On Wed, 26 Aug 2009, Lee Schermerhorn wrote:
> > I think it would probably be better to use the generic NODEMASK_ALLOC()
> > interface by requiring it to pass the entire type (including "struct") as
> > part of the first parameter. Then it automatically takes care of
> > dynamically allocating large nodemasks vs. allocating them on the stack.
> >
> > Would it work by redefining NODEMASK_ALLOC() in the NODES_SHIFT > 8 case
> > to be this:
> >
> > #define NODEMASK_ALLOC(x, m) x *m = kmalloc(sizeof(*m), GFP_KERNEL);
> >
> > and converting NODEMASK_SCRATCH(x) to NODEMASK_ALLOC(struct
> > nodemask_scratch, x), and then doing this in your code:
> >
> > NODEMASK_ALLOC(nodemask_t, nodes_allowed);
> > if (nodes_allowed)
> > *nodes_allowed = nodemask_of_node(node);
> >
> > The NODEMASK_{ALLOC,SCRATCH}() interface is in its infancy so it can
> > probably be made more general to handle cases like this.
>
> I just don't know what that would accomplish. Heck, I'm not all that
> happy with the alloc_nodemask_from_node() because it's allocating both a
> hidden nodemask_t and a pointer thereto on the stack just to return a
> pointer to a kmalloc()ed nodemask_t--which is what I want/need here.
>
> One issue I have with NODEMASK_ALLOC() [and nodemask_of_node(), et al]
> is that it declares the pointer variable as well as initializing it,
> perhaps with kmalloc(), ... Indeed, it's purpose is to replace on
> stack nodemask declarations.
>
Right, which is why I suggest we only have one such interface to
dynamically allocate nodemasks when NODES_SHIFT > 8. That's what defines
NODEMASK_ALLOC() as being special: it's taking NODES_SHIFT into
consideration just like CPUMASK_ALLOC() would take NR_CPUS into
consideration. Your use case is the intended purpose of NODEMASK_ALLOC()
and I see no reason why your code can't use the same interface with some
modification and it's in the best interest of a maintainability to not
duplicate specialized cases where pre-existing interfaces can be used (or
improved, in this case).
> So, to use it at the start of, e.g., set_max_huge_pages() where I can
> safely use it throughout the function, I'll end up allocating the
> nodes_allowed mask on every call, whether or not a node is specified or
> there is a non-default mempolicy. If it turns out that no node was
> specified and we have default policy, we need to free the mask and NULL
> out nodes_allowed up front so that we get default behavior. That seems
> uglier to me that only allocating the nodemask when we know we need one.
>
Not with my suggested code of disabling local irqs, getting a reference to
the mempolicy so it can't be freed, reenabling, and then only using
NODEMASK_ALLOC() in the switch statement on mpol->mode for MPOL_PREFERRED.
> I'm not opposed to using a generic function/macro where one exists that
> suits my purposes. I just don't see one. I tried to create
> one--alloc_nodemask_from_node(), and to keep Mel happy, I tried to reuse
> nodemask_from_node() to initialize it. I'm really not happy with the
> results--because of those extra, hidden stack variables. I could
> eliminate those by creating a out of line function, but there's no good
> place to put a generic nodemask function--no nodemask.c.
>
Using NODEMASK_ALLOC(nodes_allowed) wouldn't really be a hidden stack
variable, would it? I think most developers would assume that it is
some automatic variable called `nodes_allowed' since it's later referenced
(and only needs to be in the case of MPOL_PREFERRED if my mpol_get()
solution with disabled local irqs is used).
--
To unsubscribe from this list: send the line "unsubscribe linux-numa" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
