1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2022, Oracle and/or its affiliates.
4 * Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved
6 #include <linux/iova_bitmap.h>
8 #include <linux/slab.h>
9 #include <linux/highmem.h>
11 #define BITS_PER_PAGE (PAGE_SIZE * BITS_PER_BYTE)
14 * struct iova_bitmap_map - A bitmap representing an IOVA range
16 * Main data structure for tracking mapped user pages of bitmap data.
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
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
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:
32 * data[(iova / page_size) / 64] & (1ULL << (iova % 64))
34 struct iova_bitmap_map {
35 /* base IOVA representing bit 0 of the first page */
38 /* page size order that each bit granules to */
39 unsigned long pgshift;
41 /* page offset of the first user page pinned */
44 /* number of pages pinned */
47 /* pinned pages representing the bitmap data */
52 * struct iova_bitmap - The IOVA bitmap object
54 * Main data structure for iterating over the bitmap data.
56 * Abstracts the pinning work and iterates in IOVA ranges.
57 * It uses a windowing scheme and pins the bitmap in relatively
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.
66 * For example iterating on a total IOVA range of 4G..128G, it will walk
67 * through this set of ranges:
72 * An example of the APIs on how to use/iterate over the IOVA bitmap:
74 * bitmap = iova_bitmap_alloc(iova, length, page_size, data);
76 * return PTR_ERR(bitmap);
78 * ret = iova_bitmap_for_each(bitmap, arg, dirty_reporter_fn);
80 * iova_bitmap_free(bitmap);
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:
87 * iova_bitmap_set(bitmap, iova, iova_length);
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.
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
99 /* IOVA range representing the currently mapped bitmap data */
100 struct iova_bitmap_map mapped;
102 /* userspace address of the bitmap */
105 /* u64 index that @mapped points to */
106 unsigned long mapped_base_index;
108 /* how many u64 can we walk in total */
109 unsigned long mapped_total_index;
111 /* base IOVA of the whole bitmap */
114 /* length of the IOVA range for the whole bitmap */
119 * Converts a relative IOVA to a bitmap index.
120 * This function provides the index into the u64 array (bitmap::bitmap)
121 * for a given IOVA offset.
122 * Relative IOVA means relative to the bitmap::mapped base IOVA
123 * (stored in mapped::iova). All computations in this file are done using
124 * relative IOVAs and thus avoid an extra subtraction against mapped::iova.
125 * The user API iova_bitmap_set() always uses a regular absolute IOVAs.
127 static unsigned long iova_bitmap_offset_to_index(struct iova_bitmap *bitmap,
130 unsigned long pgsize = 1 << bitmap->mapped.pgshift;
132 return iova / (BITS_PER_TYPE(*bitmap->bitmap) * pgsize);
136 * Converts a bitmap index to a *relative* IOVA.
138 static unsigned long iova_bitmap_index_to_offset(struct iova_bitmap *bitmap,
141 unsigned long pgshift = bitmap->mapped.pgshift;
143 return (index * BITS_PER_TYPE(*bitmap->bitmap)) << pgshift;
147 * Returns the base IOVA of the mapped range.
149 static unsigned long iova_bitmap_mapped_iova(struct iova_bitmap *bitmap)
151 unsigned long skip = bitmap->mapped_base_index;
153 return bitmap->iova + iova_bitmap_index_to_offset(bitmap, skip);
157 * Pins the bitmap user pages for the current range window.
158 * This is internal to IOVA bitmap and called when advancing the
159 * index (@mapped_base_index) or allocating the bitmap.
161 static int iova_bitmap_get(struct iova_bitmap *bitmap)
163 struct iova_bitmap_map *mapped = &bitmap->mapped;
164 unsigned long npages;
169 * @mapped_base_index is the index of the currently mapped u64 words
170 * that we have access. Anything before @mapped_base_index is not
171 * mapped. The range @mapped_base_index .. @mapped_total_index-1 is
172 * mapped but capped at a maximum number of pages.
174 npages = DIV_ROUND_UP((bitmap->mapped_total_index -
175 bitmap->mapped_base_index) *
176 sizeof(*bitmap->bitmap), PAGE_SIZE);
179 * We always cap at max number of 'struct page' a base page can fit.
180 * This is, for example, on x86 means 2M of bitmap data max.
182 npages = min(npages, PAGE_SIZE / sizeof(struct page *));
185 * Bitmap address to be pinned is calculated via pointer arithmetic
186 * with bitmap u64 word index.
188 addr = bitmap->bitmap + bitmap->mapped_base_index;
190 ret = pin_user_pages_fast((unsigned long)addr, npages,
191 FOLL_WRITE, mapped->pages);
195 mapped->npages = (unsigned long)ret;
196 /* Base IOVA where @pages point to i.e. bit 0 of the first page */
197 mapped->iova = iova_bitmap_mapped_iova(bitmap);
200 * offset of the page where pinned pages bit 0 is located.
201 * This handles the case where the bitmap is not PAGE_SIZE
204 mapped->pgoff = offset_in_page(addr);
209 * Unpins the bitmap user pages and clears @npages
210 * (un)pinning is abstracted from API user and it's done when advancing
211 * the index or freeing the bitmap.
213 static void iova_bitmap_put(struct iova_bitmap *bitmap)
215 struct iova_bitmap_map *mapped = &bitmap->mapped;
217 if (mapped->npages) {
218 unpin_user_pages(mapped->pages, mapped->npages);
224 * iova_bitmap_alloc() - Allocates an IOVA bitmap object
225 * @iova: Start address of the IOVA range
226 * @length: Length of the IOVA range
227 * @page_size: Page size of the IOVA bitmap. It defines what each bit
228 * granularity represents
229 * @data: Userspace address of the bitmap
231 * Allocates an IOVA object and initializes all its fields including the
232 * first user pages of @data.
234 * Return: A pointer to a newly allocated struct iova_bitmap
235 * or ERR_PTR() on error.
237 struct iova_bitmap *iova_bitmap_alloc(unsigned long iova, size_t length,
238 unsigned long page_size, u64 __user *data)
240 struct iova_bitmap_map *mapped;
241 struct iova_bitmap *bitmap;
244 bitmap = kzalloc(sizeof(*bitmap), GFP_KERNEL);
246 return ERR_PTR(-ENOMEM);
248 mapped = &bitmap->mapped;
249 mapped->pgshift = __ffs(page_size);
250 bitmap->bitmap = data;
251 bitmap->mapped_total_index =
252 iova_bitmap_offset_to_index(bitmap, length - 1) + 1;
254 bitmap->length = length;
256 mapped->pages = (struct page **)__get_free_page(GFP_KERNEL);
257 if (!mapped->pages) {
262 rc = iova_bitmap_get(bitmap);
268 iova_bitmap_free(bitmap);
273 * iova_bitmap_free() - Frees an IOVA bitmap object
274 * @bitmap: IOVA bitmap to free
276 * It unpins and releases pages array memory and clears any leftover
279 void iova_bitmap_free(struct iova_bitmap *bitmap)
281 struct iova_bitmap_map *mapped = &bitmap->mapped;
283 iova_bitmap_put(bitmap);
286 free_page((unsigned long)mapped->pages);
287 mapped->pages = NULL;
294 * Returns the remaining bitmap indexes from mapped_total_index to process for
295 * the currently pinned bitmap pages.
297 static unsigned long iova_bitmap_mapped_remaining(struct iova_bitmap *bitmap)
299 unsigned long remaining, bytes;
301 bytes = (bitmap->mapped.npages << PAGE_SHIFT) - bitmap->mapped.pgoff;
303 remaining = bitmap->mapped_total_index - bitmap->mapped_base_index;
304 remaining = min_t(unsigned long, remaining,
305 bytes / sizeof(*bitmap->bitmap));
311 * Returns the length of the mapped IOVA range.
313 static unsigned long iova_bitmap_mapped_length(struct iova_bitmap *bitmap)
315 unsigned long max_iova = bitmap->iova + bitmap->length - 1;
316 unsigned long iova = iova_bitmap_mapped_iova(bitmap);
317 unsigned long remaining;
320 * iova_bitmap_mapped_remaining() returns a number of indexes which
321 * when converted to IOVA gives us a max length that the bitmap
322 * pinned data can cover. Afterwards, that is capped to
323 * only cover the IOVA range in @bitmap::iova .. @bitmap::length.
325 remaining = iova_bitmap_index_to_offset(bitmap,
326 iova_bitmap_mapped_remaining(bitmap));
328 if (iova + remaining - 1 > max_iova)
329 remaining -= ((iova + remaining - 1) - max_iova);
335 * Returns true if there's not more data to iterate.
337 static bool iova_bitmap_done(struct iova_bitmap *bitmap)
339 return bitmap->mapped_base_index >= bitmap->mapped_total_index;
343 * Advances to the next range, releases the current pinned
344 * pages and pins the next set of bitmap pages.
345 * Returns 0 on success or otherwise errno.
347 static int iova_bitmap_advance(struct iova_bitmap *bitmap)
349 unsigned long iova = iova_bitmap_mapped_length(bitmap) - 1;
350 unsigned long count = iova_bitmap_offset_to_index(bitmap, iova) + 1;
352 bitmap->mapped_base_index += count;
354 iova_bitmap_put(bitmap);
355 if (iova_bitmap_done(bitmap))
358 /* When advancing the index we pin the next set of bitmap pages */
359 return iova_bitmap_get(bitmap);
363 * iova_bitmap_for_each() - Iterates over the bitmap
364 * @bitmap: IOVA bitmap to iterate
365 * @opaque: Additional argument to pass to the callback
366 * @fn: Function that gets called for each IOVA range
368 * Helper function to iterate over bitmap data representing a portion of IOVA
369 * space. It hides the complexity of iterating bitmaps and translating the
370 * mapped bitmap user pages into IOVA ranges to process.
372 * Return: 0 on success, and an error on failure either upon
373 * iteration or when the callback returns an error.
375 int iova_bitmap_for_each(struct iova_bitmap *bitmap, void *opaque,
380 for (; !iova_bitmap_done(bitmap) && !ret;
381 ret = iova_bitmap_advance(bitmap)) {
382 ret = fn(bitmap, iova_bitmap_mapped_iova(bitmap),
383 iova_bitmap_mapped_length(bitmap), opaque);
392 * iova_bitmap_set() - Records an IOVA range in bitmap
393 * @bitmap: IOVA bitmap
394 * @iova: IOVA to start
395 * @length: IOVA range length
397 * Set the bits corresponding to the range [iova .. iova+length-1] in
401 void iova_bitmap_set(struct iova_bitmap *bitmap,
402 unsigned long iova, size_t length)
404 struct iova_bitmap_map *mapped = &bitmap->mapped;
405 unsigned long cur_bit = ((iova - mapped->iova) >>
406 mapped->pgshift) + mapped->pgoff * BITS_PER_BYTE;
407 unsigned long last_bit = (((iova + length - 1) - mapped->iova) >>
408 mapped->pgshift) + mapped->pgoff * BITS_PER_BYTE;
411 unsigned int page_idx = cur_bit / BITS_PER_PAGE;
412 unsigned int offset = cur_bit % BITS_PER_PAGE;
413 unsigned int nbits = min(BITS_PER_PAGE - offset,
414 last_bit - cur_bit + 1);
417 kaddr = kmap_local_page(mapped->pages[page_idx]);
418 bitmap_set(kaddr, offset, nbits);
421 } while (cur_bit <= last_bit);
423 EXPORT_SYMBOL_GPL(iova_bitmap_set);