1
2
3
4
5
6
7
8
9
10
11#include <linux/pci.h>
12#include <linux/slab.h>
13#include <linux/mutex.h>
14#include <linux/export.h>
15#include <linux/string.h>
16#include <linux/delay.h>
17#include <linux/pci-ats.h>
18#include "pci.h"
19
20#define VIRTFN_ID_LEN 16
21
22static inline u8 virtfn_bus(struct pci_dev *dev, int id)
23{
24 return dev->bus->number + ((dev->devfn + dev->sriov->offset +
25 dev->sriov->stride * id) >> 8);
26}
27
28static inline u8 virtfn_devfn(struct pci_dev *dev, int id)
29{
30 return (dev->devfn + dev->sriov->offset +
31 dev->sriov->stride * id) & 0xff;
32}
33
34static struct pci_bus *virtfn_add_bus(struct pci_bus *bus, int busnr)
35{
36 int rc;
37 struct pci_bus *child;
38
39 if (bus->number == busnr)
40 return bus;
41
42 child = pci_find_bus(pci_domain_nr(bus), busnr);
43 if (child)
44 return child;
45
46 child = pci_add_new_bus(bus, NULL, busnr);
47 if (!child)
48 return NULL;
49
50 pci_bus_insert_busn_res(child, busnr, busnr);
51 child->dev.parent = bus->bridge;
52 rc = pci_bus_add_child(child);
53 if (rc) {
54 pci_remove_bus(child);
55 return NULL;
56 }
57
58 return child;
59}
60
61static void virtfn_remove_bus(struct pci_bus *bus, int busnr)
62{
63 struct pci_bus *child;
64
65 if (bus->number == busnr)
66 return;
67
68 child = pci_find_bus(pci_domain_nr(bus), busnr);
69 BUG_ON(!child);
70
71 if (list_empty(&child->devices))
72 pci_remove_bus(child);
73}
74
75static int virtfn_add(struct pci_dev *dev, int id, int reset)
76{
77 int i;
78 int rc;
79 u64 size;
80 char buf[VIRTFN_ID_LEN];
81 struct pci_dev *virtfn;
82 struct resource *res;
83 struct pci_sriov *iov = dev->sriov;
84
85 virtfn = alloc_pci_dev();
86 if (!virtfn)
87 return -ENOMEM;
88
89 mutex_lock(&iov->dev->sriov->lock);
90 virtfn->bus = virtfn_add_bus(dev->bus, virtfn_bus(dev, id));
91 if (!virtfn->bus) {
92 kfree(virtfn);
93 mutex_unlock(&iov->dev->sriov->lock);
94 return -ENOMEM;
95 }
96 virtfn->devfn = virtfn_devfn(dev, id);
97 virtfn->vendor = dev->vendor;
98 pci_read_config_word(dev, iov->pos + PCI_SRIOV_VF_DID, &virtfn->device);
99 pci_setup_device(virtfn);
100 virtfn->dev.parent = dev->dev.parent;
101
102 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
103 res = dev->resource + PCI_IOV_RESOURCES + i;
104 if (!res->parent)
105 continue;
106 virtfn->resource[i].name = pci_name(virtfn);
107 virtfn->resource[i].flags = res->flags;
108 size = resource_size(res);
109 do_div(size, iov->total_VFs);
110 virtfn->resource[i].start = res->start + size * id;
111 virtfn->resource[i].end = virtfn->resource[i].start + size - 1;
112 rc = request_resource(res, &virtfn->resource[i]);
113 BUG_ON(rc);
114 }
115
116 if (reset)
117 __pci_reset_function(virtfn);
118
119 pci_device_add(virtfn, virtfn->bus);
120 mutex_unlock(&iov->dev->sriov->lock);
121
122 virtfn->physfn = pci_dev_get(dev);
123 virtfn->is_virtfn = 1;
124
125 rc = pci_bus_add_device(virtfn);
126 if (rc)
127 goto failed1;
128 sprintf(buf, "virtfn%u", id);
129 rc = sysfs_create_link(&dev->dev.kobj, &virtfn->dev.kobj, buf);
130 if (rc)
131 goto failed1;
132 rc = sysfs_create_link(&virtfn->dev.kobj, &dev->dev.kobj, "physfn");
133 if (rc)
134 goto failed2;
135
136 kobject_uevent(&virtfn->dev.kobj, KOBJ_CHANGE);
137
138 return 0;
139
140failed2:
141 sysfs_remove_link(&dev->dev.kobj, buf);
142failed1:
143 pci_dev_put(dev);
144 mutex_lock(&iov->dev->sriov->lock);
145 pci_stop_and_remove_bus_device(virtfn);
146 virtfn_remove_bus(dev->bus, virtfn_bus(dev, id));
147 mutex_unlock(&iov->dev->sriov->lock);
148
149 return rc;
150}
151
152static void virtfn_remove(struct pci_dev *dev, int id, int reset)
153{
154 char buf[VIRTFN_ID_LEN];
155 struct pci_bus *bus;
156 struct pci_dev *virtfn;
157 struct pci_sriov *iov = dev->sriov;
158
159 bus = pci_find_bus(pci_domain_nr(dev->bus), virtfn_bus(dev, id));
160 if (!bus)
161 return;
162
163 virtfn = pci_get_slot(bus, virtfn_devfn(dev, id));
164 if (!virtfn)
165 return;
166
167 pci_dev_put(virtfn);
168
169 if (reset) {
170 device_release_driver(&virtfn->dev);
171 __pci_reset_function(virtfn);
172 }
173
174 sprintf(buf, "virtfn%u", id);
175 sysfs_remove_link(&dev->dev.kobj, buf);
176
177
178
179
180
181 if (virtfn->dev.kobj.sd)
182 sysfs_remove_link(&virtfn->dev.kobj, "physfn");
183
184 mutex_lock(&iov->dev->sriov->lock);
185 pci_stop_and_remove_bus_device(virtfn);
186 virtfn_remove_bus(dev->bus, virtfn_bus(dev, id));
187 mutex_unlock(&iov->dev->sriov->lock);
188
189 pci_dev_put(dev);
190}
191
192static int sriov_migration(struct pci_dev *dev)
193{
194 u16 status;
195 struct pci_sriov *iov = dev->sriov;
196
197 if (!iov->num_VFs)
198 return 0;
199
200 if (!(iov->cap & PCI_SRIOV_CAP_VFM))
201 return 0;
202
203 pci_read_config_word(dev, iov->pos + PCI_SRIOV_STATUS, &status);
204 if (!(status & PCI_SRIOV_STATUS_VFM))
205 return 0;
206
207 schedule_work(&iov->mtask);
208
209 return 1;
210}
211
212static void sriov_migration_task(struct work_struct *work)
213{
214 int i;
215 u8 state;
216 u16 status;
217 struct pci_sriov *iov = container_of(work, struct pci_sriov, mtask);
218
219 for (i = iov->initial_VFs; i < iov->num_VFs; i++) {
220 state = readb(iov->mstate + i);
221 if (state == PCI_SRIOV_VFM_MI) {
222 writeb(PCI_SRIOV_VFM_AV, iov->mstate + i);
223 state = readb(iov->mstate + i);
224 if (state == PCI_SRIOV_VFM_AV)
225 virtfn_add(iov->self, i, 1);
226 } else if (state == PCI_SRIOV_VFM_MO) {
227 virtfn_remove(iov->self, i, 1);
228 writeb(PCI_SRIOV_VFM_UA, iov->mstate + i);
229 state = readb(iov->mstate + i);
230 if (state == PCI_SRIOV_VFM_AV)
231 virtfn_add(iov->self, i, 0);
232 }
233 }
234
235 pci_read_config_word(iov->self, iov->pos + PCI_SRIOV_STATUS, &status);
236 status &= ~PCI_SRIOV_STATUS_VFM;
237 pci_write_config_word(iov->self, iov->pos + PCI_SRIOV_STATUS, status);
238}
239
240static int sriov_enable_migration(struct pci_dev *dev, int nr_virtfn)
241{
242 int bir;
243 u32 table;
244 resource_size_t pa;
245 struct pci_sriov *iov = dev->sriov;
246
247 if (nr_virtfn <= iov->initial_VFs)
248 return 0;
249
250 pci_read_config_dword(dev, iov->pos + PCI_SRIOV_VFM, &table);
251 bir = PCI_SRIOV_VFM_BIR(table);
252 if (bir > PCI_STD_RESOURCE_END)
253 return -EIO;
254
255 table = PCI_SRIOV_VFM_OFFSET(table);
256 if (table + nr_virtfn > pci_resource_len(dev, bir))
257 return -EIO;
258
259 pa = pci_resource_start(dev, bir) + table;
260 iov->mstate = ioremap(pa, nr_virtfn);
261 if (!iov->mstate)
262 return -ENOMEM;
263
264 INIT_WORK(&iov->mtask, sriov_migration_task);
265
266 iov->ctrl |= PCI_SRIOV_CTRL_VFM | PCI_SRIOV_CTRL_INTR;
267 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
268
269 return 0;
270}
271
272static void sriov_disable_migration(struct pci_dev *dev)
273{
274 struct pci_sriov *iov = dev->sriov;
275
276 iov->ctrl &= ~(PCI_SRIOV_CTRL_VFM | PCI_SRIOV_CTRL_INTR);
277 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
278
279 cancel_work_sync(&iov->mtask);
280 iounmap(iov->mstate);
281}
282
283static int sriov_enable(struct pci_dev *dev, int nr_virtfn)
284{
285 int rc;
286 int i, j;
287 int nres;
288 u16 offset, stride, initial;
289 struct resource *res;
290 struct pci_dev *pdev;
291 struct pci_sriov *iov = dev->sriov;
292 int bars = 0;
293
294 if (!nr_virtfn)
295 return 0;
296
297 if (iov->num_VFs)
298 return -EINVAL;
299
300 pci_read_config_word(dev, iov->pos + PCI_SRIOV_INITIAL_VF, &initial);
301 if (initial > iov->total_VFs ||
302 (!(iov->cap & PCI_SRIOV_CAP_VFM) && (initial != iov->total_VFs)))
303 return -EIO;
304
305 if (nr_virtfn < 0 || nr_virtfn > iov->total_VFs ||
306 (!(iov->cap & PCI_SRIOV_CAP_VFM) && (nr_virtfn > initial)))
307 return -EINVAL;
308
309 pci_write_config_word(dev, iov->pos + PCI_SRIOV_NUM_VF, nr_virtfn);
310 pci_read_config_word(dev, iov->pos + PCI_SRIOV_VF_OFFSET, &offset);
311 pci_read_config_word(dev, iov->pos + PCI_SRIOV_VF_STRIDE, &stride);
312 if (!offset || (nr_virtfn > 1 && !stride))
313 return -EIO;
314
315 nres = 0;
316 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
317 bars |= (1 << (i + PCI_IOV_RESOURCES));
318 res = dev->resource + PCI_IOV_RESOURCES + i;
319 if (res->parent)
320 nres++;
321 }
322 if (nres != iov->nres) {
323 dev_err(&dev->dev, "not enough MMIO resources for SR-IOV\n");
324 return -ENOMEM;
325 }
326
327 iov->offset = offset;
328 iov->stride = stride;
329
330 if (virtfn_bus(dev, nr_virtfn - 1) > dev->bus->busn_res.end) {
331 dev_err(&dev->dev, "SR-IOV: bus number out of range\n");
332 return -ENOMEM;
333 }
334
335 if (pci_enable_resources(dev, bars)) {
336 dev_err(&dev->dev, "SR-IOV: IOV BARS not allocated\n");
337 return -ENOMEM;
338 }
339
340 if (iov->link != dev->devfn) {
341 pdev = pci_get_slot(dev->bus, iov->link);
342 if (!pdev)
343 return -ENODEV;
344
345 pci_dev_put(pdev);
346
347 if (!pdev->is_physfn)
348 return -ENODEV;
349
350 rc = sysfs_create_link(&dev->dev.kobj,
351 &pdev->dev.kobj, "dep_link");
352 if (rc)
353 return rc;
354 }
355
356 iov->ctrl |= PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE;
357 pci_cfg_access_lock(dev);
358 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
359 msleep(100);
360 pci_cfg_access_unlock(dev);
361
362 iov->initial_VFs = initial;
363 if (nr_virtfn < initial)
364 initial = nr_virtfn;
365
366 for (i = 0; i < initial; i++) {
367 rc = virtfn_add(dev, i, 0);
368 if (rc)
369 goto failed;
370 }
371
372 if (iov->cap & PCI_SRIOV_CAP_VFM) {
373 rc = sriov_enable_migration(dev, nr_virtfn);
374 if (rc)
375 goto failed;
376 }
377
378 kobject_uevent(&dev->dev.kobj, KOBJ_CHANGE);
379 iov->num_VFs = nr_virtfn;
380
381 return 0;
382
383failed:
384 for (j = 0; j < i; j++)
385 virtfn_remove(dev, j, 0);
386
387 iov->ctrl &= ~(PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE);
388 pci_cfg_access_lock(dev);
389 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
390 ssleep(1);
391 pci_cfg_access_unlock(dev);
392
393 if (iov->link != dev->devfn)
394 sysfs_remove_link(&dev->dev.kobj, "dep_link");
395
396 return rc;
397}
398
399static void sriov_disable(struct pci_dev *dev)
400{
401 int i;
402 struct pci_sriov *iov = dev->sriov;
403
404 if (!iov->num_VFs)
405 return;
406
407 if (iov->cap & PCI_SRIOV_CAP_VFM)
408 sriov_disable_migration(dev);
409
410 for (i = 0; i < iov->num_VFs; i++)
411 virtfn_remove(dev, i, 0);
412
413 iov->ctrl &= ~(PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE);
414 pci_cfg_access_lock(dev);
415 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
416 ssleep(1);
417 pci_cfg_access_unlock(dev);
418
419 if (iov->link != dev->devfn)
420 sysfs_remove_link(&dev->dev.kobj, "dep_link");
421
422 iov->num_VFs = 0;
423}
424
425static int sriov_init(struct pci_dev *dev, int pos)
426{
427 int i;
428 int rc;
429 int nres;
430 u32 pgsz;
431 u16 ctrl, total, offset, stride;
432 struct pci_sriov *iov;
433 struct resource *res;
434 struct pci_dev *pdev;
435
436 if (pci_pcie_type(dev) != PCI_EXP_TYPE_RC_END &&
437 pci_pcie_type(dev) != PCI_EXP_TYPE_ENDPOINT)
438 return -ENODEV;
439
440 pci_read_config_word(dev, pos + PCI_SRIOV_CTRL, &ctrl);
441 if (ctrl & PCI_SRIOV_CTRL_VFE) {
442 pci_write_config_word(dev, pos + PCI_SRIOV_CTRL, 0);
443 ssleep(1);
444 }
445
446 pci_read_config_word(dev, pos + PCI_SRIOV_TOTAL_VF, &total);
447 if (!total)
448 return 0;
449
450 ctrl = 0;
451 list_for_each_entry(pdev, &dev->bus->devices, bus_list)
452 if (pdev->is_physfn)
453 goto found;
454
455 pdev = NULL;
456 if (pci_ari_enabled(dev->bus))
457 ctrl |= PCI_SRIOV_CTRL_ARI;
458
459found:
460 pci_write_config_word(dev, pos + PCI_SRIOV_CTRL, ctrl);
461 pci_read_config_word(dev, pos + PCI_SRIOV_VF_OFFSET, &offset);
462 pci_read_config_word(dev, pos + PCI_SRIOV_VF_STRIDE, &stride);
463 if (!offset || (total > 1 && !stride))
464 return -EIO;
465
466 pci_read_config_dword(dev, pos + PCI_SRIOV_SUP_PGSIZE, &pgsz);
467 i = PAGE_SHIFT > 12 ? PAGE_SHIFT - 12 : 0;
468 pgsz &= ~((1 << i) - 1);
469 if (!pgsz)
470 return -EIO;
471
472 pgsz &= ~(pgsz - 1);
473 pci_write_config_dword(dev, pos + PCI_SRIOV_SYS_PGSIZE, pgsz);
474
475 nres = 0;
476 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
477 res = dev->resource + PCI_IOV_RESOURCES + i;
478 i += __pci_read_base(dev, pci_bar_unknown, res,
479 pos + PCI_SRIOV_BAR + i * 4);
480 if (!res->flags)
481 continue;
482 if (resource_size(res) & (PAGE_SIZE - 1)) {
483 rc = -EIO;
484 goto failed;
485 }
486 res->end = res->start + resource_size(res) * total - 1;
487 nres++;
488 }
489
490 iov = kzalloc(sizeof(*iov), GFP_KERNEL);
491 if (!iov) {
492 rc = -ENOMEM;
493 goto failed;
494 }
495
496 iov->pos = pos;
497 iov->nres = nres;
498 iov->ctrl = ctrl;
499 iov->total_VFs = total;
500 iov->offset = offset;
501 iov->stride = stride;
502 iov->pgsz = pgsz;
503 iov->self = dev;
504 pci_read_config_dword(dev, pos + PCI_SRIOV_CAP, &iov->cap);
505 pci_read_config_byte(dev, pos + PCI_SRIOV_FUNC_LINK, &iov->link);
506 if (pci_pcie_type(dev) == PCI_EXP_TYPE_RC_END)
507 iov->link = PCI_DEVFN(PCI_SLOT(dev->devfn), iov->link);
508
509 if (pdev)
510 iov->dev = pci_dev_get(pdev);
511 else
512 iov->dev = dev;
513
514 mutex_init(&iov->lock);
515
516 dev->sriov = iov;
517 dev->is_physfn = 1;
518
519 return 0;
520
521failed:
522 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
523 res = dev->resource + PCI_IOV_RESOURCES + i;
524 res->flags = 0;
525 }
526
527 return rc;
528}
529
530static void sriov_release(struct pci_dev *dev)
531{
532 BUG_ON(dev->sriov->num_VFs);
533
534 if (dev != dev->sriov->dev)
535 pci_dev_put(dev->sriov->dev);
536
537 mutex_destroy(&dev->sriov->lock);
538
539 kfree(dev->sriov);
540 dev->sriov = NULL;
541}
542
543static void sriov_restore_state(struct pci_dev *dev)
544{
545 int i;
546 u16 ctrl;
547 struct pci_sriov *iov = dev->sriov;
548
549 pci_read_config_word(dev, iov->pos + PCI_SRIOV_CTRL, &ctrl);
550 if (ctrl & PCI_SRIOV_CTRL_VFE)
551 return;
552
553 for (i = PCI_IOV_RESOURCES; i <= PCI_IOV_RESOURCE_END; i++)
554 pci_update_resource(dev, i);
555
556 pci_write_config_dword(dev, iov->pos + PCI_SRIOV_SYS_PGSIZE, iov->pgsz);
557 pci_write_config_word(dev, iov->pos + PCI_SRIOV_NUM_VF, iov->num_VFs);
558 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
559 if (iov->ctrl & PCI_SRIOV_CTRL_VFE)
560 msleep(100);
561}
562
563
564
565
566
567
568
569int pci_iov_init(struct pci_dev *dev)
570{
571 int pos;
572
573 if (!pci_is_pcie(dev))
574 return -ENODEV;
575
576 pos = pci_find_ext_capability(dev, PCI_EXT_CAP_ID_SRIOV);
577 if (pos)
578 return sriov_init(dev, pos);
579
580 return -ENODEV;
581}
582
583
584
585
586
587void pci_iov_release(struct pci_dev *dev)
588{
589 if (dev->is_physfn)
590 sriov_release(dev);
591}
592
593
594
595
596
597
598
599
600
601int pci_iov_resource_bar(struct pci_dev *dev, int resno,
602 enum pci_bar_type *type)
603{
604 if (resno < PCI_IOV_RESOURCES || resno > PCI_IOV_RESOURCE_END)
605 return 0;
606
607 BUG_ON(!dev->is_physfn);
608
609 *type = pci_bar_unknown;
610
611 return dev->sriov->pos + PCI_SRIOV_BAR +
612 4 * (resno - PCI_IOV_RESOURCES);
613}
614
615
616
617
618
619
620
621
622
623
624
625resource_size_t pci_sriov_resource_alignment(struct pci_dev *dev, int resno)
626{
627 struct resource tmp;
628 enum pci_bar_type type;
629 int reg = pci_iov_resource_bar(dev, resno, &type);
630
631 if (!reg)
632 return 0;
633
634 __pci_read_base(dev, type, &tmp, reg);
635 return resource_alignment(&tmp);
636}
637
638
639
640
641
642void pci_restore_iov_state(struct pci_dev *dev)
643{
644 if (dev->is_physfn)
645 sriov_restore_state(dev);
646}
647
648
649
650
651
652
653
654
655int pci_iov_bus_range(struct pci_bus *bus)
656{
657 int max = 0;
658 u8 busnr;
659 struct pci_dev *dev;
660
661 list_for_each_entry(dev, &bus->devices, bus_list) {
662 if (!dev->is_physfn)
663 continue;
664 busnr = virtfn_bus(dev, dev->sriov->total_VFs - 1);
665 if (busnr > max)
666 max = busnr;
667 }
668
669 return max ? max - bus->number : 0;
670}
671
672
673
674
675
676
677
678
679int pci_enable_sriov(struct pci_dev *dev, int nr_virtfn)
680{
681 might_sleep();
682
683 if (!dev->is_physfn)
684 return -ENODEV;
685
686 return sriov_enable(dev, nr_virtfn);
687}
688EXPORT_SYMBOL_GPL(pci_enable_sriov);
689
690
691
692
693
694void pci_disable_sriov(struct pci_dev *dev)
695{
696 might_sleep();
697
698 if (!dev->is_physfn)
699 return;
700
701 sriov_disable(dev);
702}
703EXPORT_SYMBOL_GPL(pci_disable_sriov);
704
705
706
707
708
709
710
711
712
713
714
715irqreturn_t pci_sriov_migration(struct pci_dev *dev)
716{
717 if (!dev->is_physfn)
718 return IRQ_NONE;
719
720 return sriov_migration(dev) ? IRQ_HANDLED : IRQ_NONE;
721}
722EXPORT_SYMBOL_GPL(pci_sriov_migration);
723
724
725
726
727
728
729
730int pci_num_vf(struct pci_dev *dev)
731{
732 if (!dev->is_physfn)
733 return 0;
734
735 return dev->sriov->num_VFs;
736}
737EXPORT_SYMBOL_GPL(pci_num_vf);
738
739
740
741
742
743
744
745
746
747
748
749
750
751int pci_sriov_set_totalvfs(struct pci_dev *dev, u16 numvfs)
752{
753 if (!dev->is_physfn || (numvfs > dev->sriov->total_VFs))
754 return -EINVAL;
755
756
757 if (dev->sriov->ctrl & PCI_SRIOV_CTRL_VFE)
758 return -EBUSY;
759 else
760 dev->sriov->driver_max_VFs = numvfs;
761
762 return 0;
763}
764EXPORT_SYMBOL_GPL(pci_sriov_set_totalvfs);
765
766
767
768
769
770
771
772
773
774int pci_sriov_get_totalvfs(struct pci_dev *dev)
775{
776 if (!dev->is_physfn)
777 return -EINVAL;
778
779 if (dev->sriov->driver_max_VFs)
780 return dev->sriov->driver_max_VFs;
781
782 return dev->sriov->total_VFs;
783}
784EXPORT_SYMBOL_GPL(pci_sriov_get_totalvfs);
785