45181 - NSPR assumes the FD_CLOEXEC flag is cleared.

Status: RESOLVED FIXED

(NSPR :: NSPR, defect, P3)

Description

[Wan-Teh Chang]

This bug affects all Unix platforms.

When NSPR creates a new PRFileDesc, it asserts that
the close-on-exec (FD_CLOEXEC) flag of the Unix file
descriptor is cleared.  (This assertion is in function
pt_SetMethods in ptio.c.)  While this is true for
Unix file descriptors opened by NSPR (the FD_CLOEXEC
flag is cleared by default), we cannot make this
assumption on Unix file descriptors that are not
opened by NSPR, such as stdin, stdout, stderr, and
the Unix file descriptors being imported into NSPR,
because their close-on-exec flag may have been set
by code outside NSPR.

Comment 1

[Wan-Teh Chang]

Attachment: Proposed patch.

Comment 2

[Wan-Teh Chang]

Some comments on the patch (attachment id=12054).

1. I define a new three-state boolean type _PRTriStateBool,
   which has an "unknown" state _PR_TRI_UNKNOWN (-1).  The
   type of fd->secret->inheritable is changed from PRBool to
   _PRTriStateBool to allow the 'inheritable' attribute of a
   fd to be unknown.

2. In classic NSPR, _PR_MD_INIT_FILEDESC() (called by
   PR_AllocFileDesc()) is now a no-op on Unix.  It used
   to set fd->secret->inheritable to PR_TRUE.

3. I added two new MD functions.
   In classic NSPR:
       void _PR_MD_INIT_FD_INHERITABLE(PRFileDesc *fd);
   Use this function after _PR_Getfd() or PR_AllocFileDesc()
   (which calls _PR_Getfd()).

   In both classic and pthreads NSPR:
       void _PR_MD_QUERY_FD_INHERITABLE(PRFileDesc *fd, PRBool imported);

4. In the pthreads version, pt_SetMethods() has a new 'imported'
   argument.

5. 'imported' is true for the following functions:
   PR_ImportFile, PR_ImportPipe, PR_ImportTCPSocket, PR_ImportUDPSocket,
   and the creation of the file descriptors _pr_stdin, _prstdout,
   _pr_stderr.

   'imported' is false for PR_Open, PR_OpenFile, PR_CreatePipe,
   PR_Socket, and PR_NewTCPSocketPair.

   I don't know whether 'imported' should be true or false for
   the newly accepted socket returned by PR_Accept or PR_AcceptRead,
   because I don't know whether the accepted socket has the same
   inheritability attribute as the listening socket.  I decided to
   do nothing to the code in this area so the current behavior stays
   unchanged, whether it's correct or not.

Comment 3

[Wan-Teh Chang]

I checked in the patch (attachment id=12054) on
the main trunk.  However, I found that the NT
code can be improved.  I will post another patch
for that.

Comment 4

[Wan-Teh Chang]

Attachment: Improvement for imported fd's on NT.

Comment 5

[Wan-Teh Chang]

Attachment: Revised patch. (Missed prsocket.c)

Comment 6

[Wan-Teh Chang]

I checked in the revised patch (attachment id=12060)
on the main trunk.  Marked the bug fixed.