backend runloop spins on backchannel EOF

[michaelrsweet]

Version: 1.3.9
CUPS.org User: martin.pitt.canonical

http://bugs.debian.org/489045 reported that 1.3.9 does not print anything when using HPJetDirect. Instead, the socket backend uses 100% CPU and just keeps spinning.

strace shows that it's spinning in the device_fd (backchannel) loop:

19284 select(6, [4 5], [5], NULL, NULL) = 1 (in [5])
19284 read(5, "", 1024)

Samuel Thibault samuel.thibault@ens-lyon.org tracked this down to a bug in backendRunLoop(): As stated in the manpage, select() reports an fd as "ready for reading" if it is EOF (i. e. read() returns 0). The current code does not intercept this and just keeps select()ing and read()ing. He also provided a patch to stop reading from device_fd on EOF.

The patch looks good and correct to me. It was confirmed to fix the printing (see Debian bug report trail).

[michaelrsweet]

CUPS.org User: mike

Um, if the backchannel is closed, then we should stop immediately as that is an error condition.

[michaelrsweet]

CUPS.org User: mike

Fixed in Subversion repository.

Please test the attached alternate patch, which logs the read error and then disables back-channel reads until the job is done or we fail on a write...

[michaelrsweet]

CUPS.org User: martin.pitt.canonical

The reporters confirmed that this patch works fine. Thank you!

[michaelrsweet]

"str3001.patch":

Index: backend/runloop.c

--- backend/runloop.c (revision 8127)
+++ backend/runloop.c (working copy)
@@ -283,6 +283,12 @@
CUPS_LLCAST bc_bytes);
cupsBackChannelWrite(bc_buffer, bc_bytes, 1.0);
}

•  else if (bc_bytes < 0 && errno != EAGAIN && errno != EINTR)
  

•  {
  

•    fprintf(stderr, "DEBUG: Error reading back-channel data: %s\n",
  

•       strerror(errno));
  

• use_bc = 0;

•  }
  

  }

  /*