io_uring/kbuf: vmap pinned buffer ring

This avoids needing to care about HIGHMEM, and it makes the buffer
indexing easier as both ring provided buffer methods are now virtually
mapped in a contigious fashion.

Signed-off-by: Jens Axboe <axboe@kernel.dk>

1 files changed, 15 insertions, 24 deletions

diff --git a/io_uring/kbuf.c b/io_uring/kbuf.c
index 3aa16e27f509..4289f4a92693 100644
--- a/io_uring/kbuf.c
+++ b/io_uring/kbuf.c
@@ -7,6 +7,7 @@
 #include <linux/slab.h>
 #include <linux/namei.h>
 #include <linux/poll.h>
+#include <linux/vmalloc.h>
 #include <linux/io_uring.h>
 
 #include <uapi/linux/io_uring.h>
@@ -146,15 +147,7 @@ static void __user *io_ring_buffer_select(struct io_kiocb *req, size_t *len,
 		req->flags |= REQ_F_BL_EMPTY;
 
 	head &= bl->mask;
-	/* mmaped buffers are always contig */
-	if (bl->is_mmap || head < IO_BUFFER_LIST_BUF_PER_PAGE) {
-		buf = &br->bufs[head];
-	} else {
-		int off = head & (IO_BUFFER_LIST_BUF_PER_PAGE - 1);
-		int index = head / IO_BUFFER_LIST_BUF_PER_PAGE;
-		buf = page_address(bl->buf_pages[index]);
-		buf += off;
-	}
+	buf = &br->bufs[head];
 	if (*len == 0 || *len > buf->len)
 		*len = buf->len;
 	req->flags |= REQ_F_BUFFER_RING;
@@ -241,6 +234,7 @@ static int __io_remove_buffers(struct io_ring_ctx *ctx,
 			for (j = 0; j < bl->buf_nr_pages; j++)
 				unpin_user_page(bl->buf_pages[j]);
 			kvfree(bl->buf_pages);
+			vunmap(bl->buf_ring);
 			bl->buf_pages = NULL;
 			bl->buf_nr_pages = 0;
 		}
@@ -498,9 +492,9 @@ err:
 static int io_pin_pbuf_ring(struct io_uring_buf_reg *reg,
 			    struct io_buffer_list *bl)
 {
-	struct io_uring_buf_ring *br;
+	struct io_uring_buf_ring *br = NULL;
+	int nr_pages, ret, i;
 	struct page **pages;
-	int i, nr_pages;
 
 	pages = io_pin_pages(reg->ring_addr,
 			     flex_array_size(br, bufs, reg->ring_entries),
@@ -508,18 +502,12 @@ static int io_pin_pbuf_ring(struct io_uring_buf_reg *reg,
 	if (IS_ERR(pages))
 		return PTR_ERR(pages);
 
-	/*
-	 * Apparently some 32-bit boxes (ARM) will return highmem pages,
-	 * which then need to be mapped. We could support that, but it'd
-	 * complicate the code and slowdown the common cases quite a bit.
-	 * So just error out, returning -EINVAL just like we did on kernels
-	 * that didn't support mapped buffer rings.
-	 */
-	for (i = 0; i < nr_pages; i++)
-		if (PageHighMem(pages[i]))
-			goto error_unpin;
+	br = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
+	if (!br) {
+		ret = -ENOMEM;
+		goto error_unpin;
+	}
 
-	br = page_address(pages[0]);
 #ifdef SHM_COLOUR
 	/*
 	 * On platforms that have specific aliasing requirements, SHM_COLOUR
@@ -530,8 +518,10 @@ static int io_pin_pbuf_ring(struct io_uring_buf_reg *reg,
 	 * should use IOU_PBUF_RING_MMAP instead, and liburing will handle
 	 * this transparently.
 	 */
-	if ((reg->ring_addr | (unsigned long) br) & (SHM_COLOUR - 1))
+	if ((reg->ring_addr | (unsigned long) br) & (SHM_COLOUR - 1)) {
+		ret = -EINVAL;
 		goto error_unpin;
+	}
 #endif
 	bl->buf_pages = pages;
 	bl->buf_nr_pages = nr_pages;
@@ -543,7 +533,8 @@ error_unpin:
 	for (i = 0; i < nr_pages; i++)
 		unpin_user_page(pages[i]);
 	kvfree(pages);
-	return -EINVAL;
+	vunmap(br);
+	return ret;
 }
 
 /*