Re: [dpdk-dev] [PATCH v2 3/4] eal: add synchronous multi-process communication

["Ananyev, Konstantin" <konstantin.ananyev@intel.com>]

> 
> 
> On 1/17/2018 6:50 PM, Ananyev, Konstantin wrote:
> >
> >>> Hi Jianfeng,
> >>>
> >>>> -----Original Message-----
> >>>> From: Tan, Jianfeng
> >>>> Sent: Tuesday, January 16, 2018 8:11 AM
> >>>> To: Ananyev, Konstantin <konstantin.ananyev@intel.com>; dev@dpdk.org; Burakov, Anatoly <anatoly.burakov@intel.com>
> >>>> Cc: Richardson, Bruce <bruce.richardson@intel.com>; thomas@monjalon.net
> >>>> Subject: Re: [PATCH v2 3/4] eal: add synchronous multi-process communication
> >>>>
> >>>> Thank you, Konstantin and Anatoly firstly. Other comments are well
> >>>> received and I'll send out a new version.
> >>>>
> >>>>
> >>>> On 1/16/2018 8:00 AM, Ananyev, Konstantin wrote:
> >>>>>> We need the synchronous way for multi-process communication,
> >>>>>> i.e., blockingly waiting for reply message when we send a request
> >>>>>> to the peer process.
> >>>>>>
> >>>>>> We add two APIs rte_eal_mp_request() and rte_eal_mp_reply() for
> >>>>>> such use case. By invoking rte_eal_mp_request(), a request message
> >>>>>> is sent out, and then it waits there for a reply message. The
> >>>>>> timeout is hard-coded 5 Sec. And the replied message will be copied
> >>>>>> in the parameters of this API so that the caller can decide how
> >>>>>> to translate those information (including params and fds). Note
> >>>>>> if a primary process owns multiple secondary processes, this API
> >>>>>> will fail.
> >>>>>>
> >>>>>> The API rte_eal_mp_reply() is always called by an mp action handler.
> >>>>>> Here we add another parameter for rte_eal_mp_t so that the action
> >>>>>> handler knows which peer address to reply.
> >>>>>>
> >>>>>> We use mutex in rte_eal_mp_request() to guarantee that only one
> >>>>>> request is on the fly for one pair of processes.
> >>>>> You don't need to do things in such strange and restrictive way.
> >>>>> Instead you can do something like that:
> >>>>> 1) Introduce new struct, list for it and mutex
> >>>>>     struct sync_request {
> >>>>>          int reply_received;
> >>>>>          char dst[PATH_MAX];
> >>>>>          char reply[...];
> >>>>>          LIST_ENTRY(sync_request) next;
> >>>>> };
> >>>>>
> >>>>> static struct
> >>>>>        LIST_HEAD(list, sync_request);
> >>>>>        pthread_mutex_t lock;
> >>>>>       pthead_cond_t cond;
> >>>>> } sync_requests;
> >>>>>
> >>>>> 2) then at request() call:
> >>>>>      Grab sync_requests.lock
> >>>>>      Check do we already have a pending request for that destination,
> >>>>>      If yes - the release the lock and returns with error.
> >>>>>      - allocate and init new sync_request struct, set reply_received=0
> >>>>>      - do send_msg()
> >>>>>      -then in a cycle:
> >>>>>      pthread_cond_timed_wait(&sync_requests.cond, &sync_request.lock, &timespec);
> >>>>>      - at return from it check if sync_request.reply_received == 1, if not
> >>>>> check if timeout expired and either return a failure or go to the start of the cycle.
> >>>>>
> >>>>> 3) at mp_handler() if REPLY received - grab sync_request.lock,
> >>>>>        search through sync_requests.list for dst[] ,
> >>>>>       if found, then set it's reply_received=1, copy the received message into reply
> >>>>>       and call pthread_cond_braodcast((&sync_requests.cond);
> >>>> The only benefit I can see is that now the sender can request to
> >>>> multiple receivers at the same time. And it makes things more
> >>>> complicated. Do we really need this?
> >>> The benefit is that one thread is blocked waiting for response,
> >>> your mp_handler can still receive and handle other messages.
> >> This can already be done in the original implementation. mp_handler
> >> listens for msg, request from the other peer(s), and replies the
> >> requests, which is not affected.
> >>
> >>> Plus as you said - other threads can keep sending messages.
> >> For this one, in the original implementation, other threads can still
> >> send msg, but not request. I suppose the request is not in a fast path,
> >> why we care to make it fast?
> >>
> > +int
> > +rte_eal_mp_request(const char *action_name,
> > +		   void *params,
> > +		   int len_p,
> > +		   int fds[],
> > +		   int fds_in,
> > +		   int fds_out)
> > +{
> > +	int i, j;
> > +	int sockfd;
> > +	int nprocs;
> > +	int ret = 0;
> > +	struct mp_msghdr *req;
> > +	struct timeval tv;
> > +	char buf[MAX_MSG_LENGTH];
> > +	struct mp_msghdr *hdr;
> > +
> > +	RTE_LOG(DEBUG, EAL, "request: %s\n", action_name);
> > +
> > +	if (fds_in > SCM_MAX_FD || fds_out > SCM_MAX_FD) {
> > +		RTE_LOG(ERR, EAL, "Cannot send more than %d FDs\n", SCM_MAX_FD);
> > +		rte_errno = -E2BIG;
> > +		return 0;
> > +	}
> > +
> > +	req = format_msg(action_name, params, len_p, fds_in, MP_REQ);
> > +	if (req == NULL)
> > +		return 0;
> > +
> > +	if ((sockfd = open_unix_fd(0)) < 0) {
> > +		free(req);
> > +		return 0;
> > +	}
> > +
> > +	tv.tv_sec = 5;  /* 5 Secs Timeout */
> > +	tv.tv_usec = 0;
> > +	if (setsockopt(sockfd, SOL_SOCKET, SO_RCVTIMEO,
> > +			(const void *)&tv, sizeof(struct timeval)) < 0)
> > +		RTE_LOG(INFO, EAL, "Failed to set recv timeout\n");
> >
> > I f you set it just for one call, why do you not restore it?
> 
> Yes, original code is buggy, I should have put it into the critical section.
> 
> Do you mean we just create once and use for ever? if yes, we could put
> the open and setting into mp_init().
> 
> > Also I don't think it is a good idea to change it here -
> > if you'll make timeout a parameter value - then it could be overwritten
> > by different threads.
> 
> For simplicity, I'm not inclined to put the timeout as an parameter
> exposing to caller. So if you agree, I'll put it into the mp_init() with
> open.

My preference would be to have timeout value on a per call basis.
For one request user would like to wait no more than 5sec,
for another one user would probably be ok to wait forever.

> 
> >
> > +
> > +	/* Only allow one req at a time */
> > +	pthread_mutex_lock(&mp_mutex_request);
> > +
> > +	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
> > +		nprocs = 0;
> > +		for (i = 0; i < MAX_SECONDARY_PROCS; ++i)
> > +			if (!mp_sec_sockets[i]) {
> > +				j = i;
> > +				nprocs++;
> > +			}
> > +
> > +		if (nprocs > 1) {
> > +			RTE_LOG(ERR, EAL,
> > +				"multi secondary processes not supported\n");
> > +			goto free_and_ret;
> > +		}
> > +
> > +		ret = send_msg(sockfd, mp_sec_sockets[j], req, fds);
> >
> > As I remember - sndmsg() is also blocking call, so under some conditions you can stall
> > there forever.
> 
>  From linux's unix_diagram_sendmsg(), we see:
>      timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);

Ok, but it would have effect only if (msg->msg_flags & MSG_DONTWAIT) != 0.
And for that, as I remember you need your socket in non-blocking mode, no?

> 
> I assume it will not block for datagram unix socket in Linux. But I'm
> not sure what it behaves in freebsd.
> 
> Anyway, better to add an explicit setsockopt() to make it not blocking.

You can't do that - at the same moment another thread might call your sendmsg()
and it might expect it to be blocking call.

> 
> > As mp_mutex_requestis still held - next rte_eal_mp_request(0 will also block forever here.
> >
> > +	} else
> > +		ret = send_msg(sockfd, eal_mp_unix_path(), req, fds);
> > +
> > +	if (ret == 0) {
> > +		RTE_LOG(ERR, EAL, "failed to send request: %s\n", action_name);
> > +		ret = -1;
> > +		goto free_and_ret;
> > +	}
> > +
> > +	ret = read_msg(sockfd, buf, MAX_MSG_LENGTH, fds, fds_out, NULL);
> >
> > if the message you receive is not a reply you are expecting -
> > it will be simply dropped - mp_handler() would never process it.
> 
> We cannot detect if it's the right reply absolutely correctly, but just
> check the action_name, which means, it still possibly gets a wrong reply
> if an action_name contains multiple requests.
> 
> Is just comparing the action_name acceptable?

As I can see the main issue here is that you can call recvmsg() from 2 different
points and they are not syncronised:
1. your mp_handler() doesn't aware about reply you are waiting and not 
have any handler associated with it.
So if mp_handler() will receive a reply it will just drop it.
2. your reply() is not aware about any other messages and associated actions -
so again it can't handle them properly (and probably shouldn't).

The simplest (and most common) way - always call recvmsg from one place -
mp_handler() and have a special action for reply msg.
As I wrote before that action will be just find the appropriate buffer provided
by reply() - copy message into it and signal thread waiting in reply() that
it can proceed.
  
Konstantin

> 
> >
> > +	if (ret > 0) {
> > +		hdr = (struct mp_msghdr *)buf;
> > +		if (hdr->len_params == len_p)
> > +			memcpy(params, hdr->params, len_p);
> > +		else {
> > +			RTE_LOG(ERR, EAL, "invalid reply\n");
> > +			ret = 0;
> > +		}
> > +	}
> > +
> > +free_and_ret:
> > +	free(req);
> > +	close(sockfd);
> > +	pthread_mutex_unlock(&mp_mutex_request);
> > +	return ret;
> > +}
> >
> > All of the above makes me think that current implementation is erroneous
> > and needs to be reworked.
> 
> Thank you for your review. I'll work on a new version.
> 
> Thanks,
> Jianfeng
> 
> > Konstantin
> >
> >