[dpdk-dev] [PATCH 4/4] vhost: avoid populate guest memory

[Maxime Coquelin]

On 02/22/2018 03:42 AM, Tan, Jianfeng wrote:
> Hi Maxime,
> 
>> -----Original Message-----
>> From: Maxime Coquelin [mailto:maxime.coquelin at redhat.com]
>> Sent: Tuesday, February 20, 2018 4:45 AM
>> To: Tan, Jianfeng; dev at dpdk.org
>> Cc: yliu at fridaylinux.org
>> Subject: Re: [PATCH 4/4] vhost: avoid populate guest memory
>>
>> Hi Jianfeng,
>>
>> On 02/14/2018 05:01 AM, Jianfeng Tan wrote:
>>> It's not necessary to polulate guest memory from vhost side.
>>>
>>> Cc: maxime.coquelin at redhat.com
>>> Cc: yliu at fridaylinux.org
>>>
>>> Signed-off-by: Jianfeng Tan <jianfeng.tan at intel.com>
>>> ---
>>>    lib/librte_vhost/vhost_user.c | 4 +++-
>>>    1 file changed, 3 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/lib/librte_vhost/vhost_user.c b/lib/librte_vhost/vhost_user.c
>>> index 90ed211..9bd0391 100644
>>> --- a/lib/librte_vhost/vhost_user.c
>>> +++ b/lib/librte_vhost/vhost_user.c
>>> @@ -644,6 +644,7 @@ vhost_user_set_mem_table(struct virtio_net *dev,
>> struct VhostUserMsg *pmsg)
>>>    	uint64_t mmap_offset;
>>>    	uint64_t alignment;
>>>    	uint32_t i;
>>> +	int populate;
>>>    	int fd;
>>>
>>>    	if (dev->mem && !vhost_memory_changed(&memory, dev->mem))
>> {
>>> @@ -714,8 +715,9 @@ vhost_user_set_mem_table(struct virtio_net *dev,
>> struct VhostUserMsg *pmsg)
>>>    		}
>>>    		mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
>>>
>>> +		populate = (dev->dequeue_zero_copy) ? MAP_POPULATE :
>> 0;
>>>    		mmap_addr = mmap(NULL, mmap_size, PROT_READ |
>> PROT_WRITE,
>>> -				 MAP_SHARED | MAP_POPULATE, fd, 0);
>>> +				 MAP_SHARED | populate, fd, 0);
>>>
>>>    		if (mmap_addr == MAP_FAILED) {
>>>    			RTE_LOG(ERR, VHOST_CONFIG,
>>>
>>
>> Wouldn't not populating all the guest memory have a bad impact on 0%
>> acceptable loss use-cases?
> 
> Yes, it could affect such use case; but we can address that by warming up the system a little bit, can't we?

I'm not sure this is a good idea to ask the real user to warm-up the
system.

Also, even with benchmarking, the loss happens when the queues are full,
so it is likely that it happens with buffers not used before, even if
system has been warmed-up.

>  From a good point of view, it could save the memory for VMs without pre-allocating.

What could be done is maybe to have an EAL API for mmaping, with an
associated EAL parameter to state whether it want populating or not.
This option would be disabled by default.

Does that sounds reasonable?

Cheers,
Maxime

> Thanks,
> Jianfeng
>