Extending MADV_PROTECT

[John Baldwin]

On Tuesday, July 16, 2013 1:16:48 pm Konstantin Belousov wrote:
> On Tue, Jul 16, 2013 at 11:30:40AM -0400, John Baldwin wrote:
> > On Saturday, July 13, 2013 1:58:35 pm Konstantin Belousov wrote:
> > > On Fri, Jul 12, 2013 at 05:48:57PM -0400, John Baldwin wrote:
> > > > On Friday, June 28, 2013 2:46:01 pm John Baldwin wrote:
> > > > > Ok, there isn't really a clear consensus here, but I need a system call to let
> > > > > me toggle this flag on existing processes.
> > > > > 
> > > > > One reason I don't like the procctl() approach is I am uneasy about forcing
> > > > > a certain behavior for how commands treat pgid (first-fail vs best-effort).
> > > > > I guess it can always change in the future so that isn't completely unsolvable.
> > > > > 
> > > > > I guess I am fine just making it use hardcoded sizes instead of full-blown
> > > > > ioctl encoding.
> > > > 
> > > > Ok, I have updated patches for this for HEAD.  I have not yet implemented the
> > > > inheritance bits because I'm loathe to add the first bit to a p_flag2. :-P
> > > > However, if that's the best course of action I suppose we can do that.
> > > > 
> > > > The kernel patch is at www.freebsd.org/~jhb/patches/procctl.patch
> > > > 
> > > > The patch for the protect binary is at www.freebsd.org/~jhb/patches/protect.patch
> > > > 
> > > 
> > > It seems that p_cansee() is called twice, once in kern_procctl(), and
> > > then in protect_setchild().
> > 
> > Yes, this is because protect_setchild() can descend to child processes, and
> > you might not be able to "see" a child process if it exec'd a suid binary, etc.
> > 
> > > Is AUE_WAIT6 the correct audit event id for procctl ?
> > 
> > Nope.  (It probably needs a new one?)
> Yes, probably needs a new id.

I punted on this for now.

> > > I thought proposing to use pget() for P_PID case in kern_procctl(), but
> > > indeed open coding of the process lookup is easier, since otherwise
> > > you would need to move proctree_lock acquisition to P_PGID.
> > > 
> > > Why do you need PPROT_CLEAR ?  If you do need the flag, would it be better
> > > to assign a non-zero value to it ?
> > 
> > I need it for 'protect -c' which is similar to 'ktrace -C'.  That is, to
> > allow protection to be removed from existing processes.  I added a constant
> > for it to make the code clear as the caller should pick one of PPROT_SET
> > or PPROT_CLEAR (kind of like MAP_PRIVATE vs MAP_SHARED for mmap(2)).  Otherwise
> > you would have 'procctl(..., PROC_SPROTECT, 0)' which is not as obvious
> > to me as ', PPROT_CLEAR)'.  Also, you can do 'PPROT_CLEAR | PPROT_DESCEND'
> > to clear it for all descendants.  Having 'PPROT_SET | PPROT_DESCEND' to set
> > it for descendenats and just 'PPROT_DESCEND' to clear for descendants doesn't
> > seem as readable.
> Well, I think that assigning non-zero value is justified.

I went with a non-zero value.

> > Do you have any thoughts on p_flag2 (vs some other approach, possibly I
> > should create a new field for oom-specific flags instead since we may
> > eventually store a priority there?)
> 
> p_flag2 is inevitable, it seems.  It would be added now, or, even if you
> could avoid this, in the nearest future.  I do not like mixing a flag
> and a priority-like field.

Ok, I've added p_flag2.

The updated patch (kernel bits) is at
http://people.freebsd.org/~jhb/patches/procctl2.patch

-- 
John Baldwin