[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <20201110213456.GB32222@xps13>
Date: Tue, 10 Nov 2020 16:34:56 -0500
From: Jeremy Cline <jcline@...hat.com>
To: Karol Herbst <kherbst@...hat.com>
Cc: Ben Skeggs <bskeggs@...hat.com>, David Airlie <airlied@...ux.ie>,
nouveau <nouveau@...ts.freedesktop.org>,
LKML <linux-kernel@...r.kernel.org>,
dri-devel <dri-devel@...ts.freedesktop.org>,
Daniel Vetter <daniel@...ll.ch>
Subject: Re: [Nouveau] [PATCH 2/3] drm/nouveau: manage nouveau_drm lifetime
with devres
On Fri, Nov 06, 2020 at 02:31:44PM +0100, Karol Herbst wrote:
> On Fri, Nov 6, 2020 at 3:17 AM Jeremy Cline <jcline@...hat.com> wrote:
> >
> > Make use of the devm_drm_dev_alloc() API to bind the lifetime of
> > nouveau_drm structure to the drm_device. This is important because a
> > reference to nouveau_drm is accessible from drm_device, which is
> > provided to a number of DRM layer callbacks that can run after the
> > deallocation of nouveau_drm currently occurs.
> >
> > Signed-off-by: Jeremy Cline <jcline@...hat.com>
> > ---
> > drivers/gpu/drm/nouveau/nouveau_drm.c | 44 ++++++++++++---------------
> > drivers/gpu/drm/nouveau/nouveau_drv.h | 10 ++++--
> > 2 files changed, 26 insertions(+), 28 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c b/drivers/gpu/drm/nouveau/nouveau_drm.c
> > index bc6f51bf23b7..f750c25e92f9 100644
> > --- a/drivers/gpu/drm/nouveau/nouveau_drm.c
> > +++ b/drivers/gpu/drm/nouveau/nouveau_drm.c
> > @@ -30,9 +30,11 @@
> > #include <linux/vga_switcheroo.h>
> > #include <linux/mmu_notifier.h>
> >
> > +#include <drm/drm_drv.h>
> > #include <drm/drm_crtc_helper.h>
> > #include <drm/drm_ioctl.h>
> > #include <drm/drm_vblank.h>
> > +#include <drm/drm_managed.h>
> >
> > #include <core/gpuobj.h>
> > #include <core/option.h>
> > @@ -532,13 +534,8 @@ nouveau_parent = {
> > static int
> > nouveau_drm_device_init(struct drm_device *dev)
> > {
> > - struct nouveau_drm *drm;
> > int ret;
> > -
> > - if (!(drm = kzalloc(sizeof(*drm), GFP_KERNEL)))
> > - return -ENOMEM;
> > - dev->dev_private = drm;
> > - drm->dev = dev;
> > + struct nouveau_drm *drm = nouveau_drm(dev);
> >
> > nvif_parent_ctor(&nouveau_parent, &drm->parent);
> > drm->master.base.object.parent = &drm->parent;
> > @@ -620,7 +617,6 @@ nouveau_drm_device_init(struct drm_device *dev)
> > nouveau_cli_fini(&drm->master);
> > fail_alloc:
> > nvif_parent_dtor(&drm->parent);
> > - kfree(drm);
> > return ret;
> > }
> >
> > @@ -654,7 +650,6 @@ nouveau_drm_device_fini(struct drm_device *dev)
> > nouveau_cli_fini(&drm->client);
> > nouveau_cli_fini(&drm->master);
> > nvif_parent_dtor(&drm->parent);
> > - kfree(drm);
> > }
> >
> > /*
> > @@ -720,6 +715,7 @@ static int nouveau_drm_probe(struct pci_dev *pdev,
> > {
> > struct nvkm_device *device;
> > struct drm_device *drm_dev;
> > + struct nouveau_drm *nv_dev;
> > int ret;
> >
> > if (vga_switcheroo_client_probe_defer(pdev))
> > @@ -750,15 +746,16 @@ static int nouveau_drm_probe(struct pci_dev *pdev,
> > if (nouveau_atomic)
> > driver_pci.driver_features |= DRIVER_ATOMIC;
> >
> > - drm_dev = drm_dev_alloc(&driver_pci, &pdev->dev);
> > - if (IS_ERR(drm_dev)) {
> > - ret = PTR_ERR(drm_dev);
> > + nv_dev = devm_drm_dev_alloc(&pdev->dev, &driver_stub, typeof(*nv_dev), drm_dev);
> > + if (IS_ERR(nv_dev)) {
> > + ret = PTR_ERR(nv_dev);
> > goto fail_nvkm;
> > }
> > + drm_dev = nouveau_to_drm_dev(nv_dev);
> >
> > ret = pci_enable_device(pdev);
> > if (ret)
> > - goto fail_drm;
> > + goto fail_nvkm;
> >
> > drm_dev->pdev = pdev;
> > pci_set_drvdata(pdev, drm_dev);
> > @@ -778,8 +775,6 @@ static int nouveau_drm_probe(struct pci_dev *pdev,
> > nouveau_drm_device_fini(drm_dev);
> > fail_pci:
> > pci_disable_device(pdev);
> > -fail_drm:
> > - drm_dev_put(drm_dev);
>
> it sounded like that when using devm_drm_dev_alloc we still have an
> initial refcount of 1, so at least in this regard nothing changed so I
> am wondering why this change is necessary and if the reason is
> unrelated it might make sense to move it into its own patch.
>
Okay, after a more thorough investigation and testing with
drm.debug=0x201 to trace through the allocations and de-allocations,
everything is indeed cleaned up, both here when a fake failure is
injected, and through the normal removal path.
I believe it would be problematic if nouveau was presented the device,
called devm_drm_dev_alloc(), failed later on in the probe, and then a
different driver claimed the device. It looks like that's not a problem
in practice, but I'm not familiar enough with all the layers to be 100%
confident I'm reading everything right. As far as I can tell, amdgpu
isn't explicitly dropping the reference either.
> > fail_nvkm:
> > nvkm_device_del(&device);
> > return ret;
> > @@ -799,7 +794,6 @@ nouveau_drm_device_remove(struct drm_device *dev)
> > device = nvkm_device_find(client->device);
> >
> > nouveau_drm_device_fini(dev);
> > - drm_dev_put(dev);
> > nvkm_device_del(&device);
> > }
> >
> > @@ -1285,7 +1279,8 @@ nouveau_platform_device_create(const struct nvkm_device_tegra_func *func,
> > struct platform_device *pdev,
> > struct nvkm_device **pdevice)
> > {
> > - struct drm_device *drm;
> > + struct nouveau_drm *nv_dev;
> > + struct drm_device *drm_dev;
> > int err;
> >
> > err = nvkm_device_tegra_new(func, pdev, nouveau_config, nouveau_debug,
> > @@ -1293,22 +1288,21 @@ nouveau_platform_device_create(const struct nvkm_device_tegra_func *func,
> > if (err)
> > goto err_free;
> >
> > - drm = drm_dev_alloc(&driver_platform, &pdev->dev);
> > - if (IS_ERR(drm)) {
> > - err = PTR_ERR(drm);
> > + nv_dev = devm_drm_dev_alloc(&pdev->dev, &driver_platform, typeof(*nv_dev), drm_dev);
> > + if (IS_ERR(nv_dev)) {
> > + err = PTR_ERR(nv_dev);
> > goto err_free;
> > }
> > + drm_dev = nouveau_to_drm_dev(nv_dev);
> >
> > - err = nouveau_drm_device_init(drm);
> > + err = nouveau_drm_device_init(drm_dev);
> > if (err)
> > - goto err_put;
> > + goto err_free;
> >
> > - platform_set_drvdata(pdev, drm);
> > + platform_set_drvdata(pdev, drm_dev);
> >
> > - return drm;
> > + return drm_dev;
> >
> > -err_put:
> > - drm_dev_put(drm);
> > err_free:
> > nvkm_device_del(pdevice);
> >
> > diff --git a/drivers/gpu/drm/nouveau/nouveau_drv.h b/drivers/gpu/drm/nouveau/nouveau_drv.h
> > index 3e2920a10099..cf6c33e52a5c 100644
> > --- a/drivers/gpu/drm/nouveau/nouveau_drv.h
> > +++ b/drivers/gpu/drm/nouveau/nouveau_drv.h
> > @@ -137,7 +137,11 @@ struct nouveau_drm {
> > struct nvif_parent parent;
> > struct nouveau_cli master;
> > struct nouveau_cli client;
> > - struct drm_device *dev;
> > +
> > + /**
> > + * @drm_dev: The parent DRM device object.
> > + */
> > + struct drm_device drm_dev;
> >
> > struct list_head clients;
> >
> > @@ -237,7 +241,7 @@ struct nouveau_drm {
> > static inline struct nouveau_drm *
> > nouveau_drm(struct drm_device *dev)
> > {
> > - return dev->dev_private;
> > + return container_of(dev, struct nouveau_drm, drm_dev);
> > }
> >
> > /**
> > @@ -251,7 +255,7 @@ nouveau_drm(struct drm_device *dev)
> > */
> > static inline struct drm_device *
> > nouveau_to_drm_dev(struct nouveau_drm *nv_dev) {
> > - return nv_dev->dev;
> > + return &nv_dev->drm_dev;
> > }
> >
> > /**
> > --
> > 2.28.0
> >
> > _______________________________________________
> > Nouveau mailing list
> > Nouveau@...ts.freedesktop.org
> > https://lists.freedesktop.org/mailman/listinfo/nouveau
> >
>
Powered by blists - more mailing lists