[U-Boot] [PATCH 1/7] riscv: add infrastructure for calling functions on other harts

[Auer, Lukas]

On Mon, 2019-02-18 at 10:01 +0000, Auer, Lukas wrote:
> On Mon, 2019-02-18 at 10:28 +0530, Anup Patel wrote:
> > On Tue, Feb 12, 2019 at 3:44 AM Lukas Auer
> > <lukas.auer at aisec.fraunhofer.de> wrote:
> > > Harts on RISC-V boot independently and U-Boot is responsible for
> > > managing them. Functions are called on other harts with
> > > smp_call_function(), which sends inter-processor interrupts
> > > (IPIs)
> > > to
> > > all other harts. Functions are specified with their address and
> > > two
> > > function arguments (argument 2 and 3). The first function
> > > argument
> > > is
> > > always the hart ID of the hart calling the function. On the other
> > > harts,
> > > the IPI interrupt handler handle_ipi() must be called on software
> > > interrupts to handle the request and call the specified function.
> > > 
> > > Functions are stored in the ipi_data data structure. Every hart
> > > has
> > > its
> > > own data structure in global data. While this is not required at
> > > the
> > > moment (all harts are expected to boot Linux), this does allow
> > > future
> > > expansion, where other harts may be used for monitoring or other
> > > tasks.
> > > 
> > > Signed-off-by: Lukas Auer <lukas.auer at aisec.fraunhofer.de>
> > > ---
> > > 
> > >  arch/riscv/Kconfig                   |  19 +++++
> > >  arch/riscv/include/asm/global_data.h |   5 ++
> > >  arch/riscv/include/asm/smp.h         |  53 +++++++++++++
> > >  arch/riscv/lib/Makefile              |   1 +
> > >  arch/riscv/lib/smp.c                 | 110
> > > +++++++++++++++++++++++++++
> > >  5 files changed, 188 insertions(+)
> > >  create mode 100644 arch/riscv/include/asm/smp.h
> > >  create mode 100644 arch/riscv/lib/smp.c
> > > 
> > > diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
> > > index c45e4d73a8..c0842178dd 100644
> > > --- a/arch/riscv/Kconfig
> > > +++ b/arch/riscv/Kconfig
> > > @@ -116,4 +116,23 @@ config RISCV_RDTIME
> > >  config SYS_MALLOC_F_LEN
> > >         default 0x1000
> > > 
> > > +config SMP
> > > +       bool "Symmetric Multi-Processing"
> > > +       help
> > > +         This enables support for systems with more than one
> > > CPU.
> > > If
> > > +         you say N here, U-Boot will run on single and
> > > multiprocessor
> > > +         machines, but will use only one CPU of a multiprocessor
> > > +         machine. If you say Y here, U-Boot will run on many,
> > > but
> > > not
> > > +         all, single processor machines.
> > > +
> > > +config NR_CPUS
> > > +       int "Maximum number of CPUs (2-32)"
> > > +       range 2 32
> > > +       depends on SMP
> > > +       default "8"
> > > +       help
> > > +         On multiprocessor machines, U-Boot sets up a stack for
> > > each CPU.
> > > +         Stack memory is pre-allocated. U-Boot must therefore
> > > know
> > > the
> > > +         maximum number of CPUs that may be present.
> > > +
> > >  endmenu
> > > diff --git a/arch/riscv/include/asm/global_data.h
> > > b/arch/riscv/include/asm/global_data.h
> > > index a3a342c6e1..23a5f35af5 100644
> > > --- a/arch/riscv/include/asm/global_data.h
> > > +++ b/arch/riscv/include/asm/global_data.h
> > > @@ -10,12 +10,17 @@
> > >  #ifndef        __ASM_GBL_DATA_H
> > >  #define __ASM_GBL_DATA_H
> > > 
> > > +#include <asm/smp.h>
> > > +
> > >  /* Architecture-specific global data */
> > >  struct arch_global_data {
> > >         long boot_hart;         /* boot hart id */
> > >  #ifdef CONFIG_SIFIVE_CLINT
> > >         void __iomem *clint;    /* clint base address */
> > >  #endif
> > > +#ifdef CONFIG_SMP
> > > +       struct ipi_data ipi[CONFIG_NR_CPUS];
> > 
> > The data passed by "main HART" via global data to other HARTs
> > is not visible reliably and randomly few HARTs fail to enter Linux
> > kernel on my board. I am suspecting per-HART "ipi" data in global
> > data not being cache-line aligned as the cause of behavior but
> > there
> > could be other issue too.
> > 
> > I have a hack which works reliable for me. As-per this hack, we add
> > "mdelay(10)" just before calling riscv_send_ipi() in
> > send_ipi_many().
> > This hack helped me conclude that there is some sync issue in per-
> > HART
> > "ipi" data.
> > 
> > The above issue is not seen on QEMU so we are fine there.
> > 
> > I would suggest to make per-HART "ipi" data cache-line aligned
> > (just like Linux kernel).
> > 
> 
> Interesting, this is definitely a memory coherency issue, probably
> inadequate fences. I am not sure if making the ipi data structure
> cache-line aligned will reliably fix it. I'll try to replicate the
> issue and get it fixed. Thanks for reporting it!
> 

Not sure if it is connected, but I have noticed a regression with the
current OpenSBI. Testing U-Boot with the SMP patches applied on QEMU
with 4 harts, hart 4 will not receive software interrupts. I have
bisected the issue to the following commit.

918c1354b75c74b62f67c4e929551d643f035443 is the first bad commit
commit 918c1354b75c74b62f67c4e929551d643f035443
Author: Nick Kossifidis <mickflemm at gmail.com>
Date:   Sun Feb 17 09:00:20 2019 +0200

    lib: Improve delivery of SBI_IPI_EVENT_HALT
    
    When sbi_ipi_send_many gets called with the current hartid
    included on pmask (or when pmask is NULL), and we send
    a HALT event, since the for loop works sequentially over
    all hartids on the mask, we may send a HALT event to the
    current hart before the loop finishes. So we will halt
    the current hart before it can deliver a HALT IPI to the
    rest and some harts will remain active.
    
    Make sure we send an IPI to the current hart after we've
    finished with everybody else.
    
    Signed-off-by: Nick Kossifidis <mick at ics.forth.gr>

I don't have time to look into it right now, but I will do so later.

Thanks,
Lukas