RE: [PATCH 2/4] hw/riscv: spike: Allow creating multiple sockets

[Anup Patel]

> -----Original Message-----
> From: Alistair Francis <address@hidden>
> Sent: 27 May 2020 09:00
> To: Anup Patel <address@hidden>
> Cc: Palmer Dabbelt <address@hidden>; Peter Maydell
> <address@hidden>; address@hidden;
> address@hidden; address@hidden; address@hidden;
> Atish Patra <address@hidden>; Alistair Francis
> <address@hidden>
> Subject: Re: [PATCH 2/4] hw/riscv: spike: Allow creating multiple sockets
> 
>  at all?
> 
> AlistairOn Tue, May 26, 2020 at 7:55 PM Anup Patel <address@hidden>
> wrote:
> >
> >
> >
> > > -----Original Message-----
> > > From: Alistair Francis <address@hidden>
> > > Sent: 27 May 2020 06:08
> > > To: Anup Patel <address@hidden>
> > > Cc: Palmer Dabbelt <address@hidden>; Peter Maydell
> > > <address@hidden>; address@hidden;
> > > address@hidden; address@hidden;
> > > address@hidden; Atish Patra <address@hidden>; Alistair
> > > Francis <address@hidden>
> > > Subject: Re: [PATCH 2/4] hw/riscv: spike: Allow creating multiple
> > > sockets
> > >
> > > On Fri, May 22, 2020 at 3:10 AM Anup Patel <address@hidden> wrote:
> > > >
> > > >
> > > >
> > > > > -----Original Message-----
> > > > > From: Palmer Dabbelt <address@hidden>
> > > > > Sent: 22 May 2020 01:46
> > > > > To: Anup Patel <address@hidden>
> > > > > Cc: Peter Maydell <address@hidden>; Alistair Francis
> > > > > <address@hidden>; address@hidden; Atish
> > > > > Patra <address@hidden>; address@hidden;
> > > > > address@hidden; address@hidden; Anup Patel
> > > > > <address@hidden>
> > > > > Subject: Re: [PATCH 2/4] hw/riscv: spike: Allow creating
> > > > > multiple sockets
> > > > >
> > > > > On Fri, 15 May 2020 23:37:44 PDT (-0700), Anup Patel wrote:
> > > > > > We extend RISC-V spike machine to allow creating a
> > > > > > multi-socket
> > > machine.
> > > > > > Each RISC-V spike machine socket is a set of HARTs and a CLINT
> instance.
> > > > > > Other peripherals are shared between all RISC-V spike machine
> sockets.
> > > > > > We also update RISC-V spike machine device tree to treat each
> > > > > > socket as a NUMA node.
> > > > > >
> > > > > > The number of sockets in RISC-V spike machine can be specified
> > > > > > using the "sockets=" sub-option of QEMU "-smp" command-line
> > > > > > option. By default, only one socket RISC-V spike machine will be
> created.
> > > > > >
> > > > > > Currently, we only allow creating upto maximum 4 sockets with
> > > > > > minimum
> > > > > > 2 HARTs per socket. In future, this limits can be changed.
> > > > > >
> > > > > > Signed-off-by: Anup Patel <address@hidden>
> > > > > > ---
> > > > > >  hw/riscv/spike.c         | 206 
> > > > > > ++++++++++++++++++++++++---------------
> > > > > >  include/hw/riscv/spike.h |   8 +-
> > > > > >  2 files changed, 133 insertions(+), 81 deletions(-)
> > > > > >
> > > > > > diff --git a/hw/riscv/spike.c b/hw/riscv/spike.c index
> > > > > > d5e0103d89..f63c57a87c 100644
> > > > > > --- a/hw/riscv/spike.c
> > > > > > +++ b/hw/riscv/spike.c
> > > > > > @@ -64,9 +64,11 @@ static void create_fdt(SpikeState *s, const
> > > > > > struct
> > > > > MemmapEntry *memmap,
> > > > > >      uint64_t mem_size, const char *cmdline)  {
> > > > > >      void *fdt;
> > > > > > -    int cpu;
> > > > > > -    uint32_t *cells;
> > > > > > -    char *nodename;
> > > > > > +    int cpu, socket;
> > > > > > +    uint32_t *clint_cells;
> > > > > > +    unsigned long clint_addr;
> > > > > > +    uint32_t cpu_phandle, intc_phandle, phandle = 1;
> > > > > > +    char *name, *clint_name, *clust_name, *core_name,
> > > > > > + *cpu_name, *intc_name;
> > > > > >
> > > > > >      fdt = s->fdt = create_device_tree(&s->fdt_size);
> > > > > >      if (!fdt) {
> > > > > > @@ -88,68 +90,85 @@ static void create_fdt(SpikeState *s,
> > > > > > const struct
> > > > > MemmapEntry *memmap,
> > > > > >      qemu_fdt_setprop_cell(fdt, "/soc", "#size-cells", 0x2);
> > > > > >      qemu_fdt_setprop_cell(fdt, "/soc", "#address-cells",
> > > > > > 0x2);
> > > > > >
> > > > > > -    nodename = g_strdup_printf("/memory@%lx",
> > > > > > -        (long)memmap[SPIKE_DRAM].base);
> > > > > > -    qemu_fdt_add_subnode(fdt, nodename);
> > > > > > -    qemu_fdt_setprop_cells(fdt, nodename, "reg",
> > > > > > +    name = g_strdup_printf("/memory@%lx",
> > > > > (long)memmap[SPIKE_DRAM].base);
> > > > > > +    qemu_fdt_add_subnode(fdt, name);
> > > > > > +    qemu_fdt_setprop_cells(fdt, name, "reg",
> > > > > >          memmap[SPIKE_DRAM].base >> 32,
> memmap[SPIKE_DRAM].base,
> > > > > >          mem_size >> 32, mem_size);
> > > > > > -    qemu_fdt_setprop_string(fdt, nodename, "device_type",
> "memory");
> > > > > > -    g_free(nodename);
> > > > > > +    qemu_fdt_setprop_string(fdt, name, "device_type", "memory");
> > > > > > +    g_free(name);
> > > > > >
> > > > > >      qemu_fdt_add_subnode(fdt, "/cpus");
> > > > > >      qemu_fdt_setprop_cell(fdt, "/cpus", "timebase-frequency",
> > > > > >          SIFIVE_CLINT_TIMEBASE_FREQ);
> > > > > >      qemu_fdt_setprop_cell(fdt, "/cpus", "#size-cells", 0x0);
> > > > > >      qemu_fdt_setprop_cell(fdt, "/cpus", "#address-cells",
> > > > > > 0x1);
> > > > > > +    qemu_fdt_add_subnode(fdt, "/cpus/cpu-map");
> > > > > >
> > > > > > -    for (cpu = s->soc.num_harts - 1; cpu >= 0; cpu--) {
> > > > > > -        nodename = g_strdup_printf("/cpus/cpu@%d", cpu);
> > > > > > -        char *intc = g_strdup_printf("/cpus/cpu@%d/interrupt-
> controller",
> > > cpu);
> > > > > > -        char *isa = riscv_isa_string(&s->soc.harts[cpu]);
> > > > > > -        qemu_fdt_add_subnode(fdt, nodename);
> > > > > > +    for (socket = (s->num_socs - 1); socket >= 0; socket--) {
> > > > > > +        clust_name = g_strdup_printf("/cpus/cpu-map/cluster0%d",
> socket);
> > > > > > +        qemu_fdt_add_subnode(fdt, clust_name);
> > > > > > +
> > > > > > +        clint_cells =  g_new0(uint32_t,
> > > > > > + s->soc[socket].num_harts
> > > > > > + * 4);
> > > > > > +
> > > > > > +        for (cpu = s->soc[socket].num_harts - 1; cpu >= 0; cpu--) {
> > > > > > +            cpu_phandle = phandle++;
> > > > > > +
> > > > > > +            cpu_name = g_strdup_printf("/cpus/cpu@%d",
> > > > > > +                s->soc[socket].hartid_base + cpu);
> > > > > > +            qemu_fdt_add_subnode(fdt, cpu_name);
> > > > > >  #if defined(TARGET_RISCV32)
> > > > > > -        qemu_fdt_setprop_string(fdt, nodename, "mmu-type",
> "riscv,sv32");
> > > > > > +            qemu_fdt_setprop_string(fdt, cpu_name,
> > > > > > + "mmu-type", "riscv,sv32");
> > > > > >  #else
> > > > > > -        qemu_fdt_setprop_string(fdt, nodename, "mmu-type",
> "riscv,sv48");
> > > > > > +            qemu_fdt_setprop_string(fdt, cpu_name,
> > > > > > + "mmu-type", "riscv,sv48");
> > > > > >  #endif
> > > > > > -        qemu_fdt_setprop_string(fdt, nodename, "riscv,isa", isa);
> > > > > > -        qemu_fdt_setprop_string(fdt, nodename, "compatible", 
> > > > > > "riscv");
> > > > > > -        qemu_fdt_setprop_string(fdt, nodename, "status", "okay");
> > > > > > -        qemu_fdt_setprop_cell(fdt, nodename, "reg", cpu);
> > > > > > -        qemu_fdt_setprop_string(fdt, nodename, "device_type", 
> > > > > > "cpu");
> > > > > > -        qemu_fdt_add_subnode(fdt, intc);
> > > > > > -        qemu_fdt_setprop_cell(fdt, intc, "phandle", 1);
> > > > > > -        qemu_fdt_setprop_string(fdt, intc, "compatible", 
> > > > > > "riscv,cpu-
> intc");
> > > > > > -        qemu_fdt_setprop(fdt, intc, "interrupt-controller", NULL, 
> > > > > > 0);
> > > > > > -        qemu_fdt_setprop_cell(fdt, intc, "#interrupt-cells", 1);
> > > > > > -        g_free(isa);
> > > > > > -        g_free(intc);
> > > > > > -        g_free(nodename);
> > > > > > -    }
> > > > > > +            name = riscv_isa_string(&s->soc[socket].harts[cpu]);
> > > > > > +            qemu_fdt_setprop_string(fdt, cpu_name, "riscv,isa", 
> > > > > > name);
> > > > > > +            g_free(name);
> > > > > > +            qemu_fdt_setprop_string(fdt, cpu_name, "compatible",
> "riscv");
> > > > > > +            qemu_fdt_setprop_string(fdt, cpu_name, "status", 
> > > > > > "okay");
> > > > > > +            qemu_fdt_setprop_cell(fdt, cpu_name, "reg",
> > > > > > +                s->soc[socket].hartid_base + cpu);
> > > > > > +            qemu_fdt_setprop_string(fdt, cpu_name, "device_type",
> "cpu");
> > > > > > +            qemu_fdt_setprop_cell(fdt, cpu_name, "phandle",
> > > > > > + cpu_phandle);
> > > > > > +
> > > > > > +            intc_name =
> > > > > > + g_strdup_printf("%s/interrupt-controller",
> > > cpu_name);
> > > > > > +            qemu_fdt_add_subnode(fdt, intc_name);
> > > > > > +            intc_phandle = phandle++;
> > > > > > +            qemu_fdt_setprop_cell(fdt, intc_name, "phandle",
> intc_phandle);
> > > > > > +            qemu_fdt_setprop_string(fdt, intc_name, "compatible",
> > > > > > +                "riscv,cpu-intc");
> > > > > > +            qemu_fdt_setprop(fdt, intc_name,
> > > > > > + "interrupt-controller", NULL,
> > > 0);
> > > > > > +            qemu_fdt_setprop_cell(fdt, intc_name,
> > > > > > + "#interrupt-cells", 1);
> > > > > > +
> > > > > > +            clint_cells[cpu * 4 + 0] = cpu_to_be32(intc_phandle);
> > > > > > +            clint_cells[cpu * 4 + 1] = cpu_to_be32(IRQ_M_SOFT);
> > > > > > +            clint_cells[cpu * 4 + 2] = cpu_to_be32(intc_phandle);
> > > > > > +            clint_cells[cpu * 4 + 3] =
> > > > > > + cpu_to_be32(IRQ_M_TIMER);
> > > > > > +
> > > > > > +            core_name = g_strdup_printf("%s/core%d", clust_name, 
> > > > > > cpu);
> > > > > > +            qemu_fdt_add_subnode(fdt, core_name);
> > > > > > +            qemu_fdt_setprop_cell(fdt, core_name, "cpu",
> > > > > > + cpu_phandle);
> > > > > > +
> > > > > > +            g_free(core_name);
> > > > > > +            g_free(intc_name);
> > > > > > +            g_free(cpu_name);
> > > > > > +        }
> > > > > >
> > > > > > -    cells =  g_new0(uint32_t, s->soc.num_harts * 4);
> > > > > > -    for (cpu = 0; cpu < s->soc.num_harts; cpu++) {
> > > > > > -        nodename =
> > > > > > -            g_strdup_printf("/cpus/cpu@%d/interrupt-controller", 
> > > > > > cpu);
> > > > > > -        uint32_t intc_phandle = qemu_fdt_get_phandle(fdt, 
> > > > > > nodename);
> > > > > > -        cells[cpu * 4 + 0] = cpu_to_be32(intc_phandle);
> > > > > > -        cells[cpu * 4 + 1] = cpu_to_be32(IRQ_M_SOFT);
> > > > > > -        cells[cpu * 4 + 2] = cpu_to_be32(intc_phandle);
> > > > > > -        cells[cpu * 4 + 3] = cpu_to_be32(IRQ_M_TIMER);
> > > > > > -        g_free(nodename);
> > > > > > +        clint_addr = memmap[SPIKE_CLINT].base +
> > > > > > +            (memmap[SPIKE_CLINT].size * socket);
> > > > > > +        clint_name = g_strdup_printf("/soc/clint@%lx", clint_addr);
> > > > > > +        qemu_fdt_add_subnode(fdt, clint_name);
> > > > > > +        qemu_fdt_setprop_string(fdt, clint_name,
> > > > > > + "compatible",
> > > "riscv,clint0");
> > > > > > +        qemu_fdt_setprop_cells(fdt, clint_name, "reg",
> > > > > > +            0x0, clint_addr, 0x0, memmap[SPIKE_CLINT].size);
> > > > > > +        qemu_fdt_setprop(fdt, clint_name, "interrupts-extended",
> > > > > > +            clint_cells, s->soc[socket].num_harts *
> > > > > > + sizeof(uint32_t)
> > > > > > + * 4);
> > > > > > +
> > > > > > +        g_free(clint_name);
> > > > > > +        g_free(clint_cells);
> > > > > > +        g_free(clust_name);
> > > > > >      }
> > > > > > -    nodename = g_strdup_printf("/soc/clint@%lx",
> > > > > > -        (long)memmap[SPIKE_CLINT].base);
> > > > > > -    qemu_fdt_add_subnode(fdt, nodename);
> > > > > > -    qemu_fdt_setprop_string(fdt, nodename, "compatible",
> "riscv,clint0");
> > > > > > -    qemu_fdt_setprop_cells(fdt, nodename, "reg",
> > > > > > -        0x0, memmap[SPIKE_CLINT].base,
> > > > > > -        0x0, memmap[SPIKE_CLINT].size);
> > > > > > -    qemu_fdt_setprop(fdt, nodename, "interrupts-extended",
> > > > > > -        cells, s->soc.num_harts * sizeof(uint32_t) * 4);
> > > > > > -    g_free(cells);
> > > > > > -    g_free(nodename);
> > > > > >
> > > > > >      if (cmdline) {
> > > > > >          qemu_fdt_add_subnode(fdt, "/chosen"); @@ -160,23
> > > > > > +179,51 @@ static void create_fdt(SpikeState *s, const struct
> > > > > > MemmapEntry *memmap,  static void spike_board_init(MachineState
> *machine)  {
> > > > > >      const struct MemmapEntry *memmap = spike_memmap;
> > > > > > -
> > > > > >      SpikeState *s = g_new0(SpikeState, 1);
> > > > > >      MemoryRegion *system_memory = get_system_memory();
> > > > > >      MemoryRegion *main_mem = g_new(MemoryRegion, 1);
> > > > > >      MemoryRegion *mask_rom = g_new(MemoryRegion, 1);
> > > > > >      int i;
> > > > > > +    char *soc_name;
> > > > > >      unsigned int smp_cpus = machine->smp.cpus;
> > > > > > -
> > > > > > -    /* Initialize SOC */
> > > > > > -    object_initialize_child(OBJECT(machine), "soc", &s->soc, 
> > > > > > sizeof(s-
> >soc),
> > > > > > -                            TYPE_RISCV_HART_ARRAY, &error_abort, 
> > > > > > NULL);
> > > > > > -    object_property_set_str(OBJECT(&s->soc), machine->cpu_type,
> "cpu-
> > > > > type",
> > > > > > -                            &error_abort);
> > > > > > -    object_property_set_int(OBJECT(&s->soc), smp_cpus, "num-harts",
> > > > > > -                            &error_abort);
> > > > > > -    object_property_set_bool(OBJECT(&s->soc), true, "realized",
> > > > > > -                            &error_abort);
> > > > > > +    unsigned int base_hartid, cpus_per_socket;
> > > > > > +
> > > > > > +    s->num_socs = machine->smp.sockets;
> > > > > > +
> > > > > > +    /* Ensure minumum required CPUs per socket */
> > > > > > +    if ((smp_cpus / s->num_socs) < SPIKE_CPUS_PER_SOCKET_MIN)
> > > > > > +        s->num_socs = 1;
> > > > >
> > > > > Why?  It seems like creating single-hart sockets would be a good
> > > > > test case, and I'm pretty sure it's a configuration that we had
> > > > > in embedded
> > > systems.
> > > >
> > > > Yes, single-hart sockets are sensible for testing software.
> > > >
> > > > When "sockets=" sub-option is not provided in "-smp " command line
> > > > options, the machine->smp.sockets is set same as machine->smp.cpus
> > > > by
> > > > smp_parse() function in hw/core/machine.c. This means by default
> > > > we will always get single-hart per socket. In other words, "-smp
> > > > 4" will be 4 cpus and 4 sockets. This is counter intuitive for
> > > > users because when "sockets=" is not provided we should default to
> > > > single socket irrespective to number of cpus.
> > > >
> > > > I had added SPIKE_CPUS_PER_SOCKET_MIN to handle the default case
> > > > when no "sockets=" sub-option is provided.
> > > >
> > > > Alternate approach will be:
> > > > 1. Add more members in struct CpuTopology of include/hw/boards.h
> > > >     to help us know whether "sockets=" option was passed or not 2.
> > > > Update smp_parse() for new members in struct CpuTopology 3. Assume
> > > > single-socket machine in QEMU RISC-V virt and QEMU
> > > >     RISC-V spike machines when "sockets=" option was not passed
> > > >
> > > > Suggestions ??
> > > >
> > >
> > > I think it makes sense to just stick to what smp_parse() does.
> > > That's what QEMU users are used to so we should follow that.
> > >
> > > I agree it is strange that is specifying `-smp x' you will get
> > > max_cpus number of sockets and split the CPUs via them, but that's
> > > what every other board (besides
> > > x86) does.
> >
> > So we are fine with SPIKE_CPUS_PER_SOCKET_MIN=2 for now, right ??
> 
> Why do we need SPIKE_CPUS_PER_SOCKET_MIN at all?

We are creating a separate CLINT and PLIC for each socket but the
VirtIO and other devices can only connect to one PLIC instance.

Now if we have one CPU/HART per socket (by default) then only
one CPU can take all VirtIO interrupts. We can't even change the
irq affinity.

The SPIKE_CPUS_PER_SOCKET_MIN=2 helps us ensure that
smp_parse() gives us one CPU/HART per socket (by default),
we ignore the "sockets" and force it to just one socket. That's
why SPIKE_CPUS_PER_SOCKET_MIN is a work-around for
the non-intuitive behavior of smp_parse().

My previous suggestion was to add "sockets_available" field in
"Struct CpuTopology". The other fields of "Struct CpuTopology"
will remain as-in (no change in semantics) so it will work fine
for other architectures. The "sockets_available" field will tell
use whether "sockets" sub-option was specified in command
line (or not). Using "sockets_available" field we can remove the
SPIKE_CPUS_PER_SOCKET_MIN=2 work-around. In fact, we can
totally remove SPIKE_CPUS_PER_SOCKET_MIN.

Regards,
Anup