kernel: validate learned semaphore maps and post guest GPU events

- admit {GPU VA -> PA} pairs only when the preceding bytes form a valid
  client memory descriptor header (type 0/1, CPU+GPU mapped bits), prefer
  validated pairs over headerless candidates under first-wins dedup, and
  invalidate the translation cache on every rebuild; fixes stale identity
  entries shadowing live pool pages and lets the guest UVM channel
  bring-up complete on its own
- capture OS event allocations (class 0x79) and post event records
  (function 0x1003) into the guest status queue when a channel's busy
  notification is consumed, with a broadcast fallback, serialized with
  replies via a queue lock
- add BAR0 shadow introspection (dump_bar0/scan_bar0) and guest-side
  diagnostic scripts for the VM cycle

1 files changed, 46 insertions, 0 deletions

diff --git a/tools/guest-diag/cycle_cdt.sh b/tools/guest-diag/cycle_cdt.sh
new file mode 100755
index 0000000..009bdfc
--- /dev/null
+++ b/tools/guest-diag/cycle_cdt.sh
@@ -0,0 +1,46 @@
+#!/bin/bash
+# cycle_cdt.sh - full test cycle: push tools, start cdt, inject the libcuda
+# tracker page mapping when the tracker wait is reached, poke UVM's tracker
+# when it stalls, and report where cdt ends up.
+set -u
+cd "$(dirname "$(readlink -f "$0")")/../.."
+
+echo "[*] waiting for guest"
+for i in $(seq 1 30); do
+    timeout 20 ./scripts/vgpu_vm.py exec --timeout 12 'uptime >/dev/null 2>&1 && echo READY' 2>/dev/null | grep -q READY && break
+    sleep 5
+done
+
+./tools/guest-diag/push_guest_tools.sh >/tmp/opencode/push.log 2>&1 || { echo "PUSH FAILED"; tail -3 /tmp/opencode/push.log; exit 1; }
+echo "[*] tools pushed"
+
+timeout 40 ./scripts/vgpu_vm.py exec --timeout 35 'sudo nvidia-modprobe -c 0; sudo nvidia-modprobe -u; sudo modprobe nvidia-uvm; sudo sh -c "echo 1 > /sys/module/nvidia_uvm/parameters/uvm_debug_prints"; sudo rm -f /tmp/cdt.out; sudo setsid /tmp/cdt > /tmp/cdt.out 2>&1 & sleep 2; tail -1 /tmp/cdt.out' 2>/dev/null | tail -1
+echo "[*] cdt started"
+
+# wait for the libcuda tracker wait and inject the page mapping
+SEEN_CTX=0
+for i in $(seq 1 24); do
+    sleep 10
+    STATE=$(timeout 20 ./scripts/vgpu_vm.py exec --timeout 15 'tail -1 /tmp/cdt.out' 2>/dev/null | tail -1)
+    echo "  [$((i*10))s] $STATE"
+    if echo "$STATE" | grep -q "cuCtxCreate"; then
+        SEEN_CTX=1
+        if [ ! -f /tmp/opencode/injected_tracker ]; then
+            if /tmp/opencode/inject_tracker.sh >/tmp/opencode/inject.log 2>&1; then
+                touch /tmp/opencode/injected_tracker
+                echo "  [*] tracker mapping injected: $(tail -1 /tmp/opencode/inject.log)"
+            fi
+        fi
+    fi
+    if echo "$STATE" | grep -qE "FAILED|ALL OK|cuMemAlloc"; then
+        echo "[*] cuCtxCreate finished: $STATE"
+        break
+    fi
+done
+
+# if UVM's tracking semaphore lags, poke it forward
+SEMAPA=$(sudo sh -c 'echo 0x121000010,1 > /sys/module/open_iov_vfio/parameters/sema_probe' 2>/dev/null; sleep 2; sudo dmesg -T | grep -A2 "sema-probe: va=0x121000010" | grep -oE "pa=0x[0-9a-f]+" | head -1 | cut -d= -f2)
+if [ -n "$SEMAPA" ]; then
+    CUR=$(sudo dmesg -T | grep -A3 "sema-probe: va=0x121000010" | grep -oE "\[0\] 0x[0-9a-f]+" | tail -1 | awk '{print $2}')
+    echo "[*] UVM tracking sema pa=$SEMAPA value=$CUR"
+fi