lkml.org 
[lkml]   [2020]   [Apr]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] nvme: prevent double free in nvme_alloc_ns() error handling
On 28.04.2020 07:49, Niklas Cassel wrote:
>On Tue, Apr 28, 2020 at 09:06:51AM +0200, Javier González wrote:
>> CAUTION: This email originated from outside of Western Digital. Do not click on links or open attachments unless you recognize the sender and know that the content is safe.
>>
>>
>> On 27.04.2020 18:22, Niklas Cassel wrote:
>> > On Mon, Apr 27, 2020 at 08:03:11PM +0200, Javier González 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@wdc.com>
>> > > > ---
>> > > > 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).
>
>Hello Javier,
>
>
>If I understand correctly, you are suggesting this:
>
>--- a/drivers/nvme/host/core.c
>+++ b/drivers/nvme/host/core.c
>@@ -3616,7 +3616,6 @@ static void 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);
>
>@@ -3628,6 +3627,8 @@ static void nvme_alloc_ns(struct nvme_ctrl *ctrl, unsigned nsid)
> }
> }
>
>+ ns->disk = disk;
>+
> down_write(&ctrl->namespaces_rwsem);
> list_add_tail(&ns->list, &ctrl->namespaces);
> up_write(&ctrl->namespaces_rwsem);
>@@ -3642,7 +3643,7 @@ static void nvme_alloc_ns(struct nvme_ctrl *ctrl, unsigned nsid)
>
> return;
> out_put_disk:
>- put_disk(ns->disk);
>+ put_disk(disk);
> out_unlink_ns:
> mutex_lock(&ctrl->subsys->lock);
> list_del_rcu(&ns->siblings);
>
>
>That would not solve the double free error when the registration fails,
>since disk->queue is still set, so this call will still free the queue.
>
>Later in the cleanup code blk_cleanup_queue(ns->queue);
>is called, which will still try to free what is already freed.
>
>To remove the double free by delaying assignments, we would need to
>delay the assignment of disk->queue until after the LightNVM entry point,
>but I don't think that is possible, since __nvme_revalidate_disk(), which
>is called before the LightNVM entry point uses disk->queue.
>
>Christoph said that he would clean up this mess with better block layer
>APIs. So perhaps the fix that is already queued is good enough until then.
>

Makes sense. I see some things have changed in this path since I last
looked at it.

Thanks!
Javier

\
 
 \ /
  Last update: 2020-04-28 12:34    [W:0.045 / U:1.136 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site