1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* |
3 | * Copyright (c) 2022, Oracle and/or its affiliates. |
4 | * Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved |
5 | */ |
6 | #include <linux/iova_bitmap.h> |
7 | #include <linux/mm.h> |
8 | #include <linux/slab.h> |
9 | #include <linux/highmem.h> |
10 | |
11 | #define BITS_PER_PAGE (PAGE_SIZE * BITS_PER_BYTE) |
12 | |
13 | /* |
14 | * struct iova_bitmap_map - A bitmap representing an IOVA range |
15 | * |
16 | * Main data structure for tracking mapped user pages of bitmap data. |
17 | * |
18 | * For example, for something recording dirty IOVAs, it will be provided a |
19 | * struct iova_bitmap structure, as a general structure for iterating the |
20 | * total IOVA range. The struct iova_bitmap_map, though, represents the |
21 | * subset of said IOVA space that is pinned by its parent structure (struct |
22 | * iova_bitmap). |
23 | * |
24 | * The user does not need to exact location of the bits in the bitmap. |
25 | * From user perspective the only API available is iova_bitmap_set() which |
26 | * records the IOVA *range* in the bitmap by setting the corresponding |
27 | * bits. |
28 | * |
29 | * The bitmap is an array of u64 whereas each bit represents an IOVA of |
30 | * range of (1 << pgshift). Thus formula for the bitmap data to be set is: |
31 | * |
32 | * data[(iova / page_size) / 64] & (1ULL << (iova % 64)) |
33 | */ |
34 | struct iova_bitmap_map { |
35 | /* base IOVA representing bit 0 of the first page */ |
36 | unsigned long iova; |
37 | |
38 | /* page size order that each bit granules to */ |
39 | unsigned long pgshift; |
40 | |
41 | /* page offset of the first user page pinned */ |
42 | unsigned long pgoff; |
43 | |
44 | /* number of pages pinned */ |
45 | unsigned long npages; |
46 | |
47 | /* pinned pages representing the bitmap data */ |
48 | struct page **pages; |
49 | }; |
50 | |
51 | /* |
52 | * struct iova_bitmap - The IOVA bitmap object |
53 | * |
54 | * Main data structure for iterating over the bitmap data. |
55 | * |
56 | * Abstracts the pinning work and iterates in IOVA ranges. |
57 | * It uses a windowing scheme and pins the bitmap in relatively |
58 | * big ranges e.g. |
59 | * |
60 | * The bitmap object uses one base page to store all the pinned pages |
61 | * pointers related to the bitmap. For sizeof(struct page*) == 8 it stores |
62 | * 512 struct page pointers which, if the base page size is 4K, it means |
63 | * 2M of bitmap data is pinned at a time. If the iova_bitmap page size is |
64 | * also 4K then the range window to iterate is 64G. |
65 | * |
66 | * For example iterating on a total IOVA range of 4G..128G, it will walk |
67 | * through this set of ranges: |
68 | * |
69 | * 4G - 68G-1 (64G) |
70 | * 68G - 128G-1 (64G) |
71 | * |
72 | * An example of the APIs on how to use/iterate over the IOVA bitmap: |
73 | * |
74 | * bitmap = iova_bitmap_alloc(iova, length, page_size, data); |
75 | * if (IS_ERR(bitmap)) |
76 | * return PTR_ERR(bitmap); |
77 | * |
78 | * ret = iova_bitmap_for_each(bitmap, arg, dirty_reporter_fn); |
79 | * |
80 | * iova_bitmap_free(bitmap); |
81 | * |
82 | * Each iteration of the @dirty_reporter_fn is called with a unique @iova |
83 | * and @length argument, indicating the current range available through the |
84 | * iova_bitmap. The @dirty_reporter_fn uses iova_bitmap_set() to mark dirty |
85 | * areas (@iova_length) within that provided range, as following: |
86 | * |
87 | * iova_bitmap_set(bitmap, iova, iova_length); |
88 | * |
89 | * The internals of the object uses an index @mapped_base_index that indexes |
90 | * which u64 word of the bitmap is mapped, up to @mapped_total_index. |
91 | * Those keep being incremented until @mapped_total_index is reached while |
92 | * mapping up to PAGE_SIZE / sizeof(struct page*) maximum of pages. |
93 | * |
94 | * The IOVA bitmap is usually located on what tracks DMA mapped ranges or |
95 | * some form of IOVA range tracking that co-relates to the user passed |
96 | * bitmap. |
97 | */ |
98 | struct iova_bitmap { |
99 | /* IOVA range representing the currently mapped bitmap data */ |
100 | struct iova_bitmap_map mapped; |
101 | |
102 | /* userspace address of the bitmap */ |
103 | u8 __user *bitmap; |
104 | |
105 | /* u64 index that @mapped points to */ |
106 | unsigned long mapped_base_index; |
107 | |
108 | /* how many u64 can we walk in total */ |
109 | unsigned long mapped_total_index; |
110 | |
111 | /* base IOVA of the whole bitmap */ |
112 | unsigned long iova; |
113 | |
114 | /* length of the IOVA range for the whole bitmap */ |
115 | size_t length; |
116 | |
117 | /* length of the IOVA range set ahead the pinned pages */ |
118 | unsigned long set_ahead_length; |
119 | }; |
120 | |
121 | /* |
122 | * Converts a relative IOVA to a bitmap index. |
123 | * This function provides the index into the u64 array (bitmap::bitmap) |
124 | * for a given IOVA offset. |
125 | * Relative IOVA means relative to the bitmap::mapped base IOVA |
126 | * (stored in mapped::iova). All computations in this file are done using |
127 | * relative IOVAs and thus avoid an extra subtraction against mapped::iova. |
128 | * The user API iova_bitmap_set() always uses a regular absolute IOVAs. |
129 | */ |
130 | static unsigned long iova_bitmap_offset_to_index(struct iova_bitmap *bitmap, |
131 | unsigned long iova) |
132 | { |
133 | unsigned long pgsize = 1 << bitmap->mapped.pgshift; |
134 | |
135 | return iova / (BITS_PER_TYPE(*bitmap->bitmap) * pgsize); |
136 | } |
137 | |
138 | /* |
139 | * Converts a bitmap index to a *relative* IOVA. |
140 | */ |
141 | static unsigned long iova_bitmap_index_to_offset(struct iova_bitmap *bitmap, |
142 | unsigned long index) |
143 | { |
144 | unsigned long pgshift = bitmap->mapped.pgshift; |
145 | |
146 | return (index * BITS_PER_TYPE(*bitmap->bitmap)) << pgshift; |
147 | } |
148 | |
149 | /* |
150 | * Returns the base IOVA of the mapped range. |
151 | */ |
152 | static unsigned long iova_bitmap_mapped_iova(struct iova_bitmap *bitmap) |
153 | { |
154 | unsigned long skip = bitmap->mapped_base_index; |
155 | |
156 | return bitmap->iova + iova_bitmap_index_to_offset(bitmap, index: skip); |
157 | } |
158 | |
159 | /* |
160 | * Pins the bitmap user pages for the current range window. |
161 | * This is internal to IOVA bitmap and called when advancing the |
162 | * index (@mapped_base_index) or allocating the bitmap. |
163 | */ |
164 | static int iova_bitmap_get(struct iova_bitmap *bitmap) |
165 | { |
166 | struct iova_bitmap_map *mapped = &bitmap->mapped; |
167 | unsigned long npages; |
168 | u8 __user *addr; |
169 | long ret; |
170 | |
171 | /* |
172 | * @mapped_base_index is the index of the currently mapped u64 words |
173 | * that we have access. Anything before @mapped_base_index is not |
174 | * mapped. The range @mapped_base_index .. @mapped_total_index-1 is |
175 | * mapped but capped at a maximum number of pages. |
176 | */ |
177 | npages = DIV_ROUND_UP((bitmap->mapped_total_index - |
178 | bitmap->mapped_base_index) * |
179 | sizeof(*bitmap->bitmap), PAGE_SIZE); |
180 | |
181 | /* |
182 | * Bitmap address to be pinned is calculated via pointer arithmetic |
183 | * with bitmap u64 word index. |
184 | */ |
185 | addr = bitmap->bitmap + bitmap->mapped_base_index; |
186 | |
187 | /* |
188 | * We always cap at max number of 'struct page' a base page can fit. |
189 | * This is, for example, on x86 means 2M of bitmap data max. |
190 | */ |
191 | npages = min(npages + !!offset_in_page(addr), |
192 | PAGE_SIZE / sizeof(struct page *)); |
193 | |
194 | ret = pin_user_pages_fast(start: (unsigned long)addr, nr_pages: npages, |
195 | gup_flags: FOLL_WRITE, pages: mapped->pages); |
196 | if (ret <= 0) |
197 | return -EFAULT; |
198 | |
199 | mapped->npages = (unsigned long)ret; |
200 | /* Base IOVA where @pages point to i.e. bit 0 of the first page */ |
201 | mapped->iova = iova_bitmap_mapped_iova(bitmap); |
202 | |
203 | /* |
204 | * offset of the page where pinned pages bit 0 is located. |
205 | * This handles the case where the bitmap is not PAGE_SIZE |
206 | * aligned. |
207 | */ |
208 | mapped->pgoff = offset_in_page(addr); |
209 | return 0; |
210 | } |
211 | |
212 | /* |
213 | * Unpins the bitmap user pages and clears @npages |
214 | * (un)pinning is abstracted from API user and it's done when advancing |
215 | * the index or freeing the bitmap. |
216 | */ |
217 | static void iova_bitmap_put(struct iova_bitmap *bitmap) |
218 | { |
219 | struct iova_bitmap_map *mapped = &bitmap->mapped; |
220 | |
221 | if (mapped->npages) { |
222 | unpin_user_pages(pages: mapped->pages, npages: mapped->npages); |
223 | mapped->npages = 0; |
224 | } |
225 | } |
226 | |
227 | /** |
228 | * iova_bitmap_alloc() - Allocates an IOVA bitmap object |
229 | * @iova: Start address of the IOVA range |
230 | * @length: Length of the IOVA range |
231 | * @page_size: Page size of the IOVA bitmap. It defines what each bit |
232 | * granularity represents |
233 | * @data: Userspace address of the bitmap |
234 | * |
235 | * Allocates an IOVA object and initializes all its fields including the |
236 | * first user pages of @data. |
237 | * |
238 | * Return: A pointer to a newly allocated struct iova_bitmap |
239 | * or ERR_PTR() on error. |
240 | */ |
241 | struct iova_bitmap *iova_bitmap_alloc(unsigned long iova, size_t length, |
242 | unsigned long page_size, u64 __user *data) |
243 | { |
244 | struct iova_bitmap_map *mapped; |
245 | struct iova_bitmap *bitmap; |
246 | int rc; |
247 | |
248 | bitmap = kzalloc(size: sizeof(*bitmap), GFP_KERNEL); |
249 | if (!bitmap) |
250 | return ERR_PTR(error: -ENOMEM); |
251 | |
252 | mapped = &bitmap->mapped; |
253 | mapped->pgshift = __ffs(page_size); |
254 | bitmap->bitmap = (u8 __user *)data; |
255 | bitmap->mapped_total_index = |
256 | iova_bitmap_offset_to_index(bitmap, iova: length - 1) + 1; |
257 | bitmap->iova = iova; |
258 | bitmap->length = length; |
259 | mapped->iova = iova; |
260 | mapped->pages = (struct page **)__get_free_page(GFP_KERNEL); |
261 | if (!mapped->pages) { |
262 | rc = -ENOMEM; |
263 | goto err; |
264 | } |
265 | |
266 | rc = iova_bitmap_get(bitmap); |
267 | if (rc) |
268 | goto err; |
269 | return bitmap; |
270 | |
271 | err: |
272 | iova_bitmap_free(bitmap); |
273 | return ERR_PTR(error: rc); |
274 | } |
275 | EXPORT_SYMBOL_NS_GPL(iova_bitmap_alloc, IOMMUFD); |
276 | |
277 | /** |
278 | * iova_bitmap_free() - Frees an IOVA bitmap object |
279 | * @bitmap: IOVA bitmap to free |
280 | * |
281 | * It unpins and releases pages array memory and clears any leftover |
282 | * state. |
283 | */ |
284 | void iova_bitmap_free(struct iova_bitmap *bitmap) |
285 | { |
286 | struct iova_bitmap_map *mapped = &bitmap->mapped; |
287 | |
288 | iova_bitmap_put(bitmap); |
289 | |
290 | if (mapped->pages) { |
291 | free_page((unsigned long)mapped->pages); |
292 | mapped->pages = NULL; |
293 | } |
294 | |
295 | kfree(objp: bitmap); |
296 | } |
297 | EXPORT_SYMBOL_NS_GPL(iova_bitmap_free, IOMMUFD); |
298 | |
299 | /* |
300 | * Returns the remaining bitmap indexes from mapped_total_index to process for |
301 | * the currently pinned bitmap pages. |
302 | */ |
303 | static unsigned long iova_bitmap_mapped_remaining(struct iova_bitmap *bitmap) |
304 | { |
305 | unsigned long remaining, bytes; |
306 | |
307 | bytes = (bitmap->mapped.npages << PAGE_SHIFT) - bitmap->mapped.pgoff; |
308 | |
309 | remaining = bitmap->mapped_total_index - bitmap->mapped_base_index; |
310 | remaining = min_t(unsigned long, remaining, |
311 | DIV_ROUND_UP(bytes, sizeof(*bitmap->bitmap))); |
312 | |
313 | return remaining; |
314 | } |
315 | |
316 | /* |
317 | * Returns the length of the mapped IOVA range. |
318 | */ |
319 | static unsigned long iova_bitmap_mapped_length(struct iova_bitmap *bitmap) |
320 | { |
321 | unsigned long max_iova = bitmap->iova + bitmap->length - 1; |
322 | unsigned long iova = iova_bitmap_mapped_iova(bitmap); |
323 | unsigned long remaining; |
324 | |
325 | /* |
326 | * iova_bitmap_mapped_remaining() returns a number of indexes which |
327 | * when converted to IOVA gives us a max length that the bitmap |
328 | * pinned data can cover. Afterwards, that is capped to |
329 | * only cover the IOVA range in @bitmap::iova .. @bitmap::length. |
330 | */ |
331 | remaining = iova_bitmap_index_to_offset(bitmap, |
332 | index: iova_bitmap_mapped_remaining(bitmap)); |
333 | |
334 | if (iova + remaining - 1 > max_iova) |
335 | remaining -= ((iova + remaining - 1) - max_iova); |
336 | |
337 | return remaining; |
338 | } |
339 | |
340 | /* |
341 | * Returns true if there's not more data to iterate. |
342 | */ |
343 | static bool iova_bitmap_done(struct iova_bitmap *bitmap) |
344 | { |
345 | return bitmap->mapped_base_index >= bitmap->mapped_total_index; |
346 | } |
347 | |
348 | static int iova_bitmap_set_ahead(struct iova_bitmap *bitmap, |
349 | size_t set_ahead_length) |
350 | { |
351 | int ret = 0; |
352 | |
353 | while (set_ahead_length > 0 && !iova_bitmap_done(bitmap)) { |
354 | unsigned long length = iova_bitmap_mapped_length(bitmap); |
355 | unsigned long iova = iova_bitmap_mapped_iova(bitmap); |
356 | |
357 | ret = iova_bitmap_get(bitmap); |
358 | if (ret) |
359 | break; |
360 | |
361 | length = min(length, set_ahead_length); |
362 | iova_bitmap_set(bitmap, iova, length); |
363 | |
364 | set_ahead_length -= length; |
365 | bitmap->mapped_base_index += |
366 | iova_bitmap_offset_to_index(bitmap, iova: length - 1) + 1; |
367 | iova_bitmap_put(bitmap); |
368 | } |
369 | |
370 | bitmap->set_ahead_length = 0; |
371 | return ret; |
372 | } |
373 | |
374 | /* |
375 | * Advances to the next range, releases the current pinned |
376 | * pages and pins the next set of bitmap pages. |
377 | * Returns 0 on success or otherwise errno. |
378 | */ |
379 | static int iova_bitmap_advance(struct iova_bitmap *bitmap) |
380 | { |
381 | unsigned long iova = iova_bitmap_mapped_length(bitmap) - 1; |
382 | unsigned long count = iova_bitmap_offset_to_index(bitmap, iova) + 1; |
383 | |
384 | bitmap->mapped_base_index += count; |
385 | |
386 | iova_bitmap_put(bitmap); |
387 | if (iova_bitmap_done(bitmap)) |
388 | return 0; |
389 | |
390 | /* Iterate, set and skip any bits requested for next iteration */ |
391 | if (bitmap->set_ahead_length) { |
392 | int ret; |
393 | |
394 | ret = iova_bitmap_set_ahead(bitmap, set_ahead_length: bitmap->set_ahead_length); |
395 | if (ret) |
396 | return ret; |
397 | } |
398 | |
399 | /* When advancing the index we pin the next set of bitmap pages */ |
400 | return iova_bitmap_get(bitmap); |
401 | } |
402 | |
403 | /** |
404 | * iova_bitmap_for_each() - Iterates over the bitmap |
405 | * @bitmap: IOVA bitmap to iterate |
406 | * @opaque: Additional argument to pass to the callback |
407 | * @fn: Function that gets called for each IOVA range |
408 | * |
409 | * Helper function to iterate over bitmap data representing a portion of IOVA |
410 | * space. It hides the complexity of iterating bitmaps and translating the |
411 | * mapped bitmap user pages into IOVA ranges to process. |
412 | * |
413 | * Return: 0 on success, and an error on failure either upon |
414 | * iteration or when the callback returns an error. |
415 | */ |
416 | int iova_bitmap_for_each(struct iova_bitmap *bitmap, void *opaque, |
417 | iova_bitmap_fn_t fn) |
418 | { |
419 | int ret = 0; |
420 | |
421 | for (; !iova_bitmap_done(bitmap) && !ret; |
422 | ret = iova_bitmap_advance(bitmap)) { |
423 | ret = fn(bitmap, iova_bitmap_mapped_iova(bitmap), |
424 | iova_bitmap_mapped_length(bitmap), opaque); |
425 | if (ret) |
426 | break; |
427 | } |
428 | |
429 | return ret; |
430 | } |
431 | EXPORT_SYMBOL_NS_GPL(iova_bitmap_for_each, IOMMUFD); |
432 | |
433 | /** |
434 | * iova_bitmap_set() - Records an IOVA range in bitmap |
435 | * @bitmap: IOVA bitmap |
436 | * @iova: IOVA to start |
437 | * @length: IOVA range length |
438 | * |
439 | * Set the bits corresponding to the range [iova .. iova+length-1] in |
440 | * the user bitmap. |
441 | * |
442 | */ |
443 | void iova_bitmap_set(struct iova_bitmap *bitmap, |
444 | unsigned long iova, size_t length) |
445 | { |
446 | struct iova_bitmap_map *mapped = &bitmap->mapped; |
447 | unsigned long cur_bit = ((iova - mapped->iova) >> |
448 | mapped->pgshift) + mapped->pgoff * BITS_PER_BYTE; |
449 | unsigned long last_bit = (((iova + length - 1) - mapped->iova) >> |
450 | mapped->pgshift) + mapped->pgoff * BITS_PER_BYTE; |
451 | unsigned long last_page_idx = mapped->npages - 1; |
452 | |
453 | do { |
454 | unsigned int page_idx = cur_bit / BITS_PER_PAGE; |
455 | unsigned int offset = cur_bit % BITS_PER_PAGE; |
456 | unsigned int nbits = min(BITS_PER_PAGE - offset, |
457 | last_bit - cur_bit + 1); |
458 | void *kaddr; |
459 | |
460 | if (unlikely(page_idx > last_page_idx)) |
461 | break; |
462 | |
463 | kaddr = kmap_local_page(page: mapped->pages[page_idx]); |
464 | bitmap_set(map: kaddr, start: offset, nbits); |
465 | kunmap_local(kaddr); |
466 | cur_bit += nbits; |
467 | } while (cur_bit <= last_bit); |
468 | |
469 | if (unlikely(cur_bit <= last_bit)) { |
470 | bitmap->set_ahead_length = |
471 | ((last_bit - cur_bit + 1) << bitmap->mapped.pgshift); |
472 | } |
473 | } |
474 | EXPORT_SYMBOL_NS_GPL(iova_bitmap_set, IOMMUFD); |
475 | |