[PATCH v2] lib: sbi: Allow platforms to override local TLB operations

[Xiang W]

在 2026-06-17三的 09:13 -0500，Samuel Holland写道：
> Hi,
> 
> On 2026-06-17 6:29 AM, Xiang W wrote:
> > Some T-Head based processors, for example the Sophgo SG2044, have a
> > JTLB errata that requires special handling of certain TLB maintenance
> > instructions (sfence.vma).
> > 
> > Introduce a sbi_tlb_local_operations structure so platforms can provide
> > custom local TLB flush implementations. Use it to implement the JTLB
> > workaround on affected SoCs (currently Sophgo SG2044).
> > 
> > Signed-off-by: Xiang W <wangxiang at iscas.ac.cn>
> > Signed-off-by: Han Gao <gaohan at iscas.ac.cn>
> > ---
> >  include/sbi/sbi_tlb.h                        | 31 +++++++++++++
> >  lib/sbi/sbi_tlb.c                            | 46 +++++++++++++++++---
> >  platform/generic/include/thead/c9xx_errata.h |  1 +
> >  platform/generic/thead/thead-generic.c       | 31 ++++++++++++-
> >  4 files changed, 101 insertions(+), 8 deletions(-)
> > 
> > [...]
> > +static void thead_jtlb_local_sfence_vma(struct sbi_tlb_info *tinfo)
> > +{
> > + sbi_pmu_ctr_incr_fw(SBI_PMU_FW_SFENCE_VMA_RCVD);
> > + __asm__ __volatile__("sfence.vma");
> > +}
> > +
> > +static void thead_jtlb_local_sfence_vma_asid(struct sbi_tlb_info *tinfo)
> > +{
> > + sbi_pmu_ctr_incr_fw(SBI_PMU_FW_SFENCE_VMA_ASID_RCVD);
> > + /* Flush entire MM context for a given ASID */
> > + __asm__ __volatile__("sfence.vma x0, %0"
> > +      :
> > +      : "r"(tinfo->asid)
> > +      : "memory");
> > +}
> 
> This looks similar to the bug seen on older SiFive chips, where we must always
> flush the full address space. If so, there is already a much simpler workaround
> available, using get_tlbr_flush_limit. See platform/generic/sifive/fu540.c.
> 
> Regards,
> Samuel
Hi Samuel,

Thanks for the feedback.

We did consider `get_tlbr_flush_limit`, and it can indeed solve the SG2044 JTLB issue. 
However, for SG2044 we only need to override the local `sfence.vma` and `sfence.vma` 
with ASID cases (i.e. `local_sfence_vma` and `local_sfence_vma_asid`), while leaving 
the other TLB operations on the default implementation.  

The more general local TLB operations override mechanism gives us the flexibility to 
do exactly that without forcing a full TLB flush limit for all operations.

Regards,  
Xiang W