On Wed, Sep 8, 2021 at 12:58 PM David Howells <dhowells@xxxxxxxxxx> wrote:
>
> The afs_read objects created by afs_req_issue_op() get leaked because
> afs_alloc_read() returns a ref and then afs_fetch_data() gets its own ref
> which is released when the operation completes, but the initial ref is
> never released.
>
> Fix this by discarding the initial ref at the end of afs_req_issue_op().
>
> This leak also covered another bug whereby a ref isn't got on the key
> attached to the read record by afs_req_issue_op(). This isn't a problem as
> long as the afs_read req never goes away...
>
> Fix this by calling key_get() in afs_req_issue_op().
>
> This was found by the generic/074 test. It leaks a bunch of kmalloc-192
> objects each time it is run, which can be observed by watching
> /proc/slabinfo.
>
> Fixes: f7605fa869cf ("afs: Fix leak of afs_read objects")
> Reported-by: Marc Dionne <marc.dionne@xxxxxxxxxxxx>
> Signed-off-by: David Howells <dhowells@xxxxxxxxxx>
> cc: linux-afs@xxxxxxxxxxxxxxxxxxx
> Link: https://lore.kernel.org/r/163010394740.3035676.8516846193899793357.stgit@xxxxxxxxxxxxxxxxxxxxxx/
> ---
>
> fs/afs/file.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/fs/afs/file.c b/fs/afs/file.c
> index db035ae2a134..6688fff14b0b 100644
> --- a/fs/afs/file.c
> +++ b/fs/afs/file.c
> @@ -295,7 +295,7 @@ static void afs_req_issue_op(struct netfs_read_subrequest *subreq)
> fsreq->subreq = subreq;
> fsreq->pos = subreq->start + subreq->transferred;
> fsreq->len = subreq->len - subreq->transferred;
> - fsreq->key = subreq->rreq->netfs_priv;
> + fsreq->key = key_get(subreq->rreq->netfs_priv);
> fsreq->vnode = vnode;
> fsreq->iter = &fsreq->def_iter;
>
> @@ -304,6 +304,7 @@ static void afs_req_issue_op(struct netfs_read_subrequest *subreq)
> fsreq->pos, fsreq->len);
>
> afs_fetch_data(fsreq->vnode, fsreq);
> + afs_put_read(fsreq);
> }
>
> static int afs_symlink_readpage(struct page *page)
Tested that it prevents the leak of about 49K kmalloc-192 objects for
a run of generic/074.
Reviewed-and-tested-by: Marc Dionne <marc.dionne@xxxxxxxxxxxx>
Marc
