1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/kernel.h>
3 #include <linux/init.h>
4 #include <linux/errno.h>
6 #include <linux/mman.h>
7 #include <linux/slab.h>
8 #include <linux/vmalloc.h>
9 #include <linux/io_uring.h>
10 #include <linux/io_uring_types.h>
11 #include <asm/shmparam.h>
17 static void *io_mem_alloc_compound(struct page **pages, int nr_pages,
18 size_t size, gfp_t gfp)
23 order = get_order(size);
24 if (order > MAX_PAGE_ORDER)
25 return ERR_PTR(-ENOMEM);
29 page = alloc_pages(gfp, order);
31 return ERR_PTR(-ENOMEM);
33 for (i = 0; i < nr_pages; i++)
36 return page_address(page);
39 static void *io_mem_alloc_single(struct page **pages, int nr_pages, size_t size,
45 for (i = 0; i < nr_pages; i++) {
46 pages[i] = alloc_page(gfp);
51 ret = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
57 return ERR_PTR(-ENOMEM);
60 void *io_pages_map(struct page ***out_pages, unsigned short *npages,
63 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN;
68 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
69 pages = kvmalloc_array(nr_pages, sizeof(struct page *), gfp);
71 return ERR_PTR(-ENOMEM);
73 ret = io_mem_alloc_compound(pages, nr_pages, size, gfp);
79 ret = io_mem_alloc_single(pages, nr_pages, size, gfp);
93 void io_pages_unmap(void *ptr, struct page ***pages, unsigned short *npages,
96 bool do_vunmap = false;
101 if (put_pages && *npages) {
102 struct page **to_free = *pages;
106 * Only did vmap for the non-compound multiple page case.
107 * For the compound page, we just need to put the head.
109 if (PageCompound(to_free[0]))
111 else if (*npages > 1)
113 for (i = 0; i < *npages; i++)
114 put_page(to_free[i]);
123 void io_pages_free(struct page ***pages, int npages)
125 struct page **page_array = *pages;
130 unpin_user_pages(page_array, npages);
135 struct page **io_pin_pages(unsigned long uaddr, unsigned long len, int *npages)
137 unsigned long start, end, nr_pages;
141 if (check_add_overflow(uaddr, len, &end))
142 return ERR_PTR(-EOVERFLOW);
143 if (check_add_overflow(end, PAGE_SIZE - 1, &end))
144 return ERR_PTR(-EOVERFLOW);
146 end = end >> PAGE_SHIFT;
147 start = uaddr >> PAGE_SHIFT;
148 nr_pages = end - start;
149 if (WARN_ON_ONCE(!nr_pages))
150 return ERR_PTR(-EINVAL);
151 if (WARN_ON_ONCE(nr_pages > INT_MAX))
152 return ERR_PTR(-EOVERFLOW);
154 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
156 return ERR_PTR(-ENOMEM);
158 ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
160 /* success, mapped all pages */
161 if (ret == nr_pages) {
166 /* partial map, or didn't map anything */
168 /* if we did partial map, release any pages we did get */
170 unpin_user_pages(pages, ret);
177 void *__io_uaddr_map(struct page ***pages, unsigned short *npages,
178 unsigned long uaddr, size_t size)
180 struct page **page_array;
181 unsigned int nr_pages;
186 if (uaddr & (PAGE_SIZE - 1) || !size)
187 return ERR_PTR(-EINVAL);
190 page_array = io_pin_pages(uaddr, size, &nr_pages);
191 if (IS_ERR(page_array))
194 page_addr = vmap(page_array, nr_pages, VM_MAP, PAGE_KERNEL);
201 io_pages_free(&page_array, nr_pages);
202 return ERR_PTR(-ENOMEM);
205 void io_free_region(struct io_ring_ctx *ctx, struct io_mapped_region *mr)
208 unpin_user_pages(mr->pages, mr->nr_pages);
212 vunmap(mr->vmap_ptr);
213 if (mr->nr_pages && ctx->user)
214 __io_unaccount_mem(ctx->user, mr->nr_pages);
216 memset(mr, 0, sizeof(*mr));
219 int io_create_region(struct io_ring_ctx *ctx, struct io_mapped_region *mr,
220 struct io_uring_region_desc *reg)
222 int pages_accounted = 0;
228 if (WARN_ON_ONCE(mr->pages || mr->vmap_ptr || mr->nr_pages))
230 if (memchr_inv(®->__resv, 0, sizeof(reg->__resv)))
232 if (reg->flags != IORING_MEM_REGION_TYPE_USER)
236 if (!reg->size || reg->mmap_offset || reg->id)
238 if ((reg->size >> PAGE_SHIFT) > INT_MAX)
240 if ((reg->user_addr | reg->size) & ~PAGE_MASK)
242 if (check_add_overflow(reg->user_addr, reg->size, &end))
245 pages = io_pin_pages(reg->user_addr, reg->size, &nr_pages);
247 return PTR_ERR(pages);
250 ret = __io_account_mem(ctx->user, nr_pages);
253 pages_accounted = nr_pages;
256 vptr = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
264 mr->nr_pages = nr_pages;
268 __io_unaccount_mem(ctx->user, pages_accounted);
269 io_pages_free(&pages, nr_pages);
273 static void *io_uring_validate_mmap_request(struct file *file, loff_t pgoff,
276 struct io_ring_ctx *ctx = file->private_data;
277 loff_t offset = pgoff << PAGE_SHIFT;
279 switch ((pgoff << PAGE_SHIFT) & IORING_OFF_MMAP_MASK) {
280 case IORING_OFF_SQ_RING:
281 case IORING_OFF_CQ_RING:
282 /* Don't allow mmap if the ring was setup without it */
283 if (ctx->flags & IORING_SETUP_NO_MMAP)
284 return ERR_PTR(-EINVAL);
286 return ERR_PTR(-EFAULT);
288 case IORING_OFF_SQES:
289 /* Don't allow mmap if the ring was setup without it */
290 if (ctx->flags & IORING_SETUP_NO_MMAP)
291 return ERR_PTR(-EINVAL);
293 return ERR_PTR(-EFAULT);
295 case IORING_OFF_PBUF_RING: {
296 struct io_buffer_list *bl;
300 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
301 bl = io_pbuf_get_bl(ctx, bgid);
310 return ERR_PTR(-EINVAL);
313 int io_uring_mmap_pages(struct io_ring_ctx *ctx, struct vm_area_struct *vma,
314 struct page **pages, int npages)
316 unsigned long nr_pages = npages;
318 vm_flags_set(vma, VM_DONTEXPAND);
319 return vm_insert_pages(vma, vma->vm_start, pages, &nr_pages);
324 __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
326 struct io_ring_ctx *ctx = file->private_data;
327 size_t sz = vma->vm_end - vma->vm_start;
328 long offset = vma->vm_pgoff << PAGE_SHIFT;
332 guard(mutex)(&ctx->resize_lock);
334 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
338 switch (offset & IORING_OFF_MMAP_MASK) {
339 case IORING_OFF_SQ_RING:
340 case IORING_OFF_CQ_RING:
341 npages = min(ctx->n_ring_pages, (sz + PAGE_SIZE - 1) >> PAGE_SHIFT);
342 return io_uring_mmap_pages(ctx, vma, ctx->ring_pages, npages);
343 case IORING_OFF_SQES:
344 return io_uring_mmap_pages(ctx, vma, ctx->sqe_pages,
346 case IORING_OFF_PBUF_RING:
347 return io_pbuf_mmap(file, vma);
353 unsigned long io_uring_get_unmapped_area(struct file *filp, unsigned long addr,
354 unsigned long len, unsigned long pgoff,
357 struct io_ring_ctx *ctx = filp->private_data;
361 * Do not allow to map to user-provided address to avoid breaking the
362 * aliasing rules. Userspace is not able to guess the offset address of
363 * kernel kmalloc()ed memory area.
368 guard(mutex)(&ctx->resize_lock);
370 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
375 * Some architectures have strong cache aliasing requirements.
376 * For such architectures we need a coherent mapping which aliases
377 * kernel memory *and* userspace memory. To achieve that:
378 * - use a NULL file pointer to reference physical memory, and
379 * - use the kernel virtual address of the shared io_uring context
380 * (instead of the userspace-provided address, which has to be 0UL
382 * - use the same pgoff which the get_unmapped_area() uses to
383 * calculate the page colouring.
384 * For architectures without such aliasing requirements, the
385 * architecture will return any suitable mapping because addr is 0.
389 pgoff = 0; /* has been translated to ptr above */
391 addr = (uintptr_t) ptr;
392 pgoff = addr >> PAGE_SHIFT;
396 return mm_get_unmapped_area(current->mm, filp, addr, len, pgoff, flags);
399 #else /* !CONFIG_MMU */
401 int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
403 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
406 unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
408 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
411 unsigned long io_uring_get_unmapped_area(struct file *file, unsigned long addr,
412 unsigned long len, unsigned long pgoff,
415 struct io_ring_ctx *ctx = file->private_data;
418 guard(mutex)(&ctx->resize_lock);
420 ptr = io_uring_validate_mmap_request(file, pgoff, len);
424 return (unsigned long) ptr;
427 #endif /* !CONFIG_MMU */