192797 - PR_Poll times out when it shouldn't (win9x only)

Status: NEW

(NSPR :: NSPR, defect)

Description

[Darin Fisher]

from bug 192294, PR_Poll under Win9x sometimes times out (returns 0) when passed
PR_INTERVAL_NO_TIMEOUT.  from my investigation, it appears that the bug exists
in the OS (_MD_SELECT is returning 0 when tvp == NULL).  i tried adding a
do-while loop around _MD_SELECT, like this:

  do {
      ready = _MD_SELECT(..., tvp);
  } while (ready == 0 && tvp == NULL);

but when we loop, the second _MD_SELECT call errors out with WINSOCK error 10023
"Too many open files".  i'm a bit suspicious of this error code, since i would
have expected the first call to _MD_SELECT to return this error if it is indeed
the problem.

Comment 1

[Wan-Teh Chang]

10023 does not seem to be a Winsock error.  At
least it is not defined in <winsock2.h>.  Could
you tell me where it is defined?

I suspect that the first call to _MD_SELECT
modified the fd_sets, so we can't call _MD_SELECT
on them again.  I will attach a patch for you to
try.

Comment 2

[Wan-Teh Chang]

Attachment: Experimental patch

This patch calls select() again if select() times out
when we ask it not to.	Please test this patch with
this printf statement after the _MD_SELECT() call:

    printf("select with timeout %p returned %d, error %d\n",
	tvp, ready, (ready == SOCKET_ERROR) ? WSAGetLastError() : 0);

Comment 3

[Wan-Teh Chang]

Attachment: Experimental patch v2

This version is slightly better.

Comment 4

[Wan-Teh Chang]

Attachment: Full patch

This patch addresses all the issues I found reading
the Winsock select() documentation.

1. Increase FD_SETSIZE to 1024.  (The default value of
FD_SETSIZE is only 64.)  Make PR_Poll fail with
PR_INSUFFICIENT_RESOURCES_ERROR if we try to add more
sockets to the fd_set structures than they can hold.

2. Pass a pointer to a fd_set structure to select()
only if the fd_set contains at least a socket.

3. If all three fd_set structures are empty, simply
do a sleep with the specified timeout and return 0.

Please test this patch with the printf statement (see
comment 2) after the _MD_SELECT() call.  Thanks.

Comment 5

[Darin Fisher]

i should point out that this bug always seems to occur when necko is polling
only for exceptions on a large number of sockets (~20 or so).  i'll give your
patch a try shortly ;-)

Comment 6

[Darin Fisher]

oh, i should have said... though necko is polling a bunch of sockets for
PR_POLL_EXCEPT only, there is still the "pollable event" that is being polled
for PR_POLL_READ.  so, actually my comment above is not quite correct.

Comment 7

[Mrten]

i don't know whether it has anything to do with this but i'll state it anyway:
back when i was programming the w3c wwwlib, i had to 'fix up' some loop when it
was running on Windows NT. when trying to clear some fd_set with FD_CLR it
sometimes failed to clear!

so i wrapped it up in a while loop and had one less 'bug' to worry about in the
Win32 winsock library.

see http://lists.w3.org/Archives/Public/www-lib/2000AprJun/0124.html

Comment 8

[Wan-Teh Chang]

Mrten:

Thanks for the pointer to your bug fix.  Our code is not
using FD_CLR (we use FD_ZERO, FD_SET, and FD_ISSET), so
the bug we are running into is not the same one.

Comment 9

[Chris P. Ross]

So, is this patch good?  Can it be approved and committed before 1.3 final?
This seems simple enough that it can be put in, yes?

Comment 10

[Darin Fisher]

chris: we've worked around this bug in the networking code above NSPR, so this
bug is not very critical for 1.3.

Comment 11

[Chris P. Ross]

Yeah, I saw that in bug #192294.  But I was still figuring that this might hurt
other things, and that if there was a patch that could be moved on, it may as
well be... 
Thanks...

Comment 12

[BugBot [:suhaib / :marco/ :calixte]]

The bug assignee is inactive on Bugzilla, so the assignee is being reset.