Re: [RFC v11][PATCH 05/13] Dump memory address space

[Dave Hansen <dave-23VcF4HTsmIX0ybBhKVfKdBPR1lH4CV8@public.gmane.org>]

On Thu, 2008-12-18 at 06:10 -0500, Oren Laadan wrote:
> >> +    mutex_lock(&mm->context.lock);
> >> +
> >> +    hh->ldt_entry_size = LDT_ENTRY_SIZE;
> >> +    hh->nldt = mm->context.size;
> >> +
> >> +    cr_debug("nldt %d\n", hh->nldt);
> >> +
> >> +    ret = cr_write_obj(ctx, &h, hh);
> >> +    cr_hbuf_put(ctx, sizeof(*hh));
> >> +    if (ret < 0)
> >> +        goto out;
> >> +
> >> +    ret = cr_kwrite(ctx, mm->context.ldt,
> >> +            mm->context.size * LDT_ENTRY_SIZE);
> > 
> > Do we really want to emit anything under lock?  I realize that this
> > patch goes and does a ton of writes with mmap_sem held for read -- is
> > this ok?
> 
> Because all tasks in the container must be frozen during the checkpoint,
> there is no performance penalty for keeping the locks. Although the object
> should not change in the interim anyways, the locks protects us from, e.g.
> the task unfreezing somehow, or being killed by the OOM killer, or any
> other change incurred from the "outside world" (even future code).
> 
> Put in other words - in the long run it is safer to assume that the
> underlying object may otherwise change.
> 
> (If we want to drop the lock here before cr_kwrite(), we need to copy the
> data to a temporary buffer first. If we also want to drop mmap_sem(), we
> need to be more careful with following the vma's.)
> 
> Do you see a reason to not keeping the locks ?

Mike, although we're doing writes of the checkpoint file here, the *mm*
access is read-only.  We only need really mmap_sem for write if we're
creating new VMAs, which we only do on restore.  Was there an action
taken on the mm that would require a write that we missed?

Oren, I never considered the locking overhead, either.  The fact that
the processes are frozen is very, very important here.  The code is fine
as it stands because this *is* a very simple way to do it.  But, this
probably deserves a comment. 

-- Dave