xref: /freebsd/sys/dev/pci/pci_iov.c (revision 5bb3134a8c21cb87b30e135ef168483f0333dabb)
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 __FBSDID("$FreeBSD$");
29 
30 #include "opt_bus.h"
31 
32 #include <sys/param.h>
33 #include <sys/conf.h>
34 #include <sys/kernel.h>
35 #include <sys/systm.h>
36 #include <sys/bus.h>
37 #include <sys/fcntl.h>
38 #include <sys/ioccom.h>
39 #include <sys/iov.h>
40 #include <sys/linker.h>
41 #include <sys/lock.h>
42 #include <sys/malloc.h>
43 #include <sys/module.h>
44 #include <sys/mutex.h>
45 #include <sys/pciio.h>
46 #include <sys/queue.h>
47 #include <sys/rman.h>
48 #include <sys/sysctl.h>
49 
50 #include <machine/bus.h>
51 #include <machine/stdarg.h>
52 
53 #include <sys/nv.h>
54 #include <sys/iov_schema.h>
55 
56 #include <dev/pci/pcireg.h>
57 #include <dev/pci/pcivar.h>
58 #include <dev/pci/pci_iov.h>
59 #include <dev/pci/pci_private.h>
60 #include <dev/pci/pci_iov_private.h>
61 #include <dev/pci/schema_private.h>
62 
63 #include "pcib_if.h"
64 
65 static MALLOC_DEFINE(M_SRIOV, "sr_iov", "PCI SR-IOV allocations");
66 
67 static d_ioctl_t pci_iov_ioctl;
68 
69 static struct cdevsw iov_cdevsw = {
70 	.d_version = D_VERSION,
71 	.d_name = "iov",
72 	.d_ioctl = pci_iov_ioctl
73 };
74 
75 SYSCTL_DECL(_hw_pci);
76 
77 /*
78  * The maximum amount of memory we will allocate for user configuration of an
79  * SR-IOV device.  1MB ought to be enough for anyone, but leave this
80  * configurable just in case.
81  */
82 static u_long pci_iov_max_config = 1024 * 1024;
83 SYSCTL_ULONG(_hw_pci, OID_AUTO, iov_max_config, CTLFLAG_RWTUN,
84     &pci_iov_max_config, 0, "Maximum allowed size of SR-IOV configuration.");
85 
86 #define IOV_READ(d, r, w) \
87 	pci_read_config((d)->cfg.dev, (d)->cfg.iov->iov_pos + r, w)
88 
89 #define IOV_WRITE(d, r, v, w) \
90 	pci_write_config((d)->cfg.dev, (d)->cfg.iov->iov_pos + r, v, w)
91 
92 static nvlist_t	*pci_iov_build_schema(nvlist_t **pf_schema,
93 		    nvlist_t **vf_schema);
94 static void	pci_iov_build_pf_schema(nvlist_t *schema,
95 		    nvlist_t **driver_schema);
96 static void	pci_iov_build_vf_schema(nvlist_t *schema,
97 		    nvlist_t **driver_schema);
98 static int	pci_iov_delete_iov_children(struct pci_devinfo *dinfo);
99 static nvlist_t	*pci_iov_get_pf_subsystem_schema(void);
100 static nvlist_t	*pci_iov_get_vf_subsystem_schema(void);
101 
102 int
103 pci_iov_attach_name(device_t dev, struct nvlist *pf_schema,
104     struct nvlist *vf_schema, const char *fmt, ...)
105 {
106 	char buf[NAME_MAX + 1];
107 	va_list ap;
108 
109 	va_start(ap, fmt);
110 	vsnprintf(buf, sizeof(buf), fmt, ap);
111 	va_end(ap);
112 	return (PCI_IOV_ATTACH(device_get_parent(dev), dev, pf_schema,
113 	    vf_schema, buf));
114 }
115 
116 int
117 pci_iov_attach_method(device_t bus, device_t dev, nvlist_t *pf_schema,
118     nvlist_t *vf_schema, const char *name)
119 {
120 	device_t pcib;
121 	struct pci_devinfo *dinfo;
122 	struct pcicfg_iov *iov;
123 	nvlist_t *schema;
124 	uint32_t version;
125 	int error;
126 	int iov_pos;
127 
128 	dinfo = device_get_ivars(dev);
129 	pcib = device_get_parent(bus);
130 	schema = NULL;
131 
132 	error = pci_find_extcap(dev, PCIZ_SRIOV, &iov_pos);
133 
134 	if (error != 0)
135 		return (error);
136 
137 	version = pci_read_config(dev, iov_pos, 4);
138 	if (PCI_EXTCAP_VER(version) != 1) {
139 		if (bootverbose)
140 			device_printf(dev,
141 			    "Unsupported version of SR-IOV (%d) detected\n",
142 			    PCI_EXTCAP_VER(version));
143 
144 		return (ENXIO);
145 	}
146 
147 	iov = malloc(sizeof(*dinfo->cfg.iov), M_SRIOV, M_WAITOK | M_ZERO);
148 
149 	mtx_lock(&Giant);
150 	if (dinfo->cfg.iov != NULL) {
151 		error = EBUSY;
152 		goto cleanup;
153 	}
154 	iov->iov_pf = dev;
155 	iov->iov_pos = iov_pos;
156 
157 	schema = pci_iov_build_schema(&pf_schema, &vf_schema);
158 	if (schema == NULL) {
159 		error = ENOMEM;
160 		goto cleanup;
161 	}
162 
163 	error = pci_iov_validate_schema(schema);
164 	if (error != 0)
165 		goto cleanup;
166 	iov->iov_schema = schema;
167 
168 	iov->iov_cdev = make_dev(&iov_cdevsw, device_get_unit(dev),
169 	    UID_ROOT, GID_WHEEL, 0600, "iov/%s", name);
170 
171 	if (iov->iov_cdev == NULL) {
172 		error = ENOMEM;
173 		goto cleanup;
174 	}
175 
176 	dinfo->cfg.iov = iov;
177 	iov->iov_cdev->si_drv1 = dinfo;
178 	mtx_unlock(&Giant);
179 
180 	return (0);
181 
182 cleanup:
183 	nvlist_destroy(schema);
184 	nvlist_destroy(pf_schema);
185 	nvlist_destroy(vf_schema);
186 	free(iov, M_SRIOV);
187 	mtx_unlock(&Giant);
188 	return (error);
189 }
190 
191 int
192 pci_iov_detach_method(device_t bus, device_t dev)
193 {
194 	struct pci_devinfo *dinfo;
195 	struct pcicfg_iov *iov;
196 	int error;
197 
198 	mtx_lock(&Giant);
199 	dinfo = device_get_ivars(dev);
200 	iov = dinfo->cfg.iov;
201 
202 	if (iov == NULL) {
203 		mtx_unlock(&Giant);
204 		return (0);
205 	}
206 
207 	if ((iov->iov_flags & IOV_BUSY) != 0) {
208 		mtx_unlock(&Giant);
209 		return (EBUSY);
210 	}
211 
212 	error = pci_iov_delete_iov_children(dinfo);
213 	if (error != 0) {
214 		mtx_unlock(&Giant);
215 		return (error);
216 	}
217 
218 	dinfo->cfg.iov = NULL;
219 
220 	if (iov->iov_cdev) {
221 		destroy_dev(iov->iov_cdev);
222 		iov->iov_cdev = NULL;
223 	}
224 	nvlist_destroy(iov->iov_schema);
225 
226 	free(iov, M_SRIOV);
227 	mtx_unlock(&Giant);
228 
229 	return (0);
230 }
231 
232 static nvlist_t *
233 pci_iov_build_schema(nvlist_t **pf, nvlist_t **vf)
234 {
235 	nvlist_t *schema, *pf_driver, *vf_driver;
236 
237 	/* We always take ownership of the schemas. */
238 	pf_driver = *pf;
239 	*pf = NULL;
240 	vf_driver = *vf;
241 	*vf = NULL;
242 
243 	schema = pci_iov_schema_alloc_node();
244 	if (schema == NULL)
245 		goto cleanup;
246 
247 	pci_iov_build_pf_schema(schema, &pf_driver);
248 	pci_iov_build_vf_schema(schema, &vf_driver);
249 
250 	if (nvlist_error(schema) != 0)
251 		goto cleanup;
252 
253 	return (schema);
254 
255 cleanup:
256 	nvlist_destroy(schema);
257 	nvlist_destroy(pf_driver);
258 	nvlist_destroy(vf_driver);
259 	return (NULL);
260 }
261 
262 static void
263 pci_iov_build_pf_schema(nvlist_t *schema, nvlist_t **driver_schema)
264 {
265 	nvlist_t *pf_schema, *iov_schema;
266 
267 	pf_schema = pci_iov_schema_alloc_node();
268 	if (pf_schema == NULL) {
269 		nvlist_set_error(schema, ENOMEM);
270 		return;
271 	}
272 
273 	iov_schema = pci_iov_get_pf_subsystem_schema();
274 
275 	/*
276 	 * Note that if either *driver_schema or iov_schema is NULL, then
277 	 * nvlist_move_nvlist will put the schema in the error state and
278 	 * SR-IOV will fail to initialize later, so we don't have to explicitly
279 	 * handle that case.
280 	 */
281 	nvlist_move_nvlist(pf_schema, DRIVER_CONFIG_NAME, *driver_schema);
282 	nvlist_move_nvlist(pf_schema, IOV_CONFIG_NAME, iov_schema);
283 	nvlist_move_nvlist(schema, PF_CONFIG_NAME, pf_schema);
284 	*driver_schema = NULL;
285 }
286 
287 static void
288 pci_iov_build_vf_schema(nvlist_t *schema, nvlist_t **driver_schema)
289 {
290 	nvlist_t *vf_schema, *iov_schema;
291 
292 	vf_schema = pci_iov_schema_alloc_node();
293 	if (vf_schema == NULL) {
294 		nvlist_set_error(schema, ENOMEM);
295 		return;
296 	}
297 
298 	iov_schema = pci_iov_get_vf_subsystem_schema();
299 
300 	/*
301 	 * Note that if either *driver_schema or iov_schema is NULL, then
302 	 * nvlist_move_nvlist will put the schema in the error state and
303 	 * SR-IOV will fail to initialize later, so we don't have to explicitly
304 	 * handle that case.
305 	 */
306 	nvlist_move_nvlist(vf_schema, DRIVER_CONFIG_NAME, *driver_schema);
307 	nvlist_move_nvlist(vf_schema, IOV_CONFIG_NAME, iov_schema);
308 	nvlist_move_nvlist(schema, VF_SCHEMA_NAME, vf_schema);
309 	*driver_schema = NULL;
310 }
311 
312 static nvlist_t *
313 pci_iov_get_pf_subsystem_schema(void)
314 {
315 	nvlist_t *pf;
316 
317 	pf = pci_iov_schema_alloc_node();
318 	if (pf == NULL)
319 		return (NULL);
320 
321 	pci_iov_schema_add_uint16(pf, "num_vfs", IOV_SCHEMA_REQUIRED, -1);
322 	pci_iov_schema_add_string(pf, "device", IOV_SCHEMA_REQUIRED, NULL);
323 
324 	return (pf);
325 }
326 
327 static nvlist_t *
328 pci_iov_get_vf_subsystem_schema(void)
329 {
330 	nvlist_t *vf;
331 
332 	vf = pci_iov_schema_alloc_node();
333 	if (vf == NULL)
334 		return (NULL);
335 
336 	pci_iov_schema_add_bool(vf, "passthrough", IOV_SCHEMA_HASDEFAULT, 0);
337 
338 	return (vf);
339 }
340 
341 static int
342 pci_iov_alloc_bar(struct pci_devinfo *dinfo, int bar, pci_addr_t bar_shift)
343 {
344 	struct resource *res;
345 	struct pcicfg_iov *iov;
346 	device_t dev, bus;
347 	rman_res_t start, end;
348 	pci_addr_t bar_size;
349 	int rid;
350 
351 	iov = dinfo->cfg.iov;
352 	dev = dinfo->cfg.dev;
353 	bus = device_get_parent(dev);
354 	rid = iov->iov_pos + PCIR_SRIOV_BAR(bar);
355 	bar_size = 1 << bar_shift;
356 
357 	res = pci_alloc_multi_resource(bus, dev, SYS_RES_MEMORY, &rid, 0,
358 	    ~0, 1, iov->iov_num_vfs, RF_ACTIVE);
359 
360 	if (res == NULL)
361 		return (ENXIO);
362 
363 	iov->iov_bar[bar].res = res;
364 	iov->iov_bar[bar].bar_size = bar_size;
365 	iov->iov_bar[bar].bar_shift = bar_shift;
366 
367 	start = rman_get_start(res);
368 	end = rman_get_end(res);
369 	return (rman_manage_region(&iov->rman, start, end));
370 }
371 
372 static void
373 pci_iov_add_bars(struct pcicfg_iov *iov, struct pci_devinfo *dinfo)
374 {
375 	struct pci_iov_bar *bar;
376 	uint64_t bar_start;
377 	int i;
378 
379 	for (i = 0; i <= PCIR_MAX_BAR_0; i++) {
380 		bar = &iov->iov_bar[i];
381 		if (bar->res != NULL) {
382 			bar_start = rman_get_start(bar->res) +
383 			    dinfo->cfg.vf.index * bar->bar_size;
384 
385 			pci_add_bar(dinfo->cfg.dev, PCIR_BAR(i), bar_start,
386 			    bar->bar_shift);
387 		}
388 	}
389 }
390 
391 static int
392 pci_iov_parse_config(struct pcicfg_iov *iov, struct pci_iov_arg *arg,
393     nvlist_t **ret)
394 {
395 	void *packed_config;
396 	nvlist_t *config;
397 	int error;
398 
399 	config = NULL;
400 	packed_config = NULL;
401 
402 	if (arg->len > pci_iov_max_config) {
403 		error = EMSGSIZE;
404 		goto out;
405 	}
406 
407 	packed_config = malloc(arg->len, M_SRIOV, M_WAITOK);
408 
409 	error = copyin(arg->config, packed_config, arg->len);
410 	if (error != 0)
411 		goto out;
412 
413 	config = nvlist_unpack(packed_config, arg->len, NV_FLAG_IGNORE_CASE);
414 	if (config == NULL) {
415 		error = EINVAL;
416 		goto out;
417 	}
418 
419 	error = pci_iov_schema_validate_config(iov->iov_schema, config);
420 	if (error != 0)
421 		goto out;
422 
423 	error = nvlist_error(config);
424 	if (error != 0)
425 		goto out;
426 
427 	*ret = config;
428 	config = NULL;
429 
430 out:
431 	nvlist_destroy(config);
432 	free(packed_config, M_SRIOV);
433 	return (error);
434 }
435 
436 /*
437  * Set the ARI_EN bit in the lowest-numbered PCI function with the SR-IOV
438  * capability.  This bit is only writeable on the lowest-numbered PF but
439  * affects all PFs on the device.
440  */
441 static int
442 pci_iov_set_ari(device_t bus)
443 {
444 	device_t lowest;
445 	device_t *devlist;
446 	int i, error, devcount, lowest_func, lowest_pos, iov_pos, dev_func;
447 	uint16_t iov_ctl;
448 
449 	/* If ARI is disabled on the downstream port there is nothing to do. */
450 	if (!PCIB_ARI_ENABLED(device_get_parent(bus)))
451 		return (0);
452 
453 	error = device_get_children(bus, &devlist, &devcount);
454 
455 	if (error != 0)
456 		return (error);
457 
458 	lowest = NULL;
459 	for (i = 0; i < devcount; i++) {
460 		if (pci_find_extcap(devlist[i], PCIZ_SRIOV, &iov_pos) == 0) {
461 			dev_func = pci_get_function(devlist[i]);
462 			if (lowest == NULL || dev_func < lowest_func) {
463 				lowest = devlist[i];
464 				lowest_func = dev_func;
465 				lowest_pos = iov_pos;
466 			}
467 		}
468 	}
469 	free(devlist, M_TEMP);
470 
471 	/*
472 	 * If we called this function some device must have the SR-IOV
473 	 * capability.
474 	 */
475 	KASSERT(lowest != NULL,
476 	    ("Could not find child of %s with SR-IOV capability",
477 	    device_get_nameunit(bus)));
478 
479 	iov_ctl = pci_read_config(lowest, lowest_pos + PCIR_SRIOV_CTL, 2);
480 	iov_ctl |= PCIM_SRIOV_ARI_EN;
481 	pci_write_config(lowest, lowest_pos + PCIR_SRIOV_CTL, iov_ctl, 2);
482 	if ((pci_read_config(lowest, lowest_pos + PCIR_SRIOV_CTL, 2) &
483 	    PCIM_SRIOV_ARI_EN) == 0) {
484 		device_printf(lowest, "failed to enable ARI\n");
485 		return (ENXIO);
486 	}
487 	return (0);
488 }
489 
490 static int
491 pci_iov_config_page_size(struct pci_devinfo *dinfo)
492 {
493 	uint32_t page_cap, page_size;
494 
495 	page_cap = IOV_READ(dinfo, PCIR_SRIOV_PAGE_CAP, 4);
496 
497 	/*
498 	 * If the system page size is less than the smallest SR-IOV page size
499 	 * then round up to the smallest SR-IOV page size.
500 	 */
501 	if (PAGE_SHIFT < PCI_SRIOV_BASE_PAGE_SHIFT)
502 		page_size = (1 << 0);
503 	else
504 		page_size = (1 << (PAGE_SHIFT - PCI_SRIOV_BASE_PAGE_SHIFT));
505 
506 	/* Check that the device supports the system page size. */
507 	if (!(page_size & page_cap))
508 		return (ENXIO);
509 
510 	IOV_WRITE(dinfo, PCIR_SRIOV_PAGE_SIZE, page_size, 4);
511 	return (0);
512 }
513 
514 static int
515 pci_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *config)
516 {
517 	const nvlist_t *device, *driver_config;
518 
519 	device = nvlist_get_nvlist(config, PF_CONFIG_NAME);
520 	driver_config = nvlist_get_nvlist(device, DRIVER_CONFIG_NAME);
521 	return (PCI_IOV_INIT(dev, num_vfs, driver_config));
522 }
523 
524 static int
525 pci_iov_init_rman(device_t pf, struct pcicfg_iov *iov)
526 {
527 	int error;
528 
529 	iov->rman.rm_start = 0;
530 	iov->rman.rm_end = ~0;
531 	iov->rman.rm_type = RMAN_ARRAY;
532 	snprintf(iov->rman_name, sizeof(iov->rman_name), "%s VF I/O memory",
533 	    device_get_nameunit(pf));
534 	iov->rman.rm_descr = iov->rman_name;
535 
536 	error = rman_init(&iov->rman);
537 	if (error != 0)
538 		return (error);
539 
540 	iov->iov_flags |= IOV_RMAN_INITED;
541 	return (0);
542 }
543 
544 static int
545 pci_iov_alloc_bar_ea(struct pci_devinfo *dinfo, int bar)
546 {
547 	struct pcicfg_iov *iov;
548 	rman_res_t start, end;
549 	struct resource *res;
550 	struct resource_list *rl;
551 	struct resource_list_entry *rle;
552 
553 	rl = &dinfo->resources;
554 	iov = dinfo->cfg.iov;
555 
556 	rle = resource_list_find(rl, SYS_RES_MEMORY,
557 	    iov->iov_pos + PCIR_SRIOV_BAR(bar));
558 	if (rle == NULL)
559 		rle = resource_list_find(rl, SYS_RES_IOPORT,
560 		    iov->iov_pos + PCIR_SRIOV_BAR(bar));
561 	if (rle == NULL)
562 		return (ENXIO);
563 	res = rle->res;
564 
565 	iov->iov_bar[bar].res = res;
566 	iov->iov_bar[bar].bar_size = rman_get_size(res) / iov->iov_num_vfs;
567 	iov->iov_bar[bar].bar_shift = pci_mapsize(iov->iov_bar[bar].bar_size);
568 
569 	start = rman_get_start(res);
570 	end = rman_get_end(res);
571 
572 	return (rman_manage_region(&iov->rman, start, end));
573 }
574 
575 static int
576 pci_iov_setup_bars(struct pci_devinfo *dinfo)
577 {
578 	device_t dev;
579 	struct pcicfg_iov *iov;
580 	pci_addr_t bar_value, testval;
581 	int i, last_64, error;
582 
583 	iov = dinfo->cfg.iov;
584 	dev = dinfo->cfg.dev;
585 	last_64 = 0;
586 
587 	pci_add_resources_ea(device_get_parent(dev), dev, 1);
588 
589 	for (i = 0; i <= PCIR_MAX_BAR_0; i++) {
590 		/* First, try to use BARs allocated with EA */
591 		error = pci_iov_alloc_bar_ea(dinfo, i);
592 		if (error == 0)
593 			continue;
594 
595 		/* Allocate legacy-BAR only if EA is not enabled */
596 		if (pci_ea_is_enabled(dev, iov->iov_pos + PCIR_SRIOV_BAR(i)))
597 			continue;
598 
599 		/*
600 		 * If a PCI BAR is a 64-bit wide BAR, then it spans two
601 		 * consecutive registers.  Therefore if the last BAR that
602 		 * we looked at was a 64-bit BAR, we need to skip this
603 		 * register as it's the second half of the last BAR.
604 		 */
605 		if (!last_64) {
606 			pci_read_bar(dev,
607 			    iov->iov_pos + PCIR_SRIOV_BAR(i),
608 			    &bar_value, &testval, &last_64);
609 
610 			if (testval != 0) {
611 				error = pci_iov_alloc_bar(dinfo, i,
612 				   pci_mapsize(testval));
613 				if (error != 0)
614 					return (error);
615 			}
616 		} else
617 			last_64 = 0;
618 	}
619 
620 	return (0);
621 }
622 
623 static void
624 pci_iov_enumerate_vfs(struct pci_devinfo *dinfo, const nvlist_t *config,
625     uint16_t first_rid, uint16_t rid_stride)
626 {
627 	char device_name[VF_MAX_NAME];
628 	const nvlist_t *device, *driver_config, *iov_config;
629 	device_t bus, dev, vf;
630 	struct pcicfg_iov *iov;
631 	struct pci_devinfo *vfinfo;
632 	int i, error;
633 	uint16_t vid, did, next_rid;
634 
635 	iov = dinfo->cfg.iov;
636 	dev = dinfo->cfg.dev;
637 	bus = device_get_parent(dev);
638 	next_rid = first_rid;
639 	vid = pci_get_vendor(dev);
640 	did = IOV_READ(dinfo, PCIR_SRIOV_VF_DID, 2);
641 
642 	for (i = 0; i < iov->iov_num_vfs; i++, next_rid += rid_stride) {
643 		snprintf(device_name, sizeof(device_name), VF_PREFIX"%d", i);
644 		device = nvlist_get_nvlist(config, device_name);
645 		iov_config = nvlist_get_nvlist(device, IOV_CONFIG_NAME);
646 		driver_config = nvlist_get_nvlist(device, DRIVER_CONFIG_NAME);
647 
648 		vf = PCI_CREATE_IOV_CHILD(bus, dev, next_rid, vid, did);
649 		if (vf == NULL)
650 			break;
651 
652 		/*
653 		 * If we are creating passthrough devices then force the ppt
654 		 * driver to attach to prevent a VF driver from claiming the
655 		 * VFs.
656 		 */
657 		if (nvlist_get_bool(iov_config, "passthrough"))
658 			device_set_devclass_fixed(vf, "ppt");
659 
660 		vfinfo = device_get_ivars(vf);
661 
662 		vfinfo->cfg.iov = iov;
663 		vfinfo->cfg.vf.index = i;
664 
665 		pci_iov_add_bars(iov, vfinfo);
666 
667 		error = PCI_IOV_ADD_VF(dev, i, driver_config);
668 		if (error != 0) {
669 			device_printf(dev, "Failed to add VF %d\n", i);
670 			device_delete_child(bus, vf);
671 		}
672 	}
673 
674 	bus_generic_attach(bus);
675 }
676 
677 static int
678 pci_iov_config(struct cdev *cdev, struct pci_iov_arg *arg)
679 {
680 	device_t bus, dev;
681 	struct pci_devinfo *dinfo;
682 	struct pcicfg_iov *iov;
683 	nvlist_t *config;
684 	int i, error;
685 	uint16_t rid_off, rid_stride;
686 	uint16_t first_rid, last_rid;
687 	uint16_t iov_ctl;
688 	uint16_t num_vfs, total_vfs;
689 	int iov_inited;
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);
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 		error = ENOSPC;
740 		goto out;
741 	}
742 
743 	iov_ctl = IOV_READ(dinfo, PCIR_SRIOV_CTL, 2);
744 	iov_ctl &= ~(PCIM_SRIOV_VF_EN | PCIM_SRIOV_VF_MSE);
745 	IOV_WRITE(dinfo, PCIR_SRIOV_CTL, iov_ctl, 2);
746 
747 	error = pci_iov_init_rman(dev, iov);
748 	if (error != 0)
749 		goto out;
750 
751 	iov->iov_num_vfs = num_vfs;
752 
753 	error = pci_iov_setup_bars(dinfo);
754 	if (error != 0)
755 		goto out;
756 
757 	iov_ctl = IOV_READ(dinfo, PCIR_SRIOV_CTL, 2);
758 	iov_ctl |= PCIM_SRIOV_VF_EN | PCIM_SRIOV_VF_MSE;
759 	IOV_WRITE(dinfo, PCIR_SRIOV_CTL, iov_ctl, 2);
760 
761 	/* Per specification, we must wait 100ms before accessing VFs. */
762 	pause("iov", roundup(hz, 10));
763 	pci_iov_enumerate_vfs(dinfo, config, first_rid, rid_stride);
764 
765 	nvlist_destroy(config);
766 	iov->iov_flags &= ~IOV_BUSY;
767 	mtx_unlock(&Giant);
768 
769 	return (0);
770 out:
771 	if (iov_inited)
772 		PCI_IOV_UNINIT(dev);
773 
774 	for (i = 0; i <= PCIR_MAX_BAR_0; i++) {
775 		if (iov->iov_bar[i].res != NULL) {
776 			pci_release_resource(bus, dev, SYS_RES_MEMORY,
777 			    iov->iov_pos + PCIR_SRIOV_BAR(i),
778 			    iov->iov_bar[i].res);
779 			pci_delete_resource(bus, dev, SYS_RES_MEMORY,
780 			    iov->iov_pos + PCIR_SRIOV_BAR(i));
781 			iov->iov_bar[i].res = NULL;
782 		}
783 	}
784 
785 	if (iov->iov_flags & IOV_RMAN_INITED) {
786 		rman_fini(&iov->rman);
787 		iov->iov_flags &= ~IOV_RMAN_INITED;
788 	}
789 
790 	nvlist_destroy(config);
791 	iov->iov_num_vfs = 0;
792 	iov->iov_flags &= ~IOV_BUSY;
793 	mtx_unlock(&Giant);
794 	return (error);
795 }
796 
797 void
798 pci_iov_cfg_restore(device_t dev, struct pci_devinfo *dinfo)
799 {
800 	struct pcicfg_iov *iov;
801 
802 	iov = dinfo->cfg.iov;
803 
804 	IOV_WRITE(dinfo, PCIR_SRIOV_PAGE_SIZE, iov->iov_page_size, 4);
805 	IOV_WRITE(dinfo, PCIR_SRIOV_NUM_VFS, iov->iov_num_vfs, 2);
806 	IOV_WRITE(dinfo, PCIR_SRIOV_CTL, iov->iov_ctl, 2);
807 }
808 
809 void
810 pci_iov_cfg_save(device_t dev, struct pci_devinfo *dinfo)
811 {
812 	struct pcicfg_iov *iov;
813 
814 	iov = dinfo->cfg.iov;
815 
816 	iov->iov_page_size = IOV_READ(dinfo, PCIR_SRIOV_PAGE_SIZE, 4);
817 	iov->iov_ctl = IOV_READ(dinfo, PCIR_SRIOV_CTL, 2);
818 }
819 
820 /* Return true if child is a VF of the given PF. */
821 static int
822 pci_iov_is_child_vf(struct pcicfg_iov *pf, device_t child)
823 {
824 	struct pci_devinfo *vfinfo;
825 
826 	vfinfo = device_get_ivars(child);
827 
828 	if (!(vfinfo->cfg.flags & PCICFG_VF))
829 		return (0);
830 
831 	return (pf == vfinfo->cfg.iov);
832 }
833 
834 static int
835 pci_iov_delete_iov_children(struct pci_devinfo *dinfo)
836 {
837 	device_t bus, dev, vf, *devlist;
838 	struct pcicfg_iov *iov;
839 	int i, error, devcount;
840 	uint32_t iov_ctl;
841 
842 	mtx_assert(&Giant, MA_OWNED);
843 
844 	iov = dinfo->cfg.iov;
845 	dev = dinfo->cfg.dev;
846 	bus = device_get_parent(dev);
847 	devlist = NULL;
848 
849 	iov->iov_flags |= IOV_BUSY;
850 
851 	error = device_get_children(bus, &devlist, &devcount);
852 
853 	if (error != 0)
854 		goto out;
855 
856 	for (i = 0; i < devcount; i++) {
857 		vf = devlist[i];
858 
859 		if (!pci_iov_is_child_vf(iov, vf))
860 			continue;
861 
862 		error = device_detach(vf);
863 		if (error != 0) {
864 			device_printf(dev,
865 			   "Could not disable SR-IOV: failed to detach VF %s\n",
866 			    device_get_nameunit(vf));
867 			goto out;
868 		}
869 	}
870 
871 	for (i = 0; i < devcount; i++) {
872 		vf = devlist[i];
873 
874 		if (pci_iov_is_child_vf(iov, vf))
875 			device_delete_child(bus, vf);
876 	}
877 	PCI_IOV_UNINIT(dev);
878 
879 	iov_ctl = IOV_READ(dinfo, PCIR_SRIOV_CTL, 2);
880 	iov_ctl &= ~(PCIM_SRIOV_VF_EN | PCIM_SRIOV_VF_MSE);
881 	IOV_WRITE(dinfo, PCIR_SRIOV_CTL, iov_ctl, 2);
882 	IOV_WRITE(dinfo, PCIR_SRIOV_NUM_VFS, 0, 2);
883 
884 	iov->iov_num_vfs = 0;
885 
886 	for (i = 0; i <= PCIR_MAX_BAR_0; i++) {
887 		if (iov->iov_bar[i].res != NULL) {
888 			pci_release_resource(bus, dev, SYS_RES_MEMORY,
889 			    iov->iov_pos + PCIR_SRIOV_BAR(i),
890 			    iov->iov_bar[i].res);
891 			pci_delete_resource(bus, dev, SYS_RES_MEMORY,
892 			    iov->iov_pos + PCIR_SRIOV_BAR(i));
893 			iov->iov_bar[i].res = NULL;
894 		}
895 	}
896 
897 	if (iov->iov_flags & IOV_RMAN_INITED) {
898 		rman_fini(&iov->rman);
899 		iov->iov_flags &= ~IOV_RMAN_INITED;
900 	}
901 
902 	error = 0;
903 out:
904 	free(devlist, M_TEMP);
905 	iov->iov_flags &= ~IOV_BUSY;
906 	return (error);
907 }
908 
909 static int
910 pci_iov_delete(struct cdev *cdev)
911 {
912 	struct pci_devinfo *dinfo;
913 	struct pcicfg_iov *iov;
914 	int error;
915 
916 	mtx_lock(&Giant);
917 	dinfo = cdev->si_drv1;
918 	iov = dinfo->cfg.iov;
919 
920 	if ((iov->iov_flags & IOV_BUSY) != 0) {
921 		error = EBUSY;
922 		goto out;
923 	}
924 	if (iov->iov_num_vfs == 0) {
925 		error = ECHILD;
926 		goto out;
927 	}
928 
929 	error = pci_iov_delete_iov_children(dinfo);
930 
931 out:
932 	mtx_unlock(&Giant);
933 	return (error);
934 }
935 
936 static int
937 pci_iov_get_schema_ioctl(struct cdev *cdev, struct pci_iov_schema *output)
938 {
939 	struct pci_devinfo *dinfo;
940 	void *packed;
941 	size_t output_len, size;
942 	int error;
943 
944 	packed = NULL;
945 
946 	mtx_lock(&Giant);
947 	dinfo = cdev->si_drv1;
948 	packed = nvlist_pack(dinfo->cfg.iov->iov_schema, &size);
949 	mtx_unlock(&Giant);
950 
951 	if (packed == NULL) {
952 		error = ENOMEM;
953 		goto fail;
954 	}
955 
956 	output_len = output->len;
957 	output->len = size;
958 	if (size <= output_len) {
959 		error = copyout(packed, output->schema, size);
960 
961 		if (error != 0)
962 			goto fail;
963 
964 		output->error = 0;
965 	} else
966 		/*
967 		 * If we return an error then the ioctl code won't copyout
968 		 * output back to userland, so we flag the error in the struct
969 		 * instead.
970 		 */
971 		output->error = EMSGSIZE;
972 
973 	error = 0;
974 
975 fail:
976 	free(packed, M_NVLIST);
977 
978 	return (error);
979 }
980 
981 static int
982 pci_iov_ioctl(struct cdev *dev, u_long cmd, caddr_t data, int fflag,
983     struct thread *td)
984 {
985 
986 	switch (cmd) {
987 	case IOV_CONFIG:
988 		return (pci_iov_config(dev, (struct pci_iov_arg *)data));
989 	case IOV_DELETE:
990 		return (pci_iov_delete(dev));
991 	case IOV_GET_SCHEMA:
992 		return (pci_iov_get_schema_ioctl(dev,
993 		    (struct pci_iov_schema *)data));
994 	default:
995 		return (EINVAL);
996 	}
997 }
998 
999 struct resource *
1000 pci_vf_alloc_mem_resource(device_t dev, device_t child, int *rid,
1001     rman_res_t start, rman_res_t end, rman_res_t count, u_int flags)
1002 {
1003 	struct pci_devinfo *dinfo;
1004 	struct pcicfg_iov *iov;
1005 	struct pci_map *map;
1006 	struct resource *res;
1007 	struct resource_list_entry *rle;
1008 	rman_res_t bar_start, bar_end;
1009 	pci_addr_t bar_length;
1010 	int error;
1011 
1012 	dinfo = device_get_ivars(child);
1013 	iov = dinfo->cfg.iov;
1014 
1015 	map = pci_find_bar(child, *rid);
1016 	if (map == NULL)
1017 		return (NULL);
1018 
1019 	bar_length = 1 << map->pm_size;
1020 	bar_start = map->pm_value;
1021 	bar_end = bar_start + bar_length - 1;
1022 
1023 	/* Make sure that the resource fits the constraints. */
1024 	if (bar_start >= end || bar_end <= bar_start || count != 1)
1025 		return (NULL);
1026 
1027 	/* Clamp the resource to the constraints if necessary. */
1028 	if (bar_start < start)
1029 		bar_start = start;
1030 	if (bar_end > end)
1031 		bar_end = end;
1032 	bar_length = bar_end - bar_start + 1;
1033 
1034 	res = rman_reserve_resource(&iov->rman, bar_start, bar_end,
1035 	    bar_length, flags, child);
1036 	if (res == NULL)
1037 		return (NULL);
1038 
1039 	rle = resource_list_add(&dinfo->resources, SYS_RES_MEMORY, *rid,
1040 	    bar_start, bar_end, 1);
1041 	if (rle == NULL) {
1042 		rman_release_resource(res);
1043 		return (NULL);
1044 	}
1045 
1046 	rman_set_rid(res, *rid);
1047 
1048 	if (flags & RF_ACTIVE) {
1049 		error = bus_activate_resource(child, SYS_RES_MEMORY, *rid, res);
1050 		if (error != 0) {
1051 			resource_list_delete(&dinfo->resources, SYS_RES_MEMORY,
1052 			    *rid);
1053 			rman_release_resource(res);
1054 			return (NULL);
1055 		}
1056 	}
1057 	rle->res = res;
1058 
1059 	return (res);
1060 }
1061 
1062 int
1063 pci_vf_release_mem_resource(device_t dev, device_t child, int rid,
1064     struct resource *r)
1065 {
1066 	struct pci_devinfo *dinfo;
1067 	struct resource_list_entry *rle;
1068 	int error;
1069 
1070 	dinfo = device_get_ivars(child);
1071 
1072 	if (rman_get_flags(r) & RF_ACTIVE) {
1073 		error = bus_deactivate_resource(child, SYS_RES_MEMORY, rid, r);
1074 		if (error != 0)
1075 			return (error);
1076 	}
1077 
1078 	rle = resource_list_find(&dinfo->resources, SYS_RES_MEMORY, rid);
1079 	if (rle != NULL) {
1080 		rle->res = NULL;
1081 		resource_list_delete(&dinfo->resources, SYS_RES_MEMORY,
1082 		    rid);
1083 	}
1084 
1085 	return (rman_release_resource(r));
1086 }
1087