Commit | Line | Data |
---|---|---|
5ffd229c AK |
1 | /* |
2 | * VFIO: IOMMU DMA mapping support for TCE on POWER | |
3 | * | |
4 | * Copyright (C) 2013 IBM Corp. All rights reserved. | |
5 | * Author: Alexey Kardashevskiy <aik@ozlabs.ru> | |
6 | * | |
7 | * This program is free software; you can redistribute it and/or modify | |
8 | * it under the terms of the GNU General Public License version 2 as | |
9 | * published by the Free Software Foundation. | |
10 | * | |
11 | * Derived from original vfio_iommu_type1.c: | |
12 | * Copyright (C) 2012 Red Hat, Inc. All rights reserved. | |
13 | * Author: Alex Williamson <alex.williamson@redhat.com> | |
14 | */ | |
15 | ||
16 | #include <linux/module.h> | |
17 | #include <linux/pci.h> | |
18 | #include <linux/slab.h> | |
19 | #include <linux/uaccess.h> | |
20 | #include <linux/err.h> | |
21 | #include <linux/vfio.h> | |
22 | #include <asm/iommu.h> | |
23 | #include <asm/tce.h> | |
24 | ||
25 | #define DRIVER_VERSION "0.1" | |
26 | #define DRIVER_AUTHOR "aik@ozlabs.ru" | |
27 | #define DRIVER_DESC "VFIO IOMMU SPAPR TCE" | |
28 | ||
29 | static void tce_iommu_detach_group(void *iommu_data, | |
30 | struct iommu_group *iommu_group); | |
31 | ||
32 | /* | |
33 | * VFIO IOMMU fd for SPAPR_TCE IOMMU implementation | |
34 | * | |
35 | * This code handles mapping and unmapping of user data buffers | |
36 | * into DMA'ble space using the IOMMU | |
37 | */ | |
38 | ||
39 | /* | |
40 | * The container descriptor supports only a single group per container. | |
41 | * Required by the API as the container is not supplied with the IOMMU group | |
42 | * at the moment of initialization. | |
43 | */ | |
44 | struct tce_container { | |
45 | struct mutex lock; | |
46 | struct iommu_table *tbl; | |
47 | bool enabled; | |
48 | }; | |
49 | ||
50 | static int tce_iommu_enable(struct tce_container *container) | |
51 | { | |
52 | int ret = 0; | |
53 | unsigned long locked, lock_limit, npages; | |
54 | struct iommu_table *tbl = container->tbl; | |
55 | ||
56 | if (!container->tbl) | |
57 | return -ENXIO; | |
58 | ||
59 | if (!current->mm) | |
60 | return -ESRCH; /* process exited */ | |
61 | ||
62 | if (container->enabled) | |
63 | return -EBUSY; | |
64 | ||
65 | /* | |
66 | * When userspace pages are mapped into the IOMMU, they are effectively | |
67 | * locked memory, so, theoretically, we need to update the accounting | |
68 | * of locked pages on each map and unmap. For powerpc, the map unmap | |
69 | * paths can be very hot, though, and the accounting would kill | |
70 | * performance, especially since it would be difficult to impossible | |
71 | * to handle the accounting in real mode only. | |
72 | * | |
73 | * To address that, rather than precisely accounting every page, we | |
74 | * instead account for a worst case on locked memory when the iommu is | |
75 | * enabled and disabled. The worst case upper bound on locked memory | |
76 | * is the size of the whole iommu window, which is usually relatively | |
77 | * small (compared to total memory sizes) on POWER hardware. | |
78 | * | |
79 | * Also we don't have a nice way to fail on H_PUT_TCE due to ulimits, | |
80 | * that would effectively kill the guest at random points, much better | |
81 | * enforcing the limit based on the max that the guest can map. | |
82 | */ | |
83 | down_write(¤t->mm->mmap_sem); | |
e589a440 | 84 | npages = (tbl->it_size << IOMMU_PAGE_SHIFT_4K) >> PAGE_SHIFT; |
5ffd229c AK |
85 | locked = current->mm->locked_vm + npages; |
86 | lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
87 | if (locked > lock_limit && !capable(CAP_IPC_LOCK)) { | |
88 | pr_warn("RLIMIT_MEMLOCK (%ld) exceeded\n", | |
89 | rlimit(RLIMIT_MEMLOCK)); | |
90 | ret = -ENOMEM; | |
91 | } else { | |
92 | ||
93 | current->mm->locked_vm += npages; | |
94 | container->enabled = true; | |
95 | } | |
96 | up_write(¤t->mm->mmap_sem); | |
97 | ||
98 | return ret; | |
99 | } | |
100 | ||
101 | static void tce_iommu_disable(struct tce_container *container) | |
102 | { | |
103 | if (!container->enabled) | |
104 | return; | |
105 | ||
106 | container->enabled = false; | |
107 | ||
108 | if (!container->tbl || !current->mm) | |
109 | return; | |
110 | ||
111 | down_write(¤t->mm->mmap_sem); | |
112 | current->mm->locked_vm -= (container->tbl->it_size << | |
e589a440 | 113 | IOMMU_PAGE_SHIFT_4K) >> PAGE_SHIFT; |
5ffd229c AK |
114 | up_write(¤t->mm->mmap_sem); |
115 | } | |
116 | ||
117 | static void *tce_iommu_open(unsigned long arg) | |
118 | { | |
119 | struct tce_container *container; | |
120 | ||
121 | if (arg != VFIO_SPAPR_TCE_IOMMU) { | |
122 | pr_err("tce_vfio: Wrong IOMMU type\n"); | |
123 | return ERR_PTR(-EINVAL); | |
124 | } | |
125 | ||
126 | container = kzalloc(sizeof(*container), GFP_KERNEL); | |
127 | if (!container) | |
128 | return ERR_PTR(-ENOMEM); | |
129 | ||
130 | mutex_init(&container->lock); | |
131 | ||
132 | return container; | |
133 | } | |
134 | ||
135 | static void tce_iommu_release(void *iommu_data) | |
136 | { | |
137 | struct tce_container *container = iommu_data; | |
138 | ||
139 | WARN_ON(container->tbl && !container->tbl->it_group); | |
140 | tce_iommu_disable(container); | |
141 | ||
142 | if (container->tbl && container->tbl->it_group) | |
143 | tce_iommu_detach_group(iommu_data, container->tbl->it_group); | |
144 | ||
145 | mutex_destroy(&container->lock); | |
146 | ||
147 | kfree(container); | |
148 | } | |
149 | ||
9b14a1ff AK |
150 | static int tce_iommu_clear(struct tce_container *container, |
151 | struct iommu_table *tbl, | |
152 | unsigned long entry, unsigned long pages) | |
153 | { | |
154 | unsigned long oldtce; | |
155 | struct page *page; | |
156 | ||
157 | for ( ; pages; --pages, ++entry) { | |
158 | oldtce = iommu_clear_tce(tbl, entry); | |
159 | if (!oldtce) | |
160 | continue; | |
161 | ||
162 | page = pfn_to_page(oldtce >> PAGE_SHIFT); | |
163 | WARN_ON(!page); | |
164 | if (page) { | |
165 | if (oldtce & TCE_PCI_WRITE) | |
166 | SetPageDirty(page); | |
167 | put_page(page); | |
168 | } | |
169 | } | |
170 | ||
171 | return 0; | |
172 | } | |
173 | ||
174 | static long tce_iommu_build(struct tce_container *container, | |
175 | struct iommu_table *tbl, | |
176 | unsigned long entry, unsigned long tce, unsigned long pages) | |
177 | { | |
178 | long i, ret = 0; | |
179 | struct page *page = NULL; | |
180 | unsigned long hva; | |
181 | enum dma_data_direction direction = iommu_tce_direction(tce); | |
182 | ||
183 | for (i = 0; i < pages; ++i) { | |
184 | unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK; | |
185 | ||
186 | ret = get_user_pages_fast(tce & PAGE_MASK, 1, | |
187 | direction != DMA_TO_DEVICE, &page); | |
188 | if (unlikely(ret != 1)) { | |
189 | ret = -EFAULT; | |
190 | break; | |
191 | } | |
192 | hva = (unsigned long) page_address(page) + offset; | |
193 | ||
194 | ret = iommu_tce_build(tbl, entry + i, hva, direction); | |
195 | if (ret) { | |
196 | put_page(page); | |
197 | pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n", | |
198 | __func__, entry << tbl->it_page_shift, | |
199 | tce, ret); | |
200 | break; | |
201 | } | |
202 | tce += IOMMU_PAGE_SIZE_4K; | |
203 | } | |
204 | ||
205 | if (ret) | |
206 | tce_iommu_clear(container, tbl, entry, i); | |
207 | ||
208 | return ret; | |
209 | } | |
210 | ||
5ffd229c AK |
211 | static long tce_iommu_ioctl(void *iommu_data, |
212 | unsigned int cmd, unsigned long arg) | |
213 | { | |
214 | struct tce_container *container = iommu_data; | |
215 | unsigned long minsz; | |
216 | long ret; | |
217 | ||
218 | switch (cmd) { | |
219 | case VFIO_CHECK_EXTENSION: | |
1b69be5e GS |
220 | switch (arg) { |
221 | case VFIO_SPAPR_TCE_IOMMU: | |
222 | ret = 1; | |
223 | break; | |
224 | default: | |
225 | ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg); | |
226 | break; | |
227 | } | |
228 | ||
229 | return (ret < 0) ? 0 : ret; | |
5ffd229c AK |
230 | |
231 | case VFIO_IOMMU_SPAPR_TCE_GET_INFO: { | |
232 | struct vfio_iommu_spapr_tce_info info; | |
233 | struct iommu_table *tbl = container->tbl; | |
234 | ||
235 | if (WARN_ON(!tbl)) | |
236 | return -ENXIO; | |
237 | ||
238 | minsz = offsetofend(struct vfio_iommu_spapr_tce_info, | |
239 | dma32_window_size); | |
240 | ||
241 | if (copy_from_user(&info, (void __user *)arg, minsz)) | |
242 | return -EFAULT; | |
243 | ||
244 | if (info.argsz < minsz) | |
245 | return -EINVAL; | |
246 | ||
e589a440 AP |
247 | info.dma32_window_start = tbl->it_offset << IOMMU_PAGE_SHIFT_4K; |
248 | info.dma32_window_size = tbl->it_size << IOMMU_PAGE_SHIFT_4K; | |
5ffd229c AK |
249 | info.flags = 0; |
250 | ||
251 | if (copy_to_user((void __user *)arg, &info, minsz)) | |
252 | return -EFAULT; | |
253 | ||
254 | return 0; | |
255 | } | |
256 | case VFIO_IOMMU_MAP_DMA: { | |
257 | struct vfio_iommu_type1_dma_map param; | |
258 | struct iommu_table *tbl = container->tbl; | |
9b14a1ff | 259 | unsigned long tce; |
5ffd229c AK |
260 | |
261 | if (!tbl) | |
262 | return -ENXIO; | |
263 | ||
264 | BUG_ON(!tbl->it_group); | |
265 | ||
266 | minsz = offsetofend(struct vfio_iommu_type1_dma_map, size); | |
267 | ||
268 | if (copy_from_user(¶m, (void __user *)arg, minsz)) | |
269 | return -EFAULT; | |
270 | ||
271 | if (param.argsz < minsz) | |
272 | return -EINVAL; | |
273 | ||
274 | if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ | | |
275 | VFIO_DMA_MAP_FLAG_WRITE)) | |
276 | return -EINVAL; | |
277 | ||
e589a440 AP |
278 | if ((param.size & ~IOMMU_PAGE_MASK_4K) || |
279 | (param.vaddr & ~IOMMU_PAGE_MASK_4K)) | |
5ffd229c AK |
280 | return -EINVAL; |
281 | ||
282 | /* iova is checked by the IOMMU API */ | |
283 | tce = param.vaddr; | |
284 | if (param.flags & VFIO_DMA_MAP_FLAG_READ) | |
285 | tce |= TCE_PCI_READ; | |
286 | if (param.flags & VFIO_DMA_MAP_FLAG_WRITE) | |
287 | tce |= TCE_PCI_WRITE; | |
288 | ||
289 | ret = iommu_tce_put_param_check(tbl, param.iova, tce); | |
290 | if (ret) | |
291 | return ret; | |
292 | ||
9b14a1ff AK |
293 | ret = tce_iommu_build(container, tbl, |
294 | param.iova >> IOMMU_PAGE_SHIFT_4K, | |
295 | tce, param.size >> IOMMU_PAGE_SHIFT_4K); | |
5ffd229c AK |
296 | |
297 | iommu_flush_tce(tbl); | |
298 | ||
299 | return ret; | |
300 | } | |
301 | case VFIO_IOMMU_UNMAP_DMA: { | |
302 | struct vfio_iommu_type1_dma_unmap param; | |
303 | struct iommu_table *tbl = container->tbl; | |
304 | ||
305 | if (WARN_ON(!tbl)) | |
306 | return -ENXIO; | |
307 | ||
308 | minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, | |
309 | size); | |
310 | ||
311 | if (copy_from_user(¶m, (void __user *)arg, minsz)) | |
312 | return -EFAULT; | |
313 | ||
314 | if (param.argsz < minsz) | |
315 | return -EINVAL; | |
316 | ||
317 | /* No flag is supported now */ | |
318 | if (param.flags) | |
319 | return -EINVAL; | |
320 | ||
e589a440 | 321 | if (param.size & ~IOMMU_PAGE_MASK_4K) |
5ffd229c AK |
322 | return -EINVAL; |
323 | ||
324 | ret = iommu_tce_clear_param_check(tbl, param.iova, 0, | |
e589a440 | 325 | param.size >> IOMMU_PAGE_SHIFT_4K); |
5ffd229c AK |
326 | if (ret) |
327 | return ret; | |
328 | ||
9b14a1ff | 329 | ret = tce_iommu_clear(container, tbl, |
e589a440 AP |
330 | param.iova >> IOMMU_PAGE_SHIFT_4K, |
331 | param.size >> IOMMU_PAGE_SHIFT_4K); | |
5ffd229c AK |
332 | iommu_flush_tce(tbl); |
333 | ||
334 | return ret; | |
335 | } | |
336 | case VFIO_IOMMU_ENABLE: | |
337 | mutex_lock(&container->lock); | |
338 | ret = tce_iommu_enable(container); | |
339 | mutex_unlock(&container->lock); | |
340 | return ret; | |
341 | ||
342 | ||
343 | case VFIO_IOMMU_DISABLE: | |
344 | mutex_lock(&container->lock); | |
345 | tce_iommu_disable(container); | |
346 | mutex_unlock(&container->lock); | |
347 | return 0; | |
1b69be5e GS |
348 | case VFIO_EEH_PE_OP: |
349 | if (!container->tbl || !container->tbl->it_group) | |
350 | return -ENODEV; | |
351 | ||
352 | return vfio_spapr_iommu_eeh_ioctl(container->tbl->it_group, | |
353 | cmd, arg); | |
5ffd229c AK |
354 | } |
355 | ||
356 | return -ENOTTY; | |
357 | } | |
358 | ||
359 | static int tce_iommu_attach_group(void *iommu_data, | |
360 | struct iommu_group *iommu_group) | |
361 | { | |
362 | int ret; | |
363 | struct tce_container *container = iommu_data; | |
364 | struct iommu_table *tbl = iommu_group_get_iommudata(iommu_group); | |
365 | ||
366 | BUG_ON(!tbl); | |
367 | mutex_lock(&container->lock); | |
368 | ||
369 | /* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n", | |
370 | iommu_group_id(iommu_group), iommu_group); */ | |
371 | if (container->tbl) { | |
372 | pr_warn("tce_vfio: Only one group per IOMMU container is allowed, existing id=%d, attaching id=%d\n", | |
373 | iommu_group_id(container->tbl->it_group), | |
374 | iommu_group_id(iommu_group)); | |
375 | ret = -EBUSY; | |
376 | } else if (container->enabled) { | |
377 | pr_err("tce_vfio: attaching group #%u to enabled container\n", | |
378 | iommu_group_id(iommu_group)); | |
379 | ret = -EBUSY; | |
380 | } else { | |
381 | ret = iommu_take_ownership(tbl); | |
382 | if (!ret) | |
383 | container->tbl = tbl; | |
384 | } | |
385 | ||
386 | mutex_unlock(&container->lock); | |
387 | ||
388 | return ret; | |
389 | } | |
390 | ||
391 | static void tce_iommu_detach_group(void *iommu_data, | |
392 | struct iommu_group *iommu_group) | |
393 | { | |
394 | struct tce_container *container = iommu_data; | |
395 | struct iommu_table *tbl = iommu_group_get_iommudata(iommu_group); | |
396 | ||
397 | BUG_ON(!tbl); | |
398 | mutex_lock(&container->lock); | |
399 | if (tbl != container->tbl) { | |
400 | pr_warn("tce_vfio: detaching group #%u, expected group is #%u\n", | |
401 | iommu_group_id(iommu_group), | |
402 | iommu_group_id(tbl->it_group)); | |
403 | } else { | |
404 | if (container->enabled) { | |
405 | pr_warn("tce_vfio: detaching group #%u from enabled container, forcing disable\n", | |
406 | iommu_group_id(tbl->it_group)); | |
407 | tce_iommu_disable(container); | |
408 | } | |
409 | ||
410 | /* pr_debug("tce_vfio: detaching group #%u from iommu %p\n", | |
411 | iommu_group_id(iommu_group), iommu_group); */ | |
412 | container->tbl = NULL; | |
9b14a1ff | 413 | tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size); |
5ffd229c AK |
414 | iommu_release_ownership(tbl); |
415 | } | |
416 | mutex_unlock(&container->lock); | |
417 | } | |
418 | ||
419 | const struct vfio_iommu_driver_ops tce_iommu_driver_ops = { | |
420 | .name = "iommu-vfio-powerpc", | |
421 | .owner = THIS_MODULE, | |
422 | .open = tce_iommu_open, | |
423 | .release = tce_iommu_release, | |
424 | .ioctl = tce_iommu_ioctl, | |
425 | .attach_group = tce_iommu_attach_group, | |
426 | .detach_group = tce_iommu_detach_group, | |
427 | }; | |
428 | ||
429 | static int __init tce_iommu_init(void) | |
430 | { | |
431 | return vfio_register_iommu_driver(&tce_iommu_driver_ops); | |
432 | } | |
433 | ||
434 | static void __exit tce_iommu_cleanup(void) | |
435 | { | |
436 | vfio_unregister_iommu_driver(&tce_iommu_driver_ops); | |
437 | } | |
438 | ||
439 | module_init(tce_iommu_init); | |
440 | module_exit(tce_iommu_cleanup); | |
441 | ||
442 | MODULE_VERSION(DRIVER_VERSION); | |
443 | MODULE_LICENSE("GPL v2"); | |
444 | MODULE_AUTHOR(DRIVER_AUTHOR); | |
445 | MODULE_DESCRIPTION(DRIVER_DESC); | |
446 |