io_uring: move mapping/allocation helpers to a separate file
authorJens Axboe <axboe@kernel.dk>
Wed, 27 Mar 2024 20:59:09 +0000 (14:59 -0600)
committerJens Axboe <axboe@kernel.dk>
Mon, 15 Apr 2024 14:10:26 +0000 (08:10 -0600)
Move the related code from io_uring.c into memmap.c. No functional
changes in this patch, just cleaning it up a bit now that the full
transition is done.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_uring/Makefile
io_uring/io_uring.c
io_uring/io_uring.h
io_uring/kbuf.c
io_uring/memmap.c [new file with mode: 0644]
io_uring/memmap.h [new file with mode: 0644]
io_uring/rsrc.c

index bd7c692a6a7ce737a7872d345e53752c49c66df8..fc1b23c524e83bf7de863c3c81b43a9d6c267b5b 100644 (file)
@@ -8,7 +8,8 @@ obj-$(CONFIG_IO_URING)          += io_uring.o opdef.o kbuf.o rsrc.o notif.o \
                                        xattr.o nop.o fs.o splice.o sync.o \
                                        msg_ring.o advise.o openclose.o \
                                        epoll.o statx.o timeout.o fdinfo.o \
-                                       cancel.o waitid.o register.o truncate.o
+                                       cancel.o waitid.o register.o \
+                                       truncate.o memmap.o
 obj-$(CONFIG_IO_WQ)            += io-wq.o
 obj-$(CONFIG_FUTEX)            += futex.o
 obj-$(CONFIG_NET_RX_BUSY_POLL) += napi.o
index 37066ea98f0b8b89d3d69f095544ee05a9982fb3..cc5fa4e1b34461f1e7b254547bd8413c1dc8e48b 100644 (file)
@@ -95,6 +95,7 @@
 #include "futex.h"
 #include "napi.h"
 #include "uring_cmd.h"
+#include "memmap.h"
 
 #include "timeout.h"
 #include "poll.h"
@@ -2597,111 +2598,6 @@ static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
        return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
 }
 
-void io_pages_unmap(void *ptr, struct page ***pages, unsigned short *npages,
-                   bool put_pages)
-{
-       bool do_vunmap = false;
-
-       if (!ptr)
-               return;
-
-       if (put_pages && *npages) {
-               struct page **to_free = *pages;
-               int i;
-
-               /*
-                * Only did vmap for the non-compound multiple page case.
-                * For the compound page, we just need to put the head.
-                */
-               if (PageCompound(to_free[0]))
-                       *npages = 1;
-               else if (*npages > 1)
-                       do_vunmap = true;
-               for (i = 0; i < *npages; i++)
-                       put_page(to_free[i]);
-       }
-       if (do_vunmap)
-               vunmap(ptr);
-       kvfree(*pages);
-       *pages = NULL;
-       *npages = 0;
-}
-
-static void io_pages_free(struct page ***pages, int npages)
-{
-       struct page **page_array = *pages;
-
-       if (!page_array)
-               return;
-
-       unpin_user_pages(page_array, npages);
-       kvfree(page_array);
-       *pages = NULL;
-}
-
-struct page **io_pin_pages(unsigned long uaddr, unsigned long len, int *npages)
-{
-       unsigned long start, end, nr_pages;
-       struct page **pages;
-       int ret;
-
-       end = (uaddr + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
-       start = uaddr >> PAGE_SHIFT;
-       nr_pages = end - start;
-       if (WARN_ON_ONCE(!nr_pages))
-               return ERR_PTR(-EINVAL);
-
-       pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
-       if (!pages)
-               return ERR_PTR(-ENOMEM);
-
-       ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
-                                       pages);
-       /* success, mapped all pages */
-       if (ret == nr_pages) {
-               *npages = nr_pages;
-               return pages;
-       }
-
-       /* partial map, or didn't map anything */
-       if (ret >= 0) {
-               /* if we did partial map, release any pages we did get */
-               if (ret)
-                       unpin_user_pages(pages, ret);
-               ret = -EFAULT;
-       }
-       kvfree(pages);
-       return ERR_PTR(ret);
-}
-
-static void *__io_uaddr_map(struct page ***pages, unsigned short *npages,
-                           unsigned long uaddr, size_t size)
-{
-       struct page **page_array;
-       unsigned int nr_pages;
-       void *page_addr;
-
-       *npages = 0;
-
-       if (uaddr & (PAGE_SIZE - 1) || !size)
-               return ERR_PTR(-EINVAL);
-
-       nr_pages = 0;
-       page_array = io_pin_pages(uaddr, size, &nr_pages);
-       if (IS_ERR(page_array))
-               return page_array;
-
-       page_addr = vmap(page_array, nr_pages, VM_MAP, PAGE_KERNEL);
-       if (page_addr) {
-               *pages = page_array;
-               *npages = nr_pages;
-               return page_addr;
-       }
-
-       io_pages_free(&page_array, nr_pages);
-       return ERR_PTR(-ENOMEM);
-}
-
 static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr,
                          size_t size)
 {
@@ -2736,80 +2632,6 @@ static void io_rings_free(struct io_ring_ctx *ctx)
        ctx->sq_sqes = NULL;
 }
 
-static void *io_mem_alloc_compound(struct page **pages, int nr_pages,
-                                  size_t size, gfp_t gfp)
-{
-       struct page *page;
-       int i, order;
-
-       order = get_order(size);
-       if (order > MAX_PAGE_ORDER)
-               return ERR_PTR(-ENOMEM);
-       else if (order)
-               gfp |= __GFP_COMP;
-
-       page = alloc_pages(gfp, order);
-       if (!page)
-               return ERR_PTR(-ENOMEM);
-
-       for (i = 0; i < nr_pages; i++)
-               pages[i] = page + i;
-
-       return page_address(page);
-}
-
-static void *io_mem_alloc_single(struct page **pages, int nr_pages, size_t size,
-                                gfp_t gfp)
-{
-       void *ret;
-       int i;
-
-       for (i = 0; i < nr_pages; i++) {
-               pages[i] = alloc_page(gfp);
-               if (!pages[i])
-                       goto err;
-       }
-
-       ret = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
-       if (ret)
-               return ret;
-err:
-       while (i--)
-               put_page(pages[i]);
-       return ERR_PTR(-ENOMEM);
-}
-
-void *io_pages_map(struct page ***out_pages, unsigned short *npages,
-                  size_t size)
-{
-       gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN;
-       struct page **pages;
-       int nr_pages;
-       void *ret;
-
-       nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
-       pages = kvmalloc_array(nr_pages, sizeof(struct page *), gfp);
-       if (!pages)
-               return ERR_PTR(-ENOMEM);
-
-       ret = io_mem_alloc_compound(pages, nr_pages, size, gfp);
-       if (!IS_ERR(ret))
-               goto done;
-
-       ret = io_mem_alloc_single(pages, nr_pages, size, gfp);
-       if (!IS_ERR(ret)) {
-done:
-               *out_pages = pages;
-               *npages = nr_pages;
-               return ret;
-       }
-
-       kvfree(pages);
-       *out_pages = NULL;
-       *npages = 0;
-       return ret;
-}
-
 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
                                unsigned int cq_entries, size_t *sq_offset)
 {
@@ -3370,149 +3192,6 @@ void __io_uring_cancel(bool cancel_all)
        io_uring_cancel_generic(cancel_all, NULL);
 }
 
-static void *io_uring_validate_mmap_request(struct file *file,
-                                           loff_t pgoff, size_t sz)
-{
-       struct io_ring_ctx *ctx = file->private_data;
-       loff_t offset = pgoff << PAGE_SHIFT;
-
-       switch ((pgoff << PAGE_SHIFT) & IORING_OFF_MMAP_MASK) {
-       case IORING_OFF_SQ_RING:
-       case IORING_OFF_CQ_RING:
-               /* Don't allow mmap if the ring was setup without it */
-               if (ctx->flags & IORING_SETUP_NO_MMAP)
-                       return ERR_PTR(-EINVAL);
-               return ctx->rings;
-       case IORING_OFF_SQES:
-               /* Don't allow mmap if the ring was setup without it */
-               if (ctx->flags & IORING_SETUP_NO_MMAP)
-                       return ERR_PTR(-EINVAL);
-               return ctx->sq_sqes;
-       case IORING_OFF_PBUF_RING: {
-               struct io_buffer_list *bl;
-               unsigned int bgid;
-               void *ptr;
-
-               bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
-               bl = io_pbuf_get_bl(ctx, bgid);
-               if (IS_ERR(bl))
-                       return bl;
-               ptr = bl->buf_ring;
-               io_put_bl(ctx, bl);
-               return ptr;
-               }
-       }
-
-       return ERR_PTR(-EINVAL);
-}
-
-int io_uring_mmap_pages(struct io_ring_ctx *ctx, struct vm_area_struct *vma,
-                       struct page **pages, int npages)
-{
-       unsigned long nr_pages = npages;
-
-       vm_flags_set(vma, VM_DONTEXPAND);
-       return vm_insert_pages(vma, vma->vm_start, pages, &nr_pages);
-}
-
-#ifdef CONFIG_MMU
-
-static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
-{
-       struct io_ring_ctx *ctx = file->private_data;
-       size_t sz = vma->vm_end - vma->vm_start;
-       long offset = vma->vm_pgoff << PAGE_SHIFT;
-       void *ptr;
-
-       ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
-       if (IS_ERR(ptr))
-               return PTR_ERR(ptr);
-
-       switch (offset & IORING_OFF_MMAP_MASK) {
-       case IORING_OFF_SQ_RING:
-       case IORING_OFF_CQ_RING:
-               return io_uring_mmap_pages(ctx, vma, ctx->ring_pages,
-                                               ctx->n_ring_pages);
-       case IORING_OFF_SQES:
-               return io_uring_mmap_pages(ctx, vma, ctx->sqe_pages,
-                                               ctx->n_sqe_pages);
-       case IORING_OFF_PBUF_RING:
-               return io_pbuf_mmap(file, vma);
-       }
-
-       return -EINVAL;
-}
-
-static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
-                       unsigned long addr, unsigned long len,
-                       unsigned long pgoff, unsigned long flags)
-{
-       void *ptr;
-
-       /*
-        * Do not allow to map to user-provided address to avoid breaking the
-        * aliasing rules. Userspace is not able to guess the offset address of
-        * kernel kmalloc()ed memory area.
-        */
-       if (addr)
-               return -EINVAL;
-
-       ptr = io_uring_validate_mmap_request(filp, pgoff, len);
-       if (IS_ERR(ptr))
-               return -ENOMEM;
-
-       /*
-        * Some architectures have strong cache aliasing requirements.
-        * For such architectures we need a coherent mapping which aliases
-        * kernel memory *and* userspace memory. To achieve that:
-        * - use a NULL file pointer to reference physical memory, and
-        * - use the kernel virtual address of the shared io_uring context
-        *   (instead of the userspace-provided address, which has to be 0UL
-        *   anyway).
-        * - use the same pgoff which the get_unmapped_area() uses to
-        *   calculate the page colouring.
-        * For architectures without such aliasing requirements, the
-        * architecture will return any suitable mapping because addr is 0.
-        */
-       filp = NULL;
-       flags |= MAP_SHARED;
-       pgoff = 0;      /* has been translated to ptr above */
-#ifdef SHM_COLOUR
-       addr = (uintptr_t) ptr;
-       pgoff = addr >> PAGE_SHIFT;
-#else
-       addr = 0UL;
-#endif
-       return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
-}
-
-#else /* !CONFIG_MMU */
-
-static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
-{
-       return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
-}
-
-static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
-{
-       return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
-}
-
-static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
-       unsigned long addr, unsigned long len,
-       unsigned long pgoff, unsigned long flags)
-{
-       void *ptr;
-
-       ptr = io_uring_validate_mmap_request(file, pgoff, len);
-       if (IS_ERR(ptr))
-               return PTR_ERR(ptr);
-
-       return (unsigned long) ptr;
-}
-
-#endif /* !CONFIG_MMU */
-
 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
 {
        if (flags & IORING_ENTER_EXT_ARG) {
@@ -3695,11 +3374,9 @@ out:
 static const struct file_operations io_uring_fops = {
        .release        = io_uring_release,
        .mmap           = io_uring_mmap,
+       .get_unmapped_area = io_uring_get_unmapped_area,
 #ifndef CONFIG_MMU
-       .get_unmapped_area = io_uring_nommu_get_unmapped_area,
        .mmap_capabilities = io_uring_nommu_mmap_capabilities,
-#else
-       .get_unmapped_area = io_uring_mmu_get_unmapped_area,
 #endif
        .poll           = io_uring_poll,
 #ifdef CONFIG_PROC_FS
index dec996a1c7895f152672c30600da66357a2057ee..1eb65324792a3217b27f4dfea8498dc955fcca2f 100644 (file)
@@ -69,10 +69,6 @@ bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags
 bool io_req_post_cqe(struct io_kiocb *req, s32 res, u32 cflags);
 void __io_commit_cqring_flush(struct io_ring_ctx *ctx);
 
-struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages);
-int io_uring_mmap_pages(struct io_ring_ctx *ctx, struct vm_area_struct *vma,
-                       struct page **pages, int npages);
-
 struct file *io_file_get_normal(struct io_kiocb *req, int fd);
 struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
                               unsigned issue_flags);
@@ -109,11 +105,6 @@ bool __io_alloc_req_refill(struct io_ring_ctx *ctx);
 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
                        bool cancel_all);
 
-void *io_pages_map(struct page ***out_pages, unsigned short *npages,
-                  size_t size);
-void io_pages_unmap(void *ptr, struct page ***pages, unsigned short *npages,
-                   bool put_pages);
-
 enum {
        IO_EVENTFD_OP_SIGNAL_BIT,
        IO_EVENTFD_OP_FREE_BIT,
index 10d5c98c1c562b186e3400d7d12b128e314f6f55..896770e93980bcd8a2231ea957e026da0973c508 100644 (file)
@@ -15,6 +15,7 @@
 #include "io_uring.h"
 #include "opdef.h"
 #include "kbuf.h"
+#include "memmap.h"
 
 #define IO_BUFFER_LIST_BUF_PER_PAGE (PAGE_SIZE / sizeof(struct io_uring_buf))
 
diff --git a/io_uring/memmap.c b/io_uring/memmap.c
new file mode 100644 (file)
index 0000000..523d982
--- /dev/null
@@ -0,0 +1,336 @@
+// SPDX-License-Identifier: GPL-2.0
+#include <linux/kernel.h>
+#include <linux/init.h>
+#include <linux/errno.h>
+#include <linux/mm.h>
+#include <linux/mman.h>
+#include <linux/slab.h>
+#include <linux/vmalloc.h>
+#include <linux/io_uring.h>
+#include <linux/io_uring_types.h>
+#include <asm/shmparam.h>
+
+#include "memmap.h"
+#include "kbuf.h"
+
+static void *io_mem_alloc_compound(struct page **pages, int nr_pages,
+                                  size_t size, gfp_t gfp)
+{
+       struct page *page;
+       int i, order;
+
+       order = get_order(size);
+       if (order > MAX_PAGE_ORDER)
+               return ERR_PTR(-ENOMEM);
+       else if (order)
+               gfp |= __GFP_COMP;
+
+       page = alloc_pages(gfp, order);
+       if (!page)
+               return ERR_PTR(-ENOMEM);
+
+       for (i = 0; i < nr_pages; i++)
+               pages[i] = page + i;
+
+       return page_address(page);
+}
+
+static void *io_mem_alloc_single(struct page **pages, int nr_pages, size_t size,
+                                gfp_t gfp)
+{
+       void *ret;
+       int i;
+
+       for (i = 0; i < nr_pages; i++) {
+               pages[i] = alloc_page(gfp);
+               if (!pages[i])
+                       goto err;
+       }
+
+       ret = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
+       if (ret)
+               return ret;
+err:
+       while (i--)
+               put_page(pages[i]);
+       return ERR_PTR(-ENOMEM);
+}
+
+void *io_pages_map(struct page ***out_pages, unsigned short *npages,
+                  size_t size)
+{
+       gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN;
+       struct page **pages;
+       int nr_pages;
+       void *ret;
+
+       nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
+       pages = kvmalloc_array(nr_pages, sizeof(struct page *), gfp);
+       if (!pages)
+               return ERR_PTR(-ENOMEM);
+
+       ret = io_mem_alloc_compound(pages, nr_pages, size, gfp);
+       if (!IS_ERR(ret))
+               goto done;
+
+       ret = io_mem_alloc_single(pages, nr_pages, size, gfp);
+       if (!IS_ERR(ret)) {
+done:
+               *out_pages = pages;
+               *npages = nr_pages;
+               return ret;
+       }
+
+       kvfree(pages);
+       *out_pages = NULL;
+       *npages = 0;
+       return ret;
+}
+
+void io_pages_unmap(void *ptr, struct page ***pages, unsigned short *npages,
+                   bool put_pages)
+{
+       bool do_vunmap = false;
+
+       if (!ptr)
+               return;
+
+       if (put_pages && *npages) {
+               struct page **to_free = *pages;
+               int i;
+
+               /*
+                * Only did vmap for the non-compound multiple page case.
+                * For the compound page, we just need to put the head.
+                */
+               if (PageCompound(to_free[0]))
+                       *npages = 1;
+               else if (*npages > 1)
+                       do_vunmap = true;
+               for (i = 0; i < *npages; i++)
+                       put_page(to_free[i]);
+       }
+       if (do_vunmap)
+               vunmap(ptr);
+       kvfree(*pages);
+       *pages = NULL;
+       *npages = 0;
+}
+
+void io_pages_free(struct page ***pages, int npages)
+{
+       struct page **page_array = *pages;
+
+       if (!page_array)
+               return;
+
+       unpin_user_pages(page_array, npages);
+       kvfree(page_array);
+       *pages = NULL;
+}
+
+struct page **io_pin_pages(unsigned long uaddr, unsigned long len, int *npages)
+{
+       unsigned long start, end, nr_pages;
+       struct page **pages;
+       int ret;
+
+       end = (uaddr + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
+       start = uaddr >> PAGE_SHIFT;
+       nr_pages = end - start;
+       if (WARN_ON_ONCE(!nr_pages))
+               return ERR_PTR(-EINVAL);
+
+       pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
+       if (!pages)
+               return ERR_PTR(-ENOMEM);
+
+       ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
+                                       pages);
+       /* success, mapped all pages */
+       if (ret == nr_pages) {
+               *npages = nr_pages;
+               return pages;
+       }
+
+       /* partial map, or didn't map anything */
+       if (ret >= 0) {
+               /* if we did partial map, release any pages we did get */
+               if (ret)
+                       unpin_user_pages(pages, ret);
+               ret = -EFAULT;
+       }
+       kvfree(pages);
+       return ERR_PTR(ret);
+}
+
+void *__io_uaddr_map(struct page ***pages, unsigned short *npages,
+                    unsigned long uaddr, size_t size)
+{
+       struct page **page_array;
+       unsigned int nr_pages;
+       void *page_addr;
+
+       *npages = 0;
+
+       if (uaddr & (PAGE_SIZE - 1) || !size)
+               return ERR_PTR(-EINVAL);
+
+       nr_pages = 0;
+       page_array = io_pin_pages(uaddr, size, &nr_pages);
+       if (IS_ERR(page_array))
+               return page_array;
+
+       page_addr = vmap(page_array, nr_pages, VM_MAP, PAGE_KERNEL);
+       if (page_addr) {
+               *pages = page_array;
+               *npages = nr_pages;
+               return page_addr;
+       }
+
+       io_pages_free(&page_array, nr_pages);
+       return ERR_PTR(-ENOMEM);
+}
+
+static void *io_uring_validate_mmap_request(struct file *file, loff_t pgoff,
+                                           size_t sz)
+{
+       struct io_ring_ctx *ctx = file->private_data;
+       loff_t offset = pgoff << PAGE_SHIFT;
+
+       switch ((pgoff << PAGE_SHIFT) & IORING_OFF_MMAP_MASK) {
+       case IORING_OFF_SQ_RING:
+       case IORING_OFF_CQ_RING:
+               /* Don't allow mmap if the ring was setup without it */
+               if (ctx->flags & IORING_SETUP_NO_MMAP)
+                       return ERR_PTR(-EINVAL);
+               return ctx->rings;
+       case IORING_OFF_SQES:
+               /* Don't allow mmap if the ring was setup without it */
+               if (ctx->flags & IORING_SETUP_NO_MMAP)
+                       return ERR_PTR(-EINVAL);
+               return ctx->sq_sqes;
+       case IORING_OFF_PBUF_RING: {
+               struct io_buffer_list *bl;
+               unsigned int bgid;
+               void *ptr;
+
+               bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
+               bl = io_pbuf_get_bl(ctx, bgid);
+               if (IS_ERR(bl))
+                       return bl;
+               ptr = bl->buf_ring;
+               io_put_bl(ctx, bl);
+               return ptr;
+               }
+       }
+
+       return ERR_PTR(-EINVAL);
+}
+
+int io_uring_mmap_pages(struct io_ring_ctx *ctx, struct vm_area_struct *vma,
+                       struct page **pages, int npages)
+{
+       unsigned long nr_pages = npages;
+
+       vm_flags_set(vma, VM_DONTEXPAND);
+       return vm_insert_pages(vma, vma->vm_start, pages, &nr_pages);
+}
+
+#ifdef CONFIG_MMU
+
+__cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
+{
+       struct io_ring_ctx *ctx = file->private_data;
+       size_t sz = vma->vm_end - vma->vm_start;
+       long offset = vma->vm_pgoff << PAGE_SHIFT;
+       void *ptr;
+
+       ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
+       if (IS_ERR(ptr))
+               return PTR_ERR(ptr);
+
+       switch (offset & IORING_OFF_MMAP_MASK) {
+       case IORING_OFF_SQ_RING:
+       case IORING_OFF_CQ_RING:
+               return io_uring_mmap_pages(ctx, vma, ctx->ring_pages,
+                                               ctx->n_ring_pages);
+       case IORING_OFF_SQES:
+               return io_uring_mmap_pages(ctx, vma, ctx->sqe_pages,
+                                               ctx->n_sqe_pages);
+       case IORING_OFF_PBUF_RING:
+               return io_pbuf_mmap(file, vma);
+       }
+
+       return -EINVAL;
+}
+
+unsigned long io_uring_get_unmapped_area(struct file *filp, unsigned long addr,
+                                        unsigned long len, unsigned long pgoff,
+                                        unsigned long flags)
+{
+       void *ptr;
+
+       /*
+        * Do not allow to map to user-provided address to avoid breaking the
+        * aliasing rules. Userspace is not able to guess the offset address of
+        * kernel kmalloc()ed memory area.
+        */
+       if (addr)
+               return -EINVAL;
+
+       ptr = io_uring_validate_mmap_request(filp, pgoff, len);
+       if (IS_ERR(ptr))
+               return -ENOMEM;
+
+       /*
+        * Some architectures have strong cache aliasing requirements.
+        * For such architectures we need a coherent mapping which aliases
+        * kernel memory *and* userspace memory. To achieve that:
+        * - use a NULL file pointer to reference physical memory, and
+        * - use the kernel virtual address of the shared io_uring context
+        *   (instead of the userspace-provided address, which has to be 0UL
+        *   anyway).
+        * - use the same pgoff which the get_unmapped_area() uses to
+        *   calculate the page colouring.
+        * For architectures without such aliasing requirements, the
+        * architecture will return any suitable mapping because addr is 0.
+        */
+       filp = NULL;
+       flags |= MAP_SHARED;
+       pgoff = 0;      /* has been translated to ptr above */
+#ifdef SHM_COLOUR
+       addr = (uintptr_t) ptr;
+       pgoff = addr >> PAGE_SHIFT;
+#else
+       addr = 0UL;
+#endif
+       return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
+}
+
+#else /* !CONFIG_MMU */
+
+int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
+{
+       return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
+}
+
+unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
+{
+       return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
+}
+
+unsigned long io_uring_get_unmapped_area(struct file *file, unsigned long addr,
+                                        unsigned long len, unsigned long pgoff,
+                                        unsigned long flags)
+{
+       void *ptr;
+
+       ptr = io_uring_validate_mmap_request(file, pgoff, len);
+       if (IS_ERR(ptr))
+               return PTR_ERR(ptr);
+
+       return (unsigned long) ptr;
+}
+
+#endif /* !CONFIG_MMU */
diff --git a/io_uring/memmap.h b/io_uring/memmap.h
new file mode 100644 (file)
index 0000000..5cec5b7
--- /dev/null
@@ -0,0 +1,25 @@
+#ifndef IO_URING_MEMMAP_H
+#define IO_URING_MEMMAP_H
+
+struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages);
+void io_pages_free(struct page ***pages, int npages);
+int io_uring_mmap_pages(struct io_ring_ctx *ctx, struct vm_area_struct *vma,
+                       struct page **pages, int npages);
+
+void *io_pages_map(struct page ***out_pages, unsigned short *npages,
+                  size_t size);
+void io_pages_unmap(void *ptr, struct page ***pages, unsigned short *npages,
+                   bool put_pages);
+
+void *__io_uaddr_map(struct page ***pages, unsigned short *npages,
+                    unsigned long uaddr, size_t size);
+
+#ifndef CONFIG_MMU
+unsigned int io_uring_nommu_mmap_capabilities(struct file *file);
+#endif
+unsigned long io_uring_get_unmapped_area(struct file *file, unsigned long addr,
+                                        unsigned long len, unsigned long pgoff,
+                                        unsigned long flags);
+int io_uring_mmap(struct file *file, struct vm_area_struct *vma);
+
+#endif
index 8a34181c97abe399fffbaf3cc5fbbfaa2fcb0987..65417c9553b1d87a75d333c17b3859d5a0a0ea6b 100644 (file)
@@ -16,6 +16,7 @@
 #include "alloc_cache.h"
 #include "openclose.h"
 #include "rsrc.h"
+#include "memmap.h"
 
 struct io_rsrc_update {
        struct file                     *file;