feat: VRAM booster also for drivers that name their region vidmem (or numbered), NVIDIA included once its driver registers it

TECHNICAL.md

@@ -265,9 +265,12 @@ most of the VRAM (8 GB on the Steam Machine) can be pushed into system RAM
 by the desktop and other apps, and stutter. The kernel side is the `dmem`
 cgroup controller (7.2); amdgpu and Intel's xe register their VRAM with it,
 NVIDIA's driver doesn't: with an NVIDIA card the option is shown greyed
-out, with why (`WIZARD_VRAM_FAKE_NVIDIA=1` fakes that for tests). Otherwise
+out, with why, until its driver lists a region: then it's offered like any
+other (`WIZARD_VRAM_FAKE_NVIDIA=1` fakes the grey-out, `WIZARD_VRAM_CAPACITY=<file>`
+reads the regions from a copy, for tests). Otherwise
 the **VRAM booster** option is only shown when
-`/sys/fs/cgroup/dmem.capacity` lists a VRAM region of at least 2 GB (a
+`/sys/fs/cgroup/dmem.capacity` lists a VRAM region (`vram`, or `vidmem`
+as other drivers name it, numbered with several) of at least 2 GB (a
 dedicated GPU; an integrated one registers a small carve-out) or it's already on, and is
 ticked by default; new in 2.3.0, so setups with the SteamOS conversion on
 get it ticked. CachyOS packages the userspace side, which it installs:

lib/vram-booster.sh

@@ -4,7 +4,7 @@
 # are evicted first, so a game that needs most of it doesn't spill into
 # system RAM and stutter. The kernel side is the dmem cgroup controller
 # (7.2; amdgpu and xe register their VRAM, NVIDIA's driver doesn't), so it's
-# only offered when the kernel lists a VRAM region of at least 2 GB. CachyOS packages the userspace
+# only offered when the kernel lists a VRAM region (vram or vidmem) of at least 2 GB. CachyOS packages the userspace
 # side: dmemcg-booster (a system service that enables the controller and
 # sets the limits, plus a user service) and plasma-foreground-booster, which
 # tells it which window is in front on the desktop. The latter only starts
@@ -13,14 +13,20 @@
 
 VRAM_PKGS=(dmemcg-booster plasma-foreground-booster)
 VRAM_MIN_BYTES=$((2 * 1024 * 1024 * 1024))
+# The regions the kernel's dmem controller lists (a file under /sys/fs/cgroup;
+# a variable so tests can point it at a copy).
+VRAM_CAPACITY="${WIZARD_VRAM_CAPACITY:-/sys/fs/cgroup/dmem.capacity}"
 
 vram_supported() {
     # WIZARD_VRAM_FAKE_NVIDIA=1 (tests): as with an NVIDIA card.
     [[ -n "${WIZARD_VRAM_FAKE_NVIDIA:-}" ]] && return 1
     # Only for a GPU with its own VRAM: an integrated GPU registers a small
     # carve-out too, but mostly uses system RAM, so there's little to boost.
-    awk -v min="$VRAM_MIN_BYTES" '$1 ~ /\/vram$/ && $2 >= min { found = 1 } END { exit !found }' \
+    # By what the driver registers, not the brand: amdgpu and xe name the
-        /sys/fs/cgroup/dmem.capacity 2>/dev/null
+    # region "vram", other drivers "vidmem" (numbered with several), so an
+    # NVIDIA driver that registers its memory is offered right away.
+    awk -v min="$VRAM_MIN_BYTES" '$1 ~ /\/(vram|vidmem)[0-9]*$/ && $2 >= min { found = 1 } END { exit !found }' \
+        "$VRAM_CAPACITY" 2>/dev/null
 }
 
 vram_nvidia() {
@@ -33,7 +39,7 @@ vram_nvidia() {
     return 1
 }
 
-# Shown where it works, and greyed out with an NVIDIA card, whose driver
+# Shown where it works, and greyed out with an NVIDIA card whose driver
 # doesn't register its VRAM with the kernel (yet): so it's clear why it
 # can't be turned on. Hidden elsewhere (integrated graphics, older kernel).
 vram_available() { vram_supported || vram_status || vram_nvidia; }