Re: [PATCH 1/4] hwrng: virtio - add an internal buffer

[Laurent Vivier <lvivier@redhat.com>]

On 05/10/2021 13:55, Michael S. Tsirkin wrote:
> On Thu, Sep 23, 2021 at 09:34:18AM +0200, Laurent Vivier wrote:
>> On 23/09/2021 09:04, Michael S. Tsirkin wrote:
>>> On Thu, Sep 23, 2021 at 08:26:06AM +0200, Laurent Vivier wrote:
>>>> On 22/09/2021 21:02, Michael S. Tsirkin wrote:
>>>>> On Wed, Sep 22, 2021 at 07:09:00PM +0200, Laurent Vivier wrote:
>>>>>> hwrng core uses two buffers that can be mixed in the
>>>>>> virtio-rng queue.
>>>>>>
>>>>>> If the buffer is provided with wait=0 it is enqueued in the
>>>>>> virtio-rng queue but unused by the caller.
>>>>>> On the next call, core provides another buffer but the
>>>>>> first one is filled instead and the new one queued.
>>>>>> And the caller reads the data from the new one that is not
>>>>>> updated, and the data in the first one are lost.
>>>>>>
>>>>>> To avoid this mix, virtio-rng needs to use its own unique
>>>>>> internal buffer at a cost of a data copy to the caller buffer.
>>>>>>
>>>>>> Signed-off-by: Laurent Vivier <lvivier@redhat.com>
>>>>>> ---
>>>>>>     drivers/char/hw_random/virtio-rng.c | 43 ++++++++++++++++++++++-------
>>>>>>     1 file changed, 33 insertions(+), 10 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/char/hw_random/virtio-rng.c b/drivers/char/hw_random/virtio-rng.c
>>>>>> index a90001e02bf7..208c547dcac1 100644
>>>>>> --- a/drivers/char/hw_random/virtio-rng.c
>>>>>> +++ b/drivers/char/hw_random/virtio-rng.c
>>>>>> @@ -18,13 +18,20 @@ static DEFINE_IDA(rng_index_ida);
>>>>>>     struct virtrng_info {
>>>>>>     	struct hwrng hwrng;
>>>>>>     	struct virtqueue *vq;
>>>>>> -	struct completion have_data;
>>>>>>     	char name[25];
>>>>>> -	unsigned int data_avail;
>>>>>>     	int index;
>>>>>>     	bool busy;
>>>>>>     	bool hwrng_register_done;
>>>>>>     	bool hwrng_removed;
>>>>>> +	/* data transfer */
>>>>>> +	struct completion have_data;
>>>>>> +	unsigned int data_avail;
>>>>>> +	/* minimal size returned by rng_buffer_size() */
>>>>>> +#if SMP_CACHE_BYTES < 32
>>>>>> +	u8 data[32];
>>>>>> +#else
>>>>>> +	u8 data[SMP_CACHE_BYTES];
>>>>>> +#endif
>>>>>
>>>>> Let's move this logic to a macro in hw_random.h ?
>>>>>
>>>>>>     };
>>>>>>     static void random_recv_done(struct virtqueue *vq)
>>>>>> @@ -39,14 +46,14 @@ static void random_recv_done(struct virtqueue *vq)
>>>>>>     }
>>>>>>     /* The host will fill any buffer we give it with sweet, sweet randomness. */
>>>>>> -static void register_buffer(struct virtrng_info *vi, u8 *buf, size_t size)
>>>>>> +static void register_buffer(struct virtrng_info *vi)
>>>>>>     {
>>>>>>     	struct scatterlist sg;
>>>>>> -	sg_init_one(&sg, buf, size);
>>>>>> +	sg_init_one(&sg, vi->data, sizeof(vi->data));
>>>>>
>>>>> Note that add_early_randomness requests less:
>>>>>            size_t size = min_t(size_t, 16, rng_buffer_size());
>>>>>
>>>>> maybe track how much was requested and grow up to sizeof(data)?
>>>>
>>>> I think this problem is managed by PATCH 3/4 as we reuse unused data of the buffer.
>>>
>>> the issue I'm pointing out is that we are requesting too much
>>> entropy from host - more than guest needs.
>>
>> Yes, guest asks for 16 bytes, but we request SMP_CACHE_BYTES (64 on x86_64),
>> and these 16 bytes are used with add_device_randomness(). With the following
>> patches, the remaining 48 bytes are used rapidly by hwgnd kthread or by the
>> next virtio_read.
>>
>> If there is no enough entropy the call is simply ignored as wait=0.
>>
>> At this patch level the call is always simply ignored (because wait=0) and
>> the data requested here are used by the next read that always asks for a
>> SMP_CACHE_BYTES bytes data size.
>>
>> Moreover in PATCH 4/4 we always have a pending request of size
>> SMP_CACHE_BYTES, so driver always asks a block of this size and the guest
>> takes what it needs.
>>
>> Originally I used a 16 bytes block but performance are divided by 4.
>>
>> Do you propose something else?
>>
>> Thanks,
>> Laurent
> 
> Maybe min(size, sizeof(vi->data))?
> 
But it means, in the case of mixed buffers, we will ask 16 bytes on the first call, not 
use it, and  ask SMP_CACHE_BYTES bytes on the next call to get only 16:

- add_early_randomness() asks for 16 bytes but wait = 0 and thus the request is queued but 
not used. add_early_randomness() is called when we switch from one hw_random backend to 
another (so generally only once...)

- hwrng_fillfn() and rng_dev_read() always ask rng_buffer_size() (max(32, SMP_CACHE_BYTES)).

So we can say we use SMP_CACHE_BYTES in 99% of the cases.

Moreover, this will be discarded by patch 3 and 4 as we have a loop to ask more data in a 
fixed size buffer.

I'm not sure it's worth introducing this change in this patch.

Thanks,
Laurent