On Fri, 13 Mar 2020 15:05:39 +1100
David Gibson <da...@gibson.dropbear.id.au> wrote:
> The Real Mode Area (RMA) needs to fit within the NUMA node owning memory
> at address 0. That's usually node 0, but can be a later one if there are
> some nodes which have no memory (only CPUs).
>
> This is currently handled by the spapr_node0_size() helper. It has only
> one caller, so there's not a lot of point splitting it out. It's also
> extremely easy to misread the code as clamping to the size of the smallest
> node rather than the first node with any memory.
>
> So, fold it into the caller, and add some commentary to make it a bit
> clearer exactly what it's doing.
>
> Signed-off-by: David Gibson <da...@gibson.dropbear.id.au>
> ---
> hw/ppc/spapr.c | 37 +++++++++++++++++++++----------------
> 1 file changed, 21 insertions(+), 16 deletions(-)
>
> diff --git a/hw/ppc/spapr.c b/hw/ppc/spapr.c
> index 6c32ec3c0a..6a42c0f1c9 100644
> --- a/hw/ppc/spapr.c
> +++ b/hw/ppc/spapr.c
> @@ -295,20 +295,6 @@ static void spapr_dt_pa_features(SpaprMachineState
> *spapr,
> _FDT((fdt_setprop(fdt, offset, "ibm,pa-features", pa_features,
> pa_size)));
> }
>
> -static hwaddr spapr_node0_size(MachineState *machine)
> -{
> - if (machine->numa_state->num_nodes) {
> - int i;
> - for (i = 0; i < machine->numa_state->num_nodes; ++i) {
> - if (machine->numa_state->nodes[i].node_mem) {
> - return MIN(pow2floor(machine->numa_state->nodes[i].node_mem),
> - machine->ram_size);
> - }
> - }
> - }
> - return machine->ram_size;
> -}
> -
> static void add_str(GString *s, const gchar *s1)
> {
> g_string_append_len(s, s1, strlen(s1) + 1);
> @@ -2631,10 +2617,24 @@ static hwaddr spapr_rma_size(SpaprMachineState
> *spapr, Error **errp)
> MachineState *machine = MACHINE(spapr);
> SpaprMachineClass *smc = SPAPR_MACHINE_GET_CLASS(spapr);
> hwaddr rma_size = machine->ram_size;
> - hwaddr node0_size = spapr_node0_size(machine);
>
> /* RMA has to fit in the first NUMA node */
> - rma_size = MIN(rma_size, node0_size);
> + if (machine->numa_state->num_nodes) {
> + /*
> + * It's possible for there to be some zero-memory nodes first
> + * in the list. We need the RMA to fit inside the memory of
> + * the first node which actually has some memory.
> + */
> + int i;
> +
> + for (i = 0; i < machine->numa_state->num_nodes; ++i) {
> + if (machine->numa_state->nodes[i].node_mem != 0) {
> + hwaddr node_size = machine->numa_state->nodes[i].node_mem;
> + rma_size = MIN(rma_size, pow2floor(node_size));
> + break;
> + }
> + }
> + }
>
> /*
> * VRMA access is via a special 1TiB SLB mapping, so the RMA can
> @@ -2651,6 +2651,11 @@ static hwaddr spapr_rma_size(SpaprMachineState *spapr,
> Error **errp)
> rma_size = MIN(rma_size, smc->rma_limit);
> }
>
> + /*
> + * RMA size must be a power of 2
> + */
> + rma_size = pow2floor(rma_size);
> +
The patch is identical to the last spin, for which I had
a comment already:
-----------------------------------------------------------------------
On Wed, 4 Mar 2020 12:25:55 +1100
David Gibson <da...@gibson.dropbear.id.au> wrote:
> On Tue, Mar 03, 2020 at 11:32:49AM +0100, Greg Kurz wrote:
[...]
> > In any case, it would probably help to mention somewhere
> > why the rounding is introduced by this patch.
>
> Drat. I meant to sort out your comment on the last spin better than
> this, but got part way through and forgot what I was doing.
>
-----------------------------------------------------------------------
I still think that the rounding introduced by this patch deserves
some explanations in the changelog...
> if (rma_size < MIN_RMA_SLOF) {
> error_setg(errp,
> "pSeries SLOF firmware requires >= %ldMiB guest RMA (Real Mode Area memory)",
