Re: [PATCH v6 6/6] drm/etnaviv: allow usperspace create cached coherent bo

From: Lucas Stach
Date: Wed May 31 2023 - 13:53:35 EST

Next message: Alan Stern: "Re: [PATCH v5 2/3] usb: gadget: udc: core: Invoke usb_gadget_connect only when started"
Previous message: Dmitry Baryshkov: "Re: [PATCH v1 2/3] drm/msm/dpu: retrieve DSI DSC struct at atomic_check()"
In reply to: Sui Jingfeng: "Re: [PATCH v6 6/6] drm/etnaviv: allow usperspace create cached coherent bo"
Next in thread: Sui Jingfeng: "[PATCH v6 5/6] drm/etnaviv: add driver support for the PCI devices"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

Am Donnerstag, dem 01.06.2023 um 01:29 +0800 schrieb Sui Jingfeng:
> Hi,
>
> On 2023/6/1 00:33, Lucas Stach wrote:
> > Hi Sui Jingfeng,
> >
> > Am Mittwoch, dem 31.05.2023 um 00:06 +0800 schrieb Sui Jingfeng:
> > > cached system RAM is coherent on loongson CPUs, and the GPU and DC allways
> > > snoop the CPU's cache. write-combine caching property is not suitiable for
> > > us.
> > >
> > As previously mentioned in the Mesa MR, I don't think this is the right
> > approach.
> >
> > ETNA_BO_CACHED already looks coherent to userspace, as all accesses are
> > bracketed via the ETNAVIV_GEM_CPU_PREP and ETNAVIV_GEM_CPU_FINI ioctls,
> > which will do the necessary cache maintenance on platforms where device
> > coherence isn't enforced by the hardware, so there is no need for a
> > separate ETNA_BO_CACHED_COHERENT.
>
> As far as I can see, ETNA_BO_CACHED_COHERENT could probably help to
> bypass the overhead of
>
> dma_sync_sgtable_for_cpu() and dma_sync_sgtable_for_device() brings to us.
>
>
> I have tested long time ago, there no need call this function on our
> platform.
>
> The glmark2 works as before if I comment out thoes two function.
>
> Are you serious, sir?
>
The dma_sync* functions are more or less no-ops when the device is
marked as being coherent. As the mapping is done via the DRM device
instance, you might need to propagate the coherent property from the
GPU core device to the virtual DRM device, along the lines of how we
propagate other DMA properties from the GPU device to the DRM device in
etnaviv_pdev_probe.

Other than that things should just work with minimal overhead.

> > Instead we just need a new ETNAVIV_PARAM to inform userspace about
> > hardware cache coherence being available for a specific GPU core,
>
> Ok, let me think about for a while how to implement this.
>
Simple: add new ETNAVIV_PARAM_GPU_COHERENT to
include/uapi/drm/etnaviv_drm.h, return the result from
dev_is_dma_coherent in etnaviv_gpu_get_param().

> But How about we merge this first, I create another patch to improve it
>
> with a roughly working base first? I'm just asking if the answer is No :-)
>
The answer is a firm no.

This impacts UAPI, so there is no chance to ever get rid of any wrong
decisions here, as any added UAPI needs to be supported indefinitely.
I'm not signing up for maintaining something I believe is implemented
upside down.

Please don't take this the wrong way: I'm pretty excited to see etnaviv
used on more architectures and outside of the proven platform device
paths, so I'm happy to assist in working out the design and help you
get things merged in both kernel and Mesa. But I think we are still
quite a few steps away from having things worked out enough to even
think about merging those patchsets.

Also please allow me to comment on the other patches of the series, so
I can get a better understanding of your platform/integration, before
sending another revision of those patches.

Regards,
Lucas

> > in
> > which case the userspace driver should switch to preferring
> > ETNA_BO_CACHED over ETNA_BO_WC.
>
> Yeah, ETNA_BO_CACHED is enough.
>
> ETNA_BO_CACHED_COHERENT is actually a special case of ETNA_BO_CACHED.
>
> > Regards,
> > Lucas
> >
> > > Signed-off-by: Sui Jingfeng <suijingfeng@xxxxxxxxxxx>
> > > ---
> > > drivers/gpu/drm/etnaviv/etnaviv_drv.c | 2 +-
> > > drivers/gpu/drm/etnaviv/etnaviv_gem.c | 22 +++++++++++++++++++--
> > > drivers/gpu/drm/etnaviv/etnaviv_gem_prime.c | 9 ++++++++-
> > > include/uapi/drm/etnaviv_drm.h | 11 ++++++-----
> > > 4 files changed, 35 insertions(+), 9 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/etnaviv/etnaviv_drv.c b/drivers/gpu/drm/etnaviv/etnaviv_drv.c
> > > index 052f745cecc0..2816c654c023 100644
> > > --- a/drivers/gpu/drm/etnaviv/etnaviv_drv.c
> > > +++ b/drivers/gpu/drm/etnaviv/etnaviv_drv.c
> > > @@ -274,7 +274,7 @@ static int etnaviv_ioctl_gem_new(struct drm_device *dev, void *data,
> > > struct drm_etnaviv_gem_new *args = data;
> > >
> > > if (args->flags & ~(ETNA_BO_CACHED | ETNA_BO_WC | ETNA_BO_UNCACHED |
> > > - ETNA_BO_FORCE_MMU))
> > > + ETNA_BO_CACHED_COHERENT | ETNA_BO_FORCE_MMU))
> > > return -EINVAL;
> > >
> > > return etnaviv_gem_new_handle(dev, file, args->size,
> > > diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gem.c b/drivers/gpu/drm/etnaviv/etnaviv_gem.c
> > > index b5f73502e3dd..d8b559bd33d3 100644
> > > --- a/drivers/gpu/drm/etnaviv/etnaviv_gem.c
> > > +++ b/drivers/gpu/drm/etnaviv/etnaviv_gem.c
> > > @@ -343,6 +343,7 @@ void *etnaviv_gem_vmap(struct drm_gem_object *obj)
> > > static void *etnaviv_gem_vmap_impl(struct etnaviv_gem_object *obj)
> > > {
> > > struct page **pages;
> > > + pgprot_t prot;
> > >
> > > lockdep_assert_held(&obj->lock);
> > >
> > > @@ -350,8 +351,20 @@ static void *etnaviv_gem_vmap_impl(struct etnaviv_gem_object *obj)
> > > if (IS_ERR(pages))
> > > return NULL;
> > >
> > > - return vmap(pages, obj->base.size >> PAGE_SHIFT,
> > > - VM_MAP, pgprot_writecombine(PAGE_KERNEL));
> > > + switch (obj->flags) {
> > > + case ETNA_BO_CACHED_COHERENT:
> > > + case ETNA_BO_CACHED:
> > > + prot = PAGE_KERNEL;
> > > + break;
> > > + case ETNA_BO_UNCACHED:
> > > + prot = pgprot_noncached(PAGE_KERNEL);
> > > + break;
> > > + case ETNA_BO_WC:
> > > + default:
> > > + prot = pgprot_writecombine(PAGE_KERNEL);
> > > + }
> > > +
> > > + return vmap(pages, obj->base.size >> PAGE_SHIFT, VM_MAP, prot);
> > > }
> > >
> > > static inline enum dma_data_direction etnaviv_op_to_dma_dir(u32 op)
> > > @@ -545,6 +558,7 @@ static const struct drm_gem_object_funcs etnaviv_gem_object_funcs = {
> > > static int etnaviv_gem_new_impl(struct drm_device *dev, u32 size, u32 flags,
> > > const struct etnaviv_gem_ops *ops, struct drm_gem_object **obj)
> > > {
> > > + struct etnaviv_drm_private *priv = dev->dev_private;
> > > struct etnaviv_gem_object *etnaviv_obj;
> > > unsigned sz = sizeof(*etnaviv_obj);
> > > bool valid = true;
> > > @@ -555,6 +569,10 @@ static int etnaviv_gem_new_impl(struct drm_device *dev, u32 size, u32 flags,
> > > case ETNA_BO_CACHED:
> > > case ETNA_BO_WC:
> > > break;
> > > + case ETNA_BO_CACHED_COHERENT:
> > > + if (priv->has_cached_coherent)
> > > + break;
> > > + fallthrough;
> > > default:
> > > valid = false;
> > > }
> > > diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gem_prime.c b/drivers/gpu/drm/etnaviv/etnaviv_gem_prime.c
> > > index 3524b5811682..671d91d8f1c6 100644
> > > --- a/drivers/gpu/drm/etnaviv/etnaviv_gem_prime.c
> > > +++ b/drivers/gpu/drm/etnaviv/etnaviv_gem_prime.c
> > > @@ -112,11 +112,18 @@ static const struct etnaviv_gem_ops etnaviv_gem_prime_ops = {
> > > struct drm_gem_object *etnaviv_gem_prime_import_sg_table(struct drm_device *dev,
> > > struct dma_buf_attachment *attach, struct sg_table *sgt)
> > > {
> > > + struct etnaviv_drm_private *priv = dev->dev_private;
> > > struct etnaviv_gem_object *etnaviv_obj;
> > > size_t size = PAGE_ALIGN(attach->dmabuf->size);
> > > + u32 cache_flags;
> > > int ret, npages;
> > >
> > > - ret = etnaviv_gem_new_private(dev, size, ETNA_BO_WC,
> > > + if (priv->has_cached_coherent)
> > > + cache_flags = ETNA_BO_CACHED_COHERENT;
> > > + else
> > > + cache_flags = ETNA_BO_WC;
> > > +
> > > + ret = etnaviv_gem_new_private(dev, size, cache_flags,
> > > &etnaviv_gem_prime_ops, &etnaviv_obj);
> > > if (ret < 0)
> > > return ERR_PTR(ret);
> > > diff --git a/include/uapi/drm/etnaviv_drm.h b/include/uapi/drm/etnaviv_drm.h
> > > index af024d90453d..474b0db286de 100644
> > > --- a/include/uapi/drm/etnaviv_drm.h
> > > +++ b/include/uapi/drm/etnaviv_drm.h
> > > @@ -90,13 +90,14 @@ struct drm_etnaviv_param {
> > > * GEM buffers:
> > > */
> > >
> > > -#define ETNA_BO_CACHE_MASK 0x000f0000
> > > +#define ETNA_BO_CACHE_MASK 0x000f0000
> > > /* cache modes */
> > > -#define ETNA_BO_CACHED 0x00010000
> > > -#define ETNA_BO_WC 0x00020000
> > > -#define ETNA_BO_UNCACHED 0x00040000
> > > +#define ETNA_BO_CACHED 0x00010000
> > > +#define ETNA_BO_WC 0x00020000
> > > +#define ETNA_BO_UNCACHED 0x00040000
> > > +#define ETNA_BO_CACHED_COHERENT 0x00080000
> > > /* map flags */
> > > -#define ETNA_BO_FORCE_MMU 0x00100000
> > > +#define ETNA_BO_FORCE_MMU 0x00100000
> > >
> > > struct drm_etnaviv_gem_new {
> > > __u64 size; /* in */
>

Next message: Alan Stern: "Re: [PATCH v5 2/3] usb: gadget: udc: core: Invoke usb_gadget_connect only when started"
Previous message: Dmitry Baryshkov: "Re: [PATCH v1 2/3] drm/msm/dpu: retrieve DSI DSC struct at atomic_check()"
In reply to: Sui Jingfeng: "Re: [PATCH v6 6/6] drm/etnaviv: allow usperspace create cached coherent bo"
Next in thread: Sui Jingfeng: "[PATCH v6 5/6] drm/etnaviv: add driver support for the PCI devices"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]