[PATCH v4 09/10] tools: lib: perf: Implement riscv mmap support

Alexandre Ghiti alexghiti at rivosinc.com
Mon Jul 31 09:06:25 PDT 2023


On Mon, Jul 31, 2023 at 5:10 PM Ian Rogers <irogers at google.com> wrote:
>
> On Mon, Jul 31, 2023 at 3:27 AM Alexandre Ghiti <alexghiti at rivosinc.com> wrote:
> >
> > On Mon, Jul 31, 2023 at 12:15 PM Alexandre Ghiti <alexghiti at rivosinc.com> wrote:
> > >
> > > Hi Ian,
> > >
> > > On Fri, Jul 28, 2023 at 7:53 PM Ian Rogers <irogers at google.com> wrote:
> > > >
> > > > On Thu, Jul 27, 2023 at 7:28 AM Alexandre Ghiti <alexghiti at rivosinc.com> wrote:
> > > > >
> > > > > riscv now supports mmaping hardware counters so add what's needed to
> > > > > take advantage of that in libperf.
> > > > >
> > > > > Signed-off-by: Alexandre Ghiti <alexghiti at rivosinc.com>
> > > > > Reviewed-by: Andrew Jones <ajones at ventanamicro.com>
> > > > > Reviewed-by: Atish Patra <atishp at rivosinc.com>
> > > > > ---
> > > > >  tools/lib/perf/mmap.c | 65 +++++++++++++++++++++++++++++++++++++++++++
> > > > >  1 file changed, 65 insertions(+)
> > > > >
> > > > > diff --git a/tools/lib/perf/mmap.c b/tools/lib/perf/mmap.c
> > > > > index 0d1634cedf44..378a163f0554 100644
> > > > > --- a/tools/lib/perf/mmap.c
> > > > > +++ b/tools/lib/perf/mmap.c
> > > > > @@ -392,6 +392,71 @@ static u64 read_perf_counter(unsigned int counter)
> > > > >
> > > > >  static u64 read_timestamp(void) { return read_sysreg(cntvct_el0); }
> > > > >
> > > > > +#elif __riscv_xlen == 64
> > > >
> > > > This is something of an odd guard, perhaps:
> > > > #elif defined(__riscv) && __riscv_xlen == 64
> > > >
> > > > That way it is more intention revealing that this is riscv code. Could
> > > > you add a comment relating to the __riscv_xlen ?
> > >
> > > I guess Andrew answered that already.
> > >
>
> Not sure. I still think it looks weird:
> ...
> #if defined(__i386__) || defined(__x86_64__)
> ...
> #elif defined(__aarch64__)
> ...
> #elif __riscv_xlen == 64
> ...
> #else
> static u64 read_perf_counter(unsigned int counter __maybe_unused) { return 0; }
> static u64 read_timestamp(void) { return 0; }
> #endif
>
> The first two are clearly #ifdef-ing architecture specific assembly
> code, under what conditions I get RISC-V code  ¯\(ツ)/¯ At least worth
> a comment like "csrr is only available when you have xlens of 64
> because ..."

__riscv_xlen indicates riscv64, which is the only target of this
patchset. But if you prefer, I don't mind adding back the
defined(__riscv) if I re-spin a new version.

>
> > > >
> > > > > +
> > > > > +/* TODO: implement rv32 support */
> > > > > +
> > > > > +#define CSR_CYCLE      0xc00
> > > > > +#define CSR_TIME       0xc01
> > > > > +
> > > > > +#define csr_read(csr)                                          \
> > > > > +({                                                             \
> > > > > +       register unsigned long __v;                             \
> > > > > +               __asm__ __volatile__ ("csrr %0, " #csr          \
> > > > > +                : "=r" (__v) :                                 \
> > > > > +                : "memory");                                   \
> > > >
> > > > To avoid the macro pasting that could potentially go weird, could this be:
> > > >
> > > > __asm__ __volatile__ ("csrr %0, %1",
> > > >   : "=r"(__v) /* outputs */
> > > >   : "i"(csr) /* inputs */
> > > >   : "memory" /* clobbers */)
> >
> > Forgot to answer this one: it compiles, but I have to admit that I
> > don't understand the difference and if that's correct (all macros in
> > arch/riscv/include/asm/csr.h use # to do this) and what benefits it
> > brings. Can you elaborate more on things that could "go weird"?
>
> So rather than use an input constraint for the asm block you are using
> the C preprocessor to paste in the csr argument. If csr is something
> like "1" then it looks good and you'll get "csrr %0,1". If you pass
> something like "1 << 31" then that will be pasted as "csrr %0, 1 <<
> 31" and that starts to get weird in the context of being in the
> assembler where it is unlikely the C operators work. Using the input
> constraint avoids this, causes the C compiler to check the type of the
> argument and you'll probably get more intelligible error messages as a
> consequence.
>

Thanks. So if I'm not mistaken, in this exact context, given we only
use csr_read() through the csr_read_num() function, it seems ok right?

> >
> > Thanks again,
> >
> > Alex
> >
> > > >
> > > > Also, why is this clobbering memory? Worth adding a comment.
> > >
> > > No idea, I see that it is also done this way in
> > > arch/riscv/include/asm/csr.h. @Atish Kumar Patra , @Palmer Dabbelt ?
>
> It would seem to make sense then not to have a memory constraint until
> we know why we're doing it?
>

I have just had the answer internally (thanks to @Brendan Sweeney):
csr modifications can alter how the memory is accessed (satp which
changes the address space, sum which allows/disallows userspace
access...), so we need the memory barrier to make sure the compiler
does not reorder the memory accesses.

Thanks,

Alex

> Thanks,
> Ian
>
> > >
> > > Thanks for your comments!
> > >
> > > Alex
> > >
> > > >
> > > > Thanks,
> > > > Ian
> > > >
> > > > > +                __v;                                           \
> > > > > +})
> > > > > +
> > > > > +static unsigned long csr_read_num(int csr_num)
> > > > > +{
> > > > > +#define switchcase_csr_read(__csr_num, __val)           {\
> > > > > +       case __csr_num:                                 \
> > > > > +               __val = csr_read(__csr_num);            \
> > > > > +               break; }
> > > > > +#define switchcase_csr_read_2(__csr_num, __val)         {\
> > > > > +       switchcase_csr_read(__csr_num + 0, __val)        \
> > > > > +       switchcase_csr_read(__csr_num + 1, __val)}
> > > > > +#define switchcase_csr_read_4(__csr_num, __val)         {\
> > > > > +       switchcase_csr_read_2(__csr_num + 0, __val)      \
> > > > > +       switchcase_csr_read_2(__csr_num + 2, __val)}
> > > > > +#define switchcase_csr_read_8(__csr_num, __val)         {\
> > > > > +       switchcase_csr_read_4(__csr_num + 0, __val)      \
> > > > > +       switchcase_csr_read_4(__csr_num + 4, __val)}
> > > > > +#define switchcase_csr_read_16(__csr_num, __val)        {\
> > > > > +       switchcase_csr_read_8(__csr_num + 0, __val)      \
> > > > > +       switchcase_csr_read_8(__csr_num + 8, __val)}
> > > > > +#define switchcase_csr_read_32(__csr_num, __val)        {\
> > > > > +       switchcase_csr_read_16(__csr_num + 0, __val)     \
> > > > > +       switchcase_csr_read_16(__csr_num + 16, __val)}
> > > > > +
> > > > > +       unsigned long ret = 0;
> > > > > +
> > > > > +       switch (csr_num) {
> > > > > +       switchcase_csr_read_32(CSR_CYCLE, ret)
> > > > > +       default:
> > > > > +               break;
> > > > > +       }
> > > > > +
> > > > > +       return ret;
> > > > > +#undef switchcase_csr_read_32
> > > > > +#undef switchcase_csr_read_16
> > > > > +#undef switchcase_csr_read_8
> > > > > +#undef switchcase_csr_read_4
> > > > > +#undef switchcase_csr_read_2
> > > > > +#undef switchcase_csr_read
> > > > > +}
> > > > > +
> > > > > +static u64 read_perf_counter(unsigned int counter)
> > > > > +{
> > > > > +       return csr_read_num(CSR_CYCLE + counter);
> > > > > +}
> > > > > +
> > > > > +static u64 read_timestamp(void)
> > > > > +{
> > > > > +       return csr_read_num(CSR_TIME);
> > > > > +}
> > > > > +
> > > > >  #else
> > > > >  static u64 read_perf_counter(unsigned int counter __maybe_unused) { return 0; }
> > > > >  static u64 read_timestamp(void) { return 0; }
> > > > > --
> > > > > 2.39.2
> > > > >



More information about the linux-arm-kernel mailing list