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

[Laurent Vivier <lvivier@redhat.com>]

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