XrdPollE.icc

Go to the documentation of this file.

/******************************************************************************/
/*                                                                            */
/*                          X r d P o l l E . i c c                           */
/*                                                                            */
/* (c) 2008 by the Board of Trustees of the Leland Stanford, Jr., University  */
/*                            All Rights Reserved                             */
/*   Produced by Andrew Hanushevsky for Stanford University under contract    */
/*              DE-AC02-76-SFO0515 with the Department of Energy              */
/*                                                                            */
/* This file is part of the XRootD software suite.                            */
/*                                                                            */
/* XRootD is free software: you can redistribute it and/or modify it under    */
/* the terms of the GNU Lesser General Public License as published by the     */
/* Free Software Foundation, either version 3 of the License, or (at your     */
/* option) any later version.                                                 */
/*                                                                            */
/* XRootD is distributed in the hope that it will be useful, but WITHOUT      */
/* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or      */
/* FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public       */
/* License for more details.                                                  */
/*                                                                            */
/* You should have received a copy of the GNU Lesser General Public License   */
/* along with XRootD in a file called COPYING.LESSER (LGPL license) and file  */
/* COPYING (GPL license).  If not, see <http://www.gnu.org/licenses/>.        */
/*                                                                            */
/* The copyright holder's institutional names and contributor's names may not */
/* be used to endorse or promote products derived from this software without  */
/* specific prior written permission of the institution or contributor.       */
/******************************************************************************/
 
#include <fcntl.h>
#include <cstdlib>
#include <sys/types.h>
#include <sys/stat.h>
#include <sys/epoll.h>
#include <sys/eventfd.h>
 
#include "Xrd/XrdPollE.hh"
#include "Xrd/XrdScheduler.hh"
  
/******************************************************************************/
/*                             n e w P o l l e r                              */
/******************************************************************************/
  
XrdPoll *XrdPoll::newPoller(int pollid, int maxfd)
{
   int pfd, wfd, bytes, alignment, pagsz = getpagesize();
   struct epoll_event *pp;
 
// Open the /dev/poll driver
//
#ifndef EPOLL_CLOEXEC
   if ((pfd = epoll_create(maxfd)) >= 0) fcntl(pfd, F_SETFD, FD_CLOEXEC);
      else
#else
   if ((pfd = epoll_create1(EPOLL_CLOEXEC)) < 0)
#endif
      {Log.Emsg("Poll", errno, "create epoll device"); return 0;}
 
   if ((wfd = eventfd(0, EFD_CLOEXEC)) < 0)
      {Log.Emsg("Poll", errno,
                "create an eventfd as the wait-poller descriptor");
       close(pfd);
       return 0;
      }
 
// Calculate the size of the poll table and allocate it
//
   bytes     = maxfd * sizeof(struct epoll_event);
   alignment = (bytes < pagsz ? 1024 : pagsz);
   if (posix_memalign((void **)&pp, alignment, bytes))
      {Log.Emsg("Poll", ENOMEM, "create poll table");
       close(wfd);
       close(pfd);
       return 0;
      }
 
// Create new poll object
//
   memset((void *)pp, 0, bytes);
   return (XrdPoll *)new XrdPollE(pp, maxfd, pfd, wfd);
}
 
/******************************************************************************/
/*                            D e s t r u c t o r                             */
/******************************************************************************/
  
XrdPollE::~XrdPollE()
{
   if (PollTab) free(PollTab);
   if (WaitFd >= 0) close(WaitFd);
   if (PollDfd >= 0) close(PollDfd);
}
  
/******************************************************************************/
/*                              A d d W a i t F d                             */
/******************************************************************************/
 
int XrdPollE::AddWaitFd()
{
   const unsigned int myPollEvts = EPOLLIN;
   struct epoll_event myEvent = {myPollEvts, {(void *)&WaitFd}};
 
// Add the waitfd to the poll set
//
   if (epoll_ctl(PollDfd, EPOLL_CTL_ADD, WaitFd, &myEvent) < 0)
      {int rc = errno;
       Log.Emsg("Poll", rc, "include the wait FD in the poll set");
       return rc;
      }
 
   return 0;
}
 
/******************************************************************************/
/*                               D i s a b l e                                */
/******************************************************************************/
 
void XrdPollE::Disable(XrdPollInfo &pInfo, const char *etxt)
{
 
// Simply return if the link is already disabled
//
   if (!pInfo.isEnabled) return;
 
// If Linux 2.6.9 we use EPOLLONESHOT to automatically disable a polled fd.
// So, the Disable() method need not do anything. Prior kernels did not have
// this mechanism so we need to do this manually.
//
#ifndef EPOLLONESHOT
   struct epoll_event myEvents = {0, (void *)&pInfo};
 
// Disable this fd. Unlike solaris, epoll_ctl() does not block when the pollfd
// is being waited upon by another thread.
//
   if (epoll_ctl(PollDfd, EPOLL_CTL_MOD, pInfo.FD, &myEvents))
      {Log.Emsg("Poll", errno, "disable link", lp->ID); return;}
#endif
 
// Trace this event
//
   pInfo.isEnabled = false;
   TRACEI(POLL, "Poller " <<PID <<" async disabling link " <<pInfo.FD);
 
// Check if this link needs to be rescheduled. If so, the caller better have
// the link opMutex lock held for this to work!
//
   if (etxt && Finish(pInfo, etxt)) Sched.Schedule((XrdJob *)&pInfo.Link);
}
 
/******************************************************************************/
/*                                E n a b l e                                 */
/******************************************************************************/
 
int XrdPollE::Enable(XrdPollInfo &pInfo)
{
   struct epoll_event myEvents = {ePollEvents, {(void *)&pInfo}};
 
// Simply return if the link is already enabled
//
   if (pInfo.isEnabled) return 1;
 
// Enable this fd. Unlike solaris, epoll_ctl() does not block when the pollfd
// is being waited upon by another thread.
//
   pInfo.isEnabled = true;
   if (epoll_ctl(PollDfd, EPOLL_CTL_MOD, pInfo.FD, &myEvents))
      {Log.Emsg("Poll", errno, "enable link", pInfo.Link.ID);
       pInfo.isEnabled = false;
       return 0;
      }
 
// Do final processing
//
   TRACE(POLL, "Poller " <<PID <<" enabled " <<pInfo.Link.ID);
   numEnabled++;
   return 1;
}
 
/******************************************************************************/
/*                               E x c l u d e                                */
/******************************************************************************/
  
void XrdPollE::Exclude(XrdPollInfo &pInfo)
{
 
// Make sure this link is not enabled
//
   if (pInfo.isEnabled)
      {Log.Emsg("Poll", "Detach of enabled link", pInfo.Link.ID);
       Disable(pInfo);
      }
 
// Forcibly remove the link from the poll set
//
   remFD(pInfo, 0);
 
// Wait to make sure the poll thread has completed handling the last set of
// events which may have included our Link. The handing includes examining the
// Link's PollInfo before deciding if to schedule the Link. After we return,
// the PollInfo may be reset and the Link could be subsequently reused so we
// must make sure that the there are no more events.
// This is a separate issue from ensuring that if this Link is Scheduled it
// has completed processing. In the usual flow this is ensured by calling
// Close -> Detach -> Exclude from within the Link's schedule task.
//
   Wait4Poller();
}
 
/******************************************************************************/
/*                           H a n d l e W a i t F d                          */
/******************************************************************************/
 
void XrdPollE::HandleWaitFd(const unsigned int events)
{
// Used by the polling thread to signal waiters that a polling loop has been
// completed.
//
   eventfd_t ic, cnt;
 
// We don't expect anything but EPOLLIN. But if we get an error (errors
// can be reported, despite not being selected events) abort rather than
// potentially keeping the poll thread busy repeatedly looping.
//
   if (!(events & EPOLLIN) || (events & (EPOLLERR | EPOLLHUP)))
      {char eBuff[64];
       Log.Emsg("Poll", "wait-poller handler:", x2Text(events, eBuff));
       if (events & (EPOLLERR | EPOLLHUP))
          abort();
       return;
      }
 
   if (eventfd_read(WaitFd, &cnt) < 0)
      {Log.Emsg("Poll", errno, "read from the wait-poller descriptor");
       return;
      }
 
   for (ic=0;ic<cnt;++ic) WaitFdSem.Post();
   for (ic=0;ic<cnt;++ic) WaitFdSem2.Wait();
}
 
/******************************************************************************/
/*                               I n c l u d e                                */
/******************************************************************************/
  
int XrdPollE::Include(XrdPollInfo &pInfo)
{
   struct epoll_event myEvent = {0, {(void *)&pInfo}};
   int rc;
 
// Add this fd to the poll set
//
   if ((rc = epoll_ctl(PollDfd, EPOLL_CTL_ADD, pInfo.FD, &myEvent)) < 0)
      Log.Emsg("Poll", errno, "include link", pInfo.Link.ID);
 
// All done
//
   return rc == 0;
}
 
/******************************************************************************/
/*                                 r e m F D                                  */
/******************************************************************************/
  
void XrdPollE::remFD(XrdPollInfo &pInfo, unsigned int events)
{
   struct epoll_event myEvents = {0, {(void *)&pInfo}};
 
// It works out that ONESHOT mode or even CTL_MOD requests do not necessarily
// prevent epoll_wait() from returning on an error event. So, we must manually
// remove the fd from the set and assume the logic was actually correct. If it
// wasn't then the client will eventually timeout and retry the request. This
// may cause a double remove; so we don't consider these an arror.
//
   if (pInfo.FD > 0)
      {TRACEI(POLL, "Poller " <<PID <<" removing FD " <<pInfo.FD);
       if (epoll_ctl(PollDfd, EPOLL_CTL_DEL, pInfo.FD, &myEvents)
       && (errno != ENOENT || events != 0))
          {const char *why;
           char buff[96];
           if (events & (EPOLLHUP | EPOLLRDHUP)) why = "sever";
              else if (events & EPOLLERR) why = "error";
                      else why = "disc";
           snprintf(buff, sizeof(buff), "exclude fd %d during %s (%x) event; link",
                          pInfo.FD, why, events);
           Log.Emsg("Poll", errno, buff, pInfo.Link.ID);
          }
      }
}
 
/******************************************************************************/
/*                                 S t a r t                                  */
/******************************************************************************/
  
void XrdPollE::Start(XrdSysSemaphore *syncsem, int &retcode)
{
   char eBuff[64];
   int rc, i, numpolled, num2sched;
   unsigned int waitFdEvents;
   bool haveWaiters;
   XrdJob *jfirst, *jlast;
   const short pollOK = EPOLLIN | EPOLLPRI;
   XrdLink *lp;
   XrdPollInfo *pInfo;
 
   if ((rc = AddWaitFd()))
      {retcode = rc;
       syncsem->Post();
       return;
      }
 
// Indicate to the starting thread that all went well
//
   retcode = 0;
   syncsem->Post();
 
// Now start dispatching links that are ready
//
   do {do {numpolled = epoll_wait(PollDfd, PollTab, PollMax, -1);}
          while (numpolled < 0 && errno == EINTR);
       if (numpolled == 0) continue;
       if (numpolled <  0)
          {Log.Emsg("Poll", errno, "poll for events");
           abort();
          }
       numEvents += numpolled;
 
       // Checkout which links must be dispatched (no need to lock)
       //
       jfirst = jlast = 0; num2sched = 0;
       haveWaiters = false; waitFdEvents = 0;
       for (i = 0; i < numpolled; i++)
           {if (PollTab[i].data.ptr == &WaitFd)
              {haveWaiters = true; waitFdEvents = PollTab[i].events;}
            else if ((pInfo = (XrdPollInfo *)PollTab[i].data.ptr))
              {if (!(pInfo->isEnabled) && pInfo->FD >= 0)
                  remFD(*pInfo, PollTab[i].events);
                  else {pInfo->isEnabled = 0;
                        if (!(PollTab[i].events & pollOK)
                        ||   (PollTab[i].events & POLLRDHUP))
                           Finish(*pInfo, x2Text(PollTab[i].events, eBuff));
                        lp = &(pInfo->Link);
                        lp->NextJob = jfirst; jfirst = (XrdJob *)lp;
                        if (!jlast) jlast=(XrdJob *)lp;
                        num2sched++;
#ifndef EPOLLONESHOT
                        PollTab[i].events  = 0;
                        if (epoll_ctl(PollDfd,EPOLL_CTL_MOD,pInfo.FD,&PollTab[i]))
                           Log.Emsg("Poll",errno,"disable link",pInfo.Link.ID);
#endif
                       }
              } else Log.Emsg("Poll", "null link event!!!!");
           }
 
       // Schedule the polled links
       //
       if (num2sched == 1) Sched.Schedule(jfirst);
          else if (num2sched) Sched.Schedule(num2sched, jfirst, jlast);
 
       if (haveWaiters) HandleWaitFd(waitFdEvents);
      } while(1);
}
 
/******************************************************************************/
/*                            W a i t 4 P o l l e r                           */
/******************************************************************************/
 
void XrdPollE::Wait4Poller()
{
// Triggers the polling thread to complete at least one event processing loop
// and we wait for this.
//
   if (eventfd_write(WaitFd, 1) < 0)
      {Log.Emsg("Poll", errno, "write to the wait-poller descriptor");
       return;
      }
 
// First semaphore ensures we wait for polling thread to have processed a batch
// of waitfd event and thus be outside of epoll_wait.
// Second semaphore ensures polling thread waits for all waits in its batch
// of waitfd events to have been completed.
//
   WaitFdSem.Wait();
   WaitFdSem2.Post();
}
 
/******************************************************************************/
/*                                x 2 T e x t                                 */
/******************************************************************************/
  
const char *XrdPollE::x2Text(unsigned int events, char *buff)
{
   if (events & EPOLLERR) return "socket error";
 
   if (events & (EPOLLHUP | EPOLLRDHUP)) return "hangup";
 
   sprintf(buff, "unusual event (%.4x)", events);
   return buff;
}