[PATCH 0/2] x86/sev: Per-node RMPOPT windows for >2TB systems

[Ashish Kalra]

From: Ashish Kalra <ashish.kalra@xxxxxxx>

RMPOPT_BASE is a per-core MSR whose optimization window covers at most 2 TB
of physical address space from its base.  The merged RMPOPT series programs
every core with the same base (start of RAM), so on a machine with more
than 2 TB of RAM only the first 2 TB is ever RMP-check-skipped; everything
above 2 TB pays full RMP checks for the life of the system.

This series extends the optimization across all of physical memory on >2 TB
systems by giving each NUMA node's cores a 2 TB window placed over that
node's own local memory.  A core overwhelmingly accesses its local node's
RAM, so spending each node's 2 TB budget on that memory puts the
optimization where the accesses land; whole-machine coverage is the union
of the per-node windows.  The <=2 TB case is unchanged: the window search
returns start-of-RAM for every node, so all cores share one window.

  Patch 1 generalizes the optimization scan to read each core's own
    RMPOPT_BASE (base and window size straight from the MSR) and walk only
    that window.  No behavioral change while every core shares one window.
  Patch 2 adds the per-node window selection and programming, and warms
    every window before the interrupts-disabled fan-out.

HW-validated on a ~3 TB 2-socket NPS2 box: every node optimizes its own
local memory and all populated memory is covered.

Ashish Kalra (2):
  x86/sev: Scan each core's own RMPOPT window
  x86/sev: Add per-node RMPOPT windows for >2TB systems

 arch/x86/include/asm/msr-index.h |   2 +
 arch/x86/virt/svm/sev.c          | 166 +++++++++++++++++++++++++++----
 2 files changed, 148 insertions(+), 20 deletions(-)


base-commit: 9f74092fda91ef93c2b07585785c1dc83cd7971f
-- 
2.43.0