[CRIU] Re: [PATCH 2/2] IPC: message queue stealing feature introduced

Andrew Morton akpm at linux-foundation.org
Wed Apr 4 19:12:24 EDT 2012


On Wed, 15 Feb 2012 20:54:39 +0400
Stanislav Kinsbursky <skinsbursky at parallels.com> wrote:

> v2:
> 1) compat functions added.
> 2) message slot size in array is now aligned by struct msgbuf_a.
> 3) check for enough free space in buffer before message copying added.
> 4) if MSG_STEAL flag is set, then do_msgrcv() returns number of bytes written
> to buffer.
> 5) flag MSG_NOERROR is ignored if MSG_STEAL flag is set.
> 
> This patch is required for checkpoint/restore in userspace.
> IOW, c/r requires some way to get all pending IPC messages without deleting
> them for the queue (checkpoint can fail and in this case tasks will be resumed,
> so queue have to be valid).
> To achive this, new operation flag MSG_STEAL for sys_msgrcv() system call
> introduced.
> If this flag is set, then passed struct msgbuf pointer will be used for storing
> array of structures:
> 
> struct msgbuf_a {
> 	long mtype;         /* type of message */
> 	int msize;          /* size of message */
> 	char mtext[0];      /* message text */
> };
> 
> each of which will be followed by corresponding message data.
> 

I'd be a bit more comfortable if there was some sign that other c/r
developers have reviewed and tested this and have successfully used it
in c/r operation testing?

We've been trying to isolate the c/r-specific functions inside #ifdef
CONFIG_CHECKPOINT_RESTORE, but this patch doesn't do that.  I have been
encouraging this isolation so that people who aren't using c/r don't
have to carry the overhead it adds and so that we can more easily hunt
down and remove everything if the entire c/r project doesn't work out
successfully.

This patch modifies the sys_msgrcv() API and so we should update the
manpage for that syscall.  Please work with Michael on this.

What does all the compat fiddling actually do?  I guess it's needed for
checkpoint and restore of 32-bit userspace on 64-bit kernels?  Does c/r
as a whole support that?  It should.  How well tested is this?



More information about the CRIU mailing list