Thank you for your contribution! Sashiko AI review found 1 potential
issue to consider:
- [High] drm/nouveau: resource leak and stale list entry in
  nvkm_device_ctor()
--

commit 2b74a622d800f1d2cd12ecd10013a3d94e2a156e
Author: Dave Airlie <[email protected]>

drm/nouveau: add option to leave GSP-capable GPUs to other drivers

This commit adds a DRM_NOUVEAU_BIND_GSP Kconfig option and a bind_gsp
module parameter to let nouveau skip GSP-capable (Turing and later) PCI
GPUs so other drivers such as nova-core can claim them.

> diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/device/base.c 
> b/drivers/gpu/drm/nouveau/nvkm/engine/device/base.c
> index a2ed8c4fe52a0..a027860fa8694 100644
> --- a/drivers/gpu/drm/nouveau/nvkm/engine/device/base.c
> +++ b/drivers/gpu/drm/nouveau/nvkm/engine/device/base.c
[ ... ]
> @@ -3381,6 +3381,20 @@ nvkm_device_ctor(const struct nvkm_device_func *func,
>               goto done;
>       }
>  
> +     /* Unless nouveau is the default NVIDIA driver, leave GSP-capable
> +      * (Turing and later) PCI GPUs to other drivers, such as nova-core.
> +      * Older GPUs are always bound.
> +      *
> +      * Reject before subdevs are constructed, as that would already load
> +      * GSP firmware for GPUs that are not going to be used.
> +      */
> +     if (device->func->pci && device->card_type >= TU100 && !bind_gsp) {
> +             nvdev_info(device, "leaving GSP-capable GPU to other drivers "
> +                        "(bind_gsp=0)\n");
> +             ret = -ENODEV;
> +             goto done;
> +     }
> +
>       /* read strapping information */
>       strap = nvkm_rd32(device, 0x101000);

[Severity: High]
Does rejecting a GPU here leak pdev, leave pci_dev enabled, and leave a
freed or stale entry on nv_devices?

Before reaching this check, nvkm_device_pci_new() calls
pci_enable_device() and allocates pdev, and nvkm_device_ctor() adds
&device->head to the global nv_devices list:

drivers/gpu/drm/nouveau/nvkm/engine/device/base.c:nvkm_device_ctor() {
    ...
        list_add_tail(&device->head, &nv_devices);
    ...
done:
        if (ret && device->pri) {
                iounmap(device->pri);
                device->pri = NULL;
        }
        mutex_unlock(&nv_devices_mutex);
        return ret;
}

When nvkm_device_ctor() returns -ENODEV, the done label only unmaps
device->pri, and nvkm_device_pci_new() and nouveau_drm_probe() return
ret immediately without unlinking device->head from nv_devices, calling
pci_disable_device(), or freeing pdev.  This leaves the PCI device's
enable_cnt elevated when handing off to another driver and causes any
later nouveau probe of the same device handle to fail with -EEXIST.
(Note that calling nvkm_device_del() directly on this failure path would
also hit an uninitialized device->intr.intr list in nvkm_intr_dtor(),
since nvkm_intr_ctor() has not run yet.)

-- 
Sashiko AI review ยท 
https://sashiko.dev/#/patchset/[email protected]?part=1

Reply via email to