Re: [PATCH] nvme: prevent double free in nvme_alloc_ns() error handling

[Javier GonzÃlez]

On 27.04.2020 18:22, Niklas Cassel wrote:
> On Mon, Apr 27, 2020 at 08:03:11PM +0200, Javier GonzÃlez wrote:
> > On 27.04.2020 14:34, Niklas Cassel wrote:
> > > When jumping to the out_put_disk label, we will call put_disk(), which will
> > > trigger a call to disk_release(), which calls blk_put_queue().
> > >
> > > Later in the cleanup code, we do blk_cleanup_queue(), which will also call
> > > blk_put_queue().
> > >
> > > Putting the queue twice is incorrect, and will generate a KASAN splat.
> > >
> > > Set the disk->queue pointer to NULL, before calling put_disk(), so that the
> > > first call to blk_put_queue() will not free the queue.
> > >
> > > The second call to blk_put_queue() uses another pointer to the same queue,
> > > so this call will still free the queue.
> > >
> > > Fixes: 85136c010285 ("lightnvm: simplify geometry enumeration")
> > > Signed-off-by: Niklas Cassel <niklas.cassel@xxxxxxx>
> > > ---
> > > drivers/nvme/host/core.c | 2 ++
> > > 1 file changed, 2 insertions(+)
> > >
> > > diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
> > > index 91c1bd659947..f2adea96b04c 100644
> > > --- a/drivers/nvme/host/core.c
> > > +++ b/drivers/nvme/host/core.c
> > > @@ -3642,6 +3642,8 @@ static void nvme_alloc_ns(struct nvme_ctrl *ctrl, unsigned nsid)
> > >
> > >       return;
> > >  out_put_disk:
> > > +      /* prevent double queue cleanup */
> > > +      ns->disk->queue = NULL;
> > >       put_disk(ns->disk);
> > >  out_unlink_ns:
> > >       mutex_lock(&ctrl->subsys->lock);
> > > --
> > > 2.25.3
> > >
> > What about delaying the assignment of ns->disk?
> >
> > diff --git i/drivers/nvme/host/core.c w/drivers/nvme/host/core.c
> > index a4d8c90ee7cc..6da4a9ced945 100644
> > --- i/drivers/nvme/host/core.c
> > +++ w/drivers/nvme/host/core.c
> > @@ -3541,7 +3541,6 @@ static int nvme_alloc_ns(struct nvme_ctrl *ctrl, unsigned nsid)
> >         disk->queue = ns->queue;
> >         disk->flags = flags;
> >         memcpy(disk->disk_name, disk_name, DISK_NAME_LEN);
> > -       ns->disk = disk;
> >
> >         __nvme_revalidate_disk(disk, id);
> >
> > @@ -3553,6 +3552,8 @@ static int nvme_alloc_ns(struct nvme_ctrl *ctrl, unsigned nsid)
> >                 }
> >         }
> >
> > +       ns->disk = disk;
> > +
>
> Hello Javier!
>
>
> The only case where we jump to the out_put_disk label, is if the
> nvme_nvm_register() call failed.
>
> In that case, we want to undo the alloc_disk_node() operation, i.e.,
> decrease the refcount.
>
> If we don't set "ns->disk = disk;" before the call to nvme_nvm_register(),
> then, if register fails, and we jump to the put_disk(ns->disk) label,
> ns->disk will be NULL, so the recount will not be decreased, so I assume
> that this memory would then be a memory leak.
>
>
> I think that the problem is that the block functions are a bit messy.
> Most drivers seem to do blk_cleanup_queue() first and then do put_disk(),
> but some drivers do it in the opposite way, so I think that we might have
> some more use-after-free bugs in some of these drivers that do it in the
> opposite way.

Hi Niklas,

Yes, the out_put_disk label was introduced at the same time as the
LightNVM entry point. We can do a better job at separating the cleanup
functions, but as far as I can see ns->disk is not used in the LightNVM
initialization, so delaying the initialization should be ok. Part of
this should be also changing the out_put_disk to put_disk(disk).

Note that initializing other namespace types here do not require
ns->disk either, so delaying initialization should be ok. We have been
running with this patch locally for some time.

This said, this is an alternative as your fix works.

Javier