Re: [PATCH v7] riscv: mm: Trace TLB flush requests

From: Roman Storozhenko

Date: Sat Nov 14 2026 - 04:45:07 EST


Hi,

I'd like to follow up on this patch, as it has been a few weeks
since v7 was submitted.

The patch has gone through several revisions, incorporating
review feedback and additional testing.

I'd appreciate a review from the RISC-V maintainers and any
feedback on what might still need to be addressed to move it
forward.

Thanks,
Roman

On Tue, Oct 20, 2026 at 2:52 PM Roman Storozhenko
<romeusmeister@xxxxxxxxx> wrote:
>
> On Tue, Oct 20, 2026 at 3:37 AM Jisheng Zhang <jszhang@xxxxxxxxxx> wrote:
> >
> > On Mon, Oct 19, 2026 at 07:20:51PM +0200, Roman 'Hedin' Storozhenko wrote:
> > > Make RISC-V TLB flush requests observable together with the
> > > architecture-specific context available when the request reaches
> > > the RISC-V implementation.
> > >
> > > The generic tlb:tlb_flush event provides reason and page-count
> > > information. At the RISC-V implementation level, additional
> > > context is available, including the invalidation range and stride,
> > > ASID, target CPU mask, and invalidation scope.
> > >
> > > This context is useful for debugging and workload analysis. The
> > > target CPU mask shows the address-space CPU footprint involved in
> > > an invalidation, while range and stride distinguish requests at
> > > different mapping granularities. The ASID and scope help correlate
> > > requests with MM, scheduler, IPI, and firmware activity.
> >
> > Hi Roman,
> >
> > no doubt the info is useful. But as commented in v2, the problem
> > is whether the information can only be gotten with tracepoints,
> > if no(I believe no, because the info exported by this patch are
> > all function parameters now), what't the benefit of it over kprobe
> > or function trace given the non-trivial tracepoint overhead on riscv
> > platforms until gcc14, and we still dunno whether newer gcc has
> > improved the situation or not.
> >
> > So I prefer not adding this tracepoint. But let's wait for riscv
> > maintainers decision.
> >
> > Thanks
>
> Hi Jisheng,
>
> Thanks for your feedback.
>
> I agree that much of this information can be obtained using
> kprobes or function tracing. However, I don't think that
> availability through dynamic instrumentation makes a dedicated
> tracepoint unnecessary. The same argument could be made about
> many existing kernel tracepoints.
>
> The main benefit here is not access to individual function
> parameters, but exposing a well-defined semantic event: a TLB
> invalidation request, including its scope and target CPUs,
> before it is dispatched.
>
> Reconstructing this information using kprobes requires knowledge
> of the current implementation, function signatures, and how
> different parameters relate to the invalidation scope. Such
> instrumentation can become fragile as the implementation evolves.
>
> A dedicated tracepoint provides a consistent interface that
> can be consumed directly by tools such as trace-cmd, perf,
> and BPF, without requiring users to reconstruct the request
> semantics from implementation details.
>
> Regarding overhead, I agree that it matters, particularly when
> tracing is disabled. However, I think we should distinguish
> the usefulness of the tracing interface from compiler-specific
> code generation issues. The latter can be improved independently
> and should not necessarily determine whether the former is worth
> providing.
>
> If there is still significant overhead with current compilers,
> I think it would be useful to quantify it rather than assume
> it is a permanent limitation.
>
> I believe the semantic observability provided by this tracepoint
> justifies its inclusion, but I agree that the final decision
> belongs to the RISC-V maintainers.
>
> Regards,
> Roman
>
> > >
> > > Add riscv_tlb:riscv_tlb_flush_request in flush_tlb_all() and
> > > __flush_tlb_range(). Record start, size, stride, ASID (or the
> > > no-ASID sentinel), whether an mm is associated with the request,
> > > the target CPU mask and its weight, and the derived scope.
> > >
> > > Record the complete target mask because CPU identity cannot be
> > > reconstructed from a count. Calculate its weight from the recorded
> > > bitmap so both fields agree even if mm_cpumask() changes during
> > > recording.
> > >
> > > The event describes an invalidation request before dispatch. It
> > > does not report whether a remote flush uses SBI RFENCE or Linux
> > > cross-CPU coordination, since that choice is effectively established
> > > during initialization rather than made for each request.
> > >
> > > The address field (start) may contain kernel virtual addresses
> > > for flushes involving kernel mappings. Require perf tracepoint
> > > permissions even for counting-only events, since perf filters
> > > could otherwise reveal address information through event counts.
> > >
> > > Tested on QEMU virt with OpenSBI using local and shared-mm
> > > mprotect()/munmap() workloads, exercising single, range, and
> > > address-space invalidations. A MADV_PAGEOUT reclaim workload
> > > exercised global flushes. All four scope values were observed.
> > >
> > > Confirmed that trace-cmd records the event and that perf can count
> > > it with sufficient privileges. With perf_event_paranoid=0,
> > > unprivileged attempts to count the event using perf were denied,
> > > as intended.
> > >
> > > Tested local and shared-mm invalidations with
> > > CONFIG_UBSAN_ALIGNMENT=y on QEMU virt without tracepoint-related
> > > alignment warnings.
> > >
> > > Built with and without CONFIG_PERF_EVENTS.
> > >
> > > Signed-off-by: Roman 'Hedin' Storozhenko <romeusmeister@xxxxxxxxx>
> > > ---
> > > Add a RISC-V tracepoint for observing TLB invalidation requests
> > > before they are dispatched by the architecture-specific
> > > implementation.
> > >
> > > The generic tlb:tlb_flush event describes TLB flush activity using
> > > reason and page-count information. The RISC-V event records
> > > additional context, including the invalidation range and stride,
> > > ASID, target CPU mask and its weight, and derived request scope.
> > >
> > > This context helps correlate TLB invalidation activity with MM,
> > > scheduler, IPI, and firmware activity. The event describes the
> > > request rather than the execution backend. SBI RFENCE versus Linux
> > > cross-CPU coordination is effectively determined at initialization
> > > rather than selected independently for each request.
> > >
> > > The target CPU mask weight is derived from the recorded bitmap to
> > > keep both fields consistent under concurrent mm_cpumask() updates.
> > > Since the address field (start) may contain kernel virtual
> > > addresses, the event requires perf tracepoint permissions even
> > > for counting-only access.
> > >
> > > Tested on QEMU virt with OpenSBI using local and shared-mm
> > > mprotect()/munmap() workloads. A MADV_PAGEOUT reclaim workload
> > > on QEMU virt with APLIC+IMSIC exercised global flushes. All four
> > > scope values -- single, range, address-space, and all -- were
> > > observed.
> > >
> > > Confirmed trace-cmd recording and privileged perf counting.
> > > With perf_event_paranoid=0, unprivileged attempts to count the
> > > event using perf were denied, as intended. The kernel was built
> > > with and without CONFIG_PERF_EVENTS.
> > >
> > > Tested local and shared-mm invalidations with
> > > CONFIG_UBSAN_ALIGNMENT=y on QEMU virt without tracepoint-related
> > > alignment warnings.
> > > ---
> > > Changes in v7:
> > > - Rename the tracepoint header to asm/tlb_trace.h to make its
> > > purpose explicit, as suggested by Steven Rostedt.
> > > - Link to v6: https://lore.kernel.org/r/20261018-tlb_tracepoint-v6-1-14476427870f@xxxxxxxxx
> > >
> > > Changes in v6:
> > > - Change the scope field from u8 to u32 to ensure the dynamic
> > > CPU mask is properly aligned for bitmap operations.
> > > - Verify local and shared-mm TLB flush tracing with
> > > CONFIG_UBSAN_ALIGNMENT=y on QEMU virt.
> > > - Link to v5: https://lore.kernel.org/r/20261017-tlb_tracepoint-v5-1-256fb5393a7f@xxxxxxxxx
> > >
> > > Changes in v5:
> > > - Move the architecture-specific trace event definition under
> > > arch/riscv/, as requested by Steven Rostedt.
> > > - Calculate target_mask_weight from the recorded CPU mask to
> > > keep both fields consistent under concurrent updates.
> > > - Require perf tracepoint permissions for counting-only access,
> > > since the event may expose kernel virtual addresses.
> > > - Link to v4: https://lore.kernel.org/r/20261015-tlb_tracepoint-v4-1-f238a2bd8f7e@xxxxxxxxx
> > >
> > > Changes in v4:
> > > - Drop unrelated whitespace-only changes in arch/riscv/mm/tlbflush.c.
> > > - Link to v3: https://lore.kernel.org/all/20261015-tlb_tracepoint-v3-1-eef305a3bb63@xxxxxxxxx/
> > >
> > > Changes in v3:
> > > - Rebase on current RISC-V for-next.
> > > - Thanks to Jisheng Zhang for pointing out that SBI RFENCE versus Linux
> > > cross-CPU coordination is effectively a system-wide choice established
> > > during initialization rather than a per-request decision.
> > > - Drop the local/SBI RFENCE/cross-CPU path field.
> > > - Rename the event from riscv_tlb_flush_path to
> > > riscv_tlb_flush_request and refocus it on the semantic TLB
> > > invalidation request.
> > > - Emit one request event independently of the mechanism subsequently
> > > used to perform the invalidation.
> > > - Address checkpatch.pl --strict alignment issues.
> > > - Link to v2: https://lore.kernel.org/r/20260830-tlb_tracepoint-v2-1-e3c88c24df5b@xxxxxxxxx
> > >
> > > Changes in v2:
> > > - Use trace_call__riscv_tlb_flush_path() after the explicit
> > > trace_riscv_tlb_flush_path_enabled() check to avoid a second
> > > tracepoint static-key test, as suggested by Steven Rostedt.
> > > - Link to v1: https://lore.kernel.org/r/20260829-tlb_tracepoint-v1-1-dfdaede7e741@xxxxxxxxx
> > > ---
> > > arch/riscv/include/asm/tlb_trace.h | 103 +++++++++++++++++++++++++++++++++++++
> > > arch/riscv/mm/tlbflush.c | 45 ++++++++++++++--
> > > 2 files changed, 144 insertions(+), 4 deletions(-)
> > >
> > > diff --git a/arch/riscv/include/asm/tlb_trace.h b/arch/riscv/include/asm/tlb_trace.h
> > > new file mode 100644
> > > index 000000000000..60b48eb0d355
> > > --- /dev/null
> > > +++ b/arch/riscv/include/asm/tlb_trace.h
> > > @@ -0,0 +1,103 @@
> > > +/* SPDX-License-Identifier: GPL-2.0 */
> > > +#undef TRACE_SYSTEM
> > > +#define TRACE_SYSTEM riscv_tlb
> > > +
> > > +#if !defined(_TRACE_RISCV_TLB_TRACE_H) || defined(TRACE_HEADER_MULTI_READ)
> > > +#define _TRACE_RISCV_TLB_TRACE_H
> > > +
> > > +#include <linux/cpumask.h>
> > > +#include <linux/tracepoint.h>
> > > +#include <linux/perf_event.h>
> > > +
> > > +#ifndef _TRACE_RISCV_TLB_ENUMS
> > > +#define _TRACE_RISCV_TLB_ENUMS
> > > +
> > > +enum riscv_tlb_flush_scope {
> > > + RISCV_TLB_FLUSH_SCOPE_SINGLE,
> > > + RISCV_TLB_FLUSH_SCOPE_RANGE,
> > > + RISCV_TLB_FLUSH_SCOPE_ADDRESS_SPACE,
> > > + RISCV_TLB_FLUSH_SCOPE_ALL,
> > > +};
> > > +
> > > +#endif /* _TRACE_RISCV_TLB_ENUMS */
> > > +
> > > +TRACE_DEFINE_ENUM(RISCV_TLB_FLUSH_SCOPE_SINGLE);
> > > +TRACE_DEFINE_ENUM(RISCV_TLB_FLUSH_SCOPE_RANGE);
> > > +TRACE_DEFINE_ENUM(RISCV_TLB_FLUSH_SCOPE_ADDRESS_SPACE);
> > > +TRACE_DEFINE_ENUM(RISCV_TLB_FLUSH_SCOPE_ALL);
> > > +
> > > +#define show_riscv_tlb_flush_scope(scope) \
> > > + __print_symbolic(scope, \
> > > + { RISCV_TLB_FLUSH_SCOPE_SINGLE, "single" }, \
> > > + { RISCV_TLB_FLUSH_SCOPE_RANGE, "range" }, \
> > > + { RISCV_TLB_FLUSH_SCOPE_ADDRESS_SPACE, "address-space" }, \
> > > + { RISCV_TLB_FLUSH_SCOPE_ALL, "all" })
> > > +
> > > +/*
> > > + * Record the invalidation request received by the RISC-V architecture code
> > > + *
> > > + * The target CPU mask represents the CPUs Linux intends to cover for the
> > > + * request. It can be correlated with per-CPU activity, but does not describe
> > > + * which harts ultimately performed an invalidation.
> > > + *
> > > + * The ASID is hardware-visible and may be reused. It must not be treated as a
> > > + * persistent identifier for an mm.
> > > + *
> > > + * The stride describes the invalidation granularity supplied to the RISC-V
> > > + * implementation.
> > > + *
> > > + * The event records the invalidation request before it is dispatched. It does
> > > + * not describe the mechanism used to execute the request.
> > > + *
> > > + */
> > > +TRACE_EVENT(riscv_tlb_flush_request,
> > > + TP_PROTO(unsigned long start, unsigned long size,
> > > + unsigned long stride, unsigned long asid, bool has_mm,
> > > + const struct cpumask *cmask,
> > > + enum riscv_tlb_flush_scope scope),
> > > +
> > > + TP_ARGS(start, size, stride, asid, has_mm, cmask, scope),
> > > +
> > > + TP_STRUCT__entry(__field(unsigned long, start)
> > > + __field(unsigned long, size)
> > > + __field(unsigned long, stride)
> > > + __field(unsigned long, asid)
> > > + __field(bool, has_mm)
> > > + __field(unsigned int, target_mask_weight)
> > > + __cpumask(target_cpus)
> > > + __field(u32, scope)
> > > + ),
> > > +
> > > + TP_fast_assign(__entry->start = start;
> > > + __entry->size = size;
> > > + __entry->stride = stride;
> > > + __entry->asid = asid;
> > > + __entry->has_mm = has_mm;
> > > + __assign_cpumask(target_cpus, cpumask_bits(cmask));
> > > + __entry->target_mask_weight = cpumask_weight((const struct cpumask *)
> > > + __get_cpumask(target_cpus));
> > > + __entry->scope = scope;
> > > + ),
> > > +
> > > + TP_printk("start=%#lx size=%#lx stride=%#lx asid=%#lx has_mm=%d target_mask_weight=%u target_cpus=%s scope=%s",
> > > + __entry->start, __entry->size, __entry->stride,
> > > + __entry->asid, __entry->has_mm,
> > > + __entry->target_mask_weight, __get_cpumask(target_cpus),
> > > + show_riscv_tlb_flush_scope(__entry->scope))
> > > +);
> > > +
> > > +#ifdef CONFIG_PERF_EVENTS
> > > +TRACE_EVENT_PERF_PERM(riscv_tlb_flush_request,
> > > + perf_allow_tracepoint());
> > > +#endif
> > > +
> > > +#endif /* _TRACE_RISCV_TLB_TRACE_H */
> > > +
> > > +#undef TRACE_INCLUDE_PATH
> > > +#undef TRACE_INCLUDE_FILE
> > > +
> > > +#define TRACE_INCLUDE_PATH asm
> > > +#define TRACE_INCLUDE_FILE tlb_trace
> > > +
> > > +/* This part must be outside protection. */
> > > +#include <trace/define_trace.h>
> > > diff --git a/arch/riscv/mm/tlbflush.c b/arch/riscv/mm/tlbflush.c
> > > index 962db300a166..9dcb07fb3ac0 100644
> > > --- a/arch/riscv/mm/tlbflush.c
> > > +++ b/arch/riscv/mm/tlbflush.c
> > > @@ -9,6 +9,9 @@
> > > #include <asm/mmu_context.h>
> > > #include <asm/cpufeature.h>
> > >
> > > +#define CREATE_TRACE_POINTS
> > > +#include <asm/tlb_trace.h>
> > > +
> > > #define has_svinval() riscv_has_extension_unlikely(RISCV_ISA_EXT_SVINVAL)
> > >
> > > /*
> > > @@ -63,6 +66,32 @@ void local_flush_tlb_kernel_range(unsigned long start, unsigned long end)
> > > local_flush_tlb_range_asid(start, end - start, PAGE_SIZE, FLUSH_TLB_NO_ASID);
> > > }
> > >
> > > +static enum riscv_tlb_flush_scope
> > > +riscv_tlb_get_flush_scope(unsigned long size, unsigned long stride, bool has_mm)
> > > +{
> > > + if (size == FLUSH_TLB_MAX_SIZE)
> > > + return has_mm ? RISCV_TLB_FLUSH_SCOPE_ADDRESS_SPACE :
> > > + RISCV_TLB_FLUSH_SCOPE_ALL;
> > > +
> > > + return size <= stride ? RISCV_TLB_FLUSH_SCOPE_SINGLE :
> > > + RISCV_TLB_FLUSH_SCOPE_RANGE;
> > > +}
> > > +
> > > +static __always_inline void
> > > +riscv_tlb_trace_flush_request(const struct cpumask *cmask, unsigned long start,
> > > + unsigned long size, unsigned long stride,
> > > + unsigned long asid, bool has_mm)
> > > +{
> > > + enum riscv_tlb_flush_scope scope;
> > > +
> > > + if (!trace_riscv_tlb_flush_request_enabled())
> > > + return;
> > > +
> > > + scope = riscv_tlb_get_flush_scope(size, stride, has_mm);
> > > + trace_call__riscv_tlb_flush_request(start, size, stride, asid, has_mm,
> > > + cmask, scope);
> > > +}
> > > +
> > > static void __ipi_flush_tlb_all(void *info)
> > > {
> > > local_flush_tlb_all();
> > > @@ -70,12 +99,18 @@ static void __ipi_flush_tlb_all(void *info)
> > >
> > > void flush_tlb_all(void)
> > > {
> > > - if (num_online_cpus() < 2)
> > > + riscv_tlb_trace_flush_request(cpu_online_mask, 0,
> > > + FLUSH_TLB_MAX_SIZE, 0,
> > > + FLUSH_TLB_NO_ASID, false);
> > > +
> > > + if (num_online_cpus() < 2) {
> > > local_flush_tlb_all();
> > > - else if (riscv_use_sbi_for_rfence())
> > > - sbi_remote_sfence_vma_asid(NULL, 0, FLUSH_TLB_MAX_SIZE, FLUSH_TLB_NO_ASID);
> > > - else
> > > + } else if (riscv_use_sbi_for_rfence()) {
> > > + sbi_remote_sfence_vma_asid(NULL, 0, FLUSH_TLB_MAX_SIZE,
> > > + FLUSH_TLB_NO_ASID);
> > > + } else {
> > > on_each_cpu(__ipi_flush_tlb_all, NULL, 1);
> > > + }
> > > }
> > >
> > > struct flush_tlb_range_data {
> > > @@ -105,6 +140,8 @@ static void __flush_tlb_range(struct mm_struct *mm,
> > >
> > > cpu = get_cpu();
> > >
> > > + riscv_tlb_trace_flush_request(cmask, start, size, stride, asid, !!mm);
> > > +
> > > /* Check if the TLB flush needs to be sent to other CPUs. */
> > > if (cpumask_any_but(cmask, cpu) >= nr_cpu_ids) {
> > > local_flush_tlb_range_asid(start, size, stride, asid);
> > >
> > > ---
> > > base-commit: a5ae1bf65cee84bcc121d659e7e22a767853f5f5
> > > change-id: 20260829-tlb_tracepoint-844105ab5092
> > >
> > > Best regards,
> > > --
> > > Roman 'Hedin' Storozhenko <romeusmeister@xxxxxxxxx>
> > >
>
>
>
> --
> Kind regards,
> Roman 'Hedin' Storozhenko



--
Kind regards,
Roman 'Hedin' Storozhenko