
658 struct vfio_info_cap *caps)
659 {
660 struct vfio_info_cap_header header = {
661 .id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
662 .version = 1
663 };
664
665 return vfio_info_add_capability(caps, &header, sizeof(header));
666 }
667
668 int vfio_pci_register_dev_region(struct vfio_pci_device *vdev,
669 unsigned int type, unsigned int subtype,
670 const struct vfio_pci_regops *ops,
671 size_t size, u32 flags, void *data)
672 {
673 struct vfio_pci_region *region;
674
675 region = krealloc(vdev->region,
676 (vdev->num_regions + 1) * sizeof(*region),
677 GFP_KERNEL);
678 if (!region)
679 return -ENOMEM;
680
681 vdev->region = region;
682 vdev->region[vdev->num_regions].type = type;
683 vdev->region[vdev->num_regions].subtype = subtype;
684 vdev->region[vdev->num_regions].ops = ops;
685 vdev->region[vdev->num_regions].size = size;
686 vdev->region[vdev->num_regions].flags = flags;
687 vdev->region[vdev->num_regions].data = data;
688
689 vdev->num_regions++;
690
691 return 0;
692 }
693
694 static long vfio_pci_ioctl(void *device_data,
695 unsigned int cmd, unsigned long arg)
696 {
697 struct vfio_pci_device *vdev = device_data;
698 unsigned long minsz;
699
700 if (cmd == VFIO_DEVICE_GET_INFO) {
701 struct vfio_device_info info;
702
703 minsz = offsetofend(struct vfio_device_info, num_irqs);
704
705 if (copy_from_user(&info, (void __user *)arg, minsz))
706 return -EFAULT;
707
708 if (info.argsz < minsz)
709 return -EINVAL;
710
711 info.flags = VFIO_DEVICE_FLAGS_PCI;
712
713 if (vdev->reset_works)
714 info.flags |= VFIO_DEVICE_FLAGS_RESET;
715
716 info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
717 info.num_irqs = VFIO_PCI_NUM_IRQS;
718
719 return copy_to_user((void __user *)arg, &info, minsz) ?
720 -EFAULT : 0;
721
722 } else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
723 struct pci_dev *pdev = vdev->pdev;
724 struct vfio_region_info info;
725 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
726 int i, ret;
727
728 minsz = offsetofend(struct vfio_region_info, offset);
729
730 if (copy_from_user(&info, (void __user *)arg, minsz))
731 return -EFAULT;
732
733 if (info.argsz < minsz)
734 return -EINVAL;
735
736 switch (info.index) {
737 case VFIO_PCI_CONFIG_REGION_INDEX:
738 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
739 info.size = pdev->cfg_size;
740 info.flags = VFIO_REGION_INFO_FLAG_READ |
741 VFIO_REGION_INFO_FLAG_WRITE;
742 break;
743 case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
744 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
745 info.size = pci_resource_len(pdev, info.index);
746 if (!info.size) {
747 info.flags = 0;
748 break;
749 }
750
751 info.flags = VFIO_REGION_INFO_FLAG_READ |
752 VFIO_REGION_INFO_FLAG_WRITE;
753 if (vdev->bar_mmap_supported[info.index]) {
754 info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
755 if (info.index == vdev->msix_bar) {
756 ret = msix_mmappable_cap(vdev, &caps);
757 if (ret)
758 return ret;
759 }
760 }
761
762 break;
763 case VFIO_PCI_ROM_REGION_INDEX:
764 {
765 void __iomem *io;
766 size_t size;
767 u16 orig_cmd;
768
769 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
770 info.flags = 0;
771
772 /* Report the BAR size, not the ROM size */
773 info.size = pci_resource_len(pdev, info.index);
774 if (!info.size) {
775 /* Shadow ROMs appear as PCI option ROMs */
776 if (pdev->resource[PCI_ROM_RESOURCE].flags &
777 IORESOURCE_ROM_SHADOW)
778 info.size = 0x20000;
779 else
780 break;
781 }
782
783 /*
784 * Is it really there? Enable memory decode for
785 * implicit access in pci_map_rom().
786 */
787 pci_read_config_word(pdev, PCI_COMMAND, &orig_cmd);
788 pci_write_config_word(pdev, PCI_COMMAND,
789 orig_cmd | PCI_COMMAND_MEMORY);
790
791 io = pci_map_rom(pdev, &size);
792 if (io) {
793 info.flags = VFIO_REGION_INFO_FLAG_READ;
794 pci_unmap_rom(pdev, io);
795 } else {
796 info.size = 0;
797 }
798
799 pci_write_config_word(pdev, PCI_COMMAND, orig_cmd);
800 break;
801 }
802 case VFIO_PCI_VGA_REGION_INDEX:
803 if (!vdev->has_vga)
804 return -EINVAL;
805
806 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
807 info.size = 0xc0000;
808 info.flags = VFIO_REGION_INFO_FLAG_READ |
809 VFIO_REGION_INFO_FLAG_WRITE;
810
811 break;
812 default:
813 {
814 struct vfio_region_info_cap_type cap_type = {
815 .header.id = VFIO_REGION_INFO_CAP_TYPE,
816 .header.version = 1 };
817
818 if (info.index >=
819 VFIO_PCI_NUM_REGIONS + vdev->num_regions)
820 return -EINVAL;
821 info.index = array_index_nospec(info.index,
822 VFIO_PCI_NUM_REGIONS +
823 vdev->num_regions);
824
825 i = info.index - VFIO_PCI_NUM_REGIONS;
826
827 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
828 info.size = vdev->region[i].size;
829 info.flags = vdev->region[i].flags;
830
831 cap_type.type = vdev->region[i].type;
832 cap_type.subtype = vdev->region[i].subtype;
833
834 ret = vfio_info_add_capability(&caps, &cap_type.header,
835 sizeof(cap_type));
836 if (ret)
837 return ret;
838
839 if (vdev->region[i].ops->add_capability) {
840 ret = vdev->region[i].ops->add_capability(vdev,
841 &vdev->region[i], &caps);
842 if (ret)
843 return ret;
844 }
845 }
846 }
847
848 if (caps.size) {
849 info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
850 if (info.argsz < sizeof(info) + caps.size) {
851 info.argsz = sizeof(info) + caps.size;
852 info.cap_offset = 0;
853 } else {
854 vfio_info_cap_shift(&caps, sizeof(info));
855 if (copy_to_user((void __user *)arg +
856 sizeof(info), caps.buf,
857 caps.size)) {
858 kfree(caps.buf);
859 return -EFAULT;
860 }
861 info.cap_offset = sizeof(info);
862 }
863
864 kfree(caps.buf);
865 }
866
867 return copy_to_user((void __user *)arg, &info, minsz) ?
868 -EFAULT : 0;
869
870 } else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
871 struct vfio_irq_info info;
872
873 minsz = offsetofend(struct vfio_irq_info, count);
874
875 if (copy_from_user(&info, (void __user *)arg, minsz))
876 return -EFAULT;
877
878 if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
879 return -EINVAL;
880
881 switch (info.index) {
882 case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
883 case VFIO_PCI_REQ_IRQ_INDEX:
884 break;
885 case VFIO_PCI_ERR_IRQ_INDEX:
886 if (pci_is_pcie(vdev->pdev))
887 break;
888 /* fall through */
889 default:
890 return -EINVAL;
891 }
892
893 info.flags = VFIO_IRQ_INFO_EVENTFD;
894
895 info.count = vfio_pci_get_irq_count(vdev, info.index);
896
897 if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
898 info.flags |= (VFIO_IRQ_INFO_MASKABLE |
899 VFIO_IRQ_INFO_AUTOMASKED);
900 else
901 info.flags |= VFIO_IRQ_INFO_NORESIZE;
902
903 return copy_to_user((void __user *)arg, &info, minsz) ?
904 -EFAULT : 0;
905
906 } else if (cmd == VFIO_DEVICE_SET_IRQS) {
907 struct vfio_irq_set hdr;
908 u8 *data = NULL;
909 int max, ret = 0;
910 size_t data_size = 0;
911
912 minsz = offsetofend(struct vfio_irq_set, count);
913
914 if (copy_from_user(&hdr, (void __user *)arg, minsz))
915 return -EFAULT;
916
917 max = vfio_pci_get_irq_count(vdev, hdr.index);
918
919 ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
920 VFIO_PCI_NUM_IRQS, &data_size);
921 if (ret)
922 return ret;
923
924 if (data_size) {
925 data = memdup_user((void __user *)(arg + minsz),
926 data_size);
927 if (IS_ERR(data))
928 return PTR_ERR(data);
929 }
930
931 mutex_lock(&vdev->igate);
932
933 ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
934 hdr.start, hdr.count, data);
935
936 mutex_unlock(&vdev->igate);
937 kfree(data);
938
939 return ret;
940
941 } else if (cmd == VFIO_DEVICE_RESET) {
942 return vdev->reset_works ?
943 pci_try_reset_function(vdev->pdev) : -EINVAL;
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_pci_group_entry *groups;
1022 struct vfio_pci_group_info info;
1023 bool slot = false;
1024 int i, 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 (i = 0; i < hdr.count; i++) {
1077 struct vfio_group *group;
1078 struct fd f = fdget(group_fds[i]);
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[i].group = group;
1092 groups[i].id = vfio_external_user_iommu_id(group);
1093 }
1094
1095 kfree(group_fds);
1096
1097 /* release reference to groups on error */
1098 if (ret)
1099 goto hot_reset_release;
1100
1101 info.count = hdr.count;
1102 info.groups = groups;
1103
1104 /*
1105 * Test whether all the affected devices are contained
1106 * by the set of groups provided by the user.
1107 */
1108 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1109 vfio_pci_validate_devs,
1110 &info, slot);
1111 if (!ret)
1112 /* User has access, do the reset */
1113 ret = pci_reset_bus(vdev->pdev);
1114
1115 hot_reset_release:
1116 for (i--; i >= 0; i--)
1117 vfio_group_put_external_user(groups[i].group);
1118
1119 kfree(groups);
1120 return ret;
1121 } else if (cmd == VFIO_DEVICE_IOEVENTFD) {
1122 struct vfio_device_ioeventfd ioeventfd;
1123 int count;
1124
1125 minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1126
1127 if (copy_from_user(&ioeventfd, (void __user *)arg, minsz))
1128 return -EFAULT;
1129
1130 if (ioeventfd.argsz < minsz)
1131 return -EINVAL;
1132
1133 if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1134 return -EINVAL;
1135
1136 count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1137
1138 if (hweight8(count) != 1 || ioeventfd.fd < -1)
1139 return -EINVAL;
1140
1141 return vfio_pci_ioeventfd(vdev, ioeventfd.offset,
1142 ioeventfd.data, count, ioeventfd.fd);
1143 }
1144
1145 return -ENOTTY;
1146 }
1147
1148 static ssize_t vfio_pci_rw(void *device_data, char __user *buf,
1149 size_t count, loff_t *ppos, bool iswrite)
1150 {
1151 unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1152 struct vfio_pci_device *vdev = device_data;
1153
1154 if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1155 return -EINVAL;
1156
1157 switch (index) {
1158 case VFIO_PCI_CONFIG_REGION_INDEX:
1159 return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1160
1161 case VFIO_PCI_ROM_REGION_INDEX:
1162 if (iswrite)
1163 return -EINVAL;
1164 return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1165
1166 case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1167 return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1168
1169 case VFIO_PCI_VGA_REGION_INDEX:
1170 return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1171 default:
1172 index -= VFIO_PCI_NUM_REGIONS;
1173 return vdev->region[index].ops->rw(vdev, buf,
1174 count, ppos, iswrite);
1175 }
1176
1177 return -EINVAL;
1178 }
1179
1180 static ssize_t vfio_pci_read(void *device_data, char __user *buf,
1181 size_t count, loff_t *ppos)
1182 {
1183 if (!count)
1184 return 0;
1185
1186 return vfio_pci_rw(device_data, buf, count, ppos, false);
1187 }
1188
1189 static ssize_t vfio_pci_write(void *device_data, const char __user *buf,
1190 size_t count, loff_t *ppos)
1191 {
1192 if (!count)
1193 return 0;
1194
1195 return vfio_pci_rw(device_data, (char __user *)buf, count, ppos, true);
1196 }
1197
1198 static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma)
1199 {
1200 struct vfio_pci_device *vdev = device_data;
1201 struct pci_dev *pdev = vdev->pdev;
1202 unsigned int index;
1203 u64 phys_len, req_len, pgoff, req_start;
1204 int ret;
1205
1206 index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1207
1208 if (vma->vm_end < vma->vm_start)
1209 return -EINVAL;
1210 if ((vma->vm_flags & VM_SHARED) == 0)
1211 return -EINVAL;
1212 if (index >= VFIO_PCI_NUM_REGIONS) {
1213 int regnum = index - VFIO_PCI_NUM_REGIONS;
1214 struct vfio_pci_region *region = vdev->region + regnum;
1215
1216 if (region && region->ops && region->ops->mmap &&
1217 (region->flags & VFIO_REGION_INFO_FLAG_MMAP))
1218 return region->ops->mmap(vdev, region, vma);
1219 return -EINVAL;
1220 }
1221 if (index >= VFIO_PCI_ROM_REGION_INDEX)
1222 return -EINVAL;
1223 if (!vdev->bar_mmap_supported[index])
1224 return -EINVAL;
1225
1226 phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1227 req_len = vma->vm_end - vma->vm_start;
1228 pgoff = vma->vm_pgoff &
1229 ((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1230 req_start = pgoff << PAGE_SHIFT;
1231
1232 if (req_start + req_len > phys_len)
1233 return -EINVAL;
1234
1235 /*
1236 * Even though we don't make use of the barmap for the mmap,
1237 * we need to request the region and the barmap tracks that.
1238 */
1239 if (!vdev->barmap[index]) {
1240 ret = pci_request_selected_regions(pdev,
1241 1 << index, "vfio-pci");
1242 if (ret)
1243 return ret;
1244
1245 vdev->barmap[index] = pci_iomap(pdev, index, 0);
1246 if (!vdev->barmap[index]) {
1247 pci_release_selected_regions(pdev, 1 << index);
1248 return -ENOMEM;
1249 }
1250 }
1251
1252 vma->vm_private_data = vdev;
1253 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1254 vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1255
1256 return remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1257 req_len, vma->vm_page_prot);
1258 }
1259
1260 static void vfio_pci_request(void *device_data, unsigned int count)
1261 {
1262 struct vfio_pci_device *vdev = device_data;
1263 struct pci_dev *pdev = vdev->pdev;
1264
1265 mutex_lock(&vdev->igate);
1266
1267 if (vdev->req_trigger) {
1268 if (!(count % 10))
1269 pci_notice_ratelimited(pdev,
1270 "Relaying device request to user (#%u)\n",
1271 count);
1272 eventfd_signal(vdev->req_trigger, 1);
1273 } else if (count == 0) {
1274 pci_warn(pdev,
1275 "No device request channel registered, blocked until released by user\n");
1276 }
1277
1278 mutex_unlock(&vdev->igate);
1279 }
1280
1281 static const struct vfio_device_ops vfio_pci_ops = {
1282 .name = "vfio-pci",
1283 .open = vfio_pci_open,
1284 .release = vfio_pci_release,
1285 .ioctl = vfio_pci_ioctl,
1286 .read = vfio_pci_read,
1287 .write = vfio_pci_write,
1288 .mmap = vfio_pci_mmap,
1289 .request = vfio_pci_request,
1290 };
1291
1292 static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev);
1293 static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck);
1294
1295 static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1296 {
1297 struct vfio_pci_device *vdev;
1298 struct iommu_group *group;
1299 int ret;
1300
1301 if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1302 return -EINVAL;
1303
1304 /*
1305 * Prevent binding to PFs with VFs enabled, this too easily allows
1306 * userspace instance with VFs and PFs from the same device, which
1307 * cannot work. Disabling SR-IOV here would initiate removing the
1308 * VFs, which would unbind the driver, which is prone to blocking
1309 * if that VF is also in use by vfio-pci. Just reject these PFs
1310 * and let the user sort it out.
1311 */
1312 if (pci_num_vf(pdev)) {
1313 pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
1314 return -EBUSY;
1315 }
1316
1317 group = vfio_iommu_group_get(&pdev->dev);
1318 if (!group)
1319 return -EINVAL;
1320
1321 vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
1322 if (!vdev) {
1323 vfio_iommu_group_put(group, &pdev->dev);
1324 return -ENOMEM;
1325 }
1326
1327 vdev->pdev = pdev;
1328 vdev->irq_type = VFIO_PCI_NUM_IRQS;
1329 mutex_init(&vdev->igate);
1330 spin_lock_init(&vdev->irqlock);
1331 mutex_init(&vdev->ioeventfds_lock);
1332 INIT_LIST_HEAD(&vdev->ioeventfds_list);
1333
1334 ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
1335 if (ret) {
1336 vfio_iommu_group_put(group, &pdev->dev);
1337 kfree(vdev);
1338 return ret;
1339 }
1340
1341 ret = vfio_pci_reflck_attach(vdev);
1342 if (ret) {
1343 vfio_del_group_dev(&pdev->dev);
1344 vfio_iommu_group_put(group, &pdev->dev);
1345 kfree(vdev);
1346 return ret;
1347 }
1348
1349 if (vfio_pci_is_vga(pdev)) {
1350 vga_client_register(pdev, vdev, NULL, vfio_pci_set_vga_decode);
1351 vga_set_legacy_decoding(pdev,
1352 vfio_pci_set_vga_decode(vdev, false));
1353 }
1354
1355 vfio_pci_probe_power_state(vdev);
1356
1357 if (!disable_idle_d3) {
1358 /*
1359 * pci-core sets the device power state to an unknown value at
1360 * bootup and after being removed from a driver. The only
1361 * transition it allows from this unknown state is to D0, which
1362 * typically happens when a driver calls pci_enable_device().
1363 * We're not ready to enable the device yet, but we do want to
1364 * be able to get to D3. Therefore first do a D0 transition
1365 * before going to D3.
1366 */
1367 vfio_pci_set_power_state(vdev, PCI_D0);
1368 vfio_pci_set_power_state(vdev, PCI_D3hot);
1369 }
1370
1371 return ret;
1372 }
1373
1374 static void vfio_pci_remove(struct pci_dev *pdev)
1375 {
1376 struct vfio_pci_device *vdev;
1377
1378 vdev = vfio_del_group_dev(&pdev->dev);
1379 if (!vdev)
1380 return;
1381
1382 vfio_pci_reflck_put(vdev->reflck);
1383
1384 vfio_iommu_group_put(pdev->dev.iommu_group, &pdev->dev);
1385 kfree(vdev->region);
1386 mutex_destroy(&vdev->ioeventfds_lock);
1387
1388 if (!disable_idle_d3)
1389 vfio_pci_set_power_state(vdev, PCI_D0);
1390
1391 kfree(vdev->pm_save);
1392 kfree(vdev);
1393
1394 if (vfio_pci_is_vga(pdev)) {
1395 vga_client_register(pdev, NULL, NULL, NULL);
1396 vga_set_legacy_decoding(pdev,
1397 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
1398 VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM);
1399 }
1400 }
1401
1402 static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
1403 pci_channel_state_t state)
1404 {
1405 struct vfio_pci_device *vdev;
1406 struct vfio_device *device;
1407
1408 device = vfio_device_get_from_dev(&pdev->dev);
1409 if (device == NULL)
1410 return PCI_ERS_RESULT_DISCONNECT;
1411
1412 vdev = vfio_device_data(device);
1413 if (vdev == NULL) {
1414 vfio_device_put(device);
1415 return PCI_ERS_RESULT_DISCONNECT;
1416 }
1417
1418 mutex_lock(&vdev->igate);
1419
1420 if (vdev->err_trigger)
1421 eventfd_signal(vdev->err_trigger, 1);
1422
1423 mutex_unlock(&vdev->igate);
1424
1425 vfio_device_put(device);
1426
1427 return PCI_ERS_RESULT_CAN_RECOVER;
1428 }
1429
1430 static const struct pci_error_handlers vfio_err_handlers = {
1431 .error_detected = vfio_pci_aer_err_detected,
1432 };
1433
1434 static struct pci_driver vfio_pci_driver = {
1435 .name = "vfio-pci",
1436 .id_table = NULL, /* only dynamic ids */
1437 .probe = vfio_pci_probe,
1438 .remove = vfio_pci_remove,
1439 .err_handler = &vfio_err_handlers,
1440 };
1441
1442 static DEFINE_MUTEX(reflck_lock);
1443
1444 static struct vfio_pci_reflck *vfio_pci_reflck_alloc(void)
1445 {
1446 struct vfio_pci_reflck *reflck;
1447
1448 reflck = kzalloc(sizeof(*reflck), GFP_KERNEL);
1449 if (!reflck)
1450 return ERR_PTR(-ENOMEM);
1451
1452 kref_init(&reflck->kref);
1453 mutex_init(&reflck->lock);
1454
1455 return reflck;
1456 }
1457
1458 static void vfio_pci_reflck_get(struct vfio_pci_reflck *reflck)
1459 {
1460 kref_get(&reflck->kref);
1461 }
1462
1463 static int vfio_pci_reflck_find(struct pci_dev *pdev, void *data)
1464 {
1465 struct vfio_pci_reflck **preflck = data;
1466 struct vfio_device *device;
1467 struct vfio_pci_device *vdev;
1468
1469 device = vfio_device_get_from_dev(&pdev->dev);
1470 if (!device)
1471 return 0;
1472
1473 if (pci_dev_driver(pdev) != &vfio_pci_driver) {
1474 vfio_device_put(device);
1475 return 0;
1476 }
1477
1478 vdev = vfio_device_data(device);
1479
1480 if (vdev->reflck) {
1481 vfio_pci_reflck_get(vdev->reflck);
1482 *preflck = vdev->reflck;
1483 vfio_device_put(device);
1484 return 1;
1485 }
1486
1487 vfio_device_put(device);
1488 return 0;
1489 }
1490
1491 static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev)
1492 {
1493 bool slot = !pci_probe_reset_slot(vdev->pdev->slot);
1494
1495 mutex_lock(&reflck_lock);
1496
1497 if (pci_is_root_bus(vdev->pdev->bus) ||
1498 vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_reflck_find,
1499 &vdev->reflck, slot) <= 0)
1500 vdev->reflck = vfio_pci_reflck_alloc();
1501
1502 mutex_unlock(&reflck_lock);
1503
1504 return PTR_ERR_OR_ZERO(vdev->reflck);
1505 }
1506
1507 static void vfio_pci_reflck_release(struct kref *kref)
1508 {
1509 struct vfio_pci_reflck *reflck = container_of(kref,
1510 struct vfio_pci_reflck,
1511 kref);
1512
1513 kfree(reflck);
1514 mutex_unlock(&reflck_lock);
1515 }
1516
1517 static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck)
1518 {
1519 kref_put_mutex(&reflck->kref, vfio_pci_reflck_release, &reflck_lock);
1520 }
1521
1522 struct vfio_devices {
1523 struct vfio_device **devices;
1524 int cur_index;
1525 int max_index;
1526 };
1527
1528 static int vfio_pci_get_unused_devs(struct pci_dev *pdev, void *data)
1529 {
1530 struct vfio_devices *devs = data;
1531 struct vfio_device *device;
1532 struct vfio_pci_device *vdev;
1533
1534 if (devs->cur_index == devs->max_index)
1535 return -ENOSPC;
1536
1537 device = vfio_device_get_from_dev(&pdev->dev);
1538 if (!device)
1539 return -EINVAL;
1540
1541 if (pci_dev_driver(pdev) != &vfio_pci_driver) {
1542 vfio_device_put(device);
1543 return -EBUSY;
1544 }
1545
1546 vdev = vfio_device_data(device);
1547
1548 /* Fault if the device is not unused */
1549 if (vdev->refcnt) {
1550 vfio_device_put(device);
1551 return -EBUSY;
1552 }
1553
1554 devs->devices[devs->cur_index++] = device;
1555 return 0;
1556 }
1557
1558 /*
1559 * If a bus or slot reset is available for the provided device and:
1560 * - All of the devices affected by that bus or slot reset are unused
1561 * (!refcnt)
1562 * - At least one of the affected devices is marked dirty via
1563 * needs_reset (such as by lack of FLR support)
1564 * Then attempt to perform that bus or slot reset. Callers are required
1565 * to hold vdev->reflck->lock, protecting the bus/slot reset group from
1566 * concurrent opens. A vfio_device reference is acquired for each device
1567 * to prevent unbinds during the reset operation.
1568 *
1569 * NB: vfio-core considers a group to be viable even if some devices are
1570 * bound to drivers like pci-stub or pcieport. Here we require all devices
1571 * to be bound to vfio_pci since that's the only way we can be sure they
1572 * stay put.
1573 */
1574 static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev)
1575 {
1576 struct vfio_devices devs = { .cur_index = 0 };
1577 int i = 0, ret = -EINVAL;
1578 bool slot = false;
1579 struct vfio_pci_device *tmp;
1580
1581 if (!pci_probe_reset_slot(vdev->pdev->slot))
1582 slot = true;
1583 else if (pci_probe_reset_bus(vdev->pdev->bus))
1584 return;
1585
1586 if (vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
1587 &i, slot) || !i)
1588 return;
1589
1590 devs.max_index = i;
1591 devs.devices = kcalloc(i, sizeof(struct vfio_device *), GFP_KERNEL);
1592 if (!devs.devices)
1593 return;
1594
1595 if (vfio_pci_for_each_slot_or_bus(vdev->pdev,
1596 vfio_pci_get_unused_devs,
1597 &devs, slot))
1598 goto put_devs;
1599
1600 /* Does at least one need a reset? */
1601 for (i = 0; i < devs.cur_index; i++) {
1602 tmp = vfio_device_data(devs.devices[i]);
1603 if (tmp->needs_reset) {
1604 ret = pci_reset_bus(vdev->pdev);
1605 break;
1606 }
1607 }
1608
1609 put_devs:
1610 for (i = 0; i < devs.cur_index; i++) {
1611 tmp = vfio_device_data(devs.devices[i]);
1612
1613 /*
1614 * If reset was successful, affected devices no longer need
1615 * a reset and we should return all the collateral devices
1616 * to low power. If not successful, we either didn't reset
1617 * the bus or timed out waiting for it, so let's not touch
1618 * the power state.
1619 */
1620 if (!ret) {
1621 tmp->needs_reset = false;
1622
1623 if (tmp != vdev && !disable_idle_d3)
1624 vfio_pci_set_power_state(tmp, PCI_D3hot);
1625 }
1626
1627 vfio_device_put(devs.devices[i]);
1628 }
1629
1630 kfree(devs.devices);
1631 }
1632
1633 static void __exit vfio_pci_cleanup(void)
1634 {
1635 pci_unregister_driver(&vfio_pci_driver);
1636 vfio_pci_uninit_perm_bits();
1637 }
1638
1639 static void __init vfio_pci_fill_ids(void)
1640 {
1641 char *p, *id;
1642 int rc;
1643
1644 /* no ids passed actually */
1645 if (ids[0] == '\0')
1646 return;
1647
1648 /* add ids specified in the module parameter */
1649 p = ids;
1650 while ((id = strsep(&p, ","))) {
1651 unsigned int vendor, device, subvendor = PCI_ANY_ID,
1652 subdevice = PCI_ANY_ID, class = 0, class_mask = 0;
1653 int fields;
1654
1655 if (!strlen(id))
1656 continue;
1657
1658 fields = sscanf(id, "%x:%x:%x:%x:%x:%x",
1659 &vendor, &device, &subvendor, &subdevice,
1660 &class, &class_mask);
1661
1662 if (fields < 2) {
1663 pr_warn("invalid id string \"%s\"\n", id);
1664 continue;
1665 }
1666
1667 rc = pci_add_dynid(&vfio_pci_driver, vendor, device,
1668 subvendor, subdevice, class, class_mask, 0);
1669 if (rc)
1670 pr_warn("failed to add dynamic id [%04x:%04x[%04x:%04x]] class %#08x/%08x (%d)\n",
1671 vendor, device, subvendor, subdevice,
1672 class, class_mask, rc);
1673 else
1674 pr_info("add [%04x:%04x[%04x:%04x]] class %#08x/%08x\n",
1675 vendor, device, subvendor, subdevice,
1676 class, class_mask);
1677 }
1678 }
1679
1680 static int __init vfio_pci_init(void)
1681 {
1682 int ret;
1683
1684 /* Allocate shared config space permision data used by all devices */
1685 ret = vfio_pci_init_perm_bits();
1686 if (ret)
1687 return ret;
1688
1689 /* Register and scan for devices */
1690 ret = pci_register_driver(&vfio_pci_driver);
1691 if (ret)
1692 goto out_driver;
1693
1694 vfio_pci_fill_ids();
1695
1696 return 0;
1697
1698 out_driver:
1699 vfio_pci_uninit_perm_bits();
1700 return ret;
1701 }
1702
1703 module_init(vfio_pci_init);
1704 module_exit(vfio_pci_cleanup);
1705
1706 MODULE_VERSION(DRIVER_VERSION);
1707 MODULE_LICENSE("GPL v2");
1708 MODULE_AUTHOR(DRIVER_AUTHOR);
1709 MODULE_DESCRIPTION(DRIVER_DESC);
Linux with added FPC-III support