On 03.10.2018 11:32, Ilya Maximets wrote:
> On 03.10.2018 11:02, Maxime Coquelin wrote:
>>
>>
>> On 10/03/2018 09:57 AM, Ilya Maximets wrote:
>>> On 03.10.2018 10:50, Maxime Coquelin wrote:
>>>>
>>>>
>>>> On 10/02/2018 04:15 PM, Ilya Maximets wrote:
>>>>> On 02.10.2018 12:36, Maxime Coquelin wrote:
>>>>>> Return of message handling has now changed to an enum that can
>>>>>> take non-negative value that is not zero in case a reply is
>>>>>> needed. But the code checking the variable afterwards has not
>>>>>> been updated, leading to success messages handling being
>>>>>> treated as errors.
>>>>>>
>>>>>> Fixes: 4e601952cae6 ("vhost: message handling implemented as a callback
>>>>>> array")
>>>>>>
>>>>>> Signed-off-by: Maxime Coquelin <maxime.coque...@redhat.com>
>>>>>> ---
>>>>>> lib/librte_vhost/vhost_user.c | 6 +++---
>>>>>> 1 file changed, 3 insertions(+), 3 deletions(-)
>>>>>>
>>>>>> diff --git a/lib/librte_vhost/vhost_user.c
>>>>>> b/lib/librte_vhost/vhost_user.c
>>>>>> index 7ef3fb4a4..060b41893 100644
>>>>>> --- a/lib/librte_vhost/vhost_user.c
>>>>>> +++ b/lib/librte_vhost/vhost_user.c
>>>>>> @@ -1783,7 +1783,7 @@ vhost_user_msg_handler(int vid, int fd)
>>>>>> }
>>>>>> skip_to_post_handle:
>>>>>> - if (!ret && dev->extern_ops.post_msg_handle) {
>>>>>> + if (ret != VH_RESULT_ERR && dev->extern_ops.post_msg_handle) {
>>>>>> uint32_t need_reply;
>>>>>> ret = (*dev->extern_ops.post_msg_handle)(
>>>>>> @@ -1800,10 +1800,10 @@ vhost_user_msg_handler(int vid, int fd)
>>>>>> vhost_user_unlock_all_queue_pairs(dev);
>>>>>> if (msg.flags & VHOST_USER_NEED_REPLY) {
>>>>>
>>>>> Maybe we need to reply here only if we didn't reply
>>>>> already (not VH_RESULT_REPLY) ? Otherwise, we could
>>>>> reply twice (with payload and with return code).
>>>>
>>>> Well, if the master sets this bit, it means it is waiting for
>>>> a "reply-ack", so not sending it would cause the master to wait
>>>> forever.
>>>>
>>>> It is the master responsibility to not set this bit for requests
>>>> already expecting a non "reply-ack" reply (as you fixed it for
>>>> postcopy's set mem table case).
>>>
>>> vhost-user docs in QEMU says:
>>> "
>>> For the message types that already solicit a reply from the client, the
>>> presence of VHOST_USER_PROTOCOL_F_REPLY_ACK or need_reply bit being set
>>> brings
>>> no behavioural change.
>>> "
>>> i.e. even if QEMU sets the need_reply flag, vhost should not reply twice.
>>> Am I missing something?
>>
>> Oh, right. Thanks for pointing it out.
>>
>> So coming back to the DPDK implementation, I just had a look again, and it
>> seems that we don't send a reply twice, as send_vhost_reply takes
>> care of clearing the VHOST_USER_NEED_REPLY flag.
>> Do you confirm my understanding is correct?
>
> Hmm. Yes, you're right. send_vhost_reply clears the VHOST_USER_NEED_REPLY
> flag and vhost doesn't send replies twice.
> Maybe some comment with clarifications needed here, or some more
> refactoring to make this aspect more clear.
>
Sorry. Please, ignore the text below.
> So, we have a situation where only one of the message processing stages
> is able to reply even if it's not the last stage for the message:
> 1. extern_ops.pre_msg_handle
> 2. "master"
> 3. extern_ops.post_msg_handle
> 4. result i.e. (!!ret)
>
> extern_ops API is a bit confusing from this point of view. It has
> serious restrictions for replies which are not described anywhere.
> I mean that pre and post processing stages are able to request the
> reply, but the post processing reply will be just dropped
> (or "master" reply will be dropped).
> This is, actually, not an issue until we have only one user of them,
> which uses only one of the callbacks. But maybe this API should be
> described more in comments or docs.
>
>>
>>>>
>>>>>> - msg.payload.u64 = !!ret;
>>>>>> + msg.payload.u64 = ret == VH_RESULT_ERR;
>>>>>> msg.size = sizeof(msg.payload.u64);
>>>>>> send_vhost_reply(fd, &msg);
>>>>>> - } else if (ret) {
>>>>>> + } else if (ret == VH_RESULT_ERR) {
>>>>>> RTE_LOG(ERR, VHOST_CONFIG,
>>>>>> "vhost message handling failed.\n");
>>>>>> return -1;
>>>>>>
>>>>
>>>>
>>
>>
>
>
