io_uring/kbuf: recycle freed mapped buffer ring entries

Right now we stash any potentially mmap'ed provided ring buffer range
for freeing at release time, regardless of when they get unregistered.
Since we're keeping track of these ranges anyway, keep track of their
registration state as well, and use that to recycle ranges when
appropriate rather than always allocate new ones.

The lookup is a basic scan of entries, checking for the best matching
free entry.

Fixes: c392cbecd8ec ("io_uring/kbuf: defer release of mapped buffer rings")
Signed-off-by: Jens Axboe <axboe@kernel.dk>

diff --git a/io_uring/kbuf.c b/io_uring/kbuf.c
index 85e680fc74ce2c4f5468bc02ee48e25db59399e5..325ca7f8b0a0fa848d28f33293249cc3dba2f74b 100644 (file)
--- a/io_uring/kbuf.c
+++ b/io_uring/kbuf.c
@@ -36,6 +36,8 @@ struct io_provide_buf {
 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,
@@ -216,6 +218,24 @@ static __cold int io_init_bl_list(struct io_ring_ctx *ctx)
        return 0;
 }
 
+/*
+ * 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)
 {
@@ -232,6 +252,7 @@ static int __io_remove_buffers(struct io_ring_ctx *ctx,
                         * 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) {
@@ -539,6 +560,34 @@ static int io_pin_pbuf_ring(struct io_uring_buf_reg *reg,
        return -EINVAL;
 }
 
+/*
+ * 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)
@@ -548,20 +597,26 @@ static int io_alloc_pbuf_ring(struct io_ring_ctx *ctx,
        void *ptr;
 
        ring_size = reg->ring_entries * sizeof(struct io_uring_buf_ring);
-       ptr = io_mem_alloc(ring_size);
-       if (!ptr)
-               return -ENOMEM;
 
-       /* Allocate and store deferred free entry */
-       ibf = kmalloc(sizeof(*ibf), GFP_KERNEL_ACCOUNT);
+       /* Reuse existing entry, if we can */
+       ibf = io_lookup_buf_free_entry(ctx, ring_size);
        if (!ibf) {
-               io_mem_free(ptr);
-               return -ENOMEM;
-       }
-       ibf->mem = ptr;
-       hlist_add_head(&ibf->list, &ctx->io_buf_list);
+               ptr = io_mem_alloc(ring_size);
+               if (!ptr)
+                       return -ENOMEM;
 
-       bl->buf_ring = 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->is_mapped = 1;
        bl->is_mmap = 1;
        return 0;