[bug #39934] make closes a wrong FD, results in *** INTERNAL: readdir: Bad file descriptor

[Petr Machata]

URL:
  <http://savannah.gnu.org/bugs/?39934>

                 Summary: make closes a wrong FD, results in *** INTERNAL:
readdir: Bad file descriptor
                 Project: make
            Submitted by: pmachata
            Submitted on: Втр 03 Сен 2013 17:44:02
                Severity: 3 - Normal
              Item Group: Bug
                  Status: None
                 Privacy: Public
             Assigned to: None
             Open/Closed: Open
         Discussion Lock: Any
       Component Version: SCM
        Operating System: None
           Fixed Release: None
           Triage Status: None

    _______________________________________________________

Details:

This was filed in Red Hat bugzilla here:
https://bugzilla.redhat.com/show_bug.cgi?id=885474

To reproduce this, use the following Makefile:
default:
        /home/petr/src/make/build/make -f Makefile2

And the following Makefile2 (note: the path in vpath must exist):
vpath %.c    /home/petr/
foo:

Then run make:
$ make -j 2
/home/petr/src/make/build/make -f Makefile2
make[1]: Entering directory `/home/petr/tmp/tst'
make[1]: warning: jobserver unavailable: using -j1.  Add `+' to parent make
rule.
make[1]: *** INTERNAL: readdir: Bad file descriptor
.  Stop.
make[1]: Makefile2: Field 'stem' not cached: Makefile2

make[1]: Leaving directory `/home/petr/tmp/tst'
make: *** [default] Error 2

The following excerpt from an ltrace run illustrates the problem:
18706 16:46:43.945567 opendir("/home/petr" <unfinished ...>
18706 16:46:43.945713 SYS_open("/home/petr", 591872, 0172204710) = 4
18706 16:46:43.945785 SYS_brk(0)                                 = 0x1e9d000
18706 16:46:43.945824 SYS_brk(0x1ec5000)                         = 0x1ec5000
18706 16:46:43.945870 <... opendir resumed> )                    = { 4 }
[...]
18706 16:46:44.347777 close(4 <unfinished ...>
18706 16:46:44.347882 SYS_close(4)                               = 0
18706 16:46:44.347927 <... close resumed> )                      = 0
[...]
18706 16:46:45.718922 readdir({ 4 } <unfinished ...>
18706 16:46:45.719037 SYS_getdents(4, 0x161f008, 32768)          = -9
18706 16:46:45.719103 <... readdir resumed> )                    = 0

The problem arises when parental make passes down jobserver FDs in the
environment, but the child make is not run as part of this jobsever batch. 
Child make notices JOB_FDS numbers passed down, and tries to dup, which fails,
because JOB_FDS[0] is invalid.  In then switches to -j1, and proceeds to close
JOB_FDS.  But in the meantime make managed to reuse one of those descriptors
for a directory stream for vpath.  When that is closed it later causes the
error message seen in the subject.

The immediate idea is not to close JOB_FDS if the dup fails.  That's certainly
part of the solution (because otherwise we end up calling close over invalid
file descriptors).  But that it fails is pure luck: make gets FD 3 for
Makefile2, then FD 4 for vpath dirstream, and then closes 3 after it's done
with it.  Had --jobserver-fds been 4 and 5, we would dup it without trouble.

But perhaps it's possible to move read_all_makefiles down beyond the JOB_FDS
business.  Then there would be no file operations except those in the dynamic
linker, and those don't leak descriptors.  Then we can use the dup as a
reliable test.  However I don't know whether there are interactions that I
don't see, and whether it's safe and desirable to postpone the parsing to a
later time like this.  The test suite passes with this FWIW.



    _______________________________________________________

File Attachments:


-------------------------------------------------------
Date: Втр 03 Сен 2013 17:44:02  Name:
0001-Don-t-close-JOB_FDS-if-we-know-they-are-invalid.patch  Size: 2kB   By:
pmachata
A fix
<http://savannah.gnu.org/bugs/download.php?file_id=28997>

    _______________________________________________________

Reply to this item at:

  <http://savannah.gnu.org/bugs/?39934>

_______________________________________________
  Message sent via/by Savannah
  http://savannah.gnu.org/