Message ID | 20210128151731.1333664-4-danielhb413@gmail.com |
---|---|
State | New |
Headers | show |
Series | spapr, spapr_numa: fix max-associativity-domains | expand |
On Thu, 28 Jan 2021 12:17:31 -0300 Daniel Henrique Barboza <danielhb413@gmail.com> wrote: > 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 > spapr->gpu_numa_id. > > Reported-by: Cédric Le Goater <clg@kaod.org> > Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com> > --- > hw/ppc/spapr_numa.c | 16 +++++++++++++++- > 1 file changed, 15 insertions(+), 1 deletion(-) > > diff --git a/hw/ppc/spapr_numa.c b/hw/ppc/spapr_numa.c > index f71105c783..f4d6abce87 100644 > --- a/hw/ppc/spapr_numa.c > +++ b/hw/ppc/spapr_numa.c > @@ -60,6 +60,19 @@ unsigned int spapr_numa_initial_nvgpu_NUMA_id(MachineState *machine) > return MAX(1, machine->numa_state->num_nodes); > } > > +/* > + * Note: if called before spapr_phb_pci_collect_nvgpu() finishes collecting > + * all NVGPUs, this function will not give the right number of NVGPUs NUMA > + * nodes. > + */ This helper has exactly one user : spapr_numa_write_rtas_dt(). Maybe just open-code it there, with a comment that spapr->gpu_numa_id is assumed to be correct at the time we populate the device tree ? > +static > +unsigned int spapr_numa_get_number_nvgpus_nodes(SpaprMachineState *spapr) > +{ > + MachineState *ms = MACHINE(spapr); > + > + return spapr->gpu_numa_id - spapr_numa_initial_nvgpu_NUMA_id(ms); > +} > + > /* > * This function will translate the user distances into > * what the kernel understand as possible values: 10 > @@ -311,6 +324,7 @@ 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_numa_get_number_nvgpus_nodes(spapr); > uint32_t refpoints[] = { > cpu_to_be32(0x4), > cpu_to_be32(0x3), > @@ -318,7 +332,7 @@ void spapr_numa_write_rtas_dt(SpaprMachineState *spapr, void *fdt, int rtas) > cpu_to_be32(0x1), > }; > uint32_t nr_refpoints = ARRAY_SIZE(refpoints); > - uint32_t maxdomain = ms->numa_state->num_nodes + spapr->gpu_numa_id; > + uint32_t maxdomain = ms->numa_state->num_nodes + number_nvgpus_nodes; > uint32_t maxdomains[] = { > cpu_to_be32(4), > cpu_to_be32(maxdomain),
diff --git a/hw/ppc/spapr_numa.c b/hw/ppc/spapr_numa.c index f71105c783..f4d6abce87 100644 --- a/hw/ppc/spapr_numa.c +++ b/hw/ppc/spapr_numa.c @@ -60,6 +60,19 @@ unsigned int spapr_numa_initial_nvgpu_NUMA_id(MachineState *machine) return MAX(1, machine->numa_state->num_nodes); } +/* + * Note: if called before spapr_phb_pci_collect_nvgpu() finishes collecting + * all NVGPUs, this function will not give the right number of NVGPUs NUMA + * nodes. + */ +static +unsigned int spapr_numa_get_number_nvgpus_nodes(SpaprMachineState *spapr) +{ + MachineState *ms = MACHINE(spapr); + + return spapr->gpu_numa_id - spapr_numa_initial_nvgpu_NUMA_id(ms); +} + /* * This function will translate the user distances into * what the kernel understand as possible values: 10 @@ -311,6 +324,7 @@ 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_numa_get_number_nvgpus_nodes(spapr); uint32_t refpoints[] = { cpu_to_be32(0x4), cpu_to_be32(0x3), @@ -318,7 +332,7 @@ void spapr_numa_write_rtas_dt(SpaprMachineState *spapr, void *fdt, int rtas) cpu_to_be32(0x1), }; uint32_t nr_refpoints = ARRAY_SIZE(refpoints); - uint32_t maxdomain = ms->numa_state->num_nodes + spapr->gpu_numa_id; + uint32_t maxdomain = ms->numa_state->num_nodes + number_nvgpus_nodes; uint32_t maxdomains[] = { cpu_to_be32(4), cpu_to_be32(maxdomain),
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 spapr->gpu_numa_id. Reported-by: Cédric Le Goater <clg@kaod.org> Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com> --- hw/ppc/spapr_numa.c | 16 +++++++++++++++- 1 file changed, 15 insertions(+), 1 deletion(-)