Bug 697545

Memento includes a "fast memory squeezing" feature that allows us to test 'failure paths' in gs/gpcl6/etc for behaviour after malloc exhaustion.

Running this on both gs and gpcl, it's clear that there are problem cases. We'd like to solve these.

To generate a memory squeeze report, run the following on linux:

  make memento XCFLAGS="-DMEMENTO_SQUEEZE_BUILD"
  MEMENTO_SQUEEZEAT=1 membin/gs -sDEVICE=bit -o /dev/null examples/tiger.eps > log 2>&1

This will take a while to run, and generate a large log. Killing it after a minute or so will easily give you enough to be going on with.

You can either read 'log' yourself, or (for a slightly more readable view), you can run it through the /home/robin/squeeze2html.pl script on casper:

  cat log | perl /home/robin/squeeze2html.pl > squeeze.html

If you'd rather not generate your own reports, you can see a couple of existing ones:

 https://ghostscript.com/~robin/squeeze.html
 https://ghostscript.com/~robin/squeeze2.html

The first of these comes from:

 MEMENTO_SQUEEZEAT=1 membin/gpcl6 -sDEVICE=bit -o /dev/null J11.pxl > log 2>&1

(J11.pxl can be found on casper as /home/regression/tests_private/xl/JEITA/J11.pxl)

The second comes from:

 MEMENTO_SQUEEZEAT=1 membin/gs -sDEVICE=bit -o /dev/null examples/tiger.eps > log 2>&1

In the generated reports, click on each number to show/hide the output given from failing the allocations after the nth one.

Green numbers indicate cases where the application correctly shut down with no leaked allocations. Orange cases show the application shutting down without crashing, but with some leaked blocks. Red cases are where we hit a SEGV. (I like traffic lights).

Ideally everything should be green, but as a first step, getting rid of the red cases would be good.

I had a quick prod at the early failures for PCL, and came up with:

http://git.ghostscript.com/?p=ghostpdl.git;a=commitdiff;h=92ecd60a5218eb5bd1ba5e497e5fd2550112b5ed

and

http://git.ghostscript.com/?p=ghostpdl.git;a=commitdiff;h=1392f5d2e7fe73bce8ff56cef62e4205f8a53bfd

Those fixes were enough to change 1 - 178 to being green. So small fixes like this can make a huge impact on the reports - it's hopefully not the huge daunting task that it might initially look like.

Shelly: If you'd like access to one of the cluster nodes to run the memory squeezer on, please just say. Probably best if we don't run it on casper.

Similarly, if you'd rather not run the squeezer yourself, you can reproduce the failures on Windows in MSVC.

Do a memento build, then put a breakpoint on Memento_breakpoint. Open the build properties and in the environment field put: "MEMENTO_FAILAT=100" (or whatever the actual number is. Run the build, and you'll stop in the debugger where the allocation is about to fail. You can then step manually to watch what happens.

Sadly, there may be some differences in allocation between windows and linux (because the platform dependent code may do some allocations). As such it may be necessary to adjust the exact failure numbers, but it should be a (fairly) constant small offset.

If need be I can regenerate those reports periodically for you. Let me know if you have any questions.