Possible bug in _CORE_mutex_Seize()
joel.sherrill at OARcorp.com
Mon Sep 29 21:51:17 UTC 2003
Till Straumann wrote:
> Joel Sherrill wrote:
>> Following up on my own post since I didn't notice the bottom where you
>> actually explained
>> where malloc() was called when I read this early this morning.
> Funny - I had missed that, too...
Easy enough to do even though Phil did a great job on the write-up.
>> Phil Torre wrote:
Phil.. are you able to implement a patch based upon my suggestion
or do you need me to generate an untested patch for you?
>> I think I have a solution. Move the body of the libc start hook to
>> the create hook. It is probably
>> also necessary to change cpukit/score/include/rtems/score/apimutex.h
>> so that _API_Mutex_Allocate()
>> creates the Allocator mutex as nestable rather than
>> Finally, it might not be a bad idea for it to be considered a fatal
>> RTEMS error if _API_Mutex_Locks
>> wants to block when _Thread_Dispatch_disable is non-zero. That would
>> be easier than this
>> happening again and debugging it.
>> It might also be valid to consider it a fatal error in a memory
>> allocation is attempted when
>> _Thread_Dispatch_disable is zero.
> you mean _Thread_Dispatch_disable_level is non-zero?
Yes. It is an application or RTEMS design error to get into
this situation so considering it a fatal error encourages it to
be fixed in the lab.
> Joel - are you going to put out a patch for this? We soon are
> going into production and issues like this one make me nervous ;-)
They make me nervous also. I had already cut 4.6.0 when I noticed
that Solaris did not have a complete set of tool binaries. :(
> IMHO, it would be nice to have pre5 out ASAP, as there are a
> few critical fixes which have been closed since pre4
I think this is in order. We need the fix for this in a pre5 though.
Especially with the rtems 4.6 specific
toolset implemented since pre4. You will now be able to keep 4.6
tools in parallel with snapshot tools. Plus I rewrote the Windows
chapter in the Getting Started guide.
NOTE: If anyone wants to enhance the sample application discussion
in the getting started guide to include all samples, it would be
> -- Till
>>> This pertains to rtems-4.6.0pre4 running on MPC860 with an
>>> unsubmitted BSP.
>>> The sequence of events goes like this:
>>> 1. Thread 1 (Init) is running at priority 1. It creates and
>>> starts thread 2 (notification_task) at priority 196. Since
>>> thread 2 is
>>> at a lower priority, it doesn't start executing yet.
>>> 2. Thread 1 sleeps with rtems_task_wake_after(10 ms) to wait for some
>>> external hardware to do something. As soon as it goes to sleep,
>>> thread 2 is now runnable and starts executing.
>>> 3. Thread 2 does some stuff, and then calls malloc(). Halfway
>>> rtems_region_get_segment(), the 10ms timer set by thread 1 expires.
>>> We do a context switch and thread 1 is now running.
>>> ** Before it lost the CPU, thread 2 had successfully called
>>> ** _RTEMS_Lock_allocator(). _RTEMS_Allocator_Mutex is held by **
>>> ** thread 2 when the context switch back to thread 1 occurs. **
>>> 4. Thread 1 now calls rtems_start_task(), which invokes malloc(),
>>> rtems_region_get_segment(), which calls _RTEMS_Lock_allocator().
>>> _RTEMS_Lock_allocator() returns, *without blocking*. The allocator
>>> mutex is still held by thread 2, yet thread 1 proceeds in the belief
>>> that it has the mutex.
>>> More detail:
>>> When thread 1 calls rtems_task_start() in step #4, that function
>>> calls _Thread_Get() on the task we want to start. As a side effect,
>>> _Thread_Get() increments _Thread_Dispatch_disable_level to 1.
>>> Shortly thereafter, _User_extensions_Thread_start() is called, which
>>> calls libc_start_hook(), which calls calloc()->malloc()->
>>> (Note that _Thread_Dispatch_disable_level is stil 1.)
>>> _CORE_mutex_Seize_interrupt_trylock() returns 1 (as it should), so
>>> call _Thread_Disable_dispatch() (disable level is now 2!) followed
>>> _CORE_mutex_Seize_interrupt_blocking() to block on the mutex.
>>> Because _Thread_Dispatch_disable_level is 2, the call to
>>> just decrements it to 1 and returns without calling
>>> Thread 1 now happily proceeds to corrupt the heap free block chain.
>>> I don't understand the semantics of _Thread_Dispatch_disable_level well
>>> enough to
>>> provide a patch. For now we will work around it by making sure our
>>> don't call
>>> malloc() at the same time. Hopefully those with deep kernel
>>> take a look at this and tell me if I'm smoking crack. :)
Joel Sherrill, Ph.D. Director of Research & Development
joel at OARcorp.com On-Line Applications Research
Ask me about RTEMS: a free RTOS Huntsville AL 35805
Support Available (256) 722-9985
More information about the users