diff options
Diffstat (limited to 'hw/ppc/spapr_numa.c')
| -rw-r--r-- | hw/ppc/spapr_numa.c | 379 |
1 files changed, 314 insertions, 65 deletions
diff --git a/hw/ppc/spapr_numa.c b/hw/ppc/spapr_numa.c index 779f18b994..5822938448 100644 --- a/hw/ppc/spapr_numa.c +++ b/hw/ppc/spapr_numa.c @@ -19,13 +19,51 @@ /* 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) +/* + * Retrieves max_dist_ref_points of the current NUMA affinity. + */ +static int get_max_dist_ref_points(SpaprMachineState *spapr) { - MachineState *machine = MACHINE(spapr); - SpaprMachineClass *smc = SPAPR_MACHINE_GET_CLASS(machine); + if (spapr_ovec_test(spapr->ov5_cas, OV5_FORM2_AFFINITY)) { + return FORM2_DIST_REF_POINTS; + } + + return FORM1_DIST_REF_POINTS; +} + +/* + * Retrieves numa_assoc_size of the current NUMA affinity. + */ +static int get_numa_assoc_size(SpaprMachineState *spapr) +{ + if (spapr_ovec_test(spapr->ov5_cas, OV5_FORM2_AFFINITY)) { + return FORM2_NUMA_ASSOC_SIZE; + } + + return FORM1_NUMA_ASSOC_SIZE; +} + +/* + * Retrieves vcpu_assoc_size of the current NUMA affinity. + * + * vcpu_assoc_size is the size of ibm,associativity array + * for CPUs, which has an extra element (vcpu_id) in the end. + */ +static int get_vcpu_assoc_size(SpaprMachineState *spapr) +{ + return get_numa_assoc_size(spapr) + 1; +} - return smc->pre_5_2_numa_associativity || - machine->numa_state->num_nodes <= 1; +/* + * Retrieves the ibm,associativity array of NUMA node 'node_id' + * for the current NUMA affinity. + */ +static const uint32_t *get_associativity(SpaprMachineState *spapr, int node_id) +{ + if (spapr_ovec_test(spapr->ov5_cas, OV5_FORM2_AFFINITY)) { + return spapr->FORM2_assoc_array[node_id]; + } + return spapr->FORM1_assoc_array[node_id]; } static bool spapr_numa_is_symmetrical(MachineState *ms) @@ -92,12 +130,23 @@ static uint8_t spapr_numa_get_numa_level(uint8_t distance) return 0; } -static void spapr_numa_define_associativity_domains(SpaprMachineState *spapr) +static void spapr_numa_define_FORM1_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; + int src, dst, i, j; + + /* + * 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. + */ + for (i = 1; i < nb_numa_nodes; i++) { + for (j = 1; j < FORM1_DIST_REF_POINTS; j++) { + spapr->FORM1_assoc_array[i][j] = cpu_to_be32(i); + } + } for (src = 0; src < nb_numa_nodes; src++) { for (dst = src; dst < nb_numa_nodes; dst++) { @@ -132,7 +181,7 @@ static void spapr_numa_define_associativity_domains(SpaprMachineState *spapr) * * 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 + * for each NUMA it didn't match. We have FORM1_DIST_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 @@ -147,25 +196,69 @@ static void spapr_numa_define_associativity_domains(SpaprMachineState *spapr) * 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; + assoc_src = spapr->FORM1_assoc_array[src][i]; + spapr->FORM1_assoc_array[dst][i] = assoc_src; } } } } -void spapr_numa_associativity_init(SpaprMachineState *spapr, - MachineState *machine) +static void spapr_numa_FORM1_affinity_check(MachineState *machine) +{ + int i; + + /* + * Check we don't have a memory-less/cpu-less NUMA node + * Firmware relies on the existing memory/cpu topology to provide the + * NUMA topology to the kernel. + * And the linux kernel needs to know the NUMA topology at start + * to be able to hotplug CPUs later. + */ + if (machine->numa_state->num_nodes) { + for (i = 0; i < machine->numa_state->num_nodes; ++i) { + /* check for memory-less node */ + if (machine->numa_state->nodes[i].node_mem == 0) { + CPUState *cs; + int found = 0; + /* check for cpu-less node */ + CPU_FOREACH(cs) { + PowerPCCPU *cpu = POWERPC_CPU(cs); + if (cpu->node_id == i) { + found = 1; + break; + } + } + /* memory-less and cpu-less node */ + if (!found) { + error_report( +"Memory-less/cpu-less nodes are not supported with FORM1 NUMA (node %d)", i); + exit(EXIT_FAILURE); + } + } + } + } + + if (!spapr_numa_is_symmetrical(machine)) { + error_report( +"Asymmetrical NUMA topologies aren't supported in the pSeries machine using FORM1 NUMA"); + exit(EXIT_FAILURE); + } +} + +/* + * Set NUMA machine state data based on FORM1 affinity semantics. + */ +static void spapr_numa_FORM1_affinity_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, + * position FORM1_DIST_REF_POINTS is always the numa_id, * represented by the index 'i'. * * This will break on sparse NUMA setups, when/if QEMU starts @@ -173,19 +266,8 @@ void spapr_numa_associativity_init(SpaprMachineState *spapr, * '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); - } - } + spapr->FORM1_assoc_array[i][0] = cpu_to_be32(FORM1_DIST_REF_POINTS); + spapr->FORM1_assoc_array[i][FORM1_DIST_REF_POINTS] = cpu_to_be32(i); } /* @@ -199,47 +281,95 @@ void spapr_numa_associativity_init(SpaprMachineState *spapr, 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); + spapr->FORM1_assoc_array[i][0] = cpu_to_be32(FORM1_DIST_REF_POINTS); - for (j = 1; j < MAX_DISTANCE_REF_POINTS; j++) { + for (j = 1; j < FORM1_DIST_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->FORM1_assoc_array[i][j] = gpu_assoc; } - spapr->numa_assoc_array[i][MAX_DISTANCE_REF_POINTS] = cpu_to_be32(i); + spapr->FORM1_assoc_array[i][FORM1_DIST_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. + * Guests pseries-5.1 and older uses zeroed associativity domains, + * i.e. no domain definition based on NUMA distance input. + * + * Same thing with guests that have only one NUMA node. */ - if (using_legacy_numa) { + if (smc->pre_5_2_numa_associativity || + machine->numa_state->num_nodes <= 1) { 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_FORM1_domains(spapr); +} + +/* + * Init NUMA FORM2 machine state data + */ +static void spapr_numa_FORM2_affinity_init(SpaprMachineState *spapr) +{ + int i; + + /* + * For all resources but CPUs, FORM2 associativity arrays will + * be a size 2 array with the following format: + * + * ibm,associativity = {1, numa_id} + * + * CPUs will write an additional 'vcpu_id' on top of the arrays + * being initialized here. 'numa_id' is represented by the + * index 'i' of the loop. + * + * Given that this initialization is also valid for GPU associativity + * arrays, handle everything in one single step by populating the + * arrays up to NUMA_NODES_MAX_NUM. + */ + for (i = 0; i < NUMA_NODES_MAX_NUM; i++) { + spapr->FORM2_assoc_array[i][0] = cpu_to_be32(1); + spapr->FORM2_assoc_array[i][1] = cpu_to_be32(i); } +} - spapr_numa_define_associativity_domains(spapr); +void spapr_numa_associativity_init(SpaprMachineState *spapr, + MachineState *machine) +{ + spapr_numa_FORM1_affinity_init(spapr, machine); + spapr_numa_FORM2_affinity_init(spapr); +} + +void spapr_numa_associativity_check(SpaprMachineState *spapr) +{ + /* + * FORM2 does not have any restrictions we need to handle + * at CAS time, for now. + */ + if (spapr_ovec_test(spapr->ov5_cas, OV5_FORM2_AFFINITY)) { + return; + } + + spapr_numa_FORM1_affinity_check(MACHINE(spapr)); } void spapr_numa_write_associativity_dt(SpaprMachineState *spapr, void *fdt, int offset, int nodeid) { + const uint32_t *associativity = get_associativity(spapr, nodeid); + _FDT((fdt_setprop(fdt, offset, "ibm,associativity", - spapr->numa_assoc_array[nodeid], - sizeof(spapr->numa_assoc_array[nodeid])))); + associativity, + get_numa_assoc_size(spapr) * sizeof(uint32_t)))); } static uint32_t *spapr_numa_get_vcpu_assoc(SpaprMachineState *spapr, PowerPCCPU *cpu) { - uint32_t *vcpu_assoc = g_new(uint32_t, VCPU_ASSOC_SIZE); + const uint32_t *associativity = get_associativity(spapr, cpu->node_id); + int max_distance_ref_points = get_max_dist_ref_points(spapr); + int vcpu_assoc_size = get_vcpu_assoc_size(spapr); + uint32_t *vcpu_assoc = g_new(uint32_t, vcpu_assoc_size); int index = spapr_get_vcpu_id(cpu); /* @@ -248,10 +378,10 @@ static uint32_t *spapr_numa_get_vcpu_assoc(SpaprMachineState *spapr, * 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)); + 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, associativity + 1, + (vcpu_assoc_size - 2) * sizeof(uint32_t)); return vcpu_assoc; } @@ -260,12 +390,13 @@ int spapr_numa_fixup_cpu_dt(SpaprMachineState *spapr, void *fdt, int offset, PowerPCCPU *cpu) { g_autofree uint32_t *vcpu_assoc = NULL; + int vcpu_assoc_size = get_vcpu_assoc_size(spapr); 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)); + vcpu_assoc_size * sizeof(uint32_t)); } @@ -273,27 +404,28 @@ int spapr_numa_write_assoc_lookup_arrays(SpaprMachineState *spapr, void *fdt, int offset) { MachineState *machine = MACHINE(spapr); + int max_distance_ref_points = get_max_dist_ref_points(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); + 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); + 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). + * For the lookup-array we use the ibm,associativity array of the + * current NUMA affinity, without the first element (size). */ - uint32_t *associativity = spapr->numa_assoc_array[i]; + const uint32_t *associativity = get_associativity(spapr, i); memcpy(cur_index, ++associativity, - sizeof(uint32_t) * MAX_DISTANCE_REF_POINTS); - cur_index += MAX_DISTANCE_REF_POINTS; + 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)); @@ -302,12 +434,8 @@ int spapr_numa_write_assoc_lookup_arrays(SpaprMachineState *spapr, void *fdt, 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) +static void spapr_numa_FORM1_write_rtas_dt(SpaprMachineState *spapr, + void *fdt, int rtas) { MachineState *ms = MACHINE(spapr); SpaprMachineClass *smc = SPAPR_MACHINE_GET_CLASS(spapr); @@ -329,7 +457,8 @@ void spapr_numa_write_rtas_dt(SpaprMachineState *spapr, void *fdt, int rtas) cpu_to_be32(maxdomain) }; - if (spapr_machine_using_legacy_numa(spapr)) { + if (smc->pre_5_2_numa_associativity || + ms->numa_state->num_nodes <= 1) { uint32_t legacy_refpoints[] = { cpu_to_be32(0x4), cpu_to_be32(0x4), @@ -365,6 +494,125 @@ void spapr_numa_write_rtas_dt(SpaprMachineState *spapr, void *fdt, int rtas) maxdomains, sizeof(maxdomains))); } +static void spapr_numa_FORM2_write_rtas_tables(SpaprMachineState *spapr, + void *fdt, int rtas) +{ + MachineState *ms = MACHINE(spapr); + NodeInfo *numa_info = ms->numa_state->nodes; + int nb_numa_nodes = ms->numa_state->num_nodes; + int distance_table_entries = nb_numa_nodes * nb_numa_nodes; + g_autofree uint32_t *lookup_index_table = NULL; + g_autofree uint8_t *distance_table = NULL; + int src, dst, i, distance_table_size; + + /* + * ibm,numa-lookup-index-table: array with length and a + * list of NUMA ids present in the guest. + */ + lookup_index_table = g_new0(uint32_t, nb_numa_nodes + 1); + lookup_index_table[0] = cpu_to_be32(nb_numa_nodes); + + for (i = 0; i < nb_numa_nodes; i++) { + lookup_index_table[i + 1] = cpu_to_be32(i); + } + + _FDT(fdt_setprop(fdt, rtas, "ibm,numa-lookup-index-table", + lookup_index_table, + (nb_numa_nodes + 1) * sizeof(uint32_t))); + + /* + * ibm,numa-distance-table: contains all node distances. First + * element is the size of the table as uint32, followed up + * by all the uint8 distances from the first NUMA node, then all + * distances from the second NUMA node and so on. + * + * ibm,numa-lookup-index-table is used by guest to navigate this + * array because NUMA ids can be sparse (node 0 is the first, + * node 8 is the second ...). + */ + distance_table_size = distance_table_entries * sizeof(uint8_t) + + sizeof(uint32_t); + distance_table = g_new0(uint8_t, distance_table_size); + stl_be_p(distance_table, distance_table_entries); + + /* Skip the uint32_t array length at the start */ + i = sizeof(uint32_t); + + for (src = 0; src < nb_numa_nodes; src++) { + for (dst = 0; dst < nb_numa_nodes; dst++) { + /* + * We need to be explicit with the local distance + * value to cover the case where the user didn't added any + * NUMA nodes, but QEMU adds the default NUMA node without + * adding the numa_info to retrieve distance info from. + */ + if (src == dst) { + distance_table[i++] = NUMA_DISTANCE_MIN; + continue; + } + + distance_table[i++] = numa_info[src].distance[dst]; + } + } + + _FDT(fdt_setprop(fdt, rtas, "ibm,numa-distance-table", + distance_table, distance_table_size)); +} + +/* + * This helper could be compressed in a single function with + * FORM1 logic since we're setting the same DT values, with the + * difference being a call to spapr_numa_FORM2_write_rtas_tables() + * in the end. The separation was made to avoid clogging FORM1 code + * which already has to deal with compat modes from previous + * QEMU machine types. + */ +static void spapr_numa_FORM2_write_rtas_dt(SpaprMachineState *spapr, + void *fdt, int rtas) +{ + MachineState *ms = MACHINE(spapr); + uint32_t number_nvgpus_nodes = spapr->gpu_numa_id - + spapr_numa_initial_nvgpu_numa_id(ms); + + /* + * In FORM2, ibm,associativity-reference-points will point to + * the element in the ibm,associativity array that contains the + * primary domain index (for FORM2, the first element). + * + * This value (in our case, the numa-id) is then used as an index + * to retrieve all other attributes of the node (distance, + * bandwidth, latency) via ibm,numa-lookup-index-table and other + * ibm,numa-*-table properties. + */ + uint32_t refpoints[] = { cpu_to_be32(1) }; + + uint32_t maxdomain = ms->numa_state->num_nodes + number_nvgpus_nodes; + uint32_t maxdomains[] = { cpu_to_be32(1), cpu_to_be32(maxdomain) }; + + _FDT(fdt_setprop(fdt, rtas, "ibm,associativity-reference-points", + refpoints, sizeof(refpoints))); + + _FDT(fdt_setprop(fdt, rtas, "ibm,max-associativity-domains", + maxdomains, sizeof(maxdomains))); + + spapr_numa_FORM2_write_rtas_tables(spapr, fdt, rtas); +} + +/* + * 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) +{ + if (spapr_ovec_test(spapr->ov5_cas, OV5_FORM2_AFFINITY)) { + spapr_numa_FORM2_write_rtas_dt(spapr, fdt, rtas); + return; + } + + spapr_numa_FORM1_write_rtas_dt(spapr, fdt, rtas); +} + static target_ulong h_home_node_associativity(PowerPCCPU *cpu, SpaprMachineState *spapr, target_ulong opcode, @@ -375,6 +623,7 @@ static target_ulong h_home_node_associativity(PowerPCCPU *cpu, target_ulong procno = args[1]; PowerPCCPU *tcpu; int idx, assoc_idx; + int vcpu_assoc_size = get_vcpu_assoc_size(spapr); /* only support procno from H_REGISTER_VPA */ if (flags != 0x1) { @@ -393,7 +642,7 @@ static target_ulong h_home_node_associativity(PowerPCCPU *cpu, * 12 associativity domains for vcpus. Assert and bail if that's * not the case. */ - G_STATIC_ASSERT((VCPU_ASSOC_SIZE - 1) <= 12); + g_assert((vcpu_assoc_size - 1) <= 12); vcpu_assoc = spapr_numa_get_vcpu_assoc(spapr, tcpu); /* assoc_idx starts at 1 to skip associativity size */ @@ -414,9 +663,9 @@ static target_ulong h_home_node_associativity(PowerPCCPU *cpu, * macro. The ternary will fill the remaining registers with -1 * after we went through vcpu_assoc[]. */ - a = assoc_idx < VCPU_ASSOC_SIZE ? + a = assoc_idx < vcpu_assoc_size ? be32_to_cpu(vcpu_assoc[assoc_idx++]) : -1; - b = assoc_idx < VCPU_ASSOC_SIZE ? + b = assoc_idx < vcpu_assoc_size ? be32_to_cpu(vcpu_assoc[assoc_idx++]) : -1; args[idx] = ASSOCIATIVITY(a, b); |