Commit | Line | Data |
---|---|---|
d1b054da YZ |
1 | /* |
2 | * drivers/pci/iov.c | |
3 | * | |
4 | * Copyright (C) 2009 Intel Corporation, Yu Zhao <yu.zhao@intel.com> | |
5 | * | |
6 | * PCI Express I/O Virtualization (IOV) support. | |
7 | * Single Root IOV 1.0 | |
8 | */ | |
9 | ||
10 | #include <linux/pci.h> | |
11 | #include <linux/mutex.h> | |
12 | #include <linux/string.h> | |
13 | #include <linux/delay.h> | |
14 | #include "pci.h" | |
15 | ||
dd7cc44d | 16 | #define VIRTFN_ID_LEN 16 |
d1b054da | 17 | |
a28724b0 YZ |
18 | static inline u8 virtfn_bus(struct pci_dev *dev, int id) |
19 | { | |
20 | return dev->bus->number + ((dev->devfn + dev->sriov->offset + | |
21 | dev->sriov->stride * id) >> 8); | |
22 | } | |
23 | ||
24 | static inline u8 virtfn_devfn(struct pci_dev *dev, int id) | |
25 | { | |
26 | return (dev->devfn + dev->sriov->offset + | |
27 | dev->sriov->stride * id) & 0xff; | |
28 | } | |
29 | ||
dd7cc44d YZ |
30 | static struct pci_bus *virtfn_add_bus(struct pci_bus *bus, int busnr) |
31 | { | |
32 | int rc; | |
33 | struct pci_bus *child; | |
34 | ||
35 | if (bus->number == busnr) | |
36 | return bus; | |
37 | ||
38 | child = pci_find_bus(pci_domain_nr(bus), busnr); | |
39 | if (child) | |
40 | return child; | |
41 | ||
42 | child = pci_add_new_bus(bus, NULL, busnr); | |
43 | if (!child) | |
44 | return NULL; | |
45 | ||
46 | child->subordinate = busnr; | |
47 | child->dev.parent = bus->bridge; | |
48 | rc = pci_bus_add_child(child); | |
49 | if (rc) { | |
50 | pci_remove_bus(child); | |
51 | return NULL; | |
52 | } | |
53 | ||
54 | return child; | |
55 | } | |
56 | ||
57 | static void virtfn_remove_bus(struct pci_bus *bus, int busnr) | |
58 | { | |
59 | struct pci_bus *child; | |
60 | ||
61 | if (bus->number == busnr) | |
62 | return; | |
63 | ||
64 | child = pci_find_bus(pci_domain_nr(bus), busnr); | |
65 | BUG_ON(!child); | |
66 | ||
67 | if (list_empty(&child->devices)) | |
68 | pci_remove_bus(child); | |
69 | } | |
70 | ||
71 | static int virtfn_add(struct pci_dev *dev, int id, int reset) | |
72 | { | |
73 | int i; | |
74 | int rc; | |
75 | u64 size; | |
76 | char buf[VIRTFN_ID_LEN]; | |
77 | struct pci_dev *virtfn; | |
78 | struct resource *res; | |
79 | struct pci_sriov *iov = dev->sriov; | |
80 | ||
81 | virtfn = alloc_pci_dev(); | |
82 | if (!virtfn) | |
83 | return -ENOMEM; | |
84 | ||
85 | mutex_lock(&iov->dev->sriov->lock); | |
86 | virtfn->bus = virtfn_add_bus(dev->bus, virtfn_bus(dev, id)); | |
87 | if (!virtfn->bus) { | |
88 | kfree(virtfn); | |
89 | mutex_unlock(&iov->dev->sriov->lock); | |
90 | return -ENOMEM; | |
91 | } | |
92 | virtfn->devfn = virtfn_devfn(dev, id); | |
93 | virtfn->vendor = dev->vendor; | |
94 | pci_read_config_word(dev, iov->pos + PCI_SRIOV_VF_DID, &virtfn->device); | |
95 | pci_setup_device(virtfn); | |
96 | virtfn->dev.parent = dev->dev.parent; | |
97 | ||
98 | for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) { | |
99 | res = dev->resource + PCI_IOV_RESOURCES + i; | |
100 | if (!res->parent) | |
101 | continue; | |
102 | virtfn->resource[i].name = pci_name(virtfn); | |
103 | virtfn->resource[i].flags = res->flags; | |
104 | size = resource_size(res); | |
105 | do_div(size, iov->total); | |
106 | virtfn->resource[i].start = res->start + size * id; | |
107 | virtfn->resource[i].end = virtfn->resource[i].start + size - 1; | |
108 | rc = request_resource(res, &virtfn->resource[i]); | |
109 | BUG_ON(rc); | |
110 | } | |
111 | ||
112 | if (reset) | |
113 | pci_execute_reset_function(virtfn); | |
114 | ||
115 | pci_device_add(virtfn, virtfn->bus); | |
116 | mutex_unlock(&iov->dev->sriov->lock); | |
117 | ||
118 | virtfn->physfn = pci_dev_get(dev); | |
119 | virtfn->is_virtfn = 1; | |
120 | ||
121 | rc = pci_bus_add_device(virtfn); | |
122 | if (rc) | |
123 | goto failed1; | |
124 | sprintf(buf, "virtfn%u", id); | |
125 | rc = sysfs_create_link(&dev->dev.kobj, &virtfn->dev.kobj, buf); | |
126 | if (rc) | |
127 | goto failed1; | |
128 | rc = sysfs_create_link(&virtfn->dev.kobj, &dev->dev.kobj, "physfn"); | |
129 | if (rc) | |
130 | goto failed2; | |
131 | ||
132 | kobject_uevent(&virtfn->dev.kobj, KOBJ_CHANGE); | |
133 | ||
134 | return 0; | |
135 | ||
136 | failed2: | |
137 | sysfs_remove_link(&dev->dev.kobj, buf); | |
138 | failed1: | |
139 | pci_dev_put(dev); | |
140 | mutex_lock(&iov->dev->sriov->lock); | |
141 | pci_remove_bus_device(virtfn); | |
142 | virtfn_remove_bus(dev->bus, virtfn_bus(dev, id)); | |
143 | mutex_unlock(&iov->dev->sriov->lock); | |
144 | ||
145 | return rc; | |
146 | } | |
147 | ||
148 | static void virtfn_remove(struct pci_dev *dev, int id, int reset) | |
149 | { | |
150 | char buf[VIRTFN_ID_LEN]; | |
151 | struct pci_bus *bus; | |
152 | struct pci_dev *virtfn; | |
153 | struct pci_sriov *iov = dev->sriov; | |
154 | ||
155 | bus = pci_find_bus(pci_domain_nr(dev->bus), virtfn_bus(dev, id)); | |
156 | if (!bus) | |
157 | return; | |
158 | ||
159 | virtfn = pci_get_slot(bus, virtfn_devfn(dev, id)); | |
160 | if (!virtfn) | |
161 | return; | |
162 | ||
163 | pci_dev_put(virtfn); | |
164 | ||
165 | if (reset) { | |
166 | device_release_driver(&virtfn->dev); | |
167 | pci_execute_reset_function(virtfn); | |
168 | } | |
169 | ||
170 | sprintf(buf, "virtfn%u", id); | |
171 | sysfs_remove_link(&dev->dev.kobj, buf); | |
172 | sysfs_remove_link(&virtfn->dev.kobj, "physfn"); | |
173 | ||
174 | mutex_lock(&iov->dev->sriov->lock); | |
175 | pci_remove_bus_device(virtfn); | |
176 | virtfn_remove_bus(dev->bus, virtfn_bus(dev, id)); | |
177 | mutex_unlock(&iov->dev->sriov->lock); | |
178 | ||
179 | pci_dev_put(dev); | |
180 | } | |
181 | ||
182 | static int sriov_enable(struct pci_dev *dev, int nr_virtfn) | |
183 | { | |
184 | int rc; | |
185 | int i, j; | |
186 | int nres; | |
187 | u16 offset, stride, initial; | |
188 | struct resource *res; | |
189 | struct pci_dev *pdev; | |
190 | struct pci_sriov *iov = dev->sriov; | |
191 | ||
192 | if (!nr_virtfn) | |
193 | return 0; | |
194 | ||
195 | if (iov->nr_virtfn) | |
196 | return -EINVAL; | |
197 | ||
198 | pci_read_config_word(dev, iov->pos + PCI_SRIOV_INITIAL_VF, &initial); | |
199 | if (initial > iov->total || | |
200 | (!(iov->cap & PCI_SRIOV_CAP_VFM) && (initial != iov->total))) | |
201 | return -EIO; | |
202 | ||
203 | if (nr_virtfn < 0 || nr_virtfn > iov->total || | |
204 | (!(iov->cap & PCI_SRIOV_CAP_VFM) && (nr_virtfn > initial))) | |
205 | return -EINVAL; | |
206 | ||
207 | pci_write_config_word(dev, iov->pos + PCI_SRIOV_NUM_VF, nr_virtfn); | |
208 | pci_read_config_word(dev, iov->pos + PCI_SRIOV_VF_OFFSET, &offset); | |
209 | pci_read_config_word(dev, iov->pos + PCI_SRIOV_VF_STRIDE, &stride); | |
210 | if (!offset || (nr_virtfn > 1 && !stride)) | |
211 | return -EIO; | |
212 | ||
213 | nres = 0; | |
214 | for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) { | |
215 | res = dev->resource + PCI_IOV_RESOURCES + i; | |
216 | if (res->parent) | |
217 | nres++; | |
218 | } | |
219 | if (nres != iov->nres) { | |
220 | dev_err(&dev->dev, "not enough MMIO resources for SR-IOV\n"); | |
221 | return -ENOMEM; | |
222 | } | |
223 | ||
224 | iov->offset = offset; | |
225 | iov->stride = stride; | |
226 | ||
227 | if (virtfn_bus(dev, nr_virtfn - 1) > dev->bus->subordinate) { | |
228 | dev_err(&dev->dev, "SR-IOV: bus number out of range\n"); | |
229 | return -ENOMEM; | |
230 | } | |
231 | ||
232 | if (iov->link != dev->devfn) { | |
233 | pdev = pci_get_slot(dev->bus, iov->link); | |
234 | if (!pdev) | |
235 | return -ENODEV; | |
236 | ||
237 | pci_dev_put(pdev); | |
238 | ||
239 | if (!pdev->is_physfn) | |
240 | return -ENODEV; | |
241 | ||
242 | rc = sysfs_create_link(&dev->dev.kobj, | |
243 | &pdev->dev.kobj, "dep_link"); | |
244 | if (rc) | |
245 | return rc; | |
246 | } | |
247 | ||
248 | iov->ctrl |= PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE; | |
249 | pci_block_user_cfg_access(dev); | |
250 | pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl); | |
251 | msleep(100); | |
252 | pci_unblock_user_cfg_access(dev); | |
253 | ||
254 | iov->initial = initial; | |
255 | if (nr_virtfn < initial) | |
256 | initial = nr_virtfn; | |
257 | ||
258 | for (i = 0; i < initial; i++) { | |
259 | rc = virtfn_add(dev, i, 0); | |
260 | if (rc) | |
261 | goto failed; | |
262 | } | |
263 | ||
264 | kobject_uevent(&dev->dev.kobj, KOBJ_CHANGE); | |
265 | iov->nr_virtfn = nr_virtfn; | |
266 | ||
267 | return 0; | |
268 | ||
269 | failed: | |
270 | for (j = 0; j < i; j++) | |
271 | virtfn_remove(dev, j, 0); | |
272 | ||
273 | iov->ctrl &= ~(PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE); | |
274 | pci_block_user_cfg_access(dev); | |
275 | pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl); | |
276 | ssleep(1); | |
277 | pci_unblock_user_cfg_access(dev); | |
278 | ||
279 | if (iov->link != dev->devfn) | |
280 | sysfs_remove_link(&dev->dev.kobj, "dep_link"); | |
281 | ||
282 | return rc; | |
283 | } | |
284 | ||
285 | static void sriov_disable(struct pci_dev *dev) | |
286 | { | |
287 | int i; | |
288 | struct pci_sriov *iov = dev->sriov; | |
289 | ||
290 | if (!iov->nr_virtfn) | |
291 | return; | |
292 | ||
293 | for (i = 0; i < iov->nr_virtfn; i++) | |
294 | virtfn_remove(dev, i, 0); | |
295 | ||
296 | iov->ctrl &= ~(PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE); | |
297 | pci_block_user_cfg_access(dev); | |
298 | pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl); | |
299 | ssleep(1); | |
300 | pci_unblock_user_cfg_access(dev); | |
301 | ||
302 | if (iov->link != dev->devfn) | |
303 | sysfs_remove_link(&dev->dev.kobj, "dep_link"); | |
304 | ||
305 | iov->nr_virtfn = 0; | |
306 | } | |
307 | ||
d1b054da YZ |
308 | static int sriov_init(struct pci_dev *dev, int pos) |
309 | { | |
310 | int i; | |
311 | int rc; | |
312 | int nres; | |
313 | u32 pgsz; | |
314 | u16 ctrl, total, offset, stride; | |
315 | struct pci_sriov *iov; | |
316 | struct resource *res; | |
317 | struct pci_dev *pdev; | |
318 | ||
319 | if (dev->pcie_type != PCI_EXP_TYPE_RC_END && | |
320 | dev->pcie_type != PCI_EXP_TYPE_ENDPOINT) | |
321 | return -ENODEV; | |
322 | ||
323 | pci_read_config_word(dev, pos + PCI_SRIOV_CTRL, &ctrl); | |
324 | if (ctrl & PCI_SRIOV_CTRL_VFE) { | |
325 | pci_write_config_word(dev, pos + PCI_SRIOV_CTRL, 0); | |
326 | ssleep(1); | |
327 | } | |
328 | ||
329 | pci_read_config_word(dev, pos + PCI_SRIOV_TOTAL_VF, &total); | |
330 | if (!total) | |
331 | return 0; | |
332 | ||
333 | ctrl = 0; | |
334 | list_for_each_entry(pdev, &dev->bus->devices, bus_list) | |
335 | if (pdev->is_physfn) | |
336 | goto found; | |
337 | ||
338 | pdev = NULL; | |
339 | if (pci_ari_enabled(dev->bus)) | |
340 | ctrl |= PCI_SRIOV_CTRL_ARI; | |
341 | ||
342 | found: | |
343 | pci_write_config_word(dev, pos + PCI_SRIOV_CTRL, ctrl); | |
344 | pci_write_config_word(dev, pos + PCI_SRIOV_NUM_VF, total); | |
345 | pci_read_config_word(dev, pos + PCI_SRIOV_VF_OFFSET, &offset); | |
346 | pci_read_config_word(dev, pos + PCI_SRIOV_VF_STRIDE, &stride); | |
347 | if (!offset || (total > 1 && !stride)) | |
348 | return -EIO; | |
349 | ||
350 | pci_read_config_dword(dev, pos + PCI_SRIOV_SUP_PGSIZE, &pgsz); | |
351 | i = PAGE_SHIFT > 12 ? PAGE_SHIFT - 12 : 0; | |
352 | pgsz &= ~((1 << i) - 1); | |
353 | if (!pgsz) | |
354 | return -EIO; | |
355 | ||
356 | pgsz &= ~(pgsz - 1); | |
357 | pci_write_config_dword(dev, pos + PCI_SRIOV_SYS_PGSIZE, pgsz); | |
358 | ||
359 | nres = 0; | |
360 | for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) { | |
361 | res = dev->resource + PCI_IOV_RESOURCES + i; | |
362 | i += __pci_read_base(dev, pci_bar_unknown, res, | |
363 | pos + PCI_SRIOV_BAR + i * 4); | |
364 | if (!res->flags) | |
365 | continue; | |
366 | if (resource_size(res) & (PAGE_SIZE - 1)) { | |
367 | rc = -EIO; | |
368 | goto failed; | |
369 | } | |
370 | res->end = res->start + resource_size(res) * total - 1; | |
371 | nres++; | |
372 | } | |
373 | ||
374 | iov = kzalloc(sizeof(*iov), GFP_KERNEL); | |
375 | if (!iov) { | |
376 | rc = -ENOMEM; | |
377 | goto failed; | |
378 | } | |
379 | ||
380 | iov->pos = pos; | |
381 | iov->nres = nres; | |
382 | iov->ctrl = ctrl; | |
383 | iov->total = total; | |
384 | iov->offset = offset; | |
385 | iov->stride = stride; | |
386 | iov->pgsz = pgsz; | |
387 | iov->self = dev; | |
388 | pci_read_config_dword(dev, pos + PCI_SRIOV_CAP, &iov->cap); | |
389 | pci_read_config_byte(dev, pos + PCI_SRIOV_FUNC_LINK, &iov->link); | |
390 | ||
391 | if (pdev) | |
392 | iov->dev = pci_dev_get(pdev); | |
393 | else { | |
394 | iov->dev = dev; | |
395 | mutex_init(&iov->lock); | |
396 | } | |
397 | ||
398 | dev->sriov = iov; | |
399 | dev->is_physfn = 1; | |
400 | ||
401 | return 0; | |
402 | ||
403 | failed: | |
404 | for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) { | |
405 | res = dev->resource + PCI_IOV_RESOURCES + i; | |
406 | res->flags = 0; | |
407 | } | |
408 | ||
409 | return rc; | |
410 | } | |
411 | ||
412 | static void sriov_release(struct pci_dev *dev) | |
413 | { | |
dd7cc44d YZ |
414 | BUG_ON(dev->sriov->nr_virtfn); |
415 | ||
d1b054da YZ |
416 | if (dev == dev->sriov->dev) |
417 | mutex_destroy(&dev->sriov->lock); | |
418 | else | |
419 | pci_dev_put(dev->sriov->dev); | |
420 | ||
421 | kfree(dev->sriov); | |
422 | dev->sriov = NULL; | |
423 | } | |
424 | ||
8c5cdb6a YZ |
425 | static void sriov_restore_state(struct pci_dev *dev) |
426 | { | |
427 | int i; | |
428 | u16 ctrl; | |
429 | struct pci_sriov *iov = dev->sriov; | |
430 | ||
431 | pci_read_config_word(dev, iov->pos + PCI_SRIOV_CTRL, &ctrl); | |
432 | if (ctrl & PCI_SRIOV_CTRL_VFE) | |
433 | return; | |
434 | ||
435 | for (i = PCI_IOV_RESOURCES; i <= PCI_IOV_RESOURCE_END; i++) | |
436 | pci_update_resource(dev, i); | |
437 | ||
438 | pci_write_config_dword(dev, iov->pos + PCI_SRIOV_SYS_PGSIZE, iov->pgsz); | |
dd7cc44d | 439 | pci_write_config_word(dev, iov->pos + PCI_SRIOV_NUM_VF, iov->nr_virtfn); |
8c5cdb6a YZ |
440 | pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl); |
441 | if (iov->ctrl & PCI_SRIOV_CTRL_VFE) | |
442 | msleep(100); | |
443 | } | |
444 | ||
d1b054da YZ |
445 | /** |
446 | * pci_iov_init - initialize the IOV capability | |
447 | * @dev: the PCI device | |
448 | * | |
449 | * Returns 0 on success, or negative on failure. | |
450 | */ | |
451 | int pci_iov_init(struct pci_dev *dev) | |
452 | { | |
453 | int pos; | |
454 | ||
455 | if (!dev->is_pcie) | |
456 | return -ENODEV; | |
457 | ||
458 | pos = pci_find_ext_capability(dev, PCI_EXT_CAP_ID_SRIOV); | |
459 | if (pos) | |
460 | return sriov_init(dev, pos); | |
461 | ||
462 | return -ENODEV; | |
463 | } | |
464 | ||
465 | /** | |
466 | * pci_iov_release - release resources used by the IOV capability | |
467 | * @dev: the PCI device | |
468 | */ | |
469 | void pci_iov_release(struct pci_dev *dev) | |
470 | { | |
471 | if (dev->is_physfn) | |
472 | sriov_release(dev); | |
473 | } | |
474 | ||
475 | /** | |
476 | * pci_iov_resource_bar - get position of the SR-IOV BAR | |
477 | * @dev: the PCI device | |
478 | * @resno: the resource number | |
479 | * @type: the BAR type to be filled in | |
480 | * | |
481 | * Returns position of the BAR encapsulated in the SR-IOV capability. | |
482 | */ | |
483 | int pci_iov_resource_bar(struct pci_dev *dev, int resno, | |
484 | enum pci_bar_type *type) | |
485 | { | |
486 | if (resno < PCI_IOV_RESOURCES || resno > PCI_IOV_RESOURCE_END) | |
487 | return 0; | |
488 | ||
489 | BUG_ON(!dev->is_physfn); | |
490 | ||
491 | *type = pci_bar_unknown; | |
492 | ||
493 | return dev->sriov->pos + PCI_SRIOV_BAR + | |
494 | 4 * (resno - PCI_IOV_RESOURCES); | |
495 | } | |
8c5cdb6a YZ |
496 | |
497 | /** | |
498 | * pci_restore_iov_state - restore the state of the IOV capability | |
499 | * @dev: the PCI device | |
500 | */ | |
501 | void pci_restore_iov_state(struct pci_dev *dev) | |
502 | { | |
503 | if (dev->is_physfn) | |
504 | sriov_restore_state(dev); | |
505 | } | |
a28724b0 YZ |
506 | |
507 | /** | |
508 | * pci_iov_bus_range - find bus range used by Virtual Function | |
509 | * @bus: the PCI bus | |
510 | * | |
511 | * Returns max number of buses (exclude current one) used by Virtual | |
512 | * Functions. | |
513 | */ | |
514 | int pci_iov_bus_range(struct pci_bus *bus) | |
515 | { | |
516 | int max = 0; | |
517 | u8 busnr; | |
518 | struct pci_dev *dev; | |
519 | ||
520 | list_for_each_entry(dev, &bus->devices, bus_list) { | |
521 | if (!dev->is_physfn) | |
522 | continue; | |
523 | busnr = virtfn_bus(dev, dev->sriov->total - 1); | |
524 | if (busnr > max) | |
525 | max = busnr; | |
526 | } | |
527 | ||
528 | return max ? max - bus->number : 0; | |
529 | } | |
dd7cc44d YZ |
530 | |
531 | /** | |
532 | * pci_enable_sriov - enable the SR-IOV capability | |
533 | * @dev: the PCI device | |
534 | * | |
535 | * Returns 0 on success, or negative on failure. | |
536 | */ | |
537 | int pci_enable_sriov(struct pci_dev *dev, int nr_virtfn) | |
538 | { | |
539 | might_sleep(); | |
540 | ||
541 | if (!dev->is_physfn) | |
542 | return -ENODEV; | |
543 | ||
544 | return sriov_enable(dev, nr_virtfn); | |
545 | } | |
546 | EXPORT_SYMBOL_GPL(pci_enable_sriov); | |
547 | ||
548 | /** | |
549 | * pci_disable_sriov - disable the SR-IOV capability | |
550 | * @dev: the PCI device | |
551 | */ | |
552 | void pci_disable_sriov(struct pci_dev *dev) | |
553 | { | |
554 | might_sleep(); | |
555 | ||
556 | if (!dev->is_physfn) | |
557 | return; | |
558 | ||
559 | sriov_disable(dev); | |
560 | } | |
561 | EXPORT_SYMBOL_GPL(pci_disable_sriov); |