2025-02-25 Ordinary Meeting Notes

Feb 25, 2025 | RV-LFX Self-hosted Trace TG

Attendees: Beeman Strong

Notes

• Attendees: Beeman, Andrew, Daniel, Bruce, Matthew, Iain, Bo, Greg

• Slides, video

• Agenda

  • Continue trace memory buffer address translation 

  • Trace page faults

  • Trace buffer management

• Trace memory buffer address translation

  • Recapped some of the discussion from last time

  • Propose 2 modes for trace output writes:

    • Default: use the translation regime of the mode to which trace is delegated

    • Alternate: use page tables and *IDs in new tatp/hgtatp registers

    • Both are regardless of the current priv mode, trace translation regime does not change

  • Default is sufficient for most usages (application trace, bare metal OS + app trace, root hypervisor and/or guest trace, …)

  • Alternate needed for nested hypervisor tracing guest, or for M-mode tracing other modes

    • Because in these cases none of the *atp registers hold the desired translation config

    • For nested HV tracing guest, want trace during guest execution to use the nested HV’s vsatp/hgatp.  So root HV can set tatp/hgtatp to point to them.

    • For M-mode, can use tatp to make trace use bare mode even in other priv modes

• Trace TLB entry management

  • Trace TLB entries have ASIDs/VMIDs just like loads and stores do, and those entries should be flushed by the same SFENCE/HFENCE insts 

    • Though when hgtatp is in use, trace VMID can be different from current guest’s VMID.  In such cases, HFENCE.VVMA will not flush those entries.

  • Spec should include recommendation that trace TLB entries are separate, likely only need two since pages don’t tend to be re-used once filled

• Trace page faults

  • Can skid, since trace writes are asynchronous

  • E.g., can get trace PF for VS-mode trace while in HS-mode

  • Will discuss next time

Action items

• Feb 11, 2025 - Beeman Strong - cover trace translation flush scenarios

  • Discussed via email and in this week’s meeting