39674 - Remove the Unix sigprocmask calls in prnetdb.c

Status: RESOLVED FIXED

(NSPR :: NSPR, defect, P3)

Description

[Wan-Teh Chang]

In prnetdb.c, we call sigprocmask() on Unix to block
the SIGALRM signal while calling gethostbyname() and
gethostbyaddr().  This is a vestige of classic NSPR
that can be removed.  Details follow.

The user-level thread scheduler in classic NSPR
uses setitimer() to send periodic (every 50ms) SIGALRM
signals to the process to implement thread preemption.
On some Unix flavors such as HP-UX, if the select()
system call is interrupted and SA_RESTART is set for
the signal, the select() call is restarted with the
originally specified timeout, ignoring the time that
has elapsed.  With a periodic SIGALRM, a select() can
be repeatedly interrupted and restarted and never time
out.  Since gethostbyname() and gethostbyaddr() may
call select() and we weren't sure if select() was wrapped
properly, we added these sigprocmask() calls to block
SIGALRM while calling gethostbyname() and gethostbyaddr().

These sigprocmask() calls are not needed by the
pthreads version of NSPR, which is used on most Unix
platforms now.  In fact, they are not even needed by
classic NSPR -- I verified (on HP-UX 11.00) that our
select() wrapper correctly blocks the SIGALRM signal
while it is calling the real select system call, so
it's not necessary for PR_GetHostByName() and
PR_GetHostByAddr() to block SIGALRM again. Therefore
the sigprocmask() calls in prnetdb.c can be safely
removed.

Comment 1

[Wan-Teh Chang]

Attachment: Proposed patch.

Comment 2

[Wan-Teh Chang]

Another Unix flavor that restarts an interrupted
select() call with the original timeout is SINIX.
I also verified on SINIX-N 5.43 that our select()
wrapper correctly blocks SIGALRM while calling
the real select system call.

The patch is checked in on the main trunk.
/cvsroot/mozilla/nsprpub/pr/src/misc/prnetdb.c, revision 3.14

I will add a test that verifies that the select()
wrapper blocks SIGALRM.

Comment 3

[Wan-Teh Chang]

I added a new test selintr.c that verifies that
the select() wrapper in classic NSPR properly
blocks the periodic SIGALRM clocks.

Marked the bug fixed.