Running back to back SAT runs under Linux can cause an allocation failure.

We noticed a failure while running SAT under Linux back to back. test_loop.sh 
shows how it can reproduce. The failure mode is that one of the 
SAT applications will fail because of an allocation failure. We looked into 
this and found that there are a couple of scenarios that will cause the memory 
not to be freed to the OS at process exit time frame.

1. libc refuses to free the memory back to the OS on free(). memalign() isn't 
guaranteed to work with free(). However, glibc claims it does. posix_memalign() 
is suppose to work with free(), but in our tests strace showed that glibc 
refused to give back the virtual memory space to the OS.

Once the above behavior occurs the following can happen which cause the memory 
not to be released to the OS.

a. Someone reading certain /proc files of SAT's can take a reference on the 
OS's mm structure. This can lead to a delayed free of the mm's memory.

b. There can be a thread deep in a run queue that is signalled to die but 
hasn't yet.  The memory won't be given back to the OS until this last reference 
to the mm is dropped.

My proposal is to perform the explicit mmap()/munmap() in SAT so that external 
reference counts on the mm within Linux won't cause the memory to be tied up. A 
patch is attached to do just that.

Original issue reported on code.google.com by adur...@google.com on 16 Sep 2010 at 5:39

Attachments:

sandyou / stressapptest

Running back to back SAT runs under Linux can cause an allocation failure. #11