io_uring/kbuf: use vm_insert_pages() for mmap'ed pbuf ring

Rather than use remap_pfn_range() for this and manually free later,
switch to using vm_insert_page() and have it Just Work.

This requires a bit of effort on the mmap lookup side, as the ctx
uring_lock isn't held, which  otherwise protects buffer_lists from being
torn down, and it's not safe to grab from mmap context that would
introduce an ABBA deadlock between the mmap lock and the ctx uring_lock.
Instead, lookup the buffer_list under RCU, as the the list is RCU freed
already. Use the existing reference count to determine whether it's
possible to safely grab a reference to it (eg if it's not zero already),
and drop that reference when done with the mapping. If the mmap
reference is the last one, the buffer_list and the associated memory can
go away, since the vma insertion has references to the inserted pages at
that point.

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

1 files changed, 26 insertions, 108 deletions

diff --git a/io_uring/kbuf.c b/io_uring/kbuf.c
index 4289f4a92693..820ac599d003 100644
--- a/io_uring/kbuf.c
+++ b/io_uring/kbuf.c
@@ -32,25 +32,12 @@ struct io_provide_buf {
 	__u16				bid;
 };
 
-struct io_buf_free {
-	struct hlist_node		list;
-	void				*mem;
-	size_t				size;
-	int				inuse;
-};
-
-static inline struct io_buffer_list *__io_buffer_get_list(struct io_ring_ctx *ctx,
-							  unsigned int bgid)
-{
-	return xa_load(&ctx->io_bl_xa, bgid);
-}
-
 static inline struct io_buffer_list *io_buffer_get_list(struct io_ring_ctx *ctx,
 							unsigned int bgid)
 {
 	lockdep_assert_held(&ctx->uring_lock);
 
-	return __io_buffer_get_list(ctx, bgid);
+	return xa_load(&ctx->io_bl_xa, bgid);
 }
 
 static int io_buffer_add_list(struct io_ring_ctx *ctx,
@@ -191,24 +178,6 @@ void __user *io_buffer_select(struct io_kiocb *req, size_t *len,
 	return ret;
 }
 
-/*
- * Mark the given mapped range as free for reuse
- */
-static void io_kbuf_mark_free(struct io_ring_ctx *ctx, struct io_buffer_list *bl)
-{
-	struct io_buf_free *ibf;
-
-	hlist_for_each_entry(ibf, &ctx->io_buf_list, list) {
-		if (bl->buf_ring == ibf->mem) {
-			ibf->inuse = 0;
-			return;
-		}
-	}
-
-	/* can't happen... */
-	WARN_ON_ONCE(1);
-}
-
 static int __io_remove_buffers(struct io_ring_ctx *ctx,
 			       struct io_buffer_list *bl, unsigned nbufs)
 {
@@ -220,23 +189,16 @@ static int __io_remove_buffers(struct io_ring_ctx *ctx,
 
 	if (bl->is_buf_ring) {
 		i = bl->buf_ring->tail - bl->head;
-		if (bl->is_mmap) {
-			/*
-			 * io_kbuf_list_free() will free the page(s) at
-			 * ->release() time.
-			 */
-			io_kbuf_mark_free(ctx, bl);
-			bl->buf_ring = NULL;
-			bl->is_mmap = 0;
-		} else if (bl->buf_nr_pages) {
+		if (bl->buf_nr_pages) {
 			int j;
 
-			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;
+			if (!bl->is_mmap) {
+				for (j = 0; j < bl->buf_nr_pages; j++)
+					unpin_user_page(bl->buf_pages[j]);
+			}
+			io_pages_unmap(bl->buf_ring, &bl->buf_pages,
+					&bl->buf_nr_pages, bl->is_mmap);
+			bl->is_mmap = 0;
 		}
 		/* make sure it's seen as empty */
 		INIT_LIST_HEAD(&bl->buf_list);
@@ -537,63 +499,18 @@ error_unpin:
 	return ret;
 }
 
-/*
- * See if we have a suitable region that we can reuse, rather than allocate
- * both a new io_buf_free and mem region again. We leave it on the list as
- * even a reused entry will need freeing at ring release.
- */
-static struct io_buf_free *io_lookup_buf_free_entry(struct io_ring_ctx *ctx,
-						    size_t ring_size)
-{
-	struct io_buf_free *ibf, *best = NULL;
-	size_t best_dist;
-
-	hlist_for_each_entry(ibf, &ctx->io_buf_list, list) {
-		size_t dist;
-
-		if (ibf->inuse || ibf->size < ring_size)
-			continue;
-		dist = ibf->size - ring_size;
-		if (!best || dist < best_dist) {
-			best = ibf;
-			if (!dist)
-				break;
-			best_dist = dist;
-		}
-	}
-
-	return best;
-}
-
 static int io_alloc_pbuf_ring(struct io_ring_ctx *ctx,
 			      struct io_uring_buf_reg *reg,
 			      struct io_buffer_list *bl)
 {
-	struct io_buf_free *ibf;
 	size_t ring_size;
-	void *ptr;
 
 	ring_size = reg->ring_entries * sizeof(struct io_uring_buf_ring);
 
-	/* Reuse existing entry, if we can */
-	ibf = io_lookup_buf_free_entry(ctx, ring_size);
-	if (!ibf) {
-		ptr = io_mem_alloc(ring_size);
-		if (IS_ERR(ptr))
-			return PTR_ERR(ptr);
-
-		/* Allocate and store deferred free entry */
-		ibf = kmalloc(sizeof(*ibf), GFP_KERNEL_ACCOUNT);
-		if (!ibf) {
-			io_mem_free(ptr);
-			return -ENOMEM;
-		}
-		ibf->mem = ptr;
-		ibf->size = ring_size;
-		hlist_add_head(&ibf->list, &ctx->io_buf_list);
-	}
-	ibf->inuse = 1;
-	bl->buf_ring = ibf->mem;
+	bl->buf_ring = io_pages_map(&bl->buf_pages, &bl->buf_nr_pages, ring_size);
+	if (!bl->buf_ring)
+		return -ENOMEM;
+
 	bl->is_buf_ring = 1;
 	bl->is_mmap = 1;
 	return 0;
@@ -741,18 +658,19 @@ struct io_buffer_list *io_pbuf_get_bl(struct io_ring_ctx *ctx,
 	return ERR_PTR(-EINVAL);
 }
 
-/*
- * Called at or after ->release(), free the mmap'ed buffers that we used
- * for memory mapped provided buffer rings.
- */
-void io_kbuf_mmap_list_free(struct io_ring_ctx *ctx)
+int io_pbuf_mmap(struct file *file, struct vm_area_struct *vma)
 {
-	struct io_buf_free *ibf;
-	struct hlist_node *tmp;
+	struct io_ring_ctx *ctx = file->private_data;
+	loff_t pgoff = vma->vm_pgoff << PAGE_SHIFT;
+	struct io_buffer_list *bl;
+	int bgid, ret;
 
-	hlist_for_each_entry_safe(ibf, tmp, &ctx->io_buf_list, list) {
-		hlist_del(&ibf->list);
-		io_mem_free(ibf->mem);
-		kfree(ibf);
-	}
+	bgid = (pgoff & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
+	bl = io_pbuf_get_bl(ctx, bgid);
+	if (IS_ERR(bl))
+		return PTR_ERR(bl);
+
+	ret = io_uring_mmap_pages(ctx, vma, bl->buf_pages, bl->buf_nr_pages);
+	io_put_bl(ctx, bl);
+	return ret;
 }