gsp-phys: GR engine context allocate/map/promote before the GR sem push (default off; live-unverified at promote)

1 files changed, 11 insertions, 3 deletions

diff --git a/kernel/gsp_phys_booter.h b/kernel/gsp_phys_booter.h
index 81f1803..3a19060 100644
--- a/kernel/gsp_phys_booter.h
+++ b/kernel/gsp_phys_booter.h
@@ -416,9 +416,17 @@ int gsp_phys_boot_rm_unload(void __iomem *bar0);
  *     selects the GR0 engine for the same ladder: the TSG/channel carry
  *     engineType 1, the channel walk is a single no-retry V5 attempt, a
  *     TURING_COMPUTE_A (0xc5c0) engine object is allocated on the channel,
- *     BIND uses engineType 1 and stage 4 releases a GR semaphore instead of
- *     copying (logged as REAL GR SEM OK / REAL GR SEM FAILED); all CE
- *     forwarding paths refuse once per staging. Writing 9 re-runs one GR
+ *     and the GR engine-context MAIN buffer is queried (device context
+ *     properties + subdevice MAX_SUBCONTEXT_PER_GROUP), allocated as one
+ *     contiguous zeroed sysmem buffer, mapped through fresh PDB slots above
+ *     the guest scratch pool and promoted with NV2080_CTRL_CMD_GPU_PROMOTE_CTX
+ *     carrying both its PA (bInitialize) and VAS VA (bNonmapped=0) so the
+ *     server commits ENG_CTX_VALID before BIND uses engineType 1; stage 4
+ *     then releases a GR semaphore instead of copying (logged as REAL GR SEM
+ *     OK / REAL GR SEM FAILED); all CE forwarding paths refuse once per
+ *     staging. Every context query/allocation/promote failure aborts the
+ *     ladder with no retry and the context buffer is freed by
+ *     gsp_phys_boot_staging_free() on release. Writing 9 re-runs one GR
  *     semaphore release on an already-built GR ladder without a full stage
  *     cycle (no RPC; requires the GR object and work-submit token).
  *   - "gsp-phys-guest-gpa" (write-only, mode 0200): pin the one guest RAM