OSSP CVS Repository

Ticket 67: MM_LOCK_RD mode does not work correctly with MM_SEMT_IPCSEM

Calls to mm_core_lock from mm_alloc.c with mode=MM_LOCK_RD cause deadlock sometimes (MM_SEMT_IPCSEM is defined in config.h).

To reproduce the problem I have one daemon process that created shared MM region and went to sleep (waiting for a messages) and a number of other processes which are being started repeatedly in endless script while doing the following:

• attaching to this MM region
  
• allocating a chunk of memory
  
• calling mm_display_info
  
• deallocating the memory, just allocated
  
• calling mm_display_info
  
• exiting
  
  After a short while the deadlock occurs and I see that processes are waiting forever in "while" loops in mm_core_lock.
  
  To workaround the problem I always use only mode MM_LOCK_RW while calling mm_core_lock from mm_alloc.c - this works just fine.

[Append remarks]

Remarks:

MM package was used in a multiprocess environment where communicating processes (using the same shared memory region) are NOT related, therefore in a multireader scenario we easily can get into situation when a process will try to unlock a semaphore (when number of readers goes to 0) which was locked by some other process (the first reader). Of course, it won't work. I was able to prove (with some debug printouts added) that it's exactly what happens.

In MM description one can read: "..usage of shared memory between forked (and this way strongly related) processes..." and this requirement was obviously violated. I believe that the only reasonable workaround for our environment is the one I already have - not to use MM_LOCK_RD locking mode but only MM_LOCK_RW.

[Append remarks]

Properties:

Type:	code		Version:	1.4.0
Status:	closed		Created:	2005-Sep-29 23:54
Severity:	2		Last Change:	2005-Dec-14 19:32
Priority:	4		Subsystem:	mm
Assigned To:	rse		Derived From:
Creator:	anonymous