On Tue, Jan 23, 2024 at 11:02:03AM +0800, Huang, Ying wrote:
> Gregory Price <gourry.memverge@xxxxxxxxx> writes:
>
> > + int prev_node = NUMA_NO_NODE;
>
> It appears that we should initialize prev_node with me->il_prev?
> Details are as below.
>
yeah good catch, was a rebase error from my tested code, where this is
the case. patching now.
> > + if (rem_pages <= pol->wil.cur_weight) {
> > + pol->wil.cur_weight -= rem_pages;
>
> If "pol->wil.cur_weight == 0" here, we need to change me->il_prev?
>
you are right, and also need to fetch the next cur_weight. Seems I
missed this specific case in my tests. (had this tested with a single
node but not 2, so it looked right).
Added to my test suite.
> We can replace "weight_nodes" with "i" and use a "for" loop?
>
> > + while (weight_nodes < nnodes) {
> > + node = next_node_in(prev_node, nodes);
>
> IIUC, "node" will not change in the loop, so all "weight" below will be
> the same value. To keep it simple, I think we can just copy weights
> from the global iw_table and consider the default value?
>
another rebase error here from my tested code, this should have been
node = prev_node;
while (...)
node = next_node_in(node, nodes);
I can change it to a for loop as suggested, but for more info on why I
did it this way, see the chunk below
> > + } else if (!delta_depleted) {
> > + /* if there was no delta, track last allocated node */
> > + resume_node = node;
> > + resume_weight = i < (nnodes - 1) ? weights[i+1] :
> > + weights[0];
^ this line acquires the weight of the *NEXT* node
another chunk prior to this does the same
thing. I suppose i can use next_node_in()
instead and just copy the entire weigh array
though, if that is preferable.
> > + }
>
> Can the above code be simplified as something like below?
>
> resume_node = prev_node;
> resume_weight = 0;
> for (...) {
> ...
> if (delta > weight) {
> node_pages += weight;
> delta -= weight;
> } else if (delta) {
> node_pages += delta;
> /* if delta depleted, resume from this node */
> if (delta < weight) {
> resume_node = prev_node;
> resume_weight = weight - delta;
> } else {
> resume_node = node;
> }
> delta = 0;
> }
> ...
> }
>
I'll take another look at it, but this logic is annoying because of the
corner case: me->il_prev can be NUMA_NO_NODE or an actual numa node.
If it's NUMA_NO_NODE, then the logic you have above will say "the next
node has no remaining weights assigned" and skip it on the next call to
weighted_interleave_nid or weighted_interleave_nodes.
This is incorrect - we want the weight of the first node to be
resume_weight, which is what this chunk does:
if (delta >= weight) {
/* if delta == weight, get next node weight */
resume_weight = i < (nnodes - 1) ? weights[i+1] : weights[0];
else if (delta) { /* delta < weight */
/* there's a remaining weight, use the that for resume weight */
resume_weight = weight - (node_pages % weight);
} else if (!delta_depleted) {
/* there was never a delta, track the last node and get the weight
* of the node AFTER that node, that's the resume weight */
resume_weight = i < (nnodes - 1) ? weights[i+1] : weights[0];
}
If il_prev is an actual node, and delta == 0, we want to return with
(il_prev = prev_node) but with the weight set to the weight of the
first node we're about to allocate from.
This is the reason for the annoying logic here: We have to come out of
this loop with the actual node and the actual weight.
I'll try to clean it up further and get my test suite to pass.
~Gregory
