Re: [PATCHv2] virtio_mmio: fix devm cleanup

[weiping zhang]

2017-12-12 22:45 GMT+08:00 Mark Rutland <mark.rutland@xxxxxxx>:
> On Tue, Dec 12, 2017 at 10:26:24PM +0800, weiping zhang wrote:
>> 2017-12-12 21:45 GMT+08:00 Mark Rutland <mark.rutland@xxxxxxx>:
>> Hi Mark,
>
> Hi,
>
>> thanks your patch, I dig into these three devm_xxx funciton,
>> all of them represented by a struct devres as following,
>>
>> struct devres_node {
>>         struct list_head                entry;
>>         dr_release_t                    release;
>> #ifdef CONFIG_DEBUG_DEVRES
>>         const char                      *name;
>>         size_t                          size;
>> #endif
>>
>> };
>>
>> struct devres {
>>         struct devres_node              node;
>>         /* -- 3 pointers */
>>         unsigned long long              data[]; /* guarantee ull alignment */
>> };
>
>> 2) devm_kzalloc -> devm_kmalloc
>>
>> dr = alloc_dr(devm_kmalloc_release, size, gfp, dev_to_node(dev));
>> "devm_kmalloc_release" is noop, do nothing.
>
> Please note that the release function is there to perform cleanup prior
> to the devm infrastructure releasing the memory.
>
> The devm_kmalloc_release function is a no-op since nothing has to be
> done prior to memory being freed, but the memory itself is still freed.
>
> In alloc_dr(), the struct devres is allocated together with the memory,
> since alloc_dr() does:
>
>         size_t tot_size = sizeof(struct devres) + size;
>         struct devres *dr;
>
>         dr = kmalloc_node_track_caller(tot_size, gfp, nid);
>
>         return dr->data;
>
> ... where dr->data points at the memory after the struct devres.
>
> Later, in release_nodes() we do:
>
>         list_for_each_entry_safe_reverse(dr, tmp, &todo, node.entry) {
>                 devres_log(dev, &dr->node, "REL");
>                 dr->node.release(dev, dr->data);
>                 kfree(dr);
>         }
>
> ... which will invoke the no-op devm_kmalloc_release, then free the
> devres allocation, including the dr->data memory the user requested.
>
>> so for case 2) above, we need a devm_kfree() before call
>> register_virtio_device
>
> As above, I do not believe that is the case.
>
Oh I see, thanks your detail explanation. Thanks a lot.
> Thanks,
> Mark.