blob: b06d76216599e73dfc094738c2258b1d43cbeb3a [file] [log] [blame]
David Brazdil0f672f62019-12-10 10:32:29 +00001// SPDX-License-Identifier: GPL-2.0-only
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00002/*
3 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
4 * Author: Alex Williamson <alex.williamson@redhat.com>
5 *
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00006 * 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
David Brazdil0f672f62019-12-10 10:32:29 +000012#define dev_fmt pr_fmt
Andrew Scullb4b6d4a2019-01-02 15:54:55 +000013
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>
Olivier Deprez0e641232021-09-23 10:07:05 +020030#include <linux/sched/mm.h>
Andrew Scullb4b6d4a2019-01-02 15:54:55 +000031
32#include "vfio_pci_private.h"
33
34#define DRIVER_VERSION "0.2"
35#define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
36#define DRIVER_DESC "VFIO PCI - User Level meta-driver"
37
38static char ids[1024] __initdata;
39module_param_string(ids, ids, sizeof(ids), 0);
40MODULE_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");
41
42static bool nointxmask;
43module_param_named(nointxmask, nointxmask, bool, S_IRUGO | S_IWUSR);
44MODULE_PARM_DESC(nointxmask,
45 "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.");
46
47#ifdef CONFIG_VFIO_PCI_VGA
48static bool disable_vga;
49module_param(disable_vga, bool, S_IRUGO);
50MODULE_PARM_DESC(disable_vga, "Disable VGA resource access through vfio-pci");
51#endif
52
53static bool disable_idle_d3;
54module_param(disable_idle_d3, bool, S_IRUGO | S_IWUSR);
55MODULE_PARM_DESC(disable_idle_d3,
56 "Disable using the PCI D3 low power state for idle, unused devices");
57
Andrew Scullb4b6d4a2019-01-02 15:54:55 +000058static inline bool vfio_vga_disabled(void)
59{
60#ifdef CONFIG_VFIO_PCI_VGA
61 return disable_vga;
62#else
63 return true;
64#endif
65}
66
67/*
68 * Our VGA arbiter participation is limited since we don't know anything
69 * about the device itself. However, if the device is the only VGA device
70 * downstream of a bridge and VFIO VGA support is disabled, then we can
71 * safely return legacy VGA IO and memory as not decoded since the user
72 * has no way to get to it and routing can be disabled externally at the
73 * bridge.
74 */
75static unsigned int vfio_pci_set_vga_decode(void *opaque, bool single_vga)
76{
77 struct vfio_pci_device *vdev = opaque;
78 struct pci_dev *tmp = NULL, *pdev = vdev->pdev;
79 unsigned char max_busnr;
80 unsigned int decodes;
81
82 if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
83 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
84 VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
85
86 max_busnr = pci_bus_max_busnr(pdev->bus);
87 decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
88
89 while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
90 if (tmp == pdev ||
91 pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
92 pci_is_root_bus(tmp->bus))
93 continue;
94
95 if (tmp->bus->number >= pdev->bus->number &&
96 tmp->bus->number <= max_busnr) {
97 pci_dev_put(tmp);
98 decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
99 break;
100 }
101 }
102
103 return decodes;
104}
105
106static inline bool vfio_pci_is_vga(struct pci_dev *pdev)
107{
108 return (pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA;
109}
110
111static void vfio_pci_probe_mmaps(struct vfio_pci_device *vdev)
112{
113 struct resource *res;
114 int bar;
115 struct vfio_pci_dummy_resource *dummy_res;
116
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000117 for (bar = PCI_STD_RESOURCES; bar <= PCI_STD_RESOURCE_END; bar++) {
118 res = vdev->pdev->resource + bar;
119
120 if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
121 goto no_mmap;
122
123 if (!(res->flags & IORESOURCE_MEM))
124 goto no_mmap;
125
126 /*
127 * The PCI core shouldn't set up a resource with a
128 * type but zero size. But there may be bugs that
129 * cause us to do that.
130 */
131 if (!resource_size(res))
132 goto no_mmap;
133
134 if (resource_size(res) >= PAGE_SIZE) {
135 vdev->bar_mmap_supported[bar] = true;
136 continue;
137 }
138
139 if (!(res->start & ~PAGE_MASK)) {
140 /*
141 * Add a dummy resource to reserve the remainder
142 * of the exclusive page in case that hot-add
143 * device's bar is assigned into it.
144 */
145 dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
146 if (dummy_res == NULL)
147 goto no_mmap;
148
149 dummy_res->resource.name = "vfio sub-page reserved";
150 dummy_res->resource.start = res->end + 1;
151 dummy_res->resource.end = res->start + PAGE_SIZE - 1;
152 dummy_res->resource.flags = res->flags;
153 if (request_resource(res->parent,
154 &dummy_res->resource)) {
155 kfree(dummy_res);
156 goto no_mmap;
157 }
158 dummy_res->index = bar;
159 list_add(&dummy_res->res_next,
160 &vdev->dummy_resources_list);
161 vdev->bar_mmap_supported[bar] = true;
162 continue;
163 }
164 /*
165 * Here we don't handle the case when the BAR is not page
166 * aligned because we can't expect the BAR will be
167 * assigned into the same location in a page in guest
168 * when we passthrough the BAR. And it's hard to access
169 * this BAR in userspace because we have no way to get
170 * the BAR's location in a page.
171 */
172no_mmap:
173 vdev->bar_mmap_supported[bar] = false;
174 }
175}
176
177static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev);
178static void vfio_pci_disable(struct vfio_pci_device *vdev);
Olivier Deprez0e641232021-09-23 10:07:05 +0200179static int vfio_pci_try_zap_and_vma_lock_cb(struct pci_dev *pdev, void *data);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000180
181/*
182 * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
183 * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
184 * If a device implements the former but not the latter we would typically
185 * expect broken_intx_masking be set and require an exclusive interrupt.
186 * However since we do have control of the device's ability to assert INTx,
187 * we can instead pretend that the device does not implement INTx, virtualizing
188 * the pin register to report zero and maintaining DisINTx set on the host.
189 */
190static bool vfio_pci_nointx(struct pci_dev *pdev)
191{
192 switch (pdev->vendor) {
193 case PCI_VENDOR_ID_INTEL:
194 switch (pdev->device) {
195 /* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
196 case 0x1572:
197 case 0x1574:
198 case 0x1580 ... 0x1581:
199 case 0x1583 ... 0x158b:
200 case 0x37d0 ... 0x37d2:
201 return true;
202 default:
203 return false;
204 }
205 }
206
207 return false;
208}
209
David Brazdil0f672f62019-12-10 10:32:29 +0000210static void vfio_pci_probe_power_state(struct vfio_pci_device *vdev)
211{
212 struct pci_dev *pdev = vdev->pdev;
213 u16 pmcsr;
214
215 if (!pdev->pm_cap)
216 return;
217
218 pci_read_config_word(pdev, pdev->pm_cap + PCI_PM_CTRL, &pmcsr);
219
220 vdev->needs_pm_restore = !(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
221}
222
223/*
224 * pci_set_power_state() wrapper handling devices which perform a soft reset on
225 * D3->D0 transition. Save state prior to D0/1/2->D3, stash it on the vdev,
226 * restore when returned to D0. Saved separately from pci_saved_state for use
227 * by PM capability emulation and separately from pci_dev internal saved state
228 * to avoid it being overwritten and consumed around other resets.
229 */
230int vfio_pci_set_power_state(struct vfio_pci_device *vdev, pci_power_t state)
231{
232 struct pci_dev *pdev = vdev->pdev;
233 bool needs_restore = false, needs_save = false;
234 int ret;
235
236 if (vdev->needs_pm_restore) {
237 if (pdev->current_state < PCI_D3hot && state >= PCI_D3hot) {
238 pci_save_state(pdev);
239 needs_save = true;
240 }
241
242 if (pdev->current_state >= PCI_D3hot && state <= PCI_D0)
243 needs_restore = true;
244 }
245
246 ret = pci_set_power_state(pdev, state);
247
248 if (!ret) {
249 /* D3 might be unsupported via quirk, skip unless in D3 */
250 if (needs_save && pdev->current_state >= PCI_D3hot) {
251 vdev->pm_save = pci_store_saved_state(pdev);
252 } else if (needs_restore) {
253 pci_load_and_free_saved_state(pdev, &vdev->pm_save);
254 pci_restore_state(pdev);
255 }
256 }
257
258 return ret;
259}
260
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000261static int vfio_pci_enable(struct vfio_pci_device *vdev)
262{
263 struct pci_dev *pdev = vdev->pdev;
264 int ret;
265 u16 cmd;
266 u8 msix_pos;
267
David Brazdil0f672f62019-12-10 10:32:29 +0000268 vfio_pci_set_power_state(vdev, PCI_D0);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000269
270 /* Don't allow our initial saved state to include busmaster */
271 pci_clear_master(pdev);
272
273 ret = pci_enable_device(pdev);
274 if (ret)
275 return ret;
276
277 /* If reset fails because of the device lock, fail this path entirely */
278 ret = pci_try_reset_function(pdev);
279 if (ret == -EAGAIN) {
280 pci_disable_device(pdev);
281 return ret;
282 }
283
284 vdev->reset_works = !ret;
285 pci_save_state(pdev);
286 vdev->pci_saved_state = pci_store_saved_state(pdev);
287 if (!vdev->pci_saved_state)
David Brazdil0f672f62019-12-10 10:32:29 +0000288 pci_dbg(pdev, "%s: Couldn't store saved state\n", __func__);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000289
290 if (likely(!nointxmask)) {
291 if (vfio_pci_nointx(pdev)) {
David Brazdil0f672f62019-12-10 10:32:29 +0000292 pci_info(pdev, "Masking broken INTx support\n");
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000293 vdev->nointx = true;
294 pci_intx(pdev, 0);
295 } else
296 vdev->pci_2_3 = pci_intx_mask_supported(pdev);
297 }
298
299 pci_read_config_word(pdev, PCI_COMMAND, &cmd);
300 if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
301 cmd &= ~PCI_COMMAND_INTX_DISABLE;
302 pci_write_config_word(pdev, PCI_COMMAND, cmd);
303 }
304
305 ret = vfio_config_init(vdev);
306 if (ret) {
307 kfree(vdev->pci_saved_state);
308 vdev->pci_saved_state = NULL;
309 pci_disable_device(pdev);
310 return ret;
311 }
312
313 msix_pos = pdev->msix_cap;
314 if (msix_pos) {
315 u16 flags;
316 u32 table;
317
318 pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
319 pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
320
321 vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
322 vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
323 vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
324 } else
325 vdev->msix_bar = 0xFF;
326
327 if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
328 vdev->has_vga = true;
329
330
331 if (vfio_pci_is_vga(pdev) &&
332 pdev->vendor == PCI_VENDOR_ID_INTEL &&
333 IS_ENABLED(CONFIG_VFIO_PCI_IGD)) {
334 ret = vfio_pci_igd_init(vdev);
Olivier Deprez0e641232021-09-23 10:07:05 +0200335 if (ret && ret != -ENODEV) {
David Brazdil0f672f62019-12-10 10:32:29 +0000336 pci_warn(pdev, "Failed to setup Intel IGD regions\n");
337 goto disable_exit;
338 }
339 }
340
341 if (pdev->vendor == PCI_VENDOR_ID_NVIDIA &&
342 IS_ENABLED(CONFIG_VFIO_PCI_NVLINK2)) {
343 ret = vfio_pci_nvdia_v100_nvlink2_init(vdev);
344 if (ret && ret != -ENODEV) {
345 pci_warn(pdev, "Failed to setup NVIDIA NV2 RAM region\n");
346 goto disable_exit;
347 }
348 }
349
350 if (pdev->vendor == PCI_VENDOR_ID_IBM &&
351 IS_ENABLED(CONFIG_VFIO_PCI_NVLINK2)) {
352 ret = vfio_pci_ibm_npu2_init(vdev);
353 if (ret && ret != -ENODEV) {
354 pci_warn(pdev, "Failed to setup NVIDIA NV2 ATSD region\n");
355 goto disable_exit;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000356 }
357 }
358
359 vfio_pci_probe_mmaps(vdev);
360
361 return 0;
David Brazdil0f672f62019-12-10 10:32:29 +0000362
363disable_exit:
364 vfio_pci_disable(vdev);
365 return ret;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000366}
367
368static void vfio_pci_disable(struct vfio_pci_device *vdev)
369{
370 struct pci_dev *pdev = vdev->pdev;
371 struct vfio_pci_dummy_resource *dummy_res, *tmp;
372 struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
373 int i, bar;
374
375 /* Stop the device from further DMA */
376 pci_clear_master(pdev);
377
378 vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
379 VFIO_IRQ_SET_ACTION_TRIGGER,
380 vdev->irq_type, 0, 0, NULL);
381
382 /* Device closed, don't need mutex here */
383 list_for_each_entry_safe(ioeventfd, ioeventfd_tmp,
384 &vdev->ioeventfds_list, next) {
385 vfio_virqfd_disable(&ioeventfd->virqfd);
386 list_del(&ioeventfd->next);
387 kfree(ioeventfd);
388 }
389 vdev->ioeventfds_nr = 0;
390
391 vdev->virq_disabled = false;
392
393 for (i = 0; i < vdev->num_regions; i++)
394 vdev->region[i].ops->release(vdev, &vdev->region[i]);
395
396 vdev->num_regions = 0;
397 kfree(vdev->region);
398 vdev->region = NULL; /* don't krealloc a freed pointer */
399
400 vfio_config_free(vdev);
401
402 for (bar = PCI_STD_RESOURCES; bar <= PCI_STD_RESOURCE_END; bar++) {
403 if (!vdev->barmap[bar])
404 continue;
405 pci_iounmap(pdev, vdev->barmap[bar]);
406 pci_release_selected_regions(pdev, 1 << bar);
407 vdev->barmap[bar] = NULL;
408 }
409
410 list_for_each_entry_safe(dummy_res, tmp,
411 &vdev->dummy_resources_list, res_next) {
412 list_del(&dummy_res->res_next);
413 release_resource(&dummy_res->resource);
414 kfree(dummy_res);
415 }
416
417 vdev->needs_reset = true;
418
419 /*
420 * If we have saved state, restore it. If we can reset the device,
421 * even better. Resetting with current state seems better than
422 * nothing, but saving and restoring current state without reset
423 * is just busy work.
424 */
425 if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
David Brazdil0f672f62019-12-10 10:32:29 +0000426 pci_info(pdev, "%s: Couldn't reload saved state\n", __func__);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000427
428 if (!vdev->reset_works)
429 goto out;
430
431 pci_save_state(pdev);
432 }
433
434 /*
435 * Disable INTx and MSI, presumably to avoid spurious interrupts
436 * during reset. Stolen from pci_reset_function()
437 */
438 pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
439
440 /*
David Brazdil0f672f62019-12-10 10:32:29 +0000441 * Try to get the locks ourselves to prevent a deadlock. The
442 * success of this is dependent on being able to lock the device,
443 * which is not always possible.
444 * We can not use the "try" reset interface here, which will
445 * overwrite the previously restored configuration information.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000446 */
David Brazdil0f672f62019-12-10 10:32:29 +0000447 if (vdev->reset_works && pci_cfg_access_trylock(pdev)) {
448 if (device_trylock(&pdev->dev)) {
449 if (!__pci_reset_function_locked(pdev))
450 vdev->needs_reset = false;
451 device_unlock(&pdev->dev);
452 }
453 pci_cfg_access_unlock(pdev);
454 }
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000455
456 pci_restore_state(pdev);
457out:
458 pci_disable_device(pdev);
459
460 vfio_pci_try_bus_reset(vdev);
461
462 if (!disable_idle_d3)
David Brazdil0f672f62019-12-10 10:32:29 +0000463 vfio_pci_set_power_state(vdev, PCI_D3hot);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000464}
465
466static void vfio_pci_release(void *device_data)
467{
468 struct vfio_pci_device *vdev = device_data;
469
David Brazdil0f672f62019-12-10 10:32:29 +0000470 mutex_lock(&vdev->reflck->lock);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000471
472 if (!(--vdev->refcnt)) {
473 vfio_spapr_pci_eeh_release(vdev->pdev);
474 vfio_pci_disable(vdev);
Olivier Deprez0e641232021-09-23 10:07:05 +0200475 mutex_lock(&vdev->igate);
476 if (vdev->err_trigger) {
477 eventfd_ctx_put(vdev->err_trigger);
478 vdev->err_trigger = NULL;
479 }
480 mutex_unlock(&vdev->igate);
481
482 mutex_lock(&vdev->igate);
483 if (vdev->req_trigger) {
484 eventfd_ctx_put(vdev->req_trigger);
485 vdev->req_trigger = NULL;
486 }
487 mutex_unlock(&vdev->igate);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000488 }
489
David Brazdil0f672f62019-12-10 10:32:29 +0000490 mutex_unlock(&vdev->reflck->lock);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000491
492 module_put(THIS_MODULE);
493}
494
495static int vfio_pci_open(void *device_data)
496{
497 struct vfio_pci_device *vdev = device_data;
498 int ret = 0;
499
500 if (!try_module_get(THIS_MODULE))
501 return -ENODEV;
502
David Brazdil0f672f62019-12-10 10:32:29 +0000503 mutex_lock(&vdev->reflck->lock);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000504
505 if (!vdev->refcnt) {
506 ret = vfio_pci_enable(vdev);
507 if (ret)
508 goto error;
509
510 vfio_spapr_pci_eeh_open(vdev->pdev);
511 }
512 vdev->refcnt++;
513error:
David Brazdil0f672f62019-12-10 10:32:29 +0000514 mutex_unlock(&vdev->reflck->lock);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000515 if (ret)
516 module_put(THIS_MODULE);
517 return ret;
518}
519
520static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
521{
522 if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
523 u8 pin;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000524
David Brazdil0f672f62019-12-10 10:32:29 +0000525 if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
526 vdev->nointx || vdev->pdev->is_virtfn)
527 return 0;
528
529 pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
530
531 return pin ? 1 : 0;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000532 } else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
533 u8 pos;
534 u16 flags;
535
536 pos = vdev->pdev->msi_cap;
537 if (pos) {
538 pci_read_config_word(vdev->pdev,
539 pos + PCI_MSI_FLAGS, &flags);
540 return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
541 }
542 } else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
543 u8 pos;
544 u16 flags;
545
546 pos = vdev->pdev->msix_cap;
547 if (pos) {
548 pci_read_config_word(vdev->pdev,
549 pos + PCI_MSIX_FLAGS, &flags);
550
551 return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
552 }
553 } else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
554 if (pci_is_pcie(vdev->pdev))
555 return 1;
556 } else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
557 return 1;
558 }
559
560 return 0;
561}
562
563static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
564{
565 (*(int *)data)++;
566 return 0;
567}
568
569struct vfio_pci_fill_info {
570 int max;
571 int cur;
572 struct vfio_pci_dependent_device *devices;
573};
574
575static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
576{
577 struct vfio_pci_fill_info *fill = data;
578 struct iommu_group *iommu_group;
579
580 if (fill->cur == fill->max)
581 return -EAGAIN; /* Something changed, try again */
582
583 iommu_group = iommu_group_get(&pdev->dev);
584 if (!iommu_group)
585 return -EPERM; /* Cannot reset non-isolated devices */
586
587 fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
588 fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
589 fill->devices[fill->cur].bus = pdev->bus->number;
590 fill->devices[fill->cur].devfn = pdev->devfn;
591 fill->cur++;
592 iommu_group_put(iommu_group);
593 return 0;
594}
595
596struct vfio_pci_group_entry {
597 struct vfio_group *group;
598 int id;
599};
600
601struct vfio_pci_group_info {
602 int count;
603 struct vfio_pci_group_entry *groups;
604};
605
606static int vfio_pci_validate_devs(struct pci_dev *pdev, void *data)
607{
608 struct vfio_pci_group_info *info = data;
609 struct iommu_group *group;
610 int id, i;
611
612 group = iommu_group_get(&pdev->dev);
613 if (!group)
614 return -EPERM;
615
616 id = iommu_group_id(group);
617
618 for (i = 0; i < info->count; i++)
619 if (info->groups[i].id == id)
620 break;
621
622 iommu_group_put(group);
623
624 return (i == info->count) ? -EINVAL : 0;
625}
626
627static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
628{
629 for (; pdev; pdev = pdev->bus->self)
630 if (pdev->bus == slot->bus)
631 return (pdev->slot == slot);
632 return false;
633}
634
635struct vfio_pci_walk_info {
636 int (*fn)(struct pci_dev *, void *data);
637 void *data;
638 struct pci_dev *pdev;
639 bool slot;
640 int ret;
641};
642
643static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
644{
645 struct vfio_pci_walk_info *walk = data;
646
647 if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
648 walk->ret = walk->fn(pdev, walk->data);
649
650 return walk->ret;
651}
652
653static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
654 int (*fn)(struct pci_dev *,
655 void *data), void *data,
656 bool slot)
657{
658 struct vfio_pci_walk_info walk = {
659 .fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
660 };
661
662 pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
663
664 return walk.ret;
665}
666
667static int msix_mmappable_cap(struct vfio_pci_device *vdev,
668 struct vfio_info_cap *caps)
669{
670 struct vfio_info_cap_header header = {
671 .id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
672 .version = 1
673 };
674
675 return vfio_info_add_capability(caps, &header, sizeof(header));
676}
677
678int vfio_pci_register_dev_region(struct vfio_pci_device *vdev,
679 unsigned int type, unsigned int subtype,
680 const struct vfio_pci_regops *ops,
681 size_t size, u32 flags, void *data)
682{
683 struct vfio_pci_region *region;
684
685 region = krealloc(vdev->region,
686 (vdev->num_regions + 1) * sizeof(*region),
687 GFP_KERNEL);
688 if (!region)
689 return -ENOMEM;
690
691 vdev->region = region;
692 vdev->region[vdev->num_regions].type = type;
693 vdev->region[vdev->num_regions].subtype = subtype;
694 vdev->region[vdev->num_regions].ops = ops;
695 vdev->region[vdev->num_regions].size = size;
696 vdev->region[vdev->num_regions].flags = flags;
697 vdev->region[vdev->num_regions].data = data;
698
699 vdev->num_regions++;
700
701 return 0;
702}
703
Olivier Deprez0e641232021-09-23 10:07:05 +0200704struct vfio_devices {
705 struct vfio_device **devices;
706 int cur_index;
707 int max_index;
708};
709
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000710static long vfio_pci_ioctl(void *device_data,
711 unsigned int cmd, unsigned long arg)
712{
713 struct vfio_pci_device *vdev = device_data;
714 unsigned long minsz;
715
716 if (cmd == VFIO_DEVICE_GET_INFO) {
717 struct vfio_device_info info;
718
719 minsz = offsetofend(struct vfio_device_info, num_irqs);
720
721 if (copy_from_user(&info, (void __user *)arg, minsz))
722 return -EFAULT;
723
724 if (info.argsz < minsz)
725 return -EINVAL;
726
727 info.flags = VFIO_DEVICE_FLAGS_PCI;
728
729 if (vdev->reset_works)
730 info.flags |= VFIO_DEVICE_FLAGS_RESET;
731
732 info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
733 info.num_irqs = VFIO_PCI_NUM_IRQS;
734
735 return copy_to_user((void __user *)arg, &info, minsz) ?
736 -EFAULT : 0;
737
738 } else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
739 struct pci_dev *pdev = vdev->pdev;
740 struct vfio_region_info info;
741 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
742 int i, ret;
743
744 minsz = offsetofend(struct vfio_region_info, offset);
745
746 if (copy_from_user(&info, (void __user *)arg, minsz))
747 return -EFAULT;
748
749 if (info.argsz < minsz)
750 return -EINVAL;
751
752 switch (info.index) {
753 case VFIO_PCI_CONFIG_REGION_INDEX:
754 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
755 info.size = pdev->cfg_size;
756 info.flags = VFIO_REGION_INFO_FLAG_READ |
757 VFIO_REGION_INFO_FLAG_WRITE;
758 break;
759 case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
760 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
761 info.size = pci_resource_len(pdev, info.index);
762 if (!info.size) {
763 info.flags = 0;
764 break;
765 }
766
767 info.flags = VFIO_REGION_INFO_FLAG_READ |
768 VFIO_REGION_INFO_FLAG_WRITE;
769 if (vdev->bar_mmap_supported[info.index]) {
770 info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
771 if (info.index == vdev->msix_bar) {
772 ret = msix_mmappable_cap(vdev, &caps);
773 if (ret)
774 return ret;
775 }
776 }
777
778 break;
779 case VFIO_PCI_ROM_REGION_INDEX:
780 {
781 void __iomem *io;
782 size_t size;
Olivier Deprez0e641232021-09-23 10:07:05 +0200783 u16 cmd;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000784
785 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
786 info.flags = 0;
787
788 /* Report the BAR size, not the ROM size */
789 info.size = pci_resource_len(pdev, info.index);
790 if (!info.size) {
791 /* Shadow ROMs appear as PCI option ROMs */
792 if (pdev->resource[PCI_ROM_RESOURCE].flags &
793 IORESOURCE_ROM_SHADOW)
794 info.size = 0x20000;
795 else
796 break;
797 }
798
David Brazdil0f672f62019-12-10 10:32:29 +0000799 /*
800 * Is it really there? Enable memory decode for
801 * implicit access in pci_map_rom().
802 */
Olivier Deprez0e641232021-09-23 10:07:05 +0200803 cmd = vfio_pci_memory_lock_and_enable(vdev);
David Brazdil0f672f62019-12-10 10:32:29 +0000804 io = pci_map_rom(pdev, &size);
805 if (io) {
806 info.flags = VFIO_REGION_INFO_FLAG_READ;
807 pci_unmap_rom(pdev, io);
808 } else {
809 info.size = 0;
810 }
Olivier Deprez0e641232021-09-23 10:07:05 +0200811 vfio_pci_memory_unlock_and_restore(vdev, cmd);
David Brazdil0f672f62019-12-10 10:32:29 +0000812
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000813 break;
814 }
815 case VFIO_PCI_VGA_REGION_INDEX:
816 if (!vdev->has_vga)
817 return -EINVAL;
818
819 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
820 info.size = 0xc0000;
821 info.flags = VFIO_REGION_INFO_FLAG_READ |
822 VFIO_REGION_INFO_FLAG_WRITE;
823
824 break;
825 default:
826 {
827 struct vfio_region_info_cap_type cap_type = {
828 .header.id = VFIO_REGION_INFO_CAP_TYPE,
829 .header.version = 1 };
830
831 if (info.index >=
832 VFIO_PCI_NUM_REGIONS + vdev->num_regions)
833 return -EINVAL;
834 info.index = array_index_nospec(info.index,
835 VFIO_PCI_NUM_REGIONS +
836 vdev->num_regions);
837
838 i = info.index - VFIO_PCI_NUM_REGIONS;
839
840 info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
841 info.size = vdev->region[i].size;
842 info.flags = vdev->region[i].flags;
843
844 cap_type.type = vdev->region[i].type;
845 cap_type.subtype = vdev->region[i].subtype;
846
847 ret = vfio_info_add_capability(&caps, &cap_type.header,
848 sizeof(cap_type));
849 if (ret)
850 return ret;
851
David Brazdil0f672f62019-12-10 10:32:29 +0000852 if (vdev->region[i].ops->add_capability) {
853 ret = vdev->region[i].ops->add_capability(vdev,
854 &vdev->region[i], &caps);
855 if (ret)
856 return ret;
857 }
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000858 }
859 }
860
861 if (caps.size) {
862 info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
863 if (info.argsz < sizeof(info) + caps.size) {
864 info.argsz = sizeof(info) + caps.size;
865 info.cap_offset = 0;
866 } else {
867 vfio_info_cap_shift(&caps, sizeof(info));
868 if (copy_to_user((void __user *)arg +
869 sizeof(info), caps.buf,
870 caps.size)) {
871 kfree(caps.buf);
872 return -EFAULT;
873 }
874 info.cap_offset = sizeof(info);
875 }
876
877 kfree(caps.buf);
878 }
879
880 return copy_to_user((void __user *)arg, &info, minsz) ?
881 -EFAULT : 0;
882
883 } else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
884 struct vfio_irq_info info;
885
886 minsz = offsetofend(struct vfio_irq_info, count);
887
888 if (copy_from_user(&info, (void __user *)arg, minsz))
889 return -EFAULT;
890
891 if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
892 return -EINVAL;
893
894 switch (info.index) {
895 case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
896 case VFIO_PCI_REQ_IRQ_INDEX:
897 break;
898 case VFIO_PCI_ERR_IRQ_INDEX:
899 if (pci_is_pcie(vdev->pdev))
900 break;
901 /* fall through */
902 default:
903 return -EINVAL;
904 }
905
906 info.flags = VFIO_IRQ_INFO_EVENTFD;
907
908 info.count = vfio_pci_get_irq_count(vdev, info.index);
909
910 if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
911 info.flags |= (VFIO_IRQ_INFO_MASKABLE |
912 VFIO_IRQ_INFO_AUTOMASKED);
913 else
914 info.flags |= VFIO_IRQ_INFO_NORESIZE;
915
916 return copy_to_user((void __user *)arg, &info, minsz) ?
917 -EFAULT : 0;
918
919 } else if (cmd == VFIO_DEVICE_SET_IRQS) {
920 struct vfio_irq_set hdr;
921 u8 *data = NULL;
922 int max, ret = 0;
923 size_t data_size = 0;
924
925 minsz = offsetofend(struct vfio_irq_set, count);
926
927 if (copy_from_user(&hdr, (void __user *)arg, minsz))
928 return -EFAULT;
929
930 max = vfio_pci_get_irq_count(vdev, hdr.index);
931
932 ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
933 VFIO_PCI_NUM_IRQS, &data_size);
934 if (ret)
935 return ret;
936
937 if (data_size) {
938 data = memdup_user((void __user *)(arg + minsz),
939 data_size);
940 if (IS_ERR(data))
941 return PTR_ERR(data);
942 }
943
944 mutex_lock(&vdev->igate);
945
946 ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
947 hdr.start, hdr.count, data);
948
949 mutex_unlock(&vdev->igate);
950 kfree(data);
951
952 return ret;
953
954 } else if (cmd == VFIO_DEVICE_RESET) {
Olivier Deprez0e641232021-09-23 10:07:05 +0200955 int ret;
956
957 if (!vdev->reset_works)
958 return -EINVAL;
959
960 vfio_pci_zap_and_down_write_memory_lock(vdev);
961 ret = pci_try_reset_function(vdev->pdev);
962 up_write(&vdev->memory_lock);
963
964 return ret;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000965
966 } else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
967 struct vfio_pci_hot_reset_info hdr;
968 struct vfio_pci_fill_info fill = { 0 };
969 struct vfio_pci_dependent_device *devices = NULL;
970 bool slot = false;
971 int ret = 0;
972
973 minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
974
975 if (copy_from_user(&hdr, (void __user *)arg, minsz))
976 return -EFAULT;
977
978 if (hdr.argsz < minsz)
979 return -EINVAL;
980
981 hdr.flags = 0;
982
983 /* Can we do a slot or bus reset or neither? */
984 if (!pci_probe_reset_slot(vdev->pdev->slot))
985 slot = true;
986 else if (pci_probe_reset_bus(vdev->pdev->bus))
987 return -ENODEV;
988
989 /* How many devices are affected? */
990 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
991 vfio_pci_count_devs,
992 &fill.max, slot);
993 if (ret)
994 return ret;
995
996 WARN_ON(!fill.max); /* Should always be at least one */
997
998 /*
999 * If there's enough space, fill it now, otherwise return
1000 * -ENOSPC and the number of devices affected.
1001 */
1002 if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
1003 ret = -ENOSPC;
1004 hdr.count = fill.max;
1005 goto reset_info_exit;
1006 }
1007
1008 devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
1009 if (!devices)
1010 return -ENOMEM;
1011
1012 fill.devices = devices;
1013
1014 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1015 vfio_pci_fill_devs,
1016 &fill, slot);
1017
1018 /*
1019 * If a device was removed between counting and filling,
1020 * we may come up short of fill.max. If a device was
1021 * added, we'll have a return of -EAGAIN above.
1022 */
1023 if (!ret)
1024 hdr.count = fill.cur;
1025
1026reset_info_exit:
1027 if (copy_to_user((void __user *)arg, &hdr, minsz))
1028 ret = -EFAULT;
1029
1030 if (!ret) {
1031 if (copy_to_user((void __user *)(arg + minsz), devices,
1032 hdr.count * sizeof(*devices)))
1033 ret = -EFAULT;
1034 }
1035
1036 kfree(devices);
1037 return ret;
1038
1039 } else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
1040 struct vfio_pci_hot_reset hdr;
1041 int32_t *group_fds;
1042 struct vfio_pci_group_entry *groups;
1043 struct vfio_pci_group_info info;
Olivier Deprez0e641232021-09-23 10:07:05 +02001044 struct vfio_devices devs = { .cur_index = 0 };
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001045 bool slot = false;
Olivier Deprez0e641232021-09-23 10:07:05 +02001046 int i, group_idx, mem_idx = 0, count = 0, ret = 0;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001047
1048 minsz = offsetofend(struct vfio_pci_hot_reset, count);
1049
1050 if (copy_from_user(&hdr, (void __user *)arg, minsz))
1051 return -EFAULT;
1052
1053 if (hdr.argsz < minsz || hdr.flags)
1054 return -EINVAL;
1055
1056 /* Can we do a slot or bus reset or neither? */
1057 if (!pci_probe_reset_slot(vdev->pdev->slot))
1058 slot = true;
1059 else if (pci_probe_reset_bus(vdev->pdev->bus))
1060 return -ENODEV;
1061
1062 /*
1063 * We can't let userspace give us an arbitrarily large
1064 * buffer to copy, so verify how many we think there
1065 * could be. Note groups can have multiple devices so
1066 * one group per device is the max.
1067 */
1068 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1069 vfio_pci_count_devs,
1070 &count, slot);
1071 if (ret)
1072 return ret;
1073
1074 /* Somewhere between 1 and count is OK */
1075 if (!hdr.count || hdr.count > count)
1076 return -EINVAL;
1077
1078 group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
1079 groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
1080 if (!group_fds || !groups) {
1081 kfree(group_fds);
1082 kfree(groups);
1083 return -ENOMEM;
1084 }
1085
1086 if (copy_from_user(group_fds, (void __user *)(arg + minsz),
1087 hdr.count * sizeof(*group_fds))) {
1088 kfree(group_fds);
1089 kfree(groups);
1090 return -EFAULT;
1091 }
1092
1093 /*
1094 * For each group_fd, get the group through the vfio external
1095 * user interface and store the group and iommu ID. This
1096 * ensures the group is held across the reset.
1097 */
Olivier Deprez0e641232021-09-23 10:07:05 +02001098 for (group_idx = 0; group_idx < hdr.count; group_idx++) {
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001099 struct vfio_group *group;
Olivier Deprez0e641232021-09-23 10:07:05 +02001100 struct fd f = fdget(group_fds[group_idx]);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001101 if (!f.file) {
1102 ret = -EBADF;
1103 break;
1104 }
1105
1106 group = vfio_group_get_external_user(f.file);
1107 fdput(f);
1108 if (IS_ERR(group)) {
1109 ret = PTR_ERR(group);
1110 break;
1111 }
1112
Olivier Deprez0e641232021-09-23 10:07:05 +02001113 groups[group_idx].group = group;
1114 groups[group_idx].id =
1115 vfio_external_user_iommu_id(group);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001116 }
1117
1118 kfree(group_fds);
1119
1120 /* release reference to groups on error */
1121 if (ret)
1122 goto hot_reset_release;
1123
1124 info.count = hdr.count;
1125 info.groups = groups;
1126
1127 /*
1128 * Test whether all the affected devices are contained
1129 * by the set of groups provided by the user.
1130 */
1131 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1132 vfio_pci_validate_devs,
1133 &info, slot);
Olivier Deprez0e641232021-09-23 10:07:05 +02001134 if (ret)
1135 goto hot_reset_release;
1136
1137 devs.max_index = count;
1138 devs.devices = kcalloc(count, sizeof(struct vfio_device *),
1139 GFP_KERNEL);
1140 if (!devs.devices) {
1141 ret = -ENOMEM;
1142 goto hot_reset_release;
1143 }
1144
1145 /*
1146 * We need to get memory_lock for each device, but devices
1147 * can share mmap_sem, therefore we need to zap and hold
1148 * the vma_lock for each device, and only then get each
1149 * memory_lock.
1150 */
1151 ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1152 vfio_pci_try_zap_and_vma_lock_cb,
1153 &devs, slot);
1154 if (ret)
1155 goto hot_reset_release;
1156
1157 for (; mem_idx < devs.cur_index; mem_idx++) {
1158 struct vfio_pci_device *tmp;
1159
1160 tmp = vfio_device_data(devs.devices[mem_idx]);
1161
1162 ret = down_write_trylock(&tmp->memory_lock);
1163 if (!ret) {
1164 ret = -EBUSY;
1165 goto hot_reset_release;
1166 }
1167 mutex_unlock(&tmp->vma_lock);
1168 }
1169
1170 /* User has access, do the reset */
1171 ret = pci_reset_bus(vdev->pdev);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001172
1173hot_reset_release:
Olivier Deprez0e641232021-09-23 10:07:05 +02001174 for (i = 0; i < devs.cur_index; i++) {
1175 struct vfio_device *device;
1176 struct vfio_pci_device *tmp;
1177
1178 device = devs.devices[i];
1179 tmp = vfio_device_data(device);
1180
1181 if (i < mem_idx)
1182 up_write(&tmp->memory_lock);
1183 else
1184 mutex_unlock(&tmp->vma_lock);
1185 vfio_device_put(device);
1186 }
1187 kfree(devs.devices);
1188
1189 for (group_idx--; group_idx >= 0; group_idx--)
1190 vfio_group_put_external_user(groups[group_idx].group);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001191
1192 kfree(groups);
1193 return ret;
1194 } else if (cmd == VFIO_DEVICE_IOEVENTFD) {
1195 struct vfio_device_ioeventfd ioeventfd;
1196 int count;
1197
1198 minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1199
1200 if (copy_from_user(&ioeventfd, (void __user *)arg, minsz))
1201 return -EFAULT;
1202
1203 if (ioeventfd.argsz < minsz)
1204 return -EINVAL;
1205
1206 if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1207 return -EINVAL;
1208
1209 count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1210
1211 if (hweight8(count) != 1 || ioeventfd.fd < -1)
1212 return -EINVAL;
1213
1214 return vfio_pci_ioeventfd(vdev, ioeventfd.offset,
1215 ioeventfd.data, count, ioeventfd.fd);
1216 }
1217
1218 return -ENOTTY;
1219}
1220
1221static ssize_t vfio_pci_rw(void *device_data, char __user *buf,
1222 size_t count, loff_t *ppos, bool iswrite)
1223{
1224 unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1225 struct vfio_pci_device *vdev = device_data;
1226
1227 if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1228 return -EINVAL;
1229
1230 switch (index) {
1231 case VFIO_PCI_CONFIG_REGION_INDEX:
1232 return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1233
1234 case VFIO_PCI_ROM_REGION_INDEX:
1235 if (iswrite)
1236 return -EINVAL;
1237 return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1238
1239 case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1240 return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1241
1242 case VFIO_PCI_VGA_REGION_INDEX:
1243 return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1244 default:
1245 index -= VFIO_PCI_NUM_REGIONS;
1246 return vdev->region[index].ops->rw(vdev, buf,
1247 count, ppos, iswrite);
1248 }
1249
1250 return -EINVAL;
1251}
1252
1253static ssize_t vfio_pci_read(void *device_data, char __user *buf,
1254 size_t count, loff_t *ppos)
1255{
1256 if (!count)
1257 return 0;
1258
1259 return vfio_pci_rw(device_data, buf, count, ppos, false);
1260}
1261
1262static ssize_t vfio_pci_write(void *device_data, const char __user *buf,
1263 size_t count, loff_t *ppos)
1264{
1265 if (!count)
1266 return 0;
1267
1268 return vfio_pci_rw(device_data, (char __user *)buf, count, ppos, true);
1269}
1270
Olivier Deprez0e641232021-09-23 10:07:05 +02001271/* Return 1 on zap and vma_lock acquired, 0 on contention (only with @try) */
1272static int vfio_pci_zap_and_vma_lock(struct vfio_pci_device *vdev, bool try)
1273{
1274 struct vfio_pci_mmap_vma *mmap_vma, *tmp;
1275
1276 /*
1277 * Lock ordering:
1278 * vma_lock is nested under mmap_sem for vm_ops callback paths.
1279 * The memory_lock semaphore is used by both code paths calling
1280 * into this function to zap vmas and the vm_ops.fault callback
1281 * to protect the memory enable state of the device.
1282 *
1283 * When zapping vmas we need to maintain the mmap_sem => vma_lock
1284 * ordering, which requires using vma_lock to walk vma_list to
1285 * acquire an mm, then dropping vma_lock to get the mmap_sem and
1286 * reacquiring vma_lock. This logic is derived from similar
1287 * requirements in uverbs_user_mmap_disassociate().
1288 *
1289 * mmap_sem must always be the top-level lock when it is taken.
1290 * Therefore we can only hold the memory_lock write lock when
1291 * vma_list is empty, as we'd need to take mmap_sem to clear
1292 * entries. vma_list can only be guaranteed empty when holding
1293 * vma_lock, thus memory_lock is nested under vma_lock.
1294 *
1295 * This enables the vm_ops.fault callback to acquire vma_lock,
1296 * followed by memory_lock read lock, while already holding
1297 * mmap_sem without risk of deadlock.
1298 */
1299 while (1) {
1300 struct mm_struct *mm = NULL;
1301
1302 if (try) {
1303 if (!mutex_trylock(&vdev->vma_lock))
1304 return 0;
1305 } else {
1306 mutex_lock(&vdev->vma_lock);
1307 }
1308 while (!list_empty(&vdev->vma_list)) {
1309 mmap_vma = list_first_entry(&vdev->vma_list,
1310 struct vfio_pci_mmap_vma,
1311 vma_next);
1312 mm = mmap_vma->vma->vm_mm;
1313 if (mmget_not_zero(mm))
1314 break;
1315
1316 list_del(&mmap_vma->vma_next);
1317 kfree(mmap_vma);
1318 mm = NULL;
1319 }
1320 if (!mm)
1321 return 1;
1322 mutex_unlock(&vdev->vma_lock);
1323
1324 if (try) {
1325 if (!down_read_trylock(&mm->mmap_sem)) {
1326 mmput(mm);
1327 return 0;
1328 }
1329 } else {
1330 down_read(&mm->mmap_sem);
1331 }
1332 if (mmget_still_valid(mm)) {
1333 if (try) {
1334 if (!mutex_trylock(&vdev->vma_lock)) {
1335 up_read(&mm->mmap_sem);
1336 mmput(mm);
1337 return 0;
1338 }
1339 } else {
1340 mutex_lock(&vdev->vma_lock);
1341 }
1342 list_for_each_entry_safe(mmap_vma, tmp,
1343 &vdev->vma_list, vma_next) {
1344 struct vm_area_struct *vma = mmap_vma->vma;
1345
1346 if (vma->vm_mm != mm)
1347 continue;
1348
1349 list_del(&mmap_vma->vma_next);
1350 kfree(mmap_vma);
1351
1352 zap_vma_ptes(vma, vma->vm_start,
1353 vma->vm_end - vma->vm_start);
1354 }
1355 mutex_unlock(&vdev->vma_lock);
1356 }
1357 up_read(&mm->mmap_sem);
1358 mmput(mm);
1359 }
1360}
1361
1362void vfio_pci_zap_and_down_write_memory_lock(struct vfio_pci_device *vdev)
1363{
1364 vfio_pci_zap_and_vma_lock(vdev, false);
1365 down_write(&vdev->memory_lock);
1366 mutex_unlock(&vdev->vma_lock);
1367}
1368
1369u16 vfio_pci_memory_lock_and_enable(struct vfio_pci_device *vdev)
1370{
1371 u16 cmd;
1372
1373 down_write(&vdev->memory_lock);
1374 pci_read_config_word(vdev->pdev, PCI_COMMAND, &cmd);
1375 if (!(cmd & PCI_COMMAND_MEMORY))
1376 pci_write_config_word(vdev->pdev, PCI_COMMAND,
1377 cmd | PCI_COMMAND_MEMORY);
1378
1379 return cmd;
1380}
1381
1382void vfio_pci_memory_unlock_and_restore(struct vfio_pci_device *vdev, u16 cmd)
1383{
1384 pci_write_config_word(vdev->pdev, PCI_COMMAND, cmd);
1385 up_write(&vdev->memory_lock);
1386}
1387
1388/* Caller holds vma_lock */
1389static int __vfio_pci_add_vma(struct vfio_pci_device *vdev,
1390 struct vm_area_struct *vma)
1391{
1392 struct vfio_pci_mmap_vma *mmap_vma;
1393
1394 mmap_vma = kmalloc(sizeof(*mmap_vma), GFP_KERNEL);
1395 if (!mmap_vma)
1396 return -ENOMEM;
1397
1398 mmap_vma->vma = vma;
1399 list_add(&mmap_vma->vma_next, &vdev->vma_list);
1400
1401 return 0;
1402}
1403
1404/*
1405 * Zap mmaps on open so that we can fault them in on access and therefore
1406 * our vma_list only tracks mappings accessed since last zap.
1407 */
1408static void vfio_pci_mmap_open(struct vm_area_struct *vma)
1409{
1410 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1411}
1412
1413static void vfio_pci_mmap_close(struct vm_area_struct *vma)
1414{
1415 struct vfio_pci_device *vdev = vma->vm_private_data;
1416 struct vfio_pci_mmap_vma *mmap_vma;
1417
1418 mutex_lock(&vdev->vma_lock);
1419 list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1420 if (mmap_vma->vma == vma) {
1421 list_del(&mmap_vma->vma_next);
1422 kfree(mmap_vma);
1423 break;
1424 }
1425 }
1426 mutex_unlock(&vdev->vma_lock);
1427}
1428
1429static vm_fault_t vfio_pci_mmap_fault(struct vm_fault *vmf)
1430{
1431 struct vm_area_struct *vma = vmf->vma;
1432 struct vfio_pci_device *vdev = vma->vm_private_data;
1433 struct vfio_pci_mmap_vma *mmap_vma;
1434 vm_fault_t ret = VM_FAULT_NOPAGE;
1435
1436 mutex_lock(&vdev->vma_lock);
1437 down_read(&vdev->memory_lock);
1438
1439 if (!__vfio_pci_memory_enabled(vdev)) {
1440 ret = VM_FAULT_SIGBUS;
1441 goto up_out;
1442 }
1443
1444 /*
1445 * We populate the whole vma on fault, so we need to test whether
1446 * the vma has already been mapped, such as for concurrent faults
1447 * to the same vma. io_remap_pfn_range() will trigger a BUG_ON if
1448 * we ask it to fill the same range again.
1449 */
1450 list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1451 if (mmap_vma->vma == vma)
1452 goto up_out;
1453 }
1454
1455 if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1456 vma->vm_end - vma->vm_start,
1457 vma->vm_page_prot)) {
1458 ret = VM_FAULT_SIGBUS;
1459 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1460 goto up_out;
1461 }
1462
1463 if (__vfio_pci_add_vma(vdev, vma)) {
1464 ret = VM_FAULT_OOM;
1465 zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1466 }
1467
1468up_out:
1469 up_read(&vdev->memory_lock);
1470 mutex_unlock(&vdev->vma_lock);
1471 return ret;
1472}
1473
1474static const struct vm_operations_struct vfio_pci_mmap_ops = {
1475 .open = vfio_pci_mmap_open,
1476 .close = vfio_pci_mmap_close,
1477 .fault = vfio_pci_mmap_fault,
1478};
1479
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001480static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma)
1481{
1482 struct vfio_pci_device *vdev = device_data;
1483 struct pci_dev *pdev = vdev->pdev;
1484 unsigned int index;
1485 u64 phys_len, req_len, pgoff, req_start;
1486 int ret;
1487
1488 index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1489
Olivier Deprez0e641232021-09-23 10:07:05 +02001490 if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1491 return -EINVAL;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001492 if (vma->vm_end < vma->vm_start)
1493 return -EINVAL;
1494 if ((vma->vm_flags & VM_SHARED) == 0)
1495 return -EINVAL;
David Brazdil0f672f62019-12-10 10:32:29 +00001496 if (index >= VFIO_PCI_NUM_REGIONS) {
1497 int regnum = index - VFIO_PCI_NUM_REGIONS;
1498 struct vfio_pci_region *region = vdev->region + regnum;
1499
Olivier Deprez0e641232021-09-23 10:07:05 +02001500 if (region->ops && region->ops->mmap &&
David Brazdil0f672f62019-12-10 10:32:29 +00001501 (region->flags & VFIO_REGION_INFO_FLAG_MMAP))
1502 return region->ops->mmap(vdev, region, vma);
1503 return -EINVAL;
1504 }
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001505 if (index >= VFIO_PCI_ROM_REGION_INDEX)
1506 return -EINVAL;
1507 if (!vdev->bar_mmap_supported[index])
1508 return -EINVAL;
1509
1510 phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1511 req_len = vma->vm_end - vma->vm_start;
1512 pgoff = vma->vm_pgoff &
1513 ((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1514 req_start = pgoff << PAGE_SHIFT;
1515
1516 if (req_start + req_len > phys_len)
1517 return -EINVAL;
1518
1519 /*
1520 * Even though we don't make use of the barmap for the mmap,
1521 * we need to request the region and the barmap tracks that.
1522 */
1523 if (!vdev->barmap[index]) {
1524 ret = pci_request_selected_regions(pdev,
1525 1 << index, "vfio-pci");
1526 if (ret)
1527 return ret;
1528
1529 vdev->barmap[index] = pci_iomap(pdev, index, 0);
1530 if (!vdev->barmap[index]) {
1531 pci_release_selected_regions(pdev, 1 << index);
1532 return -ENOMEM;
1533 }
1534 }
1535
1536 vma->vm_private_data = vdev;
1537 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1538 vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1539
Olivier Deprez0e641232021-09-23 10:07:05 +02001540 /*
1541 * See remap_pfn_range(), called from vfio_pci_fault() but we can't
1542 * change vm_flags within the fault handler. Set them now.
1543 */
1544 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1545 vma->vm_ops = &vfio_pci_mmap_ops;
1546
1547 return 0;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001548}
1549
1550static void vfio_pci_request(void *device_data, unsigned int count)
1551{
1552 struct vfio_pci_device *vdev = device_data;
David Brazdil0f672f62019-12-10 10:32:29 +00001553 struct pci_dev *pdev = vdev->pdev;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001554
1555 mutex_lock(&vdev->igate);
1556
1557 if (vdev->req_trigger) {
1558 if (!(count % 10))
David Brazdil0f672f62019-12-10 10:32:29 +00001559 pci_notice_ratelimited(pdev,
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001560 "Relaying device request to user (#%u)\n",
1561 count);
1562 eventfd_signal(vdev->req_trigger, 1);
1563 } else if (count == 0) {
David Brazdil0f672f62019-12-10 10:32:29 +00001564 pci_warn(pdev,
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001565 "No device request channel registered, blocked until released by user\n");
1566 }
1567
1568 mutex_unlock(&vdev->igate);
1569}
1570
1571static const struct vfio_device_ops vfio_pci_ops = {
1572 .name = "vfio-pci",
1573 .open = vfio_pci_open,
1574 .release = vfio_pci_release,
1575 .ioctl = vfio_pci_ioctl,
1576 .read = vfio_pci_read,
1577 .write = vfio_pci_write,
1578 .mmap = vfio_pci_mmap,
1579 .request = vfio_pci_request,
1580};
1581
David Brazdil0f672f62019-12-10 10:32:29 +00001582static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev);
1583static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck);
1584
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001585static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1586{
1587 struct vfio_pci_device *vdev;
1588 struct iommu_group *group;
1589 int ret;
1590
1591 if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1592 return -EINVAL;
1593
1594 /*
1595 * Prevent binding to PFs with VFs enabled, this too easily allows
1596 * userspace instance with VFs and PFs from the same device, which
1597 * cannot work. Disabling SR-IOV here would initiate removing the
1598 * VFs, which would unbind the driver, which is prone to blocking
1599 * if that VF is also in use by vfio-pci. Just reject these PFs
1600 * and let the user sort it out.
1601 */
1602 if (pci_num_vf(pdev)) {
1603 pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
1604 return -EBUSY;
1605 }
1606
1607 group = vfio_iommu_group_get(&pdev->dev);
1608 if (!group)
1609 return -EINVAL;
1610
1611 vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
1612 if (!vdev) {
1613 vfio_iommu_group_put(group, &pdev->dev);
1614 return -ENOMEM;
1615 }
1616
1617 vdev->pdev = pdev;
1618 vdev->irq_type = VFIO_PCI_NUM_IRQS;
1619 mutex_init(&vdev->igate);
1620 spin_lock_init(&vdev->irqlock);
1621 mutex_init(&vdev->ioeventfds_lock);
Olivier Deprez0e641232021-09-23 10:07:05 +02001622 INIT_LIST_HEAD(&vdev->dummy_resources_list);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001623 INIT_LIST_HEAD(&vdev->ioeventfds_list);
Olivier Deprez0e641232021-09-23 10:07:05 +02001624 mutex_init(&vdev->vma_lock);
1625 INIT_LIST_HEAD(&vdev->vma_list);
1626 init_rwsem(&vdev->memory_lock);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001627
1628 ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
1629 if (ret) {
1630 vfio_iommu_group_put(group, &pdev->dev);
1631 kfree(vdev);
1632 return ret;
1633 }
1634
David Brazdil0f672f62019-12-10 10:32:29 +00001635 ret = vfio_pci_reflck_attach(vdev);
1636 if (ret) {
1637 vfio_del_group_dev(&pdev->dev);
1638 vfio_iommu_group_put(group, &pdev->dev);
1639 kfree(vdev);
1640 return ret;
1641 }
1642
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001643 if (vfio_pci_is_vga(pdev)) {
1644 vga_client_register(pdev, vdev, NULL, vfio_pci_set_vga_decode);
1645 vga_set_legacy_decoding(pdev,
1646 vfio_pci_set_vga_decode(vdev, false));
1647 }
1648
David Brazdil0f672f62019-12-10 10:32:29 +00001649 vfio_pci_probe_power_state(vdev);
1650
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001651 if (!disable_idle_d3) {
1652 /*
1653 * pci-core sets the device power state to an unknown value at
1654 * bootup and after being removed from a driver. The only
1655 * transition it allows from this unknown state is to D0, which
1656 * typically happens when a driver calls pci_enable_device().
1657 * We're not ready to enable the device yet, but we do want to
1658 * be able to get to D3. Therefore first do a D0 transition
1659 * before going to D3.
1660 */
David Brazdil0f672f62019-12-10 10:32:29 +00001661 vfio_pci_set_power_state(vdev, PCI_D0);
1662 vfio_pci_set_power_state(vdev, PCI_D3hot);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001663 }
1664
1665 return ret;
1666}
1667
1668static void vfio_pci_remove(struct pci_dev *pdev)
1669{
1670 struct vfio_pci_device *vdev;
1671
1672 vdev = vfio_del_group_dev(&pdev->dev);
1673 if (!vdev)
1674 return;
1675
David Brazdil0f672f62019-12-10 10:32:29 +00001676 vfio_pci_reflck_put(vdev->reflck);
1677
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001678 vfio_iommu_group_put(pdev->dev.iommu_group, &pdev->dev);
1679 kfree(vdev->region);
1680 mutex_destroy(&vdev->ioeventfds_lock);
David Brazdil0f672f62019-12-10 10:32:29 +00001681
1682 if (!disable_idle_d3)
1683 vfio_pci_set_power_state(vdev, PCI_D0);
1684
1685 kfree(vdev->pm_save);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001686 kfree(vdev);
1687
1688 if (vfio_pci_is_vga(pdev)) {
1689 vga_client_register(pdev, NULL, NULL, NULL);
1690 vga_set_legacy_decoding(pdev,
1691 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
1692 VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM);
1693 }
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001694}
1695
1696static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
1697 pci_channel_state_t state)
1698{
1699 struct vfio_pci_device *vdev;
1700 struct vfio_device *device;
1701
1702 device = vfio_device_get_from_dev(&pdev->dev);
1703 if (device == NULL)
1704 return PCI_ERS_RESULT_DISCONNECT;
1705
1706 vdev = vfio_device_data(device);
1707 if (vdev == NULL) {
1708 vfio_device_put(device);
1709 return PCI_ERS_RESULT_DISCONNECT;
1710 }
1711
1712 mutex_lock(&vdev->igate);
1713
1714 if (vdev->err_trigger)
1715 eventfd_signal(vdev->err_trigger, 1);
1716
1717 mutex_unlock(&vdev->igate);
1718
1719 vfio_device_put(device);
1720
1721 return PCI_ERS_RESULT_CAN_RECOVER;
1722}
1723
1724static const struct pci_error_handlers vfio_err_handlers = {
1725 .error_detected = vfio_pci_aer_err_detected,
1726};
1727
1728static struct pci_driver vfio_pci_driver = {
1729 .name = "vfio-pci",
1730 .id_table = NULL, /* only dynamic ids */
1731 .probe = vfio_pci_probe,
1732 .remove = vfio_pci_remove,
1733 .err_handler = &vfio_err_handlers,
1734};
1735
David Brazdil0f672f62019-12-10 10:32:29 +00001736static DEFINE_MUTEX(reflck_lock);
1737
1738static struct vfio_pci_reflck *vfio_pci_reflck_alloc(void)
1739{
1740 struct vfio_pci_reflck *reflck;
1741
1742 reflck = kzalloc(sizeof(*reflck), GFP_KERNEL);
1743 if (!reflck)
1744 return ERR_PTR(-ENOMEM);
1745
1746 kref_init(&reflck->kref);
1747 mutex_init(&reflck->lock);
1748
1749 return reflck;
1750}
1751
1752static void vfio_pci_reflck_get(struct vfio_pci_reflck *reflck)
1753{
1754 kref_get(&reflck->kref);
1755}
1756
1757static int vfio_pci_reflck_find(struct pci_dev *pdev, void *data)
1758{
1759 struct vfio_pci_reflck **preflck = data;
1760 struct vfio_device *device;
1761 struct vfio_pci_device *vdev;
1762
1763 device = vfio_device_get_from_dev(&pdev->dev);
1764 if (!device)
1765 return 0;
1766
1767 if (pci_dev_driver(pdev) != &vfio_pci_driver) {
1768 vfio_device_put(device);
1769 return 0;
1770 }
1771
1772 vdev = vfio_device_data(device);
1773
1774 if (vdev->reflck) {
1775 vfio_pci_reflck_get(vdev->reflck);
1776 *preflck = vdev->reflck;
1777 vfio_device_put(device);
1778 return 1;
1779 }
1780
1781 vfio_device_put(device);
1782 return 0;
1783}
1784
1785static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev)
1786{
1787 bool slot = !pci_probe_reset_slot(vdev->pdev->slot);
1788
1789 mutex_lock(&reflck_lock);
1790
1791 if (pci_is_root_bus(vdev->pdev->bus) ||
1792 vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_reflck_find,
1793 &vdev->reflck, slot) <= 0)
1794 vdev->reflck = vfio_pci_reflck_alloc();
1795
1796 mutex_unlock(&reflck_lock);
1797
1798 return PTR_ERR_OR_ZERO(vdev->reflck);
1799}
1800
1801static void vfio_pci_reflck_release(struct kref *kref)
1802{
1803 struct vfio_pci_reflck *reflck = container_of(kref,
1804 struct vfio_pci_reflck,
1805 kref);
1806
1807 kfree(reflck);
1808 mutex_unlock(&reflck_lock);
1809}
1810
1811static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck)
1812{
1813 kref_put_mutex(&reflck->kref, vfio_pci_reflck_release, &reflck_lock);
1814}
1815
David Brazdil0f672f62019-12-10 10:32:29 +00001816static int vfio_pci_get_unused_devs(struct pci_dev *pdev, void *data)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001817{
1818 struct vfio_devices *devs = data;
1819 struct vfio_device *device;
David Brazdil0f672f62019-12-10 10:32:29 +00001820 struct vfio_pci_device *vdev;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001821
1822 if (devs->cur_index == devs->max_index)
1823 return -ENOSPC;
1824
1825 device = vfio_device_get_from_dev(&pdev->dev);
1826 if (!device)
1827 return -EINVAL;
1828
1829 if (pci_dev_driver(pdev) != &vfio_pci_driver) {
1830 vfio_device_put(device);
1831 return -EBUSY;
1832 }
1833
David Brazdil0f672f62019-12-10 10:32:29 +00001834 vdev = vfio_device_data(device);
1835
1836 /* Fault if the device is not unused */
1837 if (vdev->refcnt) {
1838 vfio_device_put(device);
1839 return -EBUSY;
1840 }
1841
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001842 devs->devices[devs->cur_index++] = device;
1843 return 0;
1844}
1845
Olivier Deprez0e641232021-09-23 10:07:05 +02001846static int vfio_pci_try_zap_and_vma_lock_cb(struct pci_dev *pdev, void *data)
1847{
1848 struct vfio_devices *devs = data;
1849 struct vfio_device *device;
1850 struct vfio_pci_device *vdev;
1851
1852 if (devs->cur_index == devs->max_index)
1853 return -ENOSPC;
1854
1855 device = vfio_device_get_from_dev(&pdev->dev);
1856 if (!device)
1857 return -EINVAL;
1858
1859 if (pci_dev_driver(pdev) != &vfio_pci_driver) {
1860 vfio_device_put(device);
1861 return -EBUSY;
1862 }
1863
1864 vdev = vfio_device_data(device);
1865
1866 /*
1867 * Locking multiple devices is prone to deadlock, runaway and
1868 * unwind if we hit contention.
1869 */
1870 if (!vfio_pci_zap_and_vma_lock(vdev, true)) {
1871 vfio_device_put(device);
1872 return -EBUSY;
1873 }
1874
1875 devs->devices[devs->cur_index++] = device;
1876 return 0;
1877}
1878
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001879/*
David Brazdil0f672f62019-12-10 10:32:29 +00001880 * If a bus or slot reset is available for the provided device and:
1881 * - All of the devices affected by that bus or slot reset are unused
1882 * (!refcnt)
1883 * - At least one of the affected devices is marked dirty via
1884 * needs_reset (such as by lack of FLR support)
1885 * Then attempt to perform that bus or slot reset. Callers are required
1886 * to hold vdev->reflck->lock, protecting the bus/slot reset group from
1887 * concurrent opens. A vfio_device reference is acquired for each device
1888 * to prevent unbinds during the reset operation.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001889 *
1890 * NB: vfio-core considers a group to be viable even if some devices are
1891 * bound to drivers like pci-stub or pcieport. Here we require all devices
1892 * to be bound to vfio_pci since that's the only way we can be sure they
1893 * stay put.
1894 */
1895static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev)
1896{
1897 struct vfio_devices devs = { .cur_index = 0 };
1898 int i = 0, ret = -EINVAL;
David Brazdil0f672f62019-12-10 10:32:29 +00001899 bool slot = false;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001900 struct vfio_pci_device *tmp;
1901
1902 if (!pci_probe_reset_slot(vdev->pdev->slot))
1903 slot = true;
1904 else if (pci_probe_reset_bus(vdev->pdev->bus))
1905 return;
1906
1907 if (vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
1908 &i, slot) || !i)
1909 return;
1910
1911 devs.max_index = i;
1912 devs.devices = kcalloc(i, sizeof(struct vfio_device *), GFP_KERNEL);
1913 if (!devs.devices)
1914 return;
1915
1916 if (vfio_pci_for_each_slot_or_bus(vdev->pdev,
David Brazdil0f672f62019-12-10 10:32:29 +00001917 vfio_pci_get_unused_devs,
1918 &devs, slot))
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001919 goto put_devs;
1920
David Brazdil0f672f62019-12-10 10:32:29 +00001921 /* Does at least one need a reset? */
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001922 for (i = 0; i < devs.cur_index; i++) {
1923 tmp = vfio_device_data(devs.devices[i]);
David Brazdil0f672f62019-12-10 10:32:29 +00001924 if (tmp->needs_reset) {
1925 ret = pci_reset_bus(vdev->pdev);
1926 break;
1927 }
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001928 }
1929
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001930put_devs:
1931 for (i = 0; i < devs.cur_index; i++) {
1932 tmp = vfio_device_data(devs.devices[i]);
David Brazdil0f672f62019-12-10 10:32:29 +00001933
1934 /*
1935 * If reset was successful, affected devices no longer need
1936 * a reset and we should return all the collateral devices
1937 * to low power. If not successful, we either didn't reset
1938 * the bus or timed out waiting for it, so let's not touch
1939 * the power state.
1940 */
1941 if (!ret) {
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001942 tmp->needs_reset = false;
1943
David Brazdil0f672f62019-12-10 10:32:29 +00001944 if (tmp != vdev && !disable_idle_d3)
1945 vfio_pci_set_power_state(tmp, PCI_D3hot);
1946 }
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001947
1948 vfio_device_put(devs.devices[i]);
1949 }
1950
1951 kfree(devs.devices);
1952}
1953
1954static void __exit vfio_pci_cleanup(void)
1955{
1956 pci_unregister_driver(&vfio_pci_driver);
1957 vfio_pci_uninit_perm_bits();
1958}
1959
1960static void __init vfio_pci_fill_ids(void)
1961{
1962 char *p, *id;
1963 int rc;
1964
1965 /* no ids passed actually */
1966 if (ids[0] == '\0')
1967 return;
1968
1969 /* add ids specified in the module parameter */
1970 p = ids;
1971 while ((id = strsep(&p, ","))) {
1972 unsigned int vendor, device, subvendor = PCI_ANY_ID,
1973 subdevice = PCI_ANY_ID, class = 0, class_mask = 0;
1974 int fields;
1975
1976 if (!strlen(id))
1977 continue;
1978
1979 fields = sscanf(id, "%x:%x:%x:%x:%x:%x",
1980 &vendor, &device, &subvendor, &subdevice,
1981 &class, &class_mask);
1982
1983 if (fields < 2) {
1984 pr_warn("invalid id string \"%s\"\n", id);
1985 continue;
1986 }
1987
1988 rc = pci_add_dynid(&vfio_pci_driver, vendor, device,
1989 subvendor, subdevice, class, class_mask, 0);
1990 if (rc)
David Brazdil0f672f62019-12-10 10:32:29 +00001991 pr_warn("failed to add dynamic id [%04x:%04x[%04x:%04x]] class %#08x/%08x (%d)\n",
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001992 vendor, device, subvendor, subdevice,
1993 class, class_mask, rc);
1994 else
David Brazdil0f672f62019-12-10 10:32:29 +00001995 pr_info("add [%04x:%04x[%04x:%04x]] class %#08x/%08x\n",
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001996 vendor, device, subvendor, subdevice,
1997 class, class_mask);
1998 }
1999}
2000
2001static int __init vfio_pci_init(void)
2002{
2003 int ret;
2004
2005 /* Allocate shared config space permision data used by all devices */
2006 ret = vfio_pci_init_perm_bits();
2007 if (ret)
2008 return ret;
2009
2010 /* Register and scan for devices */
2011 ret = pci_register_driver(&vfio_pci_driver);
2012 if (ret)
2013 goto out_driver;
2014
2015 vfio_pci_fill_ids();
2016
2017 return 0;
2018
2019out_driver:
2020 vfio_pci_uninit_perm_bits();
2021 return ret;
2022}
2023
2024module_init(vfio_pci_init);
2025module_exit(vfio_pci_cleanup);
2026
2027MODULE_VERSION(DRIVER_VERSION);
2028MODULE_LICENSE("GPL v2");
2029MODULE_AUTHOR(DRIVER_AUTHOR);
2030MODULE_DESCRIPTION(DRIVER_DESC);