Add Nvidia.md leaf file, add links in sidebar & getting started (#2029)

* Add Nvidia.md leaf file, add links in sidebar & getting started

* squash review-commits from gh to one commit

* heap reuse ratio from 1 => 0 to match currently shipped solution

* Update wiki/Nvidia.md

---------

Co-authored-by: Ivan Molodetskikh <yalterz@gmail.com>

3 files changed, 45 insertions, 0 deletions

diff --git a/wiki/Getting-Started.md b/wiki/Getting-Started.md
index 072a14ba..6a20969d 100644
--- a/wiki/Getting-Started.md
+++ b/wiki/Getting-Started.md
@@ -23,6 +23,9 @@ Finally, the [Xwayland](./Xwayland.md) page explains how to run X11 applications
 
 ### NVIDIA
 
+The NVIDIA drivers currently have an issue with high VRAM usage due to a heap reuse quirk.
+You're recommended to apply a manual fix documented [here](./Nvidia.md) if you run niri on an NVIDIA GPU.
+
 NVIDIA GPUs can have problems running niri (for example, the screen remains black upon starting from a TTY).
 Sometimes, the problems can be fixed.
 You can try the following:
diff --git a/wiki/Nvidia.md b/wiki/Nvidia.md
new file mode 100644
index 00000000..86c16319
--- /dev/null
+++ b/wiki/Nvidia.md
@@ -0,0 +1,41 @@
+### High VRAM usage fix
+
+Presently, there is a quirk in the NVIDIA drivers that affects niri's VRAM usage (the driver does not properly release VRAM back into the pool). Niri *should* use on the order of 100 MiB of VRAM (as checked in [nvtop](https://github.com/Syllo/nvtop)); if you see anywhere close to 1 GiB of VRAM in use, you are likely hitting this issue (heap not returning freed buffers to the driver).
+
+Luckily, you can mitigate this by configuring the NVIDIA drivers with a per-process application profile as follows:
+
+* `sudo mkdir -p /etc/nvidia/nvidia-application-profiles-rc.d` to make the config dir if it does not exist (it most likely does not if you are reading this)
+* write the following JSON blob to set the `GLVidHeapReuseRatio` config value for the `niri` process into the file `/etc/nvidia/nvidia-application-profiles-rc.d/50-limit-free-buffer-pool-in-wayland-compositors.json`:
+    
+    ```json
+    {
+        "rules": [
+            {
+                "pattern": {
+                    "feature": "procname",
+                    "matches": "niri"
+                },
+                "profile": "Limit Free Buffer Pool On Wayland Compositors"
+            }
+        ],
+        "profiles": [
+            {
+                "name": "Limit Free Buffer Pool On Wayland Compositors",
+                "settings": [
+                    {
+                        "key": "GLVidHeapReuseRatio",
+                        "value": 0
+                    }
+                ]
+            }
+        ]
+    }
+    ```
+    
+    (The file in `/etc/nvidia/nvidia-application-profiles-rc.d/` can be named anything, and does not actually need an extension).
+
+Restart niri after writing the config file to apply the change.
+
+The upstream issue that this solution was pulled from is [here](https://github.com/NVIDIA/egl-wayland/issues/126#issuecomment-2379945259). There is a (slim) chance that NVIDIA updates their built-in application profiles to apply this to niri automatically; it is unlikely that the underlying heuristic will see a proper fix.
+
+The fix shipped in the driver at the time of writing uses a value of 0, while the initial config posted by an Nvidia engineer approximately a year prior used a value of 1. 
diff --git a/wiki/_Sidebar.md b/wiki/_Sidebar.md
index 2a35714f..89302764 100644
--- a/wiki/_Sidebar.md
+++ b/wiki/_Sidebar.md
@@ -10,6 +10,7 @@
 * [Layer‐Shell Components](./Layer%E2%80%90Shell-Components.md)
 * [IPC, `niri msg`](./IPC.md)
 * [Application-Specific Issues](./Application-Issues.md)
+* [Nvidia](./Nvidia.md)
 * [Xwayland](./Xwayland.md)
 * [Gestures](./Gestures.md)
 * [Packaging niri](./Packaging-niri.md)