1 /*-
2 * Copyright (c) 2013-2015 Sandvine Inc.
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 *
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24 * SUCH DAMAGE.
25 */
26
27 #include <sys/cdefs.h>
28 #include "opt_bus.h"
29
30 #include <sys/param.h>
31 #include <sys/conf.h>
32 #include <sys/kernel.h>
33 #include <sys/systm.h>
34 #include <sys/bus.h>
35 #include <sys/fcntl.h>
36 #include <sys/ioccom.h>
37 #include <sys/iov.h>
38 #include <sys/linker.h>
39 #include <sys/lock.h>
40 #include <sys/malloc.h>
41 #include <sys/module.h>
42 #include <sys/mutex.h>
43 #include <sys/pciio.h>
44 #include <sys/queue.h>
45 #include <sys/rman.h>
46 #include <sys/sysctl.h>
47
48 #include <machine/bus.h>
49 #include <machine/stdarg.h>
50
51 #include <sys/nv.h>
52 #include <sys/iov_schema.h>
53
54 #include <dev/pci/pcireg.h>
55 #include <dev/pci/pcivar.h>
56 #include <dev/pci/pci_iov.h>
57 #include <dev/pci/pci_private.h>
58 #include <dev/pci/pci_iov_private.h>
59 #include <dev/pci/schema_private.h>
60
61 #include "pcib_if.h"
62
63 static MALLOC_DEFINE(M_SRIOV, "sr_iov", "PCI SR-IOV allocations");
64
65 static d_ioctl_t pci_iov_ioctl;
66
67 static struct cdevsw iov_cdevsw = {
68 .d_version = D_VERSION,
69 .d_name = "iov",
70 .d_ioctl = pci_iov_ioctl
71 };
72
73 SYSCTL_DECL(_hw_pci);
74
75 /*
76 * The maximum amount of memory we will allocate for user configuration of an
77 * SR-IOV device. 1MB ought to be enough for anyone, but leave this
78 * configurable just in case.
79 */
80 static u_long pci_iov_max_config = 1024 * 1024;
81 SYSCTL_ULONG(_hw_pci, OID_AUTO, iov_max_config, CTLFLAG_RWTUN,
82 &pci_iov_max_config, 0, "Maximum allowed size of SR-IOV configuration.");
83
84 #define IOV_READ(d, r, w) \
85 pci_read_config((d)->cfg.dev, (d)->cfg.iov->iov_pos + r, w)
86
87 #define IOV_WRITE(d, r, v, w) \
88 pci_write_config((d)->cfg.dev, (d)->cfg.iov->iov_pos + r, v, w)
89
90 static nvlist_t *pci_iov_build_schema(nvlist_t **pf_schema,
91 nvlist_t **vf_schema);
92 static void pci_iov_build_pf_schema(nvlist_t *schema,
93 nvlist_t **driver_schema);
94 static void pci_iov_build_vf_schema(nvlist_t *schema,
95 nvlist_t **driver_schema);
96 static int pci_iov_delete_iov_children(struct pci_devinfo *dinfo);
97 static nvlist_t *pci_iov_get_pf_subsystem_schema(void);
98 static nvlist_t *pci_iov_get_vf_subsystem_schema(void);
99
100 int
pci_iov_attach_name(device_t dev,struct nvlist * pf_schema,struct nvlist * vf_schema,const char * fmt,...)101 pci_iov_attach_name(device_t dev, struct nvlist *pf_schema,
102 struct nvlist *vf_schema, const char *fmt, ...)
103 {
104 char buf[NAME_MAX + 1];
105 va_list ap;
106
107 va_start(ap, fmt);
108 vsnprintf(buf, sizeof(buf), fmt, ap);
109 va_end(ap);
110 return (PCI_IOV_ATTACH(device_get_parent(dev), dev, pf_schema,
111 vf_schema, buf));
112 }
113
114 int
pci_iov_attach_method(device_t bus,device_t dev,nvlist_t * pf_schema,nvlist_t * vf_schema,const char * name)115 pci_iov_attach_method(device_t bus, device_t dev, nvlist_t *pf_schema,
116 nvlist_t *vf_schema, const char *name)
117 {
118 struct pci_devinfo *dinfo;
119 struct pcicfg_iov *iov;
120 nvlist_t *schema;
121 uint32_t version;
122 int error;
123 int iov_pos;
124
125 dinfo = device_get_ivars(dev);
126 schema = NULL;
127
128 error = pci_find_extcap(dev, PCIZ_SRIOV, &iov_pos);
129
130 if (error != 0)
131 return (error);
132
133 version = pci_read_config(dev, iov_pos, 4);
134 if (PCI_EXTCAP_VER(version) != 1) {
135 if (bootverbose)
136 device_printf(dev,
137 "Unsupported version of SR-IOV (%d) detected\n",
138 PCI_EXTCAP_VER(version));
139
140 return (ENXIO);
141 }
142
143 iov = malloc(sizeof(*dinfo->cfg.iov), M_SRIOV, M_WAITOK | M_ZERO);
144
145 mtx_lock(&Giant);
146 if (dinfo->cfg.iov != NULL) {
147 error = EBUSY;
148 goto cleanup;
149 }
150 iov->iov_pf = dev;
151 iov->iov_pos = iov_pos;
152
153 schema = pci_iov_build_schema(&pf_schema, &vf_schema);
154 if (schema == NULL) {
155 error = ENOMEM;
156 goto cleanup;
157 }
158
159 error = pci_iov_validate_schema(schema);
160 if (error != 0)
161 goto cleanup;
162 iov->iov_schema = schema;
163
164 iov->iov_cdev = make_dev(&iov_cdevsw, device_get_unit(dev),
165 UID_ROOT, GID_WHEEL, 0600, "iov/%s", name);
166
167 if (iov->iov_cdev == NULL) {
168 error = ENOMEM;
169 goto cleanup;
170 }
171
172 dinfo->cfg.iov = iov;
173 iov->iov_cdev->si_drv1 = dinfo;
174 mtx_unlock(&Giant);
175
176 return (0);
177
178 cleanup:
179 nvlist_destroy(schema);
180 nvlist_destroy(pf_schema);
181 nvlist_destroy(vf_schema);
182 free(iov, M_SRIOV);
183 mtx_unlock(&Giant);
184 return (error);
185 }
186
187 int
pci_iov_detach_method(device_t bus,device_t dev)188 pci_iov_detach_method(device_t bus, device_t dev)
189 {
190 struct pci_devinfo *dinfo;
191 struct pcicfg_iov *iov;
192 int error;
193
194 mtx_lock(&Giant);
195 dinfo = device_get_ivars(dev);
196 iov = dinfo->cfg.iov;
197
198 if (iov == NULL) {
199 mtx_unlock(&Giant);
200 return (0);
201 }
202
203 if ((iov->iov_flags & IOV_BUSY) != 0) {
204 mtx_unlock(&Giant);
205 return (EBUSY);
206 }
207
208 error = pci_iov_delete_iov_children(dinfo);
209 if (error != 0) {
210 mtx_unlock(&Giant);
211 return (error);
212 }
213
214 dinfo->cfg.iov = NULL;
215
216 if (iov->iov_cdev) {
217 destroy_dev(iov->iov_cdev);
218 iov->iov_cdev = NULL;
219 }
220 nvlist_destroy(iov->iov_schema);
221
222 free(iov, M_SRIOV);
223 mtx_unlock(&Giant);
224
225 return (0);
226 }
227
228 static nvlist_t *
pci_iov_build_schema(nvlist_t ** pf,nvlist_t ** vf)229 pci_iov_build_schema(nvlist_t **pf, nvlist_t **vf)
230 {
231 nvlist_t *schema, *pf_driver, *vf_driver;
232
233 /* We always take ownership of the schemas. */
234 pf_driver = *pf;
235 *pf = NULL;
236 vf_driver = *vf;
237 *vf = NULL;
238
239 schema = pci_iov_schema_alloc_node();
240 if (schema == NULL)
241 goto cleanup;
242
243 pci_iov_build_pf_schema(schema, &pf_driver);
244 pci_iov_build_vf_schema(schema, &vf_driver);
245
246 if (nvlist_error(schema) != 0)
247 goto cleanup;
248
249 return (schema);
250
251 cleanup:
252 nvlist_destroy(schema);
253 nvlist_destroy(pf_driver);
254 nvlist_destroy(vf_driver);
255 return (NULL);
256 }
257
258 static void
pci_iov_build_pf_schema(nvlist_t * schema,nvlist_t ** driver_schema)259 pci_iov_build_pf_schema(nvlist_t *schema, nvlist_t **driver_schema)
260 {
261 nvlist_t *pf_schema, *iov_schema;
262
263 pf_schema = pci_iov_schema_alloc_node();
264 if (pf_schema == NULL) {
265 nvlist_set_error(schema, ENOMEM);
266 return;
267 }
268
269 iov_schema = pci_iov_get_pf_subsystem_schema();
270
271 /*
272 * Note that if either *driver_schema or iov_schema is NULL, then
273 * nvlist_move_nvlist will put the schema in the error state and
274 * SR-IOV will fail to initialize later, so we don't have to explicitly
275 * handle that case.
276 */
277 nvlist_move_nvlist(pf_schema, DRIVER_CONFIG_NAME, *driver_schema);
278 nvlist_move_nvlist(pf_schema, IOV_CONFIG_NAME, iov_schema);
279 nvlist_move_nvlist(schema, PF_CONFIG_NAME, pf_schema);
280 *driver_schema = NULL;
281 }
282
283 static void
pci_iov_build_vf_schema(nvlist_t * schema,nvlist_t ** driver_schema)284 pci_iov_build_vf_schema(nvlist_t *schema, nvlist_t **driver_schema)
285 {
286 nvlist_t *vf_schema, *iov_schema;
287
288 vf_schema = pci_iov_schema_alloc_node();
289 if (vf_schema == NULL) {
290 nvlist_set_error(schema, ENOMEM);
291 return;
292 }
293
294 iov_schema = pci_iov_get_vf_subsystem_schema();
295
296 /*
297 * Note that if either *driver_schema or iov_schema is NULL, then
298 * nvlist_move_nvlist will put the schema in the error state and
299 * SR-IOV will fail to initialize later, so we don't have to explicitly
300 * handle that case.
301 */
302 nvlist_move_nvlist(vf_schema, DRIVER_CONFIG_NAME, *driver_schema);
303 nvlist_move_nvlist(vf_schema, IOV_CONFIG_NAME, iov_schema);
304 nvlist_move_nvlist(schema, VF_SCHEMA_NAME, vf_schema);
305 *driver_schema = NULL;
306 }
307
308 static nvlist_t *
pci_iov_get_pf_subsystem_schema(void)309 pci_iov_get_pf_subsystem_schema(void)
310 {
311 nvlist_t *pf;
312
313 pf = pci_iov_schema_alloc_node();
314 if (pf == NULL)
315 return (NULL);
316
317 pci_iov_schema_add_uint16(pf, "num_vfs", IOV_SCHEMA_REQUIRED, -1);
318 pci_iov_schema_add_string(pf, "device", IOV_SCHEMA_REQUIRED, NULL);
319
320 return (pf);
321 }
322
323 static nvlist_t *
pci_iov_get_vf_subsystem_schema(void)324 pci_iov_get_vf_subsystem_schema(void)
325 {
326 nvlist_t *vf;
327
328 vf = pci_iov_schema_alloc_node();
329 if (vf == NULL)
330 return (NULL);
331
332 pci_iov_schema_add_bool(vf, "passthrough", IOV_SCHEMA_HASDEFAULT, 0);
333
334 return (vf);
335 }
336
337 static int
pci_iov_alloc_bar(struct pci_devinfo * dinfo,int bar,pci_addr_t bar_shift)338 pci_iov_alloc_bar(struct pci_devinfo *dinfo, int bar, pci_addr_t bar_shift)
339 {
340 struct resource *res;
341 struct pcicfg_iov *iov;
342 device_t dev, bus;
343 rman_res_t start, end;
344 pci_addr_t bar_size;
345 int rid;
346
347 iov = dinfo->cfg.iov;
348 dev = dinfo->cfg.dev;
349 bus = device_get_parent(dev);
350 rid = iov->iov_pos + PCIR_SRIOV_BAR(bar);
351 bar_size = 1 << bar_shift;
352
353 res = pci_alloc_multi_resource(bus, dev, SYS_RES_MEMORY, &rid, 0,
354 ~0, 1, iov->iov_num_vfs, RF_ACTIVE);
355
356 if (res == NULL)
357 return (ENXIO);
358
359 iov->iov_bar[bar].res = res;
360 iov->iov_bar[bar].bar_size = bar_size;
361 iov->iov_bar[bar].bar_shift = bar_shift;
362
363 start = rman_get_start(res);
364 end = rman_get_end(res);
365 return (rman_manage_region(&iov->rman, start, end));
366 }
367
368 static void
pci_iov_add_bars(struct pcicfg_iov * iov,struct pci_devinfo * dinfo)369 pci_iov_add_bars(struct pcicfg_iov *iov, struct pci_devinfo *dinfo)
370 {
371 struct pci_iov_bar *bar;
372 uint64_t bar_start;
373 int i;
374
375 for (i = 0; i <= PCIR_MAX_BAR_0; i++) {
376 bar = &iov->iov_bar[i];
377 if (bar->res != NULL) {
378 bar_start = rman_get_start(bar->res) +
379 dinfo->cfg.vf.index * bar->bar_size;
380
381 pci_add_bar(dinfo->cfg.dev, PCIR_BAR(i), bar_start,
382 bar->bar_shift);
383 }
384 }
385 }
386
387 static int
pci_iov_parse_config(struct pcicfg_iov * iov,struct pci_iov_arg * arg,nvlist_t ** ret)388 pci_iov_parse_config(struct pcicfg_iov *iov, struct pci_iov_arg *arg,
389 nvlist_t **ret)
390 {
391 void *packed_config;
392 nvlist_t *config;
393 int error;
394
395 config = NULL;
396 packed_config = NULL;
397
398 if (arg->len > pci_iov_max_config) {
399 error = EMSGSIZE;
400 goto out;
401 }
402
403 packed_config = malloc(arg->len, M_SRIOV, M_WAITOK);
404
405 error = copyin(arg->config, packed_config, arg->len);
406 if (error != 0)
407 goto out;
408
409 config = nvlist_unpack(packed_config, arg->len, NV_FLAG_IGNORE_CASE);
410 if (config == NULL) {
411 error = EINVAL;
412 goto out;
413 }
414
415 error = pci_iov_schema_validate_config(iov->iov_schema, config);
416 if (error != 0)
417 goto out;
418
419 error = nvlist_error(config);
420 if (error != 0)
421 goto out;
422
423 *ret = config;
424 config = NULL;
425
426 out:
427 nvlist_destroy(config);
428 free(packed_config, M_SRIOV);
429 return (error);
430 }
431
432 /*
433 * Set the ARI_EN bit in the lowest-numbered PCI function with the SR-IOV
434 * capability. This bit is only writeable on the lowest-numbered PF but
435 * affects all PFs on the device.
436 */
437 static int
pci_iov_set_ari(device_t bus,bool * ari_enabled)438 pci_iov_set_ari(device_t bus, bool *ari_enabled)
439 {
440 device_t lowest;
441 device_t *devlist;
442 int i, error, devcount, lowest_func, lowest_pos, iov_pos, dev_func;
443 uint16_t iov_ctl;
444
445 /* If ARI is disabled on the downstream port there is nothing to do. */
446 if (!PCIB_ARI_ENABLED(device_get_parent(bus))) {
447 *ari_enabled = false;
448 return (0);
449 }
450
451 error = device_get_children(bus, &devlist, &devcount);
452
453 if (error != 0)
454 return (error);
455
456 lowest = NULL;
457 for (i = 0; i < devcount; i++) {
458 if (pci_find_extcap(devlist[i], PCIZ_SRIOV, &iov_pos) == 0) {
459 dev_func = pci_get_function(devlist[i]);
460 if (lowest == NULL || dev_func < lowest_func) {
461 lowest = devlist[i];
462 lowest_func = dev_func;
463 lowest_pos = iov_pos;
464 }
465 }
466 }
467 free(devlist, M_TEMP);
468
469 /*
470 * If we called this function some device must have the SR-IOV
471 * capability.
472 */
473 KASSERT(lowest != NULL,
474 ("Could not find child of %s with SR-IOV capability",
475 device_get_nameunit(bus)));
476
477 iov_ctl = pci_read_config(lowest, lowest_pos + PCIR_SRIOV_CTL, 2);
478 iov_ctl |= PCIM_SRIOV_ARI_EN;
479 pci_write_config(lowest, lowest_pos + PCIR_SRIOV_CTL, iov_ctl, 2);
480 if ((pci_read_config(lowest, lowest_pos + PCIR_SRIOV_CTL, 2) &
481 PCIM_SRIOV_ARI_EN) == 0) {
482 device_printf(lowest, "failed to enable ARI\n");
483 return (ENXIO);
484 }
485 *ari_enabled = true;
486 return (0);
487 }
488
489 static int
pci_iov_config_page_size(struct pci_devinfo * dinfo)490 pci_iov_config_page_size(struct pci_devinfo *dinfo)
491 {
492 uint32_t page_cap, page_size;
493
494 page_cap = IOV_READ(dinfo, PCIR_SRIOV_PAGE_CAP, 4);
495
496 /*
497 * If the system page size is less than the smallest SR-IOV page size
498 * then round up to the smallest SR-IOV page size.
499 */
500 if (PAGE_SHIFT < PCI_SRIOV_BASE_PAGE_SHIFT)
501 page_size = (1 << 0);
502 else
503 page_size = (1 << (PAGE_SHIFT - PCI_SRIOV_BASE_PAGE_SHIFT));
504
505 /* Check that the device supports the system page size. */
506 if (!(page_size & page_cap))
507 return (ENXIO);
508
509 IOV_WRITE(dinfo, PCIR_SRIOV_PAGE_SIZE, page_size, 4);
510 return (0);
511 }
512
513 static int
pci_iov_init(device_t dev,uint16_t num_vfs,const nvlist_t * config)514 pci_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *config)
515 {
516 const nvlist_t *device, *driver_config;
517
518 device = nvlist_get_nvlist(config, PF_CONFIG_NAME);
519 driver_config = nvlist_get_nvlist(device, DRIVER_CONFIG_NAME);
520 return (PCI_IOV_INIT(dev, num_vfs, driver_config));
521 }
522
523 static int
pci_iov_init_rman(device_t pf,struct pcicfg_iov * iov)524 pci_iov_init_rman(device_t pf, struct pcicfg_iov *iov)
525 {
526 int error;
527
528 iov->rman.rm_start = 0;
529 iov->rman.rm_end = ~0;
530 iov->rman.rm_type = RMAN_ARRAY;
531 snprintf(iov->rman_name, sizeof(iov->rman_name), "%s VF I/O memory",
532 device_get_nameunit(pf));
533 iov->rman.rm_descr = iov->rman_name;
534
535 error = rman_init(&iov->rman);
536 if (error != 0)
537 return (error);
538
539 iov->iov_flags |= IOV_RMAN_INITED;
540 return (0);
541 }
542
543 static int
pci_iov_alloc_bar_ea(struct pci_devinfo * dinfo,int bar)544 pci_iov_alloc_bar_ea(struct pci_devinfo *dinfo, int bar)
545 {
546 struct pcicfg_iov *iov;
547 rman_res_t start, end;
548 struct resource *res;
549 struct resource_list *rl;
550 struct resource_list_entry *rle;
551
552 rl = &dinfo->resources;
553 iov = dinfo->cfg.iov;
554
555 rle = resource_list_find(rl, SYS_RES_MEMORY,
556 iov->iov_pos + PCIR_SRIOV_BAR(bar));
557 if (rle == NULL)
558 rle = resource_list_find(rl, SYS_RES_IOPORT,
559 iov->iov_pos + PCIR_SRIOV_BAR(bar));
560 if (rle == NULL)
561 return (ENXIO);
562 res = rle->res;
563
564 iov->iov_bar[bar].res = res;
565 iov->iov_bar[bar].bar_size = rman_get_size(res) / iov->iov_num_vfs;
566 iov->iov_bar[bar].bar_shift = pci_mapsize(iov->iov_bar[bar].bar_size);
567
568 start = rman_get_start(res);
569 end = rman_get_end(res);
570
571 return (rman_manage_region(&iov->rman, start, end));
572 }
573
574 static int
pci_iov_setup_bars(struct pci_devinfo * dinfo)575 pci_iov_setup_bars(struct pci_devinfo *dinfo)
576 {
577 device_t dev;
578 struct pcicfg_iov *iov;
579 pci_addr_t bar_value, testval;
580 int i, last_64, error;
581
582 iov = dinfo->cfg.iov;
583 dev = dinfo->cfg.dev;
584 last_64 = 0;
585
586 pci_add_resources_ea(device_get_parent(dev), dev, 1);
587
588 for (i = 0; i <= PCIR_MAX_BAR_0; i++) {
589 /* First, try to use BARs allocated with EA */
590 error = pci_iov_alloc_bar_ea(dinfo, i);
591 if (error == 0)
592 continue;
593
594 /* Allocate legacy-BAR only if EA is not enabled */
595 if (pci_ea_is_enabled(dev, iov->iov_pos + PCIR_SRIOV_BAR(i)))
596 continue;
597
598 /*
599 * If a PCI BAR is a 64-bit wide BAR, then it spans two
600 * consecutive registers. Therefore if the last BAR that
601 * we looked at was a 64-bit BAR, we need to skip this
602 * register as it's the second half of the last BAR.
603 */
604 if (!last_64) {
605 pci_read_bar(dev,
606 iov->iov_pos + PCIR_SRIOV_BAR(i),
607 &bar_value, &testval, &last_64);
608
609 if (testval != 0) {
610 error = pci_iov_alloc_bar(dinfo, i,
611 pci_mapsize(testval));
612 if (error != 0)
613 return (error);
614 }
615 } else
616 last_64 = 0;
617 }
618
619 return (0);
620 }
621
622 static void
pci_iov_enumerate_vfs(struct pci_devinfo * dinfo,const nvlist_t * config,uint16_t first_rid,uint16_t rid_stride)623 pci_iov_enumerate_vfs(struct pci_devinfo *dinfo, const nvlist_t *config,
624 uint16_t first_rid, uint16_t rid_stride)
625 {
626 char device_name[VF_MAX_NAME];
627 const nvlist_t *device, *driver_config, *iov_config;
628 device_t bus, dev, vf;
629 struct pcicfg_iov *iov;
630 struct pci_devinfo *vfinfo;
631 int i, error;
632 uint16_t vid, did, next_rid;
633
634 iov = dinfo->cfg.iov;
635 dev = dinfo->cfg.dev;
636 bus = device_get_parent(dev);
637 next_rid = first_rid;
638 vid = pci_get_vendor(dev);
639 did = IOV_READ(dinfo, PCIR_SRIOV_VF_DID, 2);
640
641 for (i = 0; i < iov->iov_num_vfs; i++, next_rid += rid_stride) {
642 snprintf(device_name, sizeof(device_name), VF_PREFIX"%d", i);
643 device = nvlist_get_nvlist(config, device_name);
644 iov_config = nvlist_get_nvlist(device, IOV_CONFIG_NAME);
645 driver_config = nvlist_get_nvlist(device, DRIVER_CONFIG_NAME);
646
647 vf = PCI_CREATE_IOV_CHILD(bus, dev, next_rid, vid, did);
648 if (vf == NULL)
649 break;
650
651 /*
652 * If we are creating passthrough devices then force the ppt
653 * driver to attach to prevent a VF driver from claiming the
654 * VFs.
655 */
656 if (nvlist_get_bool(iov_config, "passthrough"))
657 device_set_devclass_fixed(vf, "ppt");
658
659 vfinfo = device_get_ivars(vf);
660
661 vfinfo->cfg.iov = iov;
662 vfinfo->cfg.vf.index = i;
663
664 pci_iov_add_bars(iov, vfinfo);
665
666 error = PCI_IOV_ADD_VF(dev, i, driver_config);
667 if (error != 0) {
668 device_printf(dev, "Failed to add VF %d\n", i);
669 device_delete_child(bus, vf);
670 }
671 }
672
673 bus_generic_attach(bus);
674 }
675
676 static int
pci_iov_config(struct cdev * cdev,struct pci_iov_arg * arg)677 pci_iov_config(struct cdev *cdev, struct pci_iov_arg *arg)
678 {
679 device_t bus, dev;
680 struct pci_devinfo *dinfo;
681 struct pcicfg_iov *iov;
682 nvlist_t *config;
683 int i, error;
684 uint16_t rid_off, rid_stride;
685 uint16_t first_rid, last_rid;
686 uint16_t iov_ctl;
687 uint16_t num_vfs, total_vfs;
688 int iov_inited;
689 bool ari_enabled;
690
691 mtx_lock(&Giant);
692 dinfo = cdev->si_drv1;
693 iov = dinfo->cfg.iov;
694 dev = dinfo->cfg.dev;
695 bus = device_get_parent(dev);
696 iov_inited = 0;
697 config = NULL;
698
699 if ((iov->iov_flags & IOV_BUSY) || iov->iov_num_vfs != 0) {
700 mtx_unlock(&Giant);
701 return (EBUSY);
702 }
703 iov->iov_flags |= IOV_BUSY;
704
705 error = pci_iov_parse_config(iov, arg, &config);
706 if (error != 0)
707 goto out;
708
709 num_vfs = pci_iov_config_get_num_vfs(config);
710 total_vfs = IOV_READ(dinfo, PCIR_SRIOV_TOTAL_VFS, 2);
711 if (num_vfs > total_vfs) {
712 error = EINVAL;
713 goto out;
714 }
715
716 error = pci_iov_config_page_size(dinfo);
717 if (error != 0)
718 goto out;
719
720 error = pci_iov_set_ari(bus, &ari_enabled);
721 if (error != 0)
722 goto out;
723
724 error = pci_iov_init(dev, num_vfs, config);
725 if (error != 0)
726 goto out;
727 iov_inited = 1;
728
729 IOV_WRITE(dinfo, PCIR_SRIOV_NUM_VFS, num_vfs, 2);
730
731 rid_off = IOV_READ(dinfo, PCIR_SRIOV_VF_OFF, 2);
732 rid_stride = IOV_READ(dinfo, PCIR_SRIOV_VF_STRIDE, 2);
733
734 first_rid = pci_get_rid(dev) + rid_off;
735 last_rid = first_rid + (num_vfs - 1) * rid_stride;
736
737 /* We don't yet support allocating extra bus numbers for VFs. */
738 if (pci_get_bus(dev) != PCI_RID2BUS(last_rid)) {
739 device_printf(dev, "not enough PCIe bus numbers for VFs\n");
740 error = ENOSPC;
741 goto out;
742 }
743
744 if (!ari_enabled && PCI_RID2SLOT(last_rid) != 0) {
745 error = ENOSPC;
746 goto out;
747 }
748
749 iov_ctl = IOV_READ(dinfo, PCIR_SRIOV_CTL, 2);
750 iov_ctl &= ~(PCIM_SRIOV_VF_EN | PCIM_SRIOV_VF_MSE);
751 IOV_WRITE(dinfo, PCIR_SRIOV_CTL, iov_ctl, 2);
752
753 error = pci_iov_init_rman(dev, iov);
754 if (error != 0)
755 goto out;
756
757 iov->iov_num_vfs = num_vfs;
758
759 error = pci_iov_setup_bars(dinfo);
760 if (error != 0)
761 goto out;
762
763 iov_ctl = IOV_READ(dinfo, PCIR_SRIOV_CTL, 2);
764 iov_ctl |= PCIM_SRIOV_VF_EN | PCIM_SRIOV_VF_MSE;
765 IOV_WRITE(dinfo, PCIR_SRIOV_CTL, iov_ctl, 2);
766
767 /* Per specification, we must wait 100ms before accessing VFs. */
768 pause("iov", roundup(hz, 10));
769 pci_iov_enumerate_vfs(dinfo, config, first_rid, rid_stride);
770
771 nvlist_destroy(config);
772 iov->iov_flags &= ~IOV_BUSY;
773 mtx_unlock(&Giant);
774
775 return (0);
776 out:
777 if (iov_inited)
778 PCI_IOV_UNINIT(dev);
779
780 for (i = 0; i <= PCIR_MAX_BAR_0; i++) {
781 if (iov->iov_bar[i].res != NULL) {
782 pci_release_resource(bus, dev, iov->iov_bar[i].res);
783 pci_delete_resource(bus, dev, SYS_RES_MEMORY,
784 iov->iov_pos + PCIR_SRIOV_BAR(i));
785 iov->iov_bar[i].res = NULL;
786 }
787 }
788
789 if (iov->iov_flags & IOV_RMAN_INITED) {
790 rman_fini(&iov->rman);
791 iov->iov_flags &= ~IOV_RMAN_INITED;
792 }
793
794 nvlist_destroy(config);
795 iov->iov_num_vfs = 0;
796 iov->iov_flags &= ~IOV_BUSY;
797 mtx_unlock(&Giant);
798 return (error);
799 }
800
801 void
pci_iov_cfg_restore(device_t dev,struct pci_devinfo * dinfo)802 pci_iov_cfg_restore(device_t dev, struct pci_devinfo *dinfo)
803 {
804 struct pcicfg_iov *iov;
805
806 iov = dinfo->cfg.iov;
807
808 IOV_WRITE(dinfo, PCIR_SRIOV_PAGE_SIZE, iov->iov_page_size, 4);
809 IOV_WRITE(dinfo, PCIR_SRIOV_NUM_VFS, iov->iov_num_vfs, 2);
810 IOV_WRITE(dinfo, PCIR_SRIOV_CTL, iov->iov_ctl, 2);
811 }
812
813 void
pci_iov_cfg_save(device_t dev,struct pci_devinfo * dinfo)814 pci_iov_cfg_save(device_t dev, struct pci_devinfo *dinfo)
815 {
816 struct pcicfg_iov *iov;
817
818 iov = dinfo->cfg.iov;
819
820 iov->iov_page_size = IOV_READ(dinfo, PCIR_SRIOV_PAGE_SIZE, 4);
821 iov->iov_ctl = IOV_READ(dinfo, PCIR_SRIOV_CTL, 2);
822 }
823
824 /* Return true if child is a VF of the given PF. */
825 static int
pci_iov_is_child_vf(struct pcicfg_iov * pf,device_t child)826 pci_iov_is_child_vf(struct pcicfg_iov *pf, device_t child)
827 {
828 struct pci_devinfo *vfinfo;
829
830 vfinfo = device_get_ivars(child);
831
832 if (!(vfinfo->cfg.flags & PCICFG_VF))
833 return (0);
834
835 return (pf == vfinfo->cfg.iov);
836 }
837
838 static int
pci_iov_delete_iov_children(struct pci_devinfo * dinfo)839 pci_iov_delete_iov_children(struct pci_devinfo *dinfo)
840 {
841 device_t bus, dev, vf, *devlist;
842 struct pcicfg_iov *iov;
843 int i, error, devcount;
844 uint32_t iov_ctl;
845
846 mtx_assert(&Giant, MA_OWNED);
847
848 iov = dinfo->cfg.iov;
849 dev = dinfo->cfg.dev;
850 bus = device_get_parent(dev);
851 devlist = NULL;
852
853 iov->iov_flags |= IOV_BUSY;
854
855 error = device_get_children(bus, &devlist, &devcount);
856
857 if (error != 0)
858 goto out;
859
860 for (i = 0; i < devcount; i++) {
861 vf = devlist[i];
862
863 if (!pci_iov_is_child_vf(iov, vf))
864 continue;
865
866 error = device_detach(vf);
867 if (error != 0) {
868 device_printf(dev,
869 "Could not disable SR-IOV: failed to detach VF %s\n",
870 device_get_nameunit(vf));
871 goto out;
872 }
873 }
874
875 for (i = 0; i < devcount; i++) {
876 vf = devlist[i];
877
878 if (pci_iov_is_child_vf(iov, vf))
879 device_delete_child(bus, vf);
880 }
881 PCI_IOV_UNINIT(dev);
882
883 iov_ctl = IOV_READ(dinfo, PCIR_SRIOV_CTL, 2);
884 iov_ctl &= ~(PCIM_SRIOV_VF_EN | PCIM_SRIOV_VF_MSE);
885 IOV_WRITE(dinfo, PCIR_SRIOV_CTL, iov_ctl, 2);
886 IOV_WRITE(dinfo, PCIR_SRIOV_NUM_VFS, 0, 2);
887
888 iov->iov_num_vfs = 0;
889
890 for (i = 0; i <= PCIR_MAX_BAR_0; i++) {
891 if (iov->iov_bar[i].res != NULL) {
892 pci_release_resource(bus, dev, iov->iov_bar[i].res);
893 pci_delete_resource(bus, dev, SYS_RES_MEMORY,
894 iov->iov_pos + PCIR_SRIOV_BAR(i));
895 iov->iov_bar[i].res = NULL;
896 }
897 }
898
899 if (iov->iov_flags & IOV_RMAN_INITED) {
900 rman_fini(&iov->rman);
901 iov->iov_flags &= ~IOV_RMAN_INITED;
902 }
903
904 error = 0;
905 out:
906 free(devlist, M_TEMP);
907 iov->iov_flags &= ~IOV_BUSY;
908 return (error);
909 }
910
911 static int
pci_iov_delete(struct cdev * cdev)912 pci_iov_delete(struct cdev *cdev)
913 {
914 struct pci_devinfo *dinfo;
915 struct pcicfg_iov *iov;
916 int error;
917
918 mtx_lock(&Giant);
919 dinfo = cdev->si_drv1;
920 iov = dinfo->cfg.iov;
921
922 if ((iov->iov_flags & IOV_BUSY) != 0) {
923 error = EBUSY;
924 goto out;
925 }
926 if (iov->iov_num_vfs == 0) {
927 error = ECHILD;
928 goto out;
929 }
930
931 error = pci_iov_delete_iov_children(dinfo);
932
933 out:
934 mtx_unlock(&Giant);
935 return (error);
936 }
937
938 static int
pci_iov_get_schema_ioctl(struct cdev * cdev,struct pci_iov_schema * output)939 pci_iov_get_schema_ioctl(struct cdev *cdev, struct pci_iov_schema *output)
940 {
941 struct pci_devinfo *dinfo;
942 void *packed;
943 size_t output_len, size;
944 int error;
945
946 packed = NULL;
947
948 mtx_lock(&Giant);
949 dinfo = cdev->si_drv1;
950 packed = nvlist_pack(dinfo->cfg.iov->iov_schema, &size);
951 mtx_unlock(&Giant);
952
953 if (packed == NULL) {
954 error = ENOMEM;
955 goto fail;
956 }
957
958 output_len = output->len;
959 output->len = size;
960 if (size <= output_len) {
961 error = copyout(packed, output->schema, size);
962
963 if (error != 0)
964 goto fail;
965
966 output->error = 0;
967 } else
968 /*
969 * If we return an error then the ioctl code won't copyout
970 * output back to userland, so we flag the error in the struct
971 * instead.
972 */
973 output->error = EMSGSIZE;
974
975 error = 0;
976
977 fail:
978 free(packed, M_NVLIST);
979
980 return (error);
981 }
982
983 static int
pci_iov_ioctl(struct cdev * dev,u_long cmd,caddr_t data,int fflag,struct thread * td)984 pci_iov_ioctl(struct cdev *dev, u_long cmd, caddr_t data, int fflag,
985 struct thread *td)
986 {
987
988 switch (cmd) {
989 case IOV_CONFIG:
990 return (pci_iov_config(dev, (struct pci_iov_arg *)data));
991 case IOV_DELETE:
992 return (pci_iov_delete(dev));
993 case IOV_GET_SCHEMA:
994 return (pci_iov_get_schema_ioctl(dev,
995 (struct pci_iov_schema *)data));
996 default:
997 return (EINVAL);
998 }
999 }
1000
1001 struct resource *
pci_vf_alloc_mem_resource(device_t dev,device_t child,int * rid,rman_res_t start,rman_res_t end,rman_res_t count,u_int flags)1002 pci_vf_alloc_mem_resource(device_t dev, device_t child, int *rid,
1003 rman_res_t start, rman_res_t end, rman_res_t count, u_int flags)
1004 {
1005 struct pci_devinfo *dinfo;
1006 struct pcicfg_iov *iov;
1007 struct pci_map *map;
1008 struct resource *res;
1009 struct resource_list_entry *rle;
1010 rman_res_t bar_start, bar_end;
1011 pci_addr_t bar_length;
1012 int error;
1013
1014 dinfo = device_get_ivars(child);
1015 iov = dinfo->cfg.iov;
1016
1017 map = pci_find_bar(child, *rid);
1018 if (map == NULL)
1019 return (NULL);
1020
1021 bar_length = 1 << map->pm_size;
1022 bar_start = map->pm_value;
1023 bar_end = bar_start + bar_length - 1;
1024
1025 /* Make sure that the resource fits the constraints. */
1026 if (bar_start >= end || bar_end <= bar_start || count != 1)
1027 return (NULL);
1028
1029 /* Clamp the resource to the constraints if necessary. */
1030 if (bar_start < start)
1031 bar_start = start;
1032 if (bar_end > end)
1033 bar_end = end;
1034 bar_length = bar_end - bar_start + 1;
1035
1036 res = rman_reserve_resource(&iov->rman, bar_start, bar_end,
1037 bar_length, flags, child);
1038 if (res == NULL)
1039 return (NULL);
1040
1041 rle = resource_list_add(&dinfo->resources, SYS_RES_MEMORY, *rid,
1042 bar_start, bar_end, 1);
1043 if (rle == NULL) {
1044 rman_release_resource(res);
1045 return (NULL);
1046 }
1047
1048 rman_set_rid(res, *rid);
1049 rman_set_type(res, SYS_RES_MEMORY);
1050
1051 if (flags & RF_ACTIVE) {
1052 error = bus_activate_resource(child, SYS_RES_MEMORY, *rid, res);
1053 if (error != 0) {
1054 resource_list_delete(&dinfo->resources, SYS_RES_MEMORY,
1055 *rid);
1056 rman_release_resource(res);
1057 return (NULL);
1058 }
1059 }
1060 rle->res = res;
1061
1062 return (res);
1063 }
1064
1065 int
pci_vf_release_mem_resource(device_t dev,device_t child,struct resource * r)1066 pci_vf_release_mem_resource(device_t dev, device_t child, struct resource *r)
1067 {
1068 struct pci_devinfo *dinfo;
1069 struct resource_list_entry *rle;
1070 int error, rid;
1071
1072 dinfo = device_get_ivars(child);
1073
1074 KASSERT(rman_get_type(r) == SYS_RES_MEMORY,
1075 ("%s: invalid resource %p", __func__, r));
1076 KASSERT(rman_is_region_manager(r, &dinfo->cfg.iov->rman),
1077 ("%s: rman %p doesn't match for resource %p", __func__,
1078 &dinfo->cfg.iov->rman, r));
1079
1080 if (rman_get_flags(r) & RF_ACTIVE) {
1081 error = bus_deactivate_resource(child, r);
1082 if (error != 0)
1083 return (error);
1084 }
1085
1086 rid = rman_get_rid(r);
1087 rle = resource_list_find(&dinfo->resources, SYS_RES_MEMORY, rid);
1088 if (rle != NULL) {
1089 rle->res = NULL;
1090 resource_list_delete(&dinfo->resources, SYS_RES_MEMORY,
1091 rid);
1092 }
1093
1094 return (rman_release_resource(r));
1095 }
1096
1097 int
pci_vf_activate_mem_resource(device_t dev,device_t child,struct resource * r)1098 pci_vf_activate_mem_resource(device_t dev, device_t child, struct resource *r)
1099 {
1100 #ifdef INVARIANTS
1101 struct pci_devinfo *dinfo = device_get_ivars(child);
1102 #endif
1103 struct resource_map map;
1104 int error;
1105
1106 KASSERT(rman_get_type(r) == SYS_RES_MEMORY,
1107 ("%s: invalid resource %p", __func__, r));
1108 KASSERT(rman_is_region_manager(r, &dinfo->cfg.iov->rman),
1109 ("%s: rman %p doesn't match for resource %p", __func__,
1110 &dinfo->cfg.iov->rman, r));
1111
1112 error = rman_activate_resource(r);
1113 if (error != 0)
1114 return (error);
1115
1116 if ((rman_get_flags(r) & RF_UNMAPPED) == 0) {
1117 error = BUS_MAP_RESOURCE(dev, child, r, NULL, &map);
1118 if (error != 0) {
1119 rman_deactivate_resource(r);
1120 return (error);
1121 }
1122
1123 rman_set_mapping(r, &map);
1124 }
1125 return (0);
1126 }
1127
1128 int
pci_vf_deactivate_mem_resource(device_t dev,device_t child,struct resource * r)1129 pci_vf_deactivate_mem_resource(device_t dev, device_t child, struct resource *r)
1130 {
1131 #ifdef INVARIANTS
1132 struct pci_devinfo *dinfo = device_get_ivars(child);
1133 #endif
1134 struct resource_map map;
1135 int error;
1136
1137 KASSERT(rman_get_type(r) == SYS_RES_MEMORY,
1138 ("%s: invalid resource %p", __func__, r));
1139 KASSERT(rman_is_region_manager(r, &dinfo->cfg.iov->rman),
1140 ("%s: rman %p doesn't match for resource %p", __func__,
1141 &dinfo->cfg.iov->rman, r));
1142
1143 error = rman_deactivate_resource(r);
1144 if (error != 0)
1145 return (error);
1146
1147 if ((rman_get_flags(r) & RF_UNMAPPED) == 0) {
1148 rman_get_mapping(r, &map);
1149 BUS_UNMAP_RESOURCE(dev, child, r, &map);
1150 }
1151 return (0);
1152 }
1153
1154 int
pci_vf_adjust_mem_resource(device_t dev,device_t child,struct resource * r,rman_res_t start,rman_res_t end)1155 pci_vf_adjust_mem_resource(device_t dev, device_t child, struct resource *r,
1156 rman_res_t start, rman_res_t end)
1157 {
1158 #ifdef INVARIANTS
1159 struct pci_devinfo *dinfo = device_get_ivars(child);
1160 #endif
1161
1162 KASSERT(rman_get_type(r) == SYS_RES_MEMORY,
1163 ("%s: invalid resource %p", __func__, r));
1164 KASSERT(rman_is_region_manager(r, &dinfo->cfg.iov->rman),
1165 ("%s: rman %p doesn't match for resource %p", __func__,
1166 &dinfo->cfg.iov->rman, r));
1167
1168 return (rman_adjust_resource(r, start, end));
1169 }
1170
1171 static struct resource *
pci_vf_find_parent_resource(struct pcicfg_iov * iov,struct resource * r)1172 pci_vf_find_parent_resource(struct pcicfg_iov *iov, struct resource *r)
1173 {
1174 struct resource *pres;
1175
1176 for (u_int i = 0; i <= PCIR_MAX_BAR_0; i++) {
1177 pres = iov->iov_bar[i].res;
1178 if (pres != NULL) {
1179 if (rman_get_start(pres) <= rman_get_start(r) &&
1180 rman_get_end(pres) >= rman_get_end(r))
1181 return (pres);
1182 }
1183 }
1184 return (NULL);
1185 }
1186
1187 int
pci_vf_map_mem_resource(device_t dev,device_t child,struct resource * r,struct resource_map_request * argsp,struct resource_map * map)1188 pci_vf_map_mem_resource(device_t dev, device_t child, struct resource *r,
1189 struct resource_map_request *argsp, struct resource_map *map)
1190 {
1191 struct pci_devinfo *dinfo = device_get_ivars(child);
1192 struct pcicfg_iov *iov = dinfo->cfg.iov;
1193 struct resource_map_request args;
1194 struct resource *pres;
1195 rman_res_t length, start;
1196 int error;
1197
1198 KASSERT(rman_get_type(r) == SYS_RES_MEMORY,
1199 ("%s: invalid resource %p", __func__, r));
1200 KASSERT(rman_is_region_manager(r, &iov->rman),
1201 ("%s: rman %p doesn't match for resource %p", __func__,
1202 &dinfo->cfg.iov->rman, r));
1203
1204 /* Resources must be active to be mapped. */
1205 if (!(rman_get_flags(r) & RF_ACTIVE))
1206 return (ENXIO);
1207
1208 resource_init_map_request(&args);
1209 error = resource_validate_map_request(r, argsp, &args, &start, &length);
1210 if (error)
1211 return (error);
1212
1213 pres = pci_vf_find_parent_resource(dinfo->cfg.iov, r);
1214 if (pres == NULL)
1215 return (ENOENT);
1216
1217 args.offset = start - rman_get_start(pres);
1218 args.length = length;
1219 return (bus_map_resource(iov->iov_pf, pres, &args, map));
1220 }
1221
1222 int
pci_vf_unmap_mem_resource(device_t dev,device_t child,struct resource * r,struct resource_map * map)1223 pci_vf_unmap_mem_resource(device_t dev, device_t child, struct resource *r,
1224 struct resource_map *map)
1225 {
1226 struct pci_devinfo *dinfo = device_get_ivars(child);
1227 struct pcicfg_iov *iov = dinfo->cfg.iov;
1228 struct resource *pres;
1229
1230 KASSERT(rman_get_type(r) == SYS_RES_MEMORY,
1231 ("%s: invalid resource %p", __func__, r));
1232 KASSERT(rman_is_region_manager(r, &iov->rman),
1233 ("%s: rman %p doesn't match for resource %p", __func__,
1234 &dinfo->cfg.iov->rman, r));
1235
1236 pres = pci_vf_find_parent_resource(iov, r);
1237 if (pres == NULL)
1238 return (ENOENT);
1239 return (bus_unmap_resource(iov->iov_pf, pres, map));
1240 }
1241