Re: [PATCH v3 6/6] RISC-V: Do not use cpumask data structure for hartid bitmap

[Anup Patel <anup@brainfault.org>]

On Fri, Jan 28, 2022 at 2:09 PM Geert Uytterhoeven <geert@linux-m68k.org> wrote:
>
> Hi Atish,
>
> On Fri, Jan 28, 2022 at 1:13 AM Atish Patra <atishp@atishpatra.org> wrote:
> > On Thu, Jan 27, 2022 at 12:48 AM Geert Uytterhoeven <geert@linux-m68k.org> wrote:
> >> On Thu, Jan 27, 2022 at 2:02 AM Atish Patra <atishp@atishpatra.org> wrote:
> >> > On Wed, Jan 26, 2022 at 1:10 AM Geert Uytterhoeven <geert@linux-m68k.org> wrote:
> >> > > On Wed, Jan 26, 2022 at 9:28 AM Geert Uytterhoeven <geert@linux-m68k.org> wrote:
> >> > > > On Wed, Jan 26, 2022 at 3:21 AM Atish Patra <atishp@atishpatra.org> wrote:
> >> > > > > On Tue, Jan 25, 2022 at 2:26 PM Jessica Clarke <jrtc27@jrtc27.com> wrote:
> >> > > > > > On 20 Jan 2022, at 09:09, Atish Patra <atishp@rivosinc.com> wrote:
> >> > > > > > > Currently, SBI APIs accept a hartmask that is generated from struct
> >> > > > > > > cpumask. Cpumask data structure can hold upto NR_CPUs value. Thus, it
> >> > > > > > > is not the correct data structure for hartids as it can be higher
> >> > > > > > > than NR_CPUs for platforms with sparse or discontguous hartids.
> >> > > > > > >
> >> > > > > > > Remove all association between hartid mask and struct cpumask.
> >> > > > > > >
> >> > > > > > > Reviewed-by: Anup Patel <anup@brainfault.org> (For Linux RISC-V changes)
> >> > > > > > > Acked-by: Anup Patel <anup@brainfault.org> (For KVM RISC-V changes)
> >> > > > > > > Signed-off-by: Atish Patra <atishp@rivosinc.com>
> >> > > >
> >> > > > > I am yet to reproduce it on my end.
> >> > > > > @Geert Uytterhoeven: can you please try the below diff on your end.
> >> > > >
> >> > > > Unfortunately it doesn't fix the issue for me.
> >> > > >
> >> > > > /me debugging...
> >> > >
> >> > > Found it: after this commit, the SBI_EXT_RFENCE_REMOTE_FENCE_I and
> >> > > SBI_EXT_RFENCE_REMOTE_SFENCE_VMA ecalls are now called with
> >> > > hmask = 0x8000000000000001 and hbase = 1 instead of hmask = 3 and
> >> > > hbase = 0.
> >> > >
> >> > > cpuid 1 maps to  hartid 0
> >> > > cpuid 0 maps to hartid 1
> >> > >
> >> > >     __sbi_rfence_v02:364: cpuid 1 hartid 0
> >> > >     __sbi_rfence_v02:377: hartid 0 hbase 1
> >> > >     hmask |= 1UL << (hartid - hbase);
> >> > >
> >> > > oops
> >> > >
> >> > >     __sbi_rfence_v02_call:303: SBI_EXT_RFENCE_REMOTE_FENCE_I hmask
> >> > > 8000000000000001 hbase 1
> >> > >
> >> >
> >> > Ahh yes. hmask will be incorrect if the bootcpu(cpu 0) is a higher
> >> > hartid and it is trying to do a remote tlb flush/IPI
> >> > to lower the hartid. We should generate the hartid array before the loop.
> >> >
> >> > Can you try this diff ? It seems to work for me during multiple boot
> >> > cycle on the unleashed.
> >> >
> >> > You can find the patch here as well
> >> > https://github.com/atishp04/linux/commits/v5.17-rc1
>
> >> > @@ -345,13 +368,21 @@ static int __sbi_rfence_v02(int fid, const
> >> > struct cpumask *cpu_mask,
> >> >       unsigned long arg4, unsigned long arg5)
> >> >  {
> >> >   unsigned long hartid, cpuid, hmask = 0, hbase = 0;
> >> > - int result;
> >> > + int result, index = 0, max_index = 0;
> >> > + unsigned long hartid_arr[NR_CPUS] = {0};
> >>
> >> That's up to 256 bytes on the stack. And more if the maximum
> >> number of cores is increased.
> >>
> >
> > Yeah. We can switch to dynamic allocation using kmalloc based on
> > the number of bits set in the cpumask.
>
> Even more overhead...
>
> >> > - if (!cpu_mask)
> >> > + if (!cpu_mask || cpumask_empty(cpu_mask))
> >> >   cpu_mask = cpu_online_mask;
> >> >
> >> >   for_each_cpu(cpuid, cpu_mask) {
> >> >   hartid = cpuid_to_hartid_map(cpuid);
> >> > + hartid_arr[index] = hartid;
> >> > + index++;
> >> > + }
> >> > + max_index = index;
> >> > + sort(hartid_arr, max_index, sizeof(unsigned long), cmp_ulong, NULL);
> >> > + for (index = 0; index < max_index; index++) {
> >> > + hartid = hartid_arr[index];
> >>
> >> That looks expensive to me.
> >>
> >> What about shifting hmask and adjusting hbase if a hartid is
> >> lower than the current hbase?
> >
> > That will probably work for current systems but it will fail when we have hartid > 64.
> > The below logic as it assumes that the hartids are in order. We can have a situation
> > where a two consecutive cpuid belong to hartids that require two invocations of sbi call
> > because the number of harts exceeds BITS_PER_LONG.
>
> If the number of harts exceeds BITS_PER_LONG, you always need multiple
> calls, right?
>
> I think the below (gmail-whitespace-damaged diff) should work:
>
> --- a/arch/riscv/kernel/sbi.c
> +++ b/arch/riscv/kernel/sbi.c
> @@ -249,7 +249,7 @@ static void __sbi_set_timer_v02(uint64_t stime_value)
>
>  static int __sbi_send_ipi_v02(const struct cpumask *cpu_mask)
>  {
> -       unsigned long hartid, cpuid, hmask = 0, hbase = 0;
> +       unsigned long hartid, cpuid, hmask = 0, hbase = 0, htop = 0;
>         struct sbiret ret = {0};
>         int result;
>
> @@ -258,16 +258,27 @@ static int __sbi_send_ipi_v02(const struct
> cpumask *cpu_mask)
>
>         for_each_cpu(cpuid, cpu_mask) {
>                 hartid = cpuid_to_hartid_map(cpuid);
> -               if (hmask &&
> -                   (hartid < hbase || hartid >= hbase + BITS_PER_LONG)) {
> -                       ret = sbi_ecall(SBI_EXT_IPI, SBI_EXT_IPI_SEND_IPI,
> -                                       hmask, hbase, 0, 0, 0, 0);
> -                       if (ret.error)
> -                               goto ecall_failed;
> -                       hmask = 0;
> +               if (hmask) {
> +                       if (hartid + BITS_PER_LONG <= htop ||
> +                           hartid >= hbase + BITS_PER_LONG) {
> +                               ret = sbi_ecall(SBI_EXT_IPI,
> +                                               SBI_EXT_IPI_SEND_IPI, hmask,
> +                                               hbase, 0, 0, 0, 0);
> +                               if (ret.error)
> +                                       goto ecall_failed;
> +                               hmask = 0;
> +                       } else if (hartid < hbase) {
> +                               /* shift the mask to fit lower hartid */
> +                               hmask <<= hbase - hartid;
> +                               hbase = hartid;
> +                       }
>                 }
> -               if (!hmask)
> +               if (!hmask) {
>                         hbase = hartid & -BITS_PER_LONG;
> +                       htop = hartid;
> +               } else if (hartid > htop) {
> +                       htop = hartid;
> +               }
>                 hmask |= 1UL << (hartid - hbase);
>         }
>
> @@ -344,7 +355,7 @@ static int __sbi_rfence_v02(int fid, const struct
> cpumask *cpu_mask,
>                             unsigned long start, unsigned long size,
>                             unsigned long arg4, unsigned long arg5)
>  {
> -       unsigned long hartid, cpuid, hmask = 0, hbase = 0;
> +       unsigned long hartid, cpuid, hmask = 0, hbase = 0, htop = 0;
>         int result;
>
>         if (!cpu_mask || cpumask_empty(cpu_mask))
> @@ -352,16 +363,26 @@ static int __sbi_rfence_v02(int fid, const
> struct cpumask *cpu_mask,
>
>         for_each_cpu(cpuid, cpu_mask) {
>                 hartid = cpuid_to_hartid_map(cpuid);
> -               if (hmask &&
> -                   (hartid < hbase || hartid >= hbase + BITS_PER_LONG)) {
> -                       result = __sbi_rfence_v02_call(fid, hmask, hbase,
> -                                                      start, size, arg4, arg5);
> -                       if (result)
> -                               return result;
> -                       hmask = 0;
> +               if (hmask) {
> +                       if (hartid + BITS_PER_LONG <= htop ||
> +                           hartid >= hbase + BITS_PER_LONG) {
> +                               result = __sbi_rfence_v02_call(fid, hmask,
> +                                               hbase, start, size, arg4, arg5);
> +                               if (result)
> +                                       return result;
> +                               hmask = 0;
> +                       } else if (hartid < hbase) {
> +                               /* shift the mask to fit lower hartid */
> +                               hmask <<= hbase - hartid;
> +                               hbase = hartid;
> +                       }
> +               }
> +               if (!hmask) {
> +                       hbase = hartid;
> +                       htop = hartid;
> +               } else if (hartid > htop) {
> +                       htop = hartid;
>                 }
> -               if (!hmask)
> -                       hbase = hartid & -BITS_PER_LONG;
>                 hmask |= 1UL << (hartid - hbase);
>         }
>
> Another simpler solution would be to just round hbase down to a
> multiple of 32/64 (gmail-whitespace-damaged diff):
>
> --- a/arch/riscv/kernel/sbi.c
> +++ b/arch/riscv/kernel/sbi.c
> @@ -258,16 +258,16 @@ static int __sbi_send_ipi_v02(const struct
> cpumask *cpu_mask)
>
>         for_each_cpu(cpuid, cpu_mask) {
>                 hartid = cpuid_to_hartid_map(cpuid);
> -               if (hmask && ((hbase + BITS_PER_LONG) <= hartid)) {
> +               if (hmask &&
> +                   (hartid < hbase || hartid >= hbase + BITS_PER_LONG)) {
>                         ret = sbi_ecall(SBI_EXT_IPI, SBI_EXT_IPI_SEND_IPI,
>                                         hmask, hbase, 0, 0, 0, 0);
>                         if (ret.error)
>                                 goto ecall_failed;
>                         hmask = 0;
> -                       hbase = 0;
>                 }
>                 if (!hmask)
> -                       hbase = hartid;
> +                       hbase = hartid & -BITS_PER_LONG;
>                 hmask |= 1UL << (hartid - hbase);
>         }
>
> @@ -352,16 +352,16 @@ static int __sbi_rfence_v02(int fid, const
> struct cpumask *cpu_mask,
>
>         for_each_cpu(cpuid, cpu_mask) {
>                 hartid = cpuid_to_hartid_map(cpuid);
> -               if (hmask && ((hbase + BITS_PER_LONG) <= hartid)) {
> +               if (hmask &&
> +                   (hartid < hbase || hartid >= hbase + BITS_PER_LONG)) {
>                         result = __sbi_rfence_v02_call(fid, hmask, hbase,
>                                                        start, size, arg4, arg5);
>                         if (result)
>                                 return result;
>                         hmask = 0;
> -                       hbase = 0;
>                 }
>                 if (!hmask)
> -                       hbase = hartid;
> +                       hbase = hartid & -BITS_PER_LONG;
>                 hmask |= 1UL << (hartid - hbase);
>         }
>
> But that means multiple SBI calls if you have e.g. hartids 1-64.
> The shifted mask solution doesn't suffer from that.
> Both solutions don't sort the CPUs, so they are suboptimal in case of
> hartid numberings like 0, 64, 1, 65, ...

In most cases, the hartids will be in sorted order under /cpus DT node
but it is not guaranteed that boot_cpu will be having smallest hartid

This means hartid numbering will be something like:
0, 1, 2, .....,
64, 0, 1, 2, ....
31, 0, 1, 2, .....

>
> What do you think?

Assuming hartids under /cpus DT node are ordered, I think your
approach will only have one additional SBI call compared to Atish's
approach but Atish's approach will require more memory with
increasing NR_CPUS so I suggest we go with your approach.

Can you send a patch with your approach ?

Regards,
Anup


> Thanks!
>
> Gr{oetje,eeting}s,
>
>                         Geert
>
> --
> Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org
>
> In personal conversations with technical people, I call myself a hacker. But
> when I'm talking to journalists I just say "programmer" or something like that.
>                                 -- Linus Torvalds