SysV IPC message queues

SysV IPC message queues

[Kyle Hubert]

Hi, I recently found myself using SysV IPC message queues in a Linux
specific application. However, I have one concern, I'm using
IPC_PRIVATE for the key to instantiate the message queue (then doing
pthread_creates with the open file descriptor accessed from children
threads).

Anyways, upon reception of signals, the message queue lingers around
taking up memory. I wouldn't mind this if I was using a key_t which I
could use to reopen the message queue at a later instantiation.
However, IPC_PRIVATE just grabs an available queue, and I don't have
access to the key, so I can't exactly open and remove it later. It's
my opinion that IPC_PRIVATE should cause the message queue to get
destroyed when all the open file descriptors are closed, not solely
when msgctl is used to remove it. Also, I don't want to capture
signals, since that's a slippery slope and SIGKILL will still leak
message queues. I also don't think ipcs/ipcrm are long term solutions,
since I'm handing the library off as an abstraction.

Would it be worth it to generate a patch to the Linux kernel to add
this functionality? Or, is there an undocumented ioctl I can call to
set that feature on?

IE: IPC_PRIVATE message queues should be freed when ref count drops to zero.

Thank you very much for your time,

-Kyle Hubert

Re: SysV IPC message queues

[Kyle Hubert]

OK, so the Linux kernel implementation does syscalls for SysV IPC
message queues, hence the lack of freeing resources when IPC_PRIVATE
is used. If IPC_PRIVATE is used, can't we open a file descriptor for
the user process, and it's file operation on close would just call
freequeue?

Thanks,

-Kyle Hubert

On Sun, Mar 21, 2010 at 7:28 PM, Kyle Hubert <khubert@gmail.com> wrote:
> Hi, I recently found myself using SysV IPC message queues in a Linux
> specific application. However, I have one concern, I'm using
> IPC_PRIVATE for the key to instantiate the message queue (then doing
> pthread_creates with the open file descriptor accessed from children
> threads).
>
> Anyways, upon reception of signals, the message queue lingers around
> taking up memory. I wouldn't mind this if I was using a key_t which I
> could use to reopen the message queue at a later instantiation.
> However, IPC_PRIVATE just grabs an available queue, and I don't have
> access to the key, so I can't exactly open and remove it later. It's
> my opinion that IPC_PRIVATE should cause the message queue to get
> destroyed when all the open file descriptors are closed, not solely
> when msgctl is used to remove it. Also, I don't want to capture
> signals, since that's a slippery slope and SIGKILL will still leak
> message queues. I also don't think ipcs/ipcrm are long term solutions,
> since I'm handing the library off as an abstraction.
>
> Would it be worth it to generate a patch to the Linux kernel to add
> this functionality? Or, is there an undocumented ioctl I can call to
> set that feature on?
>
> IE: IPC_PRIVATE message queues should be freed when ref count drops to zero.
>
> Thank you very much for your time,
>
> -Kyle Hubert
>

Re: SysV IPC message queues

[David Newall]

Kyle,

Processes do not open nor close message queues, rather they place 
messages on and fetch them from queues.  Let me say this again: you do 
not open a queue like you must for a file, but go directly to the 
equivalent of read and write.  There is no refcount to indicate how many 
processes are "using" a queue, although there may be one indicating how 
many are actively waiting to receive a message.  Removing a queue when 
that count reaches zero would be quite wrong because a process stops 
waiting as soon as it receives a message; so the count would be zero 
more often than not.

Do not allow IPC_PRIVATE to lead you into false understanding of what a 
message queue is.  IPC_PRIVATE does not cause creation of a private 
queue, as indicated in msgget(2): "The name choice IPC_PRIVATE was 
perhaps unfortunate, IPC_NEW would more clearly show  its  function."  
Queues are not accessed by key but by id.  You do not need to know the 
key for a queue to access it, neither do you need to use msgget(2) to 
find a queue's id.

So, no, sorry, but your idea must be rejected.

David