POWER9 and POWER10 machines come in two variants, big-core and
small-core. Big-core machines are SMT8 from software's point of view,
but the low level platform topology ("xscom registers and pervasive
addressing"), these look more like a pair of small cores ganged
together.
Presently the way this is modelled is to create one SMT8 PnvCore and add
special cases to xscom and pervasive for big-core mode that tries to
split this into two small cores, but this is becoming too complicated to
manage.
A better approach is to create 2 core structures and ganging them
together to look like an SMT8 core in TCG. Then the xscom and pervasive
models mostly do not need to differentiate big and small core modes.
This change adds initial mode bits and QEMU topology handling to
split SMT8 cores into 2xSMT4 cores.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---
include/hw/ppc/pnv.h | 2 +
include/hw/ppc/pnv_core.h | 1 +
hw/ppc/pnv.c | 79 ++++++++++++++++++++++++++++++++-------
hw/ppc/pnv_core.c | 7 +++-
4 files changed, 75 insertions(+), 14 deletions(-)
diff --git a/include/hw/ppc/pnv.h b/include/hw/ppc/pnv.h
index 1993dededf..283ddd50e7 100644
--- a/include/hw/ppc/pnv.h
+++ b/include/hw/ppc/pnv.h
@@ -101,6 +101,8 @@ struct PnvMachineState {
PnvPnor *pnor;
hwaddr fw_load_addr;
+
+ bool big_core;
};
PnvChip *pnv_get_chip(PnvMachineState *pnv, uint32_t chip_id);
diff --git a/include/hw/ppc/pnv_core.h b/include/hw/ppc/pnv_core.h
index 693acb189b..50164e9e1f 100644
--- a/include/hw/ppc/pnv_core.h
+++ b/include/hw/ppc/pnv_core.h
@@ -49,6 +49,7 @@ struct PnvCore {
/*< public >*/
PowerPCCPU **threads;
+ bool big_core;
uint32_t pir;
uint32_t hwid;
uint64_t hrmor;
diff --git a/hw/ppc/pnv.c b/hw/ppc/pnv.c
index b593a41f7c..d2ddc49142 100644
--- a/hw/ppc/pnv.c
+++ b/hw/ppc/pnv.c
@@ -1002,14 +1002,39 @@ static void pnv_init(MachineState *machine)
pnv->num_chips =
machine->smp.max_cpus / (machine->smp.cores * machine->smp.threads);
+ if (pnv->big_core) {
+ if (machine->smp.threads % 2 == 1) {
+ error_report("Cannot support %d threads with big-core option "
+ "because it must be an even number",
+ machine->smp.threads);
+ exit(1);
+ }
+ max_smt_threads *= 2;
+ }
+
if (machine->smp.threads > max_smt_threads) {
error_report("Cannot support more than %d threads/core "
"on %s machine", max_smt_threads, mc->desc);
+ if (pmc->max_smt_threads == 4) {
+ error_report("(use big-core=on for 8 threads per core)");
+ }
exit(1);
}
+ if (pnv->big_core) {
+ /*
+ * powernv models PnvCore as a SMT4 core. Big-core requires 2xPnvCore
+ * per core, so adjust topology here. pnv_dt_core() processor
+ * device-tree and TCG SMT code make the 2 cores appear as one big core
+ * from software point of view. pnv pervasive models and xscoms tend to
+ * see the big core as 2 small core halves.
+ */
+ machine->smp.cores *= 2;
+ machine->smp.threads /= 2;
+ }
+
if (!is_power_of_2(machine->smp.threads)) {
- error_report("Cannot support %d threads/core on a powernv"
+ error_report("Cannot support %d threads/core on a powernv "
"machine because it must be a power of 2",
machine->smp.threads);
exit(1);
@@ -1175,11 +1200,19 @@ static void pnv_get_pir_tir_p9(PnvChip *chip,
uint32_t core_id, uint32_t thread_id,
uint32_t *pir, uint32_t *tir)
{
- if (pir) {
- if (chip->nr_threads == 8) {
- *pir = (chip->chip_id << 8) | ((thread_id & 1) << 2) |
- (core_id << 3) | (thread_id >> 1);
- } else {
+ PnvMachineState *pnv = PNV_MACHINE(chip->pnv_machine);
+ if (pnv->big_core) {
+ /* Big-core interleaves thread ID between small-cores */
+ thread_id <<= 1;
+ thread_id |= core_id & 1;
+ core_id >>= 1;
+
+ if (pir) {
+ *pir = (chip->chip_id << 8) | (core_id << 3) | thread_id;
+ }
+ } else {
+ if (pir) {
*pir = (chip->chip_id << 8) | (core_id << 2) | thread_id;
}
}
@@ -1203,11 +1236,19 @@ static void pnv_get_pir_tir_p10(PnvChip *chip,
uint32_t core_id, uint32_t thread_id,
uint32_t *pir, uint32_t *tir)
{
- if (pir) {
- if (chip->nr_threads == 8) {
- *pir = (chip->chip_id << 8) | ((core_id / 4) << 4) |
- ((core_id % 2) << 3) | thread_id;
- } else {
+ PnvMachineState *pnv = PNV_MACHINE(chip->pnv_machine);
+
+ if (pnv->big_core) {
+ /* Big-core interleaves thread ID between small-cores */
+ thread_id <<= 1;
+ thread_id |= core_id & 1;
+ core_id >>= 1;
+
+ if (pir) {
+ *pir = (chip->chip_id << 8) | (core_id << 3) | thread_id;
+ }
+ } else {
+ if (pir) {
*pir = (chip->chip_id << 8) | (core_id << 2) | thread_id;
}
}
@@ -2180,7 +2221,8 @@ static void pnv_chip_power10_class_init(ObjectClass
*klass, void *data)
&k->parent_realize);
}
-static void pnv_chip_core_sanitize(PnvChip *chip, Error **errp)
+static void pnv_chip_core_sanitize(PnvMachineState *pnv, PnvChip *chip,
+ Error **errp)
{
PnvChipClass *pcc = PNV_CHIP_GET_CLASS(chip);
int cores_max;
@@ -2201,6 +2243,17 @@ static void pnv_chip_core_sanitize(PnvChip *chip, Error
**errp)
}
chip->cores_mask &= pcc->cores_mask;
+ /* Ensure small-cores a paired up in big-core mode */
+ if (pnv->big_core) {
+ uint64_t even_cores = chip->cores_mask & 0x5555555555555555ULL;
+ uint64_t odd_cores = chip->cores_mask & 0xaaaaaaaaaaaaaaaaULL;
+
+ if (even_cores ^ (odd_cores >> 1)) {
+ error_setg(errp, "warning: unpaired cores in big-core mode !");
+ return;
+ }
+ }
+
/* now that we have a sane layout, let check the number of cores */
cores_max = ctpop64(chip->cores_mask);
if (chip->nr_cores > cores_max) {
@@ -2224,7 +2277,7 @@ static void pnv_chip_core_realize(PnvChip *chip, Error
**errp)
}
/* Cores */
- pnv_chip_core_sanitize(chip, &error);
+ pnv_chip_core_sanitize(pnv, chip, &error);
if (error) {
error_propagate(errp, error);
return;
diff --git a/hw/ppc/pnv_core.c b/hw/ppc/pnv_core.c
index 8cfa94fbfa..16d40392db 100644
--- a/hw/ppc/pnv_core.c
+++ b/hw/ppc/pnv_core.c
@@ -249,7 +249,12 @@ static void pnv_core_cpu_realize(PnvCore *pc, PowerPCCPU
*cpu, Error **errp,
pir_spr->default_value = pir;
tir_spr->default_value = tir;
- env->core_index = core_hwid;
+ if (pc->big_core) {
+ /* 2 "small cores" get the same core index for SMT operations */
+ env->core_index = core_hwid >> 1;
+ } else {
+ env->core_index = core_hwid;
+ }
/* Set time-base frequency to 512 MHz */
cpu_ppc_tb_init(env, PNV_TIMEBASE_FREQ);