linux-kernel - Re: [PATCH] RISC-V: Don't use IPIs in flush_icache

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
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <87y1bzch0u.fsf@all.your.base.are.belong.to.us>
Date: Mon, 05 Feb 2024 12:00:33 +0100
From: Björn Töpel <bjorn@...nel.org>
To: Anup Patel <apatel@...tanamicro.com>, Alexandre Ghiti <alex@...ti.fr>
Cc: Palmer Dabbelt <palmer@...belt.com>, Paul Walmsley
 <paul.walmsley@...ive.com>, Anup Patel <anup@...infault.org>,
 linux-kernel@...r.kernel.org, Atish Patra <atishp@...shpatra.org>,
 linux-riscv@...ts.infradead.org, Andrew Jones <ajones@...tanamicro.com>
Subject: Re: [PATCH] RISC-V: Don't use IPIs in flush_icache_all() when
 patching text

Anup Patel <apatel@...tanamicro.com> writes:

> On Mon, Feb 5, 2024 at 11:52 AM Alexandre Ghiti <alex@...ti.fr> wrote:
>>
>> Hi Anup,
>>
>> On 05/02/2024 05:29, Anup Patel wrote:
>> > If some of the HARTs are parked by stop machine then IPI-based
>> > flushing in flush_icache_all() will hang. This hang is observed
>> > when text patching is invoked by various debug and BPF features.
>> >
>> > To avoid this hang, we force use of SBI-based icache flushing
>> > when patching text.
>> >
>> > Fixes: 627922843235 ("RISC-V: Use IPIs for remote icache flush when possible")
>> > Reported-by: Bjorn Topel <bjorn@...nel.org>
>> > Closes: https://gist.github.com/bjoto/04a580568378f3b5483af07cd9d22501
>> > Signed-off-by: Anup Patel <apatel@...tanamicro.com>
>> > ---
>> >   arch/riscv/include/asm/cacheflush.h | 7 ++++---
>> >   arch/riscv/kernel/hibernate.c       | 2 +-
>> >   arch/riscv/kernel/patch.c           | 4 ++--
>> >   arch/riscv/mm/cacheflush.c          | 7 ++++---
>> >   4 files changed, 11 insertions(+), 9 deletions(-)
>> >
>> > diff --git a/arch/riscv/include/asm/cacheflush.h b/arch/riscv/include/asm/cacheflush.h
>> > index a129dac4521d..561e079f34af 100644
>> > --- a/arch/riscv/include/asm/cacheflush.h
>> > +++ b/arch/riscv/include/asm/cacheflush.h
>> > @@ -32,7 +32,8 @@ static inline void flush_dcache_page(struct page *page)
>> >    * RISC-V doesn't have an instruction to flush parts of the instruction cache,
>> >    * so instead we just flush the whole thing.
>> >    */
>> > -#define flush_icache_range(start, end) flush_icache_all()
>> > +#define flush_icache_range(start, end) flush_icache_all(true)
>> > +#define flush_icache_patch_range(start, end) flush_icache_all(false)
>> >   #define flush_icache_user_page(vma, pg, addr, len) \
>> >       flush_icache_mm(vma->vm_mm, 0)
>> >
>> > @@ -43,12 +44,12 @@ static inline void flush_dcache_page(struct page *page)
>> >
>> >   #ifndef CONFIG_SMP
>> >
>> > -#define flush_icache_all() local_flush_icache_all()
>> > +#define flush_icache_all(want_ipi) local_flush_icache_all()
>> >   #define flush_icache_mm(mm, local) flush_icache_all()
>> >
>> >   #else /* CONFIG_SMP */
>> >
>> > -void flush_icache_all(void);
>> > +void flush_icache_all(bool want_ipi);
>> >   void flush_icache_mm(struct mm_struct *mm, bool local);
>> >
>> >   #endif /* CONFIG_SMP */
>> > diff --git a/arch/riscv/kernel/hibernate.c b/arch/riscv/kernel/hibernate.c
>> > index 671b686c0158..388f10e187ba 100644
>> > --- a/arch/riscv/kernel/hibernate.c
>> > +++ b/arch/riscv/kernel/hibernate.c
>> > @@ -153,7 +153,7 @@ int swsusp_arch_suspend(void)
>> >       } else {
>> >               suspend_restore_csrs(hibernate_cpu_context);
>> >               flush_tlb_all();
>> > -             flush_icache_all();
>> > +             flush_icache_all(true);
>> >
>> >               /*
>> >                * Tell the hibernation core that we've just restored the memory.
>> > diff --git a/arch/riscv/kernel/patch.c b/arch/riscv/kernel/patch.c
>> > index 37e87fdcf6a0..721e144a7847 100644
>> > --- a/arch/riscv/kernel/patch.c
>> > +++ b/arch/riscv/kernel/patch.c
>> > @@ -182,7 +182,7 @@ int patch_text_set_nosync(void *addr, u8 c, size_t len)
>> >       ret = patch_insn_set(tp, c, len);
>> >
>> >       if (!ret)
>> > -             flush_icache_range((uintptr_t)tp, (uintptr_t)tp + len);
>> > +             flush_icache_patch_range((uintptr_t)tp, (uintptr_t)tp + len);
>> >
>> >       return ret;
>> >   }
>> > @@ -217,7 +217,7 @@ int patch_text_nosync(void *addr, const void *insns, size_t len)
>> >       ret = patch_insn_write(tp, insns, len);
>> >
>> >       if (!ret)
>> > -             flush_icache_range((uintptr_t) tp, (uintptr_t) tp + len);
>> > +             flush_icache_patch_range((uintptr_t) tp, (uintptr_t) tp + len);
>> >
>> >       return ret;
>> >   }
>> > diff --git a/arch/riscv/mm/cacheflush.c b/arch/riscv/mm/cacheflush.c
>> > index 55a34f2020a8..03cd3d4831ef 100644
>> > --- a/arch/riscv/mm/cacheflush.c
>> > +++ b/arch/riscv/mm/cacheflush.c
>> > @@ -17,11 +17,12 @@ static void ipi_remote_fence_i(void *info)
>> >       return local_flush_icache_all();
>> >   }
>> >
>> > -void flush_icache_all(void)
>> > +void flush_icache_all(bool want_ipi)
>> >   {
>> >       local_flush_icache_all();
>> >
>> > -     if (IS_ENABLED(CONFIG_RISCV_SBI) && !riscv_use_ipi_for_rfence())
>> > +     if (IS_ENABLED(CONFIG_RISCV_SBI) &&
>> > +         (!want_ipi || !riscv_use_ipi_for_rfence()))
>> >               sbi_remote_fence_i(NULL);
>> >       else
>> >               on_each_cpu(ipi_remote_fence_i, NULL, 1);
>> > @@ -87,7 +88,7 @@ void flush_icache_pte(pte_t pte)
>> >       struct folio *folio = page_folio(pte_page(pte));
>> >
>> >       if (!test_bit(PG_dcache_clean, &folio->flags)) {
>> > -             flush_icache_all();
>> > +             flush_icache_all(true);
>> >               set_bit(PG_dcache_clean, &folio->flags);
>> >       }
>> >   }
>>
>>
>> Since patch_text_cb() is run on all cpus, couldn't we completely avoid
>> any remote icache flush by slightly changing patch_text_cb() instead as
>> follows?
>
> Unfortunately patch_text_cb() is not the only user of patch_text_nosync
> since patch_text_nosync() and patch_text_set_nosync() are called directly
> from other places as well.

Yeah. There is one more stop_machine() text patching user, and that's
ftrace. ftrace is using stop_machine() with the last argument set to
NULL, so only patching on *any* hart. Continuing on Alex' idea would be
to place an IPI flush in ftrace_arch_code_modify_post_process(),
unfortately that's too late since we're already moved on from
stop_machine().

> We have to update all users of patch_text_nosync() and
> patch_text_set_nosync() to move to local icache flushes which
> is a much bigger change.

Only the ftrace stop_machine() user, right? Alex solution is sufficient
for patch_text(). I'm not a super fan of conditionally calling into SBI
and passing around boolean context flags as a workaround... :-( Any
other alternatives?

The obvious fixing text patching not to be completly useless on RISC-V,
but that's an even bigger patch...


Björn