Hi Atish, On Thu, Jan 27, 2022 at 2:02 AM Atish Patra <atishp@xxxxxxxxxxxxxx> wrote: > On Wed, Jan 26, 2022 at 1:10 AM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote: > > On Wed, Jan 26, 2022 at 9:28 AM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote: > > > On Wed, Jan 26, 2022 at 3:21 AM Atish Patra <atishp@xxxxxxxxxxxxxx> wrote: > > > > On Tue, Jan 25, 2022 at 2:26 PM Jessica Clarke <jrtc27@xxxxxxxxxx> wrote: > > > > > On 20 Jan 2022, at 09:09, Atish Patra <atishp@xxxxxxxxxxxx> 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@xxxxxxxxxxxxxx> (For Linux RISC-V changes) > > > > > > Acked-by: Anup Patel <anup@xxxxxxxxxxxxxx> (For KVM RISC-V changes) > > > > > > Signed-off-by: Atish Patra <atishp@xxxxxxxxxxxx> > > > > > > > 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 Thanks, that fixes the issue for me. > --- a/arch/riscv/kernel/sbi.c > +++ b/arch/riscv/kernel/sbi.c > @@ -8,6 +8,8 @@ > #include <linux/init.h> > #include <linux/pm.h> > #include <linux/reboot.h> > +#include <linux/sort.h> > + > #include <asm/sbi.h> > #include <asm/smp.h> > > @@ -85,7 +87,7 @@ static unsigned long > __sbi_v01_cpumask_to_hartmask(const struct cpumask *cpu_mas > pr_warn("Unable to send any request to hartid > BITS_PER_LONG for > SBI v0.1\n"); > break; > } > - hmask |= 1 << hartid; > + hmask |= 1UL << hartid; > } > > return hmask; > @@ -160,7 +162,7 @@ static int __sbi_send_ipi_v01(const struct cpumask > *cpu_mask) > { > unsigned long hart_mask; > > - if (!cpu_mask) > + if (!cpu_mask || cpumask_empty(cpu_mask)) > cpu_mask = cpu_online_mask; > hart_mask = __sbi_v01_cpumask_to_hartmask(cpu_mask); > > @@ -176,7 +178,7 @@ static int __sbi_rfence_v01(int fid, const struct > cpumask *cpu_mask, > int result = 0; > unsigned long hart_mask; > > - if (!cpu_mask) > + if (!cpu_mask || cpumask_empty(cpu_mask)) > cpu_mask = cpu_online_mask; > hart_mask = __sbi_v01_cpumask_to_hartmask(cpu_mask); > > @@ -236,6 +238,18 @@ static int __sbi_rfence_v01(int fid, const struct > cpumask *cpu_mask, > static void sbi_set_power_off(void) {} > #endif /* CONFIG_RISCV_SBI_V01 */ > > +static int cmp_ulong(const void *A, const void *B) > +{ > + const unsigned long *a = A, *b = B; > + > + if (*a < *b) > + return -1; > + else if (*a > *b) > + return 1; > + else > + return 0; > +} > + > static void __sbi_set_timer_v02(uint64_t stime_value) > { > #if __riscv_xlen == 32 > @@ -251,13 +265,22 @@ static int __sbi_send_ipi_v02(const struct > cpumask *cpu_mask) > { > unsigned long hartid, cpuid, hmask = 0, hbase = 0; > struct sbiret ret = {0}; > - int result; > + int result, index = 0, max_index = 0; > + unsigned long hartid_arr[NR_CPUS] = {0}; > > - 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]; > if (hmask && ((hbase + BITS_PER_LONG) <= hartid)) { > ret = sbi_ecall(SBI_EXT_IPI, SBI_EXT_IPI_SEND_IPI, > hmask, hbase, 0, 0, 0, 0); > @@ -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. > > - 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? > if (hmask && ((hbase + BITS_PER_LONG) <= hartid)) { > result = __sbi_rfence_v02_call(fid, hmask, hbase, > start, size, arg4, arg5); Gr{oetje,eeting}s, Geert -- Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx 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
