iommu.c

  1// SPDX-License-Identifier: GPL-2.0-only
  2#include <dirent.h>
  3#include <fcntl.h>
  4#include <libgen.h>
  5#include <stdint.h>
  6#include <stdlib.h>
  7#include <string.h>
  8#include <unistd.h>
  9
 10#include <sys/eventfd.h>
 11#include <sys/ioctl.h>
 12#include <sys/mman.h>
 13
 14#include <linux/limits.h>
 15#include <linux/mman.h>
 16#include <linux/types.h>
 17#include <linux/vfio.h>
 18#include <linux/iommufd.h>
 19
 20#include "../../../kselftest.h"
 21#include <libvfio.h>
 22
 23const char *default_iommu_mode = "iommufd";
 24
 25/* Reminder: Keep in sync with FIXTURE_VARIANT_ADD_ALL_IOMMU_MODES(). */
 26static const struct iommu_mode iommu_modes[] = {
 27	{
 28		.name = "vfio_type1_iommu",
 29		.container_path = "/dev/vfio/vfio",
 30		.iommu_type = VFIO_TYPE1_IOMMU,
 31	},
 32	{
 33		.name = "vfio_type1v2_iommu",
 34		.container_path = "/dev/vfio/vfio",
 35		.iommu_type = VFIO_TYPE1v2_IOMMU,
 36	},
 37	{
 38		.name = "iommufd_compat_type1",
 39		.container_path = "/dev/iommu",
 40		.iommu_type = VFIO_TYPE1_IOMMU,
 41	},
 42	{
 43		.name = "iommufd_compat_type1v2",
 44		.container_path = "/dev/iommu",
 45		.iommu_type = VFIO_TYPE1v2_IOMMU,
 46	},
 47	{
 48		.name = "iommufd",
 49	},
 50};
 51
 52static const struct iommu_mode *lookup_iommu_mode(const char *iommu_mode)
 53{
 54	int i;
 55
 56	if (!iommu_mode)
 57		iommu_mode = default_iommu_mode;
 58
 59	for (i = 0; i < ARRAY_SIZE(iommu_modes); i++) {
 60		if (strcmp(iommu_mode, iommu_modes[i].name))
 61			continue;
 62
 63		return &iommu_modes[i];
 64	}
 65
 66	VFIO_FAIL("Unrecognized IOMMU mode: %s\n", iommu_mode);
 67}
 68
 69int __iommu_hva2iova(struct iommu *iommu, void *vaddr, iova_t *iova)
 70{
 71	struct dma_region *region;
 72
 73	list_for_each_entry(region, &iommu->dma_regions, link) {
 74		if (vaddr < region->vaddr)
 75			continue;
 76
 77		if (vaddr >= region->vaddr + region->size)
 78			continue;
 79
 80		if (iova)
 81			*iova = region->iova + (vaddr - region->vaddr);
 82
 83		return 0;
 84	}
 85
 86	return -ENOENT;
 87}
 88
 89iova_t iommu_hva2iova(struct iommu *iommu, void *vaddr)
 90{
 91	iova_t iova;
 92	int ret;
 93
 94	ret = __iommu_hva2iova(iommu, vaddr, &iova);
 95	VFIO_ASSERT_EQ(ret, 0, "%p is not mapped into the iommu\n", vaddr);
 96
 97	return iova;
 98}
 99
100static int vfio_iommu_map(struct iommu *iommu, struct dma_region *region)
101{
102	struct vfio_iommu_type1_dma_map args = {
103		.argsz = sizeof(args),
104		.flags = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE,
105		.vaddr = (u64)region->vaddr,
106		.iova = region->iova,
107		.size = region->size,
108	};
109
110	if (ioctl(iommu->container_fd, VFIO_IOMMU_MAP_DMA, &args))
111		return -errno;
112
113	return 0;
114}
115
116static int iommufd_map(struct iommu *iommu, struct dma_region *region)
117{
118	struct iommu_ioas_map args = {
119		.size = sizeof(args),
120		.flags = IOMMU_IOAS_MAP_READABLE |
121			 IOMMU_IOAS_MAP_WRITEABLE |
122			 IOMMU_IOAS_MAP_FIXED_IOVA,
123		.user_va = (u64)region->vaddr,
124		.iova = region->iova,
125		.length = region->size,
126		.ioas_id = iommu->ioas_id,
127	};
128
129	if (ioctl(iommu->iommufd, IOMMU_IOAS_MAP, &args))
130		return -errno;
131
132	return 0;
133}
134
135int __iommu_map(struct iommu *iommu, struct dma_region *region)
136{
137	int ret;
138
139	if (iommu->iommufd)
140		ret = iommufd_map(iommu, region);
141	else
142		ret = vfio_iommu_map(iommu, region);
143
144	if (ret)
145		return ret;
146
147	list_add(&region->link, &iommu->dma_regions);
148
149	return 0;
150}
151
152static int __vfio_iommu_unmap(int fd, u64 iova, u64 size, u32 flags, u64 *unmapped)
153{
154	struct vfio_iommu_type1_dma_unmap args = {
155		.argsz = sizeof(args),
156		.iova = iova,
157		.size = size,
158		.flags = flags,
159	};
160
161	if (ioctl(fd, VFIO_IOMMU_UNMAP_DMA, &args))
162		return -errno;
163
164	if (unmapped)
165		*unmapped = args.size;
166
167	return 0;
168}
169
170static int vfio_iommu_unmap(struct iommu *iommu, struct dma_region *region,
171			    u64 *unmapped)
172{
173	return __vfio_iommu_unmap(iommu->container_fd, region->iova,
174				  region->size, 0, unmapped);
175}
176
177static int __iommufd_unmap(int fd, u64 iova, u64 length, u32 ioas_id, u64 *unmapped)
178{
179	struct iommu_ioas_unmap args = {
180		.size = sizeof(args),
181		.iova = iova,
182		.length = length,
183		.ioas_id = ioas_id,
184	};
185
186	if (ioctl(fd, IOMMU_IOAS_UNMAP, &args))
187		return -errno;
188
189	if (unmapped)
190		*unmapped = args.length;
191
192	return 0;
193}
194
195static int iommufd_unmap(struct iommu *iommu, struct dma_region *region,
196			 u64 *unmapped)
197{
198	return __iommufd_unmap(iommu->iommufd, region->iova, region->size,
199			       iommu->ioas_id, unmapped);
200}
201
202int __iommu_unmap(struct iommu *iommu, struct dma_region *region, u64 *unmapped)
203{
204	int ret;
205
206	if (iommu->iommufd)
207		ret = iommufd_unmap(iommu, region, unmapped);
208	else
209		ret = vfio_iommu_unmap(iommu, region, unmapped);
210
211	if (ret)
212		return ret;
213
214	list_del_init(&region->link);
215
216	return 0;
217}
218
219int __iommu_unmap_all(struct iommu *iommu, u64 *unmapped)
220{
221	int ret;
222	struct dma_region *curr, *next;
223
224	if (iommu->iommufd)
225		ret = __iommufd_unmap(iommu->iommufd, 0, UINT64_MAX,
226				      iommu->ioas_id, unmapped);
227	else
228		ret = __vfio_iommu_unmap(iommu->container_fd, 0, 0,
229					 VFIO_DMA_UNMAP_FLAG_ALL, unmapped);
230
231	if (ret)
232		return ret;
233
234	list_for_each_entry_safe(curr, next, &iommu->dma_regions, link)
235		list_del_init(&curr->link);
236
237	return 0;
238}
239
240static struct vfio_info_cap_header *next_cap_hdr(void *buf, u32 bufsz,
241						 u32 *cap_offset)
242{
243	struct vfio_info_cap_header *hdr;
244
245	if (!*cap_offset)
246		return NULL;
247
248	VFIO_ASSERT_LT(*cap_offset, bufsz);
249	VFIO_ASSERT_GE(bufsz - *cap_offset, sizeof(*hdr));
250
251	hdr = (struct vfio_info_cap_header *)((u8 *)buf + *cap_offset);
252	*cap_offset = hdr->next;
253
254	return hdr;
255}
256
257static struct vfio_info_cap_header *vfio_iommu_info_cap_hdr(struct vfio_iommu_type1_info *info,
258							    u16 cap_id)
259{
260	struct vfio_info_cap_header *hdr;
261	u32 cap_offset = info->cap_offset;
262	u32 max_depth;
263	u32 depth = 0;
264
265	if (!(info->flags & VFIO_IOMMU_INFO_CAPS))
266		return NULL;
267
268	if (cap_offset)
269		VFIO_ASSERT_GE(cap_offset, sizeof(*info));
270
271	max_depth = (info->argsz - sizeof(*info)) / sizeof(*hdr);
272
273	while ((hdr = next_cap_hdr(info, info->argsz, &cap_offset))) {
274		depth++;
275		VFIO_ASSERT_LE(depth, max_depth, "Capability chain contains a cycle\n");
276
277		if (hdr->id == cap_id)
278			return hdr;
279	}
280
281	return NULL;
282}
283
284/* Return buffer including capability chain, if present. Free with free() */
285static struct vfio_iommu_type1_info *vfio_iommu_get_info(int container_fd)
286{
287	struct vfio_iommu_type1_info *info;
288
289	info = malloc(sizeof(*info));
290	VFIO_ASSERT_NOT_NULL(info);
291
292	*info = (struct vfio_iommu_type1_info) {
293		.argsz = sizeof(*info),
294	};
295
296	ioctl_assert(container_fd, VFIO_IOMMU_GET_INFO, info);
297	VFIO_ASSERT_GE(info->argsz, sizeof(*info));
298
299	info = realloc(info, info->argsz);
300	VFIO_ASSERT_NOT_NULL(info);
301
302	ioctl_assert(container_fd, VFIO_IOMMU_GET_INFO, info);
303	VFIO_ASSERT_GE(info->argsz, sizeof(*info));
304
305	return info;
306}
307
308/*
309 * Return iova ranges for the device's container. Normalize vfio_iommu_type1 to
310 * report iommufd's iommu_iova_range. Free with free().
311 */
312static struct iommu_iova_range *vfio_iommu_iova_ranges(struct iommu *iommu,
313						       u32 *nranges)
314{
315	struct vfio_iommu_type1_info_cap_iova_range *cap_range;
316	struct vfio_iommu_type1_info *info;
317	struct vfio_info_cap_header *hdr;
318	struct iommu_iova_range *ranges = NULL;
319
320	info = vfio_iommu_get_info(iommu->container_fd);
321	hdr = vfio_iommu_info_cap_hdr(info, VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE);
322	VFIO_ASSERT_NOT_NULL(hdr);
323
324	cap_range = container_of(hdr, struct vfio_iommu_type1_info_cap_iova_range, header);
325	VFIO_ASSERT_GT(cap_range->nr_iovas, 0);
326
327	ranges = calloc(cap_range->nr_iovas, sizeof(*ranges));
328	VFIO_ASSERT_NOT_NULL(ranges);
329
330	for (u32 i = 0; i < cap_range->nr_iovas; i++) {
331		ranges[i] = (struct iommu_iova_range){
332			.start = cap_range->iova_ranges[i].start,
333			.last = cap_range->iova_ranges[i].end,
334		};
335	}
336
337	*nranges = cap_range->nr_iovas;
338
339	free(info);
340	return ranges;
341}
342
343/* Return iova ranges of the device's IOAS. Free with free() */
344static struct iommu_iova_range *iommufd_iova_ranges(struct iommu *iommu,
345						    u32 *nranges)
346{
347	struct iommu_iova_range *ranges;
348	int ret;
349
350	struct iommu_ioas_iova_ranges query = {
351		.size = sizeof(query),
352		.ioas_id = iommu->ioas_id,
353	};
354
355	ret = ioctl(iommu->iommufd, IOMMU_IOAS_IOVA_RANGES, &query);
356	VFIO_ASSERT_EQ(ret, -1);
357	VFIO_ASSERT_EQ(errno, EMSGSIZE);
358	VFIO_ASSERT_GT(query.num_iovas, 0);
359
360	ranges = calloc(query.num_iovas, sizeof(*ranges));
361	VFIO_ASSERT_NOT_NULL(ranges);
362
363	query.allowed_iovas = (uintptr_t)ranges;
364
365	ioctl_assert(iommu->iommufd, IOMMU_IOAS_IOVA_RANGES, &query);
366	*nranges = query.num_iovas;
367
368	return ranges;
369}
370
371static int iova_range_comp(const void *a, const void *b)
372{
373	const struct iommu_iova_range *ra = a, *rb = b;
374
375	if (ra->start < rb->start)
376		return -1;
377
378	if (ra->start > rb->start)
379		return 1;
380
381	return 0;
382}
383
384/* Return sorted IOVA ranges of the device. Free with free(). */
385struct iommu_iova_range *iommu_iova_ranges(struct iommu *iommu, u32 *nranges)
386{
387	struct iommu_iova_range *ranges;
388
389	if (iommu->iommufd)
390		ranges = iommufd_iova_ranges(iommu, nranges);
391	else
392		ranges = vfio_iommu_iova_ranges(iommu, nranges);
393
394	if (!ranges)
395		return NULL;
396
397	VFIO_ASSERT_GT(*nranges, 0);
398
399	/* Sort and check that ranges are sane and non-overlapping */
400	qsort(ranges, *nranges, sizeof(*ranges), iova_range_comp);
401	VFIO_ASSERT_LT(ranges[0].start, ranges[0].last);
402
403	for (u32 i = 1; i < *nranges; i++) {
404		VFIO_ASSERT_LT(ranges[i].start, ranges[i].last);
405		VFIO_ASSERT_LT(ranges[i - 1].last, ranges[i].start);
406	}
407
408	return ranges;
409}
410
411static u32 iommufd_ioas_alloc(int iommufd)
412{
413	struct iommu_ioas_alloc args = {
414		.size = sizeof(args),
415	};
416
417	ioctl_assert(iommufd, IOMMU_IOAS_ALLOC, &args);
418	return args.out_ioas_id;
419}
420
421struct iommu *iommu_init(const char *iommu_mode)
422{
423	const char *container_path;
424	struct iommu *iommu;
425	int version;
426
427	iommu = calloc(1, sizeof(*iommu));
428	VFIO_ASSERT_NOT_NULL(iommu);
429
430	INIT_LIST_HEAD(&iommu->dma_regions);
431
432	iommu->mode = lookup_iommu_mode(iommu_mode);
433
434	container_path = iommu->mode->container_path;
435	if (container_path) {
436		iommu->container_fd = open(container_path, O_RDWR);
437		VFIO_ASSERT_GE(iommu->container_fd, 0, "open(%s) failed\n", container_path);
438
439		version = ioctl(iommu->container_fd, VFIO_GET_API_VERSION);
440		VFIO_ASSERT_EQ(version, VFIO_API_VERSION, "Unsupported version: %d\n", version);
441	} else {
442		/*
443		 * Require device->iommufd to be >0 so that a simple non-0 check can be
444		 * used to check if iommufd is enabled. In practice open() will never
445		 * return 0 unless stdin is closed.
446		 */
447		iommu->iommufd = open("/dev/iommu", O_RDWR);
448		VFIO_ASSERT_GT(iommu->iommufd, 0);
449
450		iommu->ioas_id = iommufd_ioas_alloc(iommu->iommufd);
451	}
452
453	return iommu;
454}
455
456void iommu_cleanup(struct iommu *iommu)
457{
458	if (iommu->iommufd)
459		VFIO_ASSERT_EQ(close(iommu->iommufd), 0);
460	else
461		VFIO_ASSERT_EQ(close(iommu->container_fd), 0);
462
463	free(iommu);
464}