[GHC] #4934: threadWaitRead works incorrectly on nonthreaded RTS

[GHC]

#4934: threadWaitRead works incorrectly on nonthreaded RTS
------------------------------------------------+--------------------------
        Reporter:  slyfox                       |            Owner:
            Type:  bug                          |           Status:  patch
        Priority:  normal                       |        Milestone:  ⊥
       Component:  Runtime System               |          Version:  7.0.1
      Resolution:                               |         Keywords:
Operating System:  Linux                        |     Architecture:  x86_64
 Type of failure:  Incorrect result at runtime  |  (amd64)
       Test Case:                               |       Difficulty:
        Blocking:                               |  Unknown
                                                |       Blocked By:
                                                |  Related Tickets:
------------------------------------------------+--------------------------

Comment (by slyfox):

 > Yes, please squash the patches and then I'll review.

 Will do.

 > Not sure what you mean by "slightly change polling cycle to resume only
 descriptors with data available", don't we already do that?

 I was a bit unclear. I meant: only in a case of '''select() == EBADF'''
 failure we
 always resumed all the blocked threads:
 https://ghc.haskell.org/trac/ghc/browser/ghc/rts/posix/Select.c#L295
 {{{
   if ( errno == EBADF )
     unblock_all = rtsTrue;
     ...
   ready = unblock_all || FD_ISSET(tso->block_info.fd, &rfd); // [1]
   ...
   if (ready)
     pushOnRunQueue(&MainCapability,tso);
 }}}

 Current patchset only propagates exceptions early,
 but does not fix spurious resumes (as it was before).

 I propose to fix resume problem:
 {{{
 ---    ready = unblock_all || FD_ISSET(tso->block_info.fd, &rfd);
 +++    if (unblock_all) ready = fd_is_really_ready (fd);
 +++    else             ready = FD_ISSET(tso->block_info.fd, &rfd);
 }}}

 It' won't be hard to implement on top of current patchset.
 Secondary '''select()''' can already provide all the info.

 Should I do it as a) a separate patch/ticket, b) merge into final result,
 c) don't touch it at all?

 Thanks!

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/4934#comment:23>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler