b01fec3659
The current logic for calculating 'maxdomain' making it a sum of numa_state->num_nodes with spapr->gpu_numa_id. spapr->gpu_numa_id is used as a index to determine the next available NUMA id that a given NVGPU can use. The problem is that the initial value of gpu_numa_id, for any topology that has more than one NUMA node, is equal to numa_state->num_nodes. This means that our maxdomain will always be, at least, twice the amount of existing NUMA nodes. This means that a guest with 4 NUMA nodes will end up with the following max-associativity-domains: rtas/ibm,max-associativity-domains 00000004 00000008 00000008 00000008 00000008 This overtuning of maxdomains doesn't go unnoticed in the guest, being detected in SLUB during boot: dmesg | grep SLUB [ 0.000000] SLUB: HWalign=128, Order=0-3, MinObjects=0, CPUs=4, Nodes=8 SLUB is detecting 8 total nodes, with 4 nodes being online. This patch fixes ibm,max-associativity-domains by considering the amount of NVGPUs NUMA nodes presented in the guest, instead of just spapr->gpu_numa_id. Reported-by: Cédric Le Goater <clg@kaod.org> Tested-by: Cédric Le Goater <clg@kaod.org> Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com> Message-Id: <20210128174213.1349181-4-danielhb413@gmail.com> Reviewed-by: Greg Kurz <groug@kaod.org> Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
437 lines
15 KiB
C
437 lines
15 KiB
C
/*
|
|
* QEMU PowerPC pSeries Logical Partition NUMA associativity handling
|
|
*
|
|
* Copyright IBM Corp. 2020
|
|
*
|
|
* Authors:
|
|
* Daniel Henrique Barboza <danielhb413@gmail.com>
|
|
*
|
|
* This work is licensed under the terms of the GNU GPL, version 2 or later.
|
|
* See the COPYING file in the top-level directory.
|
|
*/
|
|
|
|
#include "qemu/osdep.h"
|
|
#include "qemu-common.h"
|
|
#include "hw/ppc/spapr_numa.h"
|
|
#include "hw/pci-host/spapr.h"
|
|
#include "hw/ppc/fdt.h"
|
|
|
|
/* Moved from hw/ppc/spapr_pci_nvlink2.c */
|
|
#define SPAPR_GPU_NUMA_ID (cpu_to_be32(1))
|
|
|
|
static bool spapr_machine_using_legacy_numa(SpaprMachineState *spapr)
|
|
{
|
|
MachineState *machine = MACHINE(spapr);
|
|
SpaprMachineClass *smc = SPAPR_MACHINE_GET_CLASS(machine);
|
|
|
|
return smc->pre_5_2_numa_associativity ||
|
|
machine->numa_state->num_nodes <= 1;
|
|
}
|
|
|
|
static bool spapr_numa_is_symmetrical(MachineState *ms)
|
|
{
|
|
int src, dst;
|
|
int nb_numa_nodes = ms->numa_state->num_nodes;
|
|
NodeInfo *numa_info = ms->numa_state->nodes;
|
|
|
|
for (src = 0; src < nb_numa_nodes; src++) {
|
|
for (dst = src; dst < nb_numa_nodes; dst++) {
|
|
if (numa_info[src].distance[dst] !=
|
|
numa_info[dst].distance[src]) {
|
|
return false;
|
|
}
|
|
}
|
|
}
|
|
|
|
return true;
|
|
}
|
|
|
|
/*
|
|
* NVLink2-connected GPU RAM needs to be placed on a separate NUMA node.
|
|
* We assign a new numa ID per GPU in spapr_pci_collect_nvgpu() which is
|
|
* called from vPHB reset handler so we initialize the counter here.
|
|
* If no NUMA is configured from the QEMU side, we start from 1 as GPU RAM
|
|
* must be equally distant from any other node.
|
|
* The final value of spapr->gpu_numa_id is going to be written to
|
|
* max-associativity-domains in spapr_build_fdt().
|
|
*/
|
|
unsigned int spapr_numa_initial_nvgpu_numa_id(MachineState *machine)
|
|
{
|
|
return MAX(1, machine->numa_state->num_nodes);
|
|
}
|
|
|
|
/*
|
|
* This function will translate the user distances into
|
|
* what the kernel understand as possible values: 10
|
|
* (local distance), 20, 40, 80 and 160, and return the equivalent
|
|
* NUMA level for each. Current heuristic is:
|
|
* - local distance (10) returns numa_level = 0x4, meaning there is
|
|
* no rounding for local distance
|
|
* - distances between 11 and 30 inclusive -> rounded to 20,
|
|
* numa_level = 0x3
|
|
* - distances between 31 and 60 inclusive -> rounded to 40,
|
|
* numa_level = 0x2
|
|
* - distances between 61 and 120 inclusive -> rounded to 80,
|
|
* numa_level = 0x1
|
|
* - everything above 120 returns numa_level = 0 to indicate that
|
|
* there is no match. This will be calculated as disntace = 160
|
|
* by the kernel (as of v5.9)
|
|
*/
|
|
static uint8_t spapr_numa_get_numa_level(uint8_t distance)
|
|
{
|
|
if (distance == 10) {
|
|
return 0x4;
|
|
} else if (distance > 11 && distance <= 30) {
|
|
return 0x3;
|
|
} else if (distance > 31 && distance <= 60) {
|
|
return 0x2;
|
|
} else if (distance > 61 && distance <= 120) {
|
|
return 0x1;
|
|
}
|
|
|
|
return 0;
|
|
}
|
|
|
|
static void spapr_numa_define_associativity_domains(SpaprMachineState *spapr)
|
|
{
|
|
MachineState *ms = MACHINE(spapr);
|
|
NodeInfo *numa_info = ms->numa_state->nodes;
|
|
int nb_numa_nodes = ms->numa_state->num_nodes;
|
|
int src, dst, i;
|
|
|
|
for (src = 0; src < nb_numa_nodes; src++) {
|
|
for (dst = src; dst < nb_numa_nodes; dst++) {
|
|
/*
|
|
* This is how the associativity domain between A and B
|
|
* is calculated:
|
|
*
|
|
* - get the distance D between them
|
|
* - get the correspondent NUMA level 'n_level' for D
|
|
* - all associativity arrays were initialized with their own
|
|
* numa_ids, and we're calculating the distance in node_id
|
|
* ascending order, starting from node id 0 (the first node
|
|
* retrieved by numa_state). This will have a cascade effect in
|
|
* the algorithm because the associativity domains that node 0
|
|
* defines will be carried over to other nodes, and node 1
|
|
* associativities will be carried over after taking node 0
|
|
* associativities into account, and so on. This happens because
|
|
* we'll assign assoc_src as the associativity domain of dst
|
|
* as well, for all NUMA levels beyond and including n_level.
|
|
*
|
|
* The PPC kernel expects the associativity domains of node 0 to
|
|
* be always 0, and this algorithm will grant that by default.
|
|
*/
|
|
uint8_t distance = numa_info[src].distance[dst];
|
|
uint8_t n_level = spapr_numa_get_numa_level(distance);
|
|
uint32_t assoc_src;
|
|
|
|
/*
|
|
* n_level = 0 means that the distance is greater than our last
|
|
* rounded value (120). In this case there is no NUMA level match
|
|
* between src and dst and we can skip the remaining of the loop.
|
|
*
|
|
* The Linux kernel will assume that the distance between src and
|
|
* dst, in this case of no match, is 10 (local distance) doubled
|
|
* for each NUMA it didn't match. We have MAX_DISTANCE_REF_POINTS
|
|
* levels (4), so this gives us 10*2*2*2*2 = 160.
|
|
*
|
|
* This logic can be seen in the Linux kernel source code, as of
|
|
* v5.9, in arch/powerpc/mm/numa.c, function __node_distance().
|
|
*/
|
|
if (n_level == 0) {
|
|
continue;
|
|
}
|
|
|
|
/*
|
|
* We must assign all assoc_src to dst, starting from n_level
|
|
* and going up to 0x1.
|
|
*/
|
|
for (i = n_level; i > 0; i--) {
|
|
assoc_src = spapr->numa_assoc_array[src][i];
|
|
spapr->numa_assoc_array[dst][i] = assoc_src;
|
|
}
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void spapr_numa_associativity_init(SpaprMachineState *spapr,
|
|
MachineState *machine)
|
|
{
|
|
SpaprMachineClass *smc = SPAPR_MACHINE_GET_CLASS(spapr);
|
|
int nb_numa_nodes = machine->numa_state->num_nodes;
|
|
int i, j, max_nodes_with_gpus;
|
|
bool using_legacy_numa = spapr_machine_using_legacy_numa(spapr);
|
|
|
|
/*
|
|
* For all associativity arrays: first position is the size,
|
|
* position MAX_DISTANCE_REF_POINTS is always the numa_id,
|
|
* represented by the index 'i'.
|
|
*
|
|
* This will break on sparse NUMA setups, when/if QEMU starts
|
|
* to support it, because there will be no more guarantee that
|
|
* 'i' will be a valid node_id set by the user.
|
|
*/
|
|
for (i = 0; i < nb_numa_nodes; i++) {
|
|
spapr->numa_assoc_array[i][0] = cpu_to_be32(MAX_DISTANCE_REF_POINTS);
|
|
spapr->numa_assoc_array[i][MAX_DISTANCE_REF_POINTS] = cpu_to_be32(i);
|
|
|
|
/*
|
|
* Fill all associativity domains of non-zero NUMA nodes with
|
|
* node_id. This is required because the default value (0) is
|
|
* considered a match with associativity domains of node 0.
|
|
*/
|
|
if (!using_legacy_numa && i != 0) {
|
|
for (j = 1; j < MAX_DISTANCE_REF_POINTS; j++) {
|
|
spapr->numa_assoc_array[i][j] = cpu_to_be32(i);
|
|
}
|
|
}
|
|
}
|
|
|
|
/*
|
|
* Initialize NVLink GPU associativity arrays. We know that
|
|
* the first GPU will take the first available NUMA id, and
|
|
* we'll have a maximum of NVGPU_MAX_NUM GPUs in the machine.
|
|
* At this point we're not sure if there are GPUs or not, but
|
|
* let's initialize the associativity arrays and allow NVLink
|
|
* GPUs to be handled like regular NUMA nodes later on.
|
|
*/
|
|
max_nodes_with_gpus = nb_numa_nodes + NVGPU_MAX_NUM;
|
|
|
|
for (i = nb_numa_nodes; i < max_nodes_with_gpus; i++) {
|
|
spapr->numa_assoc_array[i][0] = cpu_to_be32(MAX_DISTANCE_REF_POINTS);
|
|
|
|
for (j = 1; j < MAX_DISTANCE_REF_POINTS; j++) {
|
|
uint32_t gpu_assoc = smc->pre_5_1_assoc_refpoints ?
|
|
SPAPR_GPU_NUMA_ID : cpu_to_be32(i);
|
|
spapr->numa_assoc_array[i][j] = gpu_assoc;
|
|
}
|
|
|
|
spapr->numa_assoc_array[i][MAX_DISTANCE_REF_POINTS] = cpu_to_be32(i);
|
|
}
|
|
|
|
/*
|
|
* Legacy NUMA guests (pseries-5.1 and older, or guests with only
|
|
* 1 NUMA node) will not benefit from anything we're going to do
|
|
* after this point.
|
|
*/
|
|
if (using_legacy_numa) {
|
|
return;
|
|
}
|
|
|
|
if (!spapr_numa_is_symmetrical(machine)) {
|
|
error_report("Asymmetrical NUMA topologies aren't supported "
|
|
"in the pSeries machine");
|
|
exit(EXIT_FAILURE);
|
|
}
|
|
|
|
spapr_numa_define_associativity_domains(spapr);
|
|
}
|
|
|
|
void spapr_numa_write_associativity_dt(SpaprMachineState *spapr, void *fdt,
|
|
int offset, int nodeid)
|
|
{
|
|
_FDT((fdt_setprop(fdt, offset, "ibm,associativity",
|
|
spapr->numa_assoc_array[nodeid],
|
|
sizeof(spapr->numa_assoc_array[nodeid]))));
|
|
}
|
|
|
|
static uint32_t *spapr_numa_get_vcpu_assoc(SpaprMachineState *spapr,
|
|
PowerPCCPU *cpu)
|
|
{
|
|
uint32_t *vcpu_assoc = g_new(uint32_t, VCPU_ASSOC_SIZE);
|
|
int index = spapr_get_vcpu_id(cpu);
|
|
|
|
/*
|
|
* VCPUs have an extra 'cpu_id' value in ibm,associativity
|
|
* compared to other resources. Increment the size at index
|
|
* 0, put cpu_id last, then copy the remaining associativity
|
|
* domains.
|
|
*/
|
|
vcpu_assoc[0] = cpu_to_be32(MAX_DISTANCE_REF_POINTS + 1);
|
|
vcpu_assoc[VCPU_ASSOC_SIZE - 1] = cpu_to_be32(index);
|
|
memcpy(vcpu_assoc + 1, spapr->numa_assoc_array[cpu->node_id] + 1,
|
|
(VCPU_ASSOC_SIZE - 2) * sizeof(uint32_t));
|
|
|
|
return vcpu_assoc;
|
|
}
|
|
|
|
int spapr_numa_fixup_cpu_dt(SpaprMachineState *spapr, void *fdt,
|
|
int offset, PowerPCCPU *cpu)
|
|
{
|
|
g_autofree uint32_t *vcpu_assoc = NULL;
|
|
|
|
vcpu_assoc = spapr_numa_get_vcpu_assoc(spapr, cpu);
|
|
|
|
/* Advertise NUMA via ibm,associativity */
|
|
return fdt_setprop(fdt, offset, "ibm,associativity", vcpu_assoc,
|
|
VCPU_ASSOC_SIZE * sizeof(uint32_t));
|
|
}
|
|
|
|
|
|
int spapr_numa_write_assoc_lookup_arrays(SpaprMachineState *spapr, void *fdt,
|
|
int offset)
|
|
{
|
|
MachineState *machine = MACHINE(spapr);
|
|
int nb_numa_nodes = machine->numa_state->num_nodes;
|
|
int nr_nodes = nb_numa_nodes ? nb_numa_nodes : 1;
|
|
uint32_t *int_buf, *cur_index, buf_len;
|
|
int ret, i;
|
|
|
|
/* ibm,associativity-lookup-arrays */
|
|
buf_len = (nr_nodes * MAX_DISTANCE_REF_POINTS + 2) * sizeof(uint32_t);
|
|
cur_index = int_buf = g_malloc0(buf_len);
|
|
int_buf[0] = cpu_to_be32(nr_nodes);
|
|
/* Number of entries per associativity list */
|
|
int_buf[1] = cpu_to_be32(MAX_DISTANCE_REF_POINTS);
|
|
cur_index += 2;
|
|
for (i = 0; i < nr_nodes; i++) {
|
|
/*
|
|
* For the lookup-array we use the ibm,associativity array,
|
|
* from numa_assoc_array. without the first element (size).
|
|
*/
|
|
uint32_t *associativity = spapr->numa_assoc_array[i];
|
|
memcpy(cur_index, ++associativity,
|
|
sizeof(uint32_t) * MAX_DISTANCE_REF_POINTS);
|
|
cur_index += MAX_DISTANCE_REF_POINTS;
|
|
}
|
|
ret = fdt_setprop(fdt, offset, "ibm,associativity-lookup-arrays", int_buf,
|
|
(cur_index - int_buf) * sizeof(uint32_t));
|
|
g_free(int_buf);
|
|
|
|
return ret;
|
|
}
|
|
|
|
/*
|
|
* Helper that writes ibm,associativity-reference-points and
|
|
* max-associativity-domains in the RTAS pointed by @rtas
|
|
* in the DT @fdt.
|
|
*/
|
|
void spapr_numa_write_rtas_dt(SpaprMachineState *spapr, void *fdt, int rtas)
|
|
{
|
|
MachineState *ms = MACHINE(spapr);
|
|
SpaprMachineClass *smc = SPAPR_MACHINE_GET_CLASS(spapr);
|
|
uint32_t number_nvgpus_nodes = spapr->gpu_numa_id -
|
|
spapr_numa_initial_nvgpu_numa_id(ms);
|
|
uint32_t refpoints[] = {
|
|
cpu_to_be32(0x4),
|
|
cpu_to_be32(0x3),
|
|
cpu_to_be32(0x2),
|
|
cpu_to_be32(0x1),
|
|
};
|
|
uint32_t nr_refpoints = ARRAY_SIZE(refpoints);
|
|
uint32_t maxdomain = ms->numa_state->num_nodes + number_nvgpus_nodes;
|
|
uint32_t maxdomains[] = {
|
|
cpu_to_be32(4),
|
|
cpu_to_be32(maxdomain),
|
|
cpu_to_be32(maxdomain),
|
|
cpu_to_be32(maxdomain),
|
|
cpu_to_be32(maxdomain)
|
|
};
|
|
|
|
if (spapr_machine_using_legacy_numa(spapr)) {
|
|
uint32_t legacy_refpoints[] = {
|
|
cpu_to_be32(0x4),
|
|
cpu_to_be32(0x4),
|
|
cpu_to_be32(0x2),
|
|
};
|
|
uint32_t legacy_maxdomain = spapr->gpu_numa_id > 1 ? 1 : 0;
|
|
uint32_t legacy_maxdomains[] = {
|
|
cpu_to_be32(4),
|
|
cpu_to_be32(legacy_maxdomain),
|
|
cpu_to_be32(legacy_maxdomain),
|
|
cpu_to_be32(legacy_maxdomain),
|
|
cpu_to_be32(spapr->gpu_numa_id),
|
|
};
|
|
|
|
G_STATIC_ASSERT(sizeof(legacy_refpoints) <= sizeof(refpoints));
|
|
G_STATIC_ASSERT(sizeof(legacy_maxdomains) <= sizeof(maxdomains));
|
|
|
|
nr_refpoints = 3;
|
|
|
|
memcpy(refpoints, legacy_refpoints, sizeof(legacy_refpoints));
|
|
memcpy(maxdomains, legacy_maxdomains, sizeof(legacy_maxdomains));
|
|
|
|
/* pseries-5.0 and older reference-points array is {0x4, 0x4} */
|
|
if (smc->pre_5_1_assoc_refpoints) {
|
|
nr_refpoints = 2;
|
|
}
|
|
}
|
|
|
|
_FDT(fdt_setprop(fdt, rtas, "ibm,associativity-reference-points",
|
|
refpoints, nr_refpoints * sizeof(refpoints[0])));
|
|
|
|
_FDT(fdt_setprop(fdt, rtas, "ibm,max-associativity-domains",
|
|
maxdomains, sizeof(maxdomains)));
|
|
}
|
|
|
|
static target_ulong h_home_node_associativity(PowerPCCPU *cpu,
|
|
SpaprMachineState *spapr,
|
|
target_ulong opcode,
|
|
target_ulong *args)
|
|
{
|
|
g_autofree uint32_t *vcpu_assoc = NULL;
|
|
target_ulong flags = args[0];
|
|
target_ulong procno = args[1];
|
|
PowerPCCPU *tcpu;
|
|
int idx, assoc_idx;
|
|
|
|
/* only support procno from H_REGISTER_VPA */
|
|
if (flags != 0x1) {
|
|
return H_FUNCTION;
|
|
}
|
|
|
|
tcpu = spapr_find_cpu(procno);
|
|
if (tcpu == NULL) {
|
|
return H_P2;
|
|
}
|
|
|
|
/*
|
|
* Given that we want to be flexible with the sizes and indexes,
|
|
* we must consider that there is a hard limit of how many
|
|
* associativities domain we can fit in R4 up to R9, which would be
|
|
* 12 associativity domains for vcpus. Assert and bail if that's
|
|
* not the case.
|
|
*/
|
|
G_STATIC_ASSERT((VCPU_ASSOC_SIZE - 1) <= 12);
|
|
|
|
vcpu_assoc = spapr_numa_get_vcpu_assoc(spapr, tcpu);
|
|
/* assoc_idx starts at 1 to skip associativity size */
|
|
assoc_idx = 1;
|
|
|
|
#define ASSOCIATIVITY(a, b) (((uint64_t)(a) << 32) | \
|
|
((uint64_t)(b) & 0xffffffff))
|
|
|
|
for (idx = 0; idx < 6; idx++) {
|
|
int32_t a, b;
|
|
|
|
/*
|
|
* vcpu_assoc[] will contain the associativity domains for tcpu,
|
|
* including tcpu->node_id and procno, meaning that we don't
|
|
* need to use these variables here.
|
|
*
|
|
* We'll read 2 values at a time to fill up the ASSOCIATIVITY()
|
|
* macro. The ternary will fill the remaining registers with -1
|
|
* after we went through vcpu_assoc[].
|
|
*/
|
|
a = assoc_idx < VCPU_ASSOC_SIZE ?
|
|
be32_to_cpu(vcpu_assoc[assoc_idx++]) : -1;
|
|
b = assoc_idx < VCPU_ASSOC_SIZE ?
|
|
be32_to_cpu(vcpu_assoc[assoc_idx++]) : -1;
|
|
|
|
args[idx] = ASSOCIATIVITY(a, b);
|
|
}
|
|
#undef ASSOCIATIVITY
|
|
|
|
return H_SUCCESS;
|
|
}
|
|
|
|
static void spapr_numa_register_types(void)
|
|
{
|
|
/* Virtual Processor Home Node */
|
|
spapr_register_hypercall(H_HOME_NODE_ASSOCIATIVITY,
|
|
h_home_node_associativity);
|
|
}
|
|
|
|
type_init(spapr_numa_register_types)
|