| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Date: Thu, 27 Jan 2022 09:48:09 +0100
From: Geert Uytterhoeven <geert@...ux-m68k.org>
To: Atish Patra <atishp@...shpatra.org>
Cc: Jessica Clarke <jrtc27@...c27.com>,
Atish Patra <atishp@...osinc.com>,
Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
Anup Patel <anup@...infault.org>,
Albert Ou <aou@...s.berkeley.edu>,
Damien Le Moal <damien.lemoal@....com>,
devicetree <devicetree@...r.kernel.org>,
Jisheng Zhang <jszhang@...nel.org>,
Krzysztof Kozlowski <krzysztof.kozlowski@...onical.com>,
linux-riscv <linux-riscv@...ts.infradead.org>,
Palmer Dabbelt <palmer@...belt.com>,
Paul Walmsley <paul.walmsley@...ive.com>,
Rob Herring <robh+dt@...nel.org>
Subject: Re: [PATCH v3 6/6] RISC-V: Do not use cpumask data structure for
hartid bitmap
Hi Atish,
On Thu, Jan 27, 2022 at 2:02 AM Atish Patra <atishp@...shpatra.org> wrote:
> On Wed, Jan 26, 2022 at 1:10 AM Geert Uytterhoeven <geert@...ux-m68k.org> wrote:
> > On Wed, Jan 26, 2022 at 9:28 AM Geert Uytterhoeven <geert@...ux-m68k.org> wrote:
> > > On Wed, Jan 26, 2022 at 3:21 AM Atish Patra <atishp@...shpatra.org> wrote:
> > > > On Tue, Jan 25, 2022 at 2:26 PM Jessica Clarke <jrtc27@...c27.com> wrote:
> > > > > On 20 Jan 2022, at 09:09, Atish Patra <atishp@...osinc.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@...infault.org> (For Linux RISC-V changes)
> > > > > > Acked-by: Anup Patel <anup@...infault.org> (For KVM RISC-V changes)
> > > > > > Signed-off-by: Atish Patra <atishp@...osinc.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
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@...ux-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
Powered by blists - more mailing lists