linux-kernel - Re: [PATCH 1/4] IB/core: introduce ->release() callback

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20180918144441.GH11367@ziepe.ca>
Date:   Tue, 18 Sep 2018 08:44:41 -0600
From:   Jason Gunthorpe <jgg@...pe.ca>
To:     Jan Dakinevich <jan.dakinevich@...tuozzo.com>
Cc:     Doug Ledford <dledford@...hat.com>,
        Yishai Hadas <yishaih@...lanox.com>,
        Leon Romanovsky <leon@...nel.org>,
        Parav Pandit <parav@...lanox.com>,
        Mark Bloch <markb@...lanox.com>,
        Daniel Jurgens <danielj@...lanox.com>,
        Kees Cook <keescook@...omium.org>,
        Kamal Heib <kamalheib1@...il.com>,
        Bart Van Assche <bvanassche@....org>,
        linux-rdma@...r.kernel.org, linux-kernel@...r.kernel.org,
        Denis Lunev <den@...tuozzo.com>,
        Konstantin Khorenko <khorenko@...tuozzo.com>
Subject: Re: [PATCH 1/4] IB/core: introduce ->release() callback

On Tue, Sep 18, 2018 at 04:03:43PM +0300, Jan Dakinevich wrote:
> IB infrastructure shares common device instance constructor with
> reference counting, and it uses kzalloc() to allocate memory
> for device specific instance with incapsulated ib_device field as one
> contigous memory block.
> 
> The issue is that the device specific instances tend to be too large
> and require high page order memory allocation. Unfortunately, kzalloc()
> in ib_alloc_device() can not be replaced with kvzalloc() since it would
> require a lot of review in all IB driver to prove correctness of the
> replacement.
> 
> The driver can allocate some heavy partes of their instance for itself
> and keep pointers for them in own instance. For this it is important
> that the alocated parts have the same life time as ib_device, thus
> their deallocation should be based on the same reference counting.
> 
> Let suppose:
> 
> struct foo_ib_device {
> 	struct ib_device device;
> 
> 	void *part;
> 
> 	...
> };
> 
> To properly free memory from .foo_ib_part the driver should provide
> function for ->release() callback:
> 
> void foo_ib_release(struct ib_device *device)
> {
> 	struct foo_ib_device *foo = container_of(device,  struct foo_ib_device,
> 						 device);
> 
> 	kvfree(foo->part);
> }
> 
> ...and initialiaze this callback immediately after foo_ib_device
> instance allocation.
> 
> 	struct foo_ib_device *foo;
> 
> 	foo = ib_alloc_device(sizeof(struct foo_ib_device));
> 
> 	foo->device.release = foo_ib_release;
> 
> 	/* allocate parts */
> 	foo->part = kvmalloc(65536, GFP_KERNEL);
> 
> Signed-off-by: Jan Dakinevich <jan.dakinevich@...tuozzo.com>
>  drivers/infiniband/core/device.c | 2 ++
>  include/rdma/ib_verbs.h          | 2 ++
>  2 files changed, 4 insertions(+)
> 
> diff --git a/drivers/infiniband/core/device.c b/drivers/infiniband/core/device.c
> index db3b627..a8c8b0d 100644
> +++ b/drivers/infiniband/core/device.c
> @@ -215,6 +215,8 @@ static void ib_device_release(struct device *device)
>  		ib_cache_release_one(dev);
>  		kfree(dev->port_immutable);
>  	}
> +	if (dev->release)
> +		dev->release(dev);
>  	kfree(dev);
>  }

Nope, the driver module could be unloaded at this point.

The driver should free memory after its call to ib_unregister_device
returns.

Jason