[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [libmicrohttpd] Suspend/resume with single thread and external epoll
From: |
Christian Grothoff |
Subject: |
Re: [libmicrohttpd] Suspend/resume with single thread and external epoll not sending response |
Date: |
Wed, 14 Mar 2018 05:24:02 +0100 |
User-agent: |
Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101 Thunderbird/52.6.0 |
Done: 85913afa..7d7ccbcd
Thanks for the suggestion!
-Christian
On 03/13/2018 07:09 PM, Robert D Kocisko wrote:
> Hi Christian,
>
> Thank you! Very cool to see the example working. Funny, I
> implemented MHD_get_timeout in the main loop in my project but in all
> my tests it always returned MHD_NO, so I assumed it was not applicable
> to my use case and didn't include it in the example I wrote. But I
> didn't make the connection that it *also* had to be called after
> MHD_resume_connection until your response here. Would you consider
> the attached patch? It adds one line to the docs that I'm quite sure
> would have pointed me in the right direction had it been there before.
>
> Thanks!
> Bob
>
> On Sun, Mar 11, 2018 at 1:11 PM, Christian Grothoff <address@hidden> wrote:
>>
>> Dear Bob,
>>
>> I've analyzed your code, and the issue is on your end: you simply didn't
>> set the timeout correctly. When using an external event loop, it is
>> mandatory that you ask MHD for the timeout using MHD_get_timeout() and
>> use that with select/poll/epoll. Then, you must call MHD_run() once
>> whenever epoll() returns (including timeouts!).
>>
>> In your case, MHD would have given you a timeout of 0, but you used
>> infinity instead, with predictable results...
>>
>> I've attached a corrected version of the code.
>>
>> Happy hacking!
>>
>> Christian
>>
>> On 03/02/2018 08:26 PM, Robert D Kocisko wrote:
>>> First, thanks for your amazing work on MHD!
>>>
>>> This question is a near duplicate of the 2014 message thread from Tom
>>> Cornell entitled "Trouble getting a response sent from a separate worker
>>> thread (with external select)". However, I am not using separate worker
>>> threads--everything is in one thread and so I don't think the
>>> recommendations found in that thread apply to my scenario.
>>>
>>> Basically after receiving a request from an HTTP client, I want to be
>>> able to do some asynchronous 'work' which is really just waiting on
>>> another process such as a database engine to calculate and return the
>>> result which, when complete, I will forward back to the client. This is
>>> all done in one thread using epoll, so I don't want any blocking and I
>>> don't want any busy loops. MHD's external epoll support combined with
>>> suspend/resume fits into this architecture perfectly, but there's a
>>> problem: after resuming the connection and queueing the data, the
>>> headers are sent to the client immediately, but the body of the response
>>> does not get sent until another client request arrives.
>>>
>>> Anyway, to make this all concrete, I've put together a small working
>>> example (below) which shows the problem. This is built against the
>>> latest dev rev (7f1dbb2) on elementaryOS (which is Ubuntu 16.04). Every
>>> time a request comes in it suspends the connection and starts a 1 second
>>> timer which, when it expires, resumes the connection. When the
>>> connection is resumed the response is queued (simply echos the request
>>> url). I realize this example leaks timer fds and doesn't clean up
>>> properly but it successfully demonstrates the problem.
>>>
>>> I have experimented with calling MHD_run() twice after
>>> MHD_resume_connection() rather than the once required by the docs, and
>>> that does seem to work, but that seems extremely hacky and I'm not sure
>>> if twice is enough (why twice and not three times?). I've skimmed the
>>> source code looking for obvious answers but none are readily apparent to me.
>>>
>>> At this point I'm pretty sure that this is a bug with MHD but am I
>>> missing something?
>>>
>>> Thanks!
>>> Bob Kocisko
>>>
>>> -------------------------
>>>
>>> #include "platform.h"
>>> #include <microhttpd.h>
>>> #include <sys/epoll.h>
>>> #include <sys/timerfd.h>
>>>
>>> #define TIMEOUT_INFINITE -1
>>>
>>> struct Request {
>>> struct MHD_Connection *connection;
>>> int timerfd;
>>> };
>>>
>>> int epfd;
>>> struct epoll_event evt;
>>>
>>> static int
>>> ahc_echo (void *cls,
>>> struct MHD_Connection *connection,
>>> const char *url,
>>> const char *method,
>>> const char *version,
>>> const char *upload_data, size_t *upload_data_size, void **ptr)
>>> {
>>> struct MHD_Response *response;
>>> int ret;
>>> struct Request* req;
>>> struct itimerspec ts;
>>> (void)url; /* Unused. Silent compiler warning. */
>>> (void)version; /* Unused. Silent compiler warning. */
>>> (void)upload_data; /* Unused. Silent compiler warning. */
>>> (void)upload_data_size; /* Unused. Silent compiler warning. */
>>>
>>> req = *ptr;
>>> if (!req)
>>> {
>>>
>>> req = malloc(sizeof(struct Request));
>>> req->connection = connection;
>>> req->timerfd = 0;
>>> *ptr = req;
>>> return MHD_YES;
>>> }
>>>
>>> if (req->timerfd)
>>> {
>>> // send response (echo request url)
>>> response = MHD_create_response_from_buffer (strlen (url),
>>> (void *) url,
>>> MHD_RESPMEM_MUST_COPY);
>>> ret = MHD_queue_response (connection, MHD_HTTP_OK, response);
>>> MHD_destroy_response (response);
>>> return ret;
>>> }
>>> else
>>> {
>>> // create timer and suspend connection
>>> req->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
>>> if (-1 == req->timerfd)
>>> {
>>> printf("timerfd_create: %s", strerror(errno));
>>> return MHD_NO;
>>> }
>>> evt.events = EPOLLIN;
>>> evt.data.ptr = req;
>>> if (-1 == epoll_ctl(epfd, EPOLL_CTL_ADD, req->timerfd, &evt))
>>> {
>>> printf("epoll_ctl: %s", strerror(errno));
>>> return MHD_NO;
>>> }
>>> ts.it_value.tv_sec = 1;
>>> ts.it_value.tv_nsec = 0;
>>> ts.it_interval.tv_sec = 0;
>>> ts.it_interval.tv_nsec = 0;
>>> if (-1 == timerfd_settime(req->timerfd, 0, &ts, NULL))
>>> {
>>> printf("timerfd_settime: %s", strerror(errno));
>>> return MHD_NO;
>>> }
>>>
>>> MHD_suspend_connection(connection);
>>> return MHD_YES;
>>> }
>>> }
>>>
>>> static int
>>> connection_done(struct MHD_Connection *connection,
>>> void **con_cls,
>>> enum MHD_RequestTerminationCode toe)
>>> {
>>> free(*con_cls);
>>> }
>>>
>>> int
>>> main (int argc, char *const *argv)
>>> {
>>> struct MHD_Daemon *d;
>>> const union MHD_DaemonInfo * info;
>>> int current_event_count;
>>> struct epoll_event events_list[1];
>>> struct Request *req;
>>> uint64_t timer_expirations;
>>>
>>> if (argc != 2)
>>> {
>>> printf ("%s PORT\n", argv[0]);
>>> return 1;
>>> }
>>> d = MHD_start_daemon (MHD_USE_EPOLL | MHD_ALLOW_SUSPEND_RESUME,
>>> atoi (argv[1]),
>>> NULL, NULL, &ahc_echo, NULL,
>>> MHD_OPTION_NOTIFY_COMPLETED, &connection_done, NULL,
>>> MHD_OPTION_END);
>>> if (d == NULL)
>>> return 1;
>>>
>>> info = MHD_get_daemon_info(d, MHD_DAEMON_INFO_EPOLL_FD);
>>> if (info == NULL)
>>> return 1;
>>>
>>> epfd = epoll_create1(EPOLL_CLOEXEC);
>>> if (-1 == epfd)
>>> return 1;
>>>
>>> evt.events = EPOLLIN;
>>> evt.data.ptr = NULL;
>>> if (-1 == epoll_ctl(epfd, EPOLL_CTL_ADD, info->epoll_fd, &evt))
>>> return 1;
>>>
>>> while (1)
>>> {
>>> current_event_count = epoll_wait(epfd, events_list, 1,
>>> TIMEOUT_INFINITE);
>>>
>>> if (1 == current_event_count)
>>> {
>>> if (events_list[0].data.ptr)
>>> {
>>> // A timer has timed out
>>> req = events_list[0].data.ptr;
>>> // read from the fd so the system knows we heard the notice
>>> if (-1 == read(req->timerfd, &timer_expirations,
>>> sizeof(timer_expirations)))
>>> {
>>> return 1;
>>> }
>>> // Now resume the connection
>>> MHD_resume_connection(req->connection);
>>> if (!MHD_run(d))
>>> return 1;
>>> }
>>> else
>>> {
>>> // MHD is ready
>>> if (!MHD_run(d))
>>> return 1;
>>> }
>>> }
>>> else if (0 == current_event_count)
>>> {
>>> // no events: continue
>>> }
>>> else
>>> {
>>> // error
>>> return 1;
>>> }
>>> }
>>>
>>> return 0;
>>> }
>>>
>>>
signature.asc
Description: OpenPGP digital signature