[Lima] [PATCH v8] drm/lima: driver for ARM Mali4xx GPUs

[Rob Herring]

On Sat, Mar 9, 2019 at 6:21 AM Qiang Yu <yuq825 at gmail.com> wrote:
>
> - Mali 4xx GPUs have two kinds of processors GP and PP. GP is for
>   OpenGL vertex shader processing and PP is for fragment shader
>   processing. Each processor has its own MMU so prcessors work in
>   virtual address space.
> - There's only one GP but multiple PP (max 4 for mali 400 and 8
>   for mali 450) in the same mali 4xx GPU. All PPs are grouped
>   togather to handle a single fragment shader task divided by
>   FB output tiled pixels. Mali 400 user space driver is
>   responsible for assign target tiled pixels to each PP, but mali
>   450 has a HW module called DLBU to dynamically balance each
>   PP's load.
> - User space driver allocate buffer object and map into GPU
>   virtual address space, upload command stream and draw data with
>   CPU mmap of the buffer object, then submit task to GP/PP with
>   a register frame indicating where is the command stream and misc
>   settings.
> - There's no command stream validation/relocation due to each user
>   process has its own GPU virtual address space. GP/PP's MMU switch
>   virtual address space before running two tasks from different
>   user process. Error or evil user space code just get MMU fault
>   or GP/PP error IRQ, then the HW/SW will be recovered.
> - Use GEM+shmem for MM. Currently just alloc and pin memory when
>   gem object creation. GPU vm map of the buffer is also done in
>   the alloc stage in kernel space. We may delay the memory
>   allocation and real GPU vm map to command submission stage in the
>   furture as improvement.
> - Use drm_sched for GPU task schedule. Each OpenGL context should
>   have a lima context object in the kernel to distinguish tasks
>   from different user. drm_sched gets task from each lima context
>   in a fair way.
>
> mesa driver can be found here before upstreamed:
> https://gitlab.freedesktop.org/lima/mesa
>
> v8:
> - add comments for in_sync
> - fix ctx free miss mutex unlock
>
> v7:
> - remove lima_fence_ops with default value
> - move fence slab create to device probe
> - check pad ioctl args to be zero
> - add comments for user/kernel interface
>
> v6:
> - fix comments by checkpatch.pl
>
> v5:
> - export gp/pp version to userspace
> - rebase on drm-misc-next
>
> v4:
> - use get param interface to get info
> - separate context create/free ioctl
> - remove unused max sched task param
> - update copyright time
> - use xarray instead of idr
> - stop using drmP.h
>
> v3:
> - fix comments from kbuild robot
> - restrict supported arch to tested ones
>
> v2:
> - fix syscall argument check
> - fix job finish fence leak since kernel 5.0
> - use drm syncobj to replace native fence
> - move buffer object GPU va map into kernel
> - reserve syscall argument space for future info
> - remove kernel gem modifier
> - switch TTM back to GEM+shmem MM
> - use time based io poll
> - use whole register name
> - adopt gem reservation obj integration
> - use drm_timeout_abs_to_jiffies
>
> Cc: Eric Anholt <eric at anholt.net>
> Cc: Rob Herring <robh at kernel.org>
> Cc: Christian König <ckoenig.leichtzumerken at gmail.com>
> Cc: Daniel Vetter <daniel at ffwll.ch>
> Cc: Alex Deucher <alexdeucher at gmail.com>
> Cc: Sam Ravnborg <sam at ravnborg.org>
> Cc: Rob Clark <robdclark at gmail.com>
> Cc: Dave Airlie <airlied at gmail.com>
> Signed-off-by: Andreas Baierl <ichgeh at imkreisrum.de>
> Signed-off-by: Erico Nunes <nunes.erico at gmail.com>
> Signed-off-by: Heiko Stuebner <heiko at sntech.de>
> Signed-off-by: Marek Vasut <marex at denx.de>
> Signed-off-by: Neil Armstrong <narmstrong at baylibre.com>
> Signed-off-by: Simon Shields <simon at lineageos.org>
> Signed-off-by: Vasily Khoruzhick <anarsoul at gmail.com>
> Signed-off-by: Qiang Yu <yuq825 at gmail.com>
> Reviewed-by: Eric Anholt <eric at anholt.net>
> Reviewed-by: Rob Herring <robh at kerrnel.org>
> ---

[...]

> +static int lima_gem_lock_bos(struct lima_bo **bos, u32 nr_bos,
> +                            struct ww_acquire_ctx *ctx)
> +{
> +       int i, ret = 0, contended, slow_locked = -1;
> +
> +       ww_acquire_init(ctx, &reservation_ww_class);
> +
> +retry:
> +       for (i = 0; i < nr_bos; i++) {
> +               if (i == slow_locked) {
> +                       slow_locked = -1;
> +                       continue;
> +               }
> +
> +               ret = ww_mutex_lock_interruptible(&bos[i]->gem.resv->lock, ctx);
> +               if (ret < 0) {
> +                       contended = i;
> +                       goto err;
> +               }
> +       }
> +
> +       ww_acquire_done(ctx);
> +       return 0;
> +
> +err:
> +       for (i--; i >= 0; i--)
> +               ww_mutex_unlock(&bos[i]->gem.resv->lock);
> +
> +       if (slow_locked >= 0)
> +               ww_mutex_unlock(&bos[slow_locked]->gem.resv->lock);
> +
> +       if (ret == -EDEADLK) {
> +               /* we lost out in a seqno race, lock and retry.. */
> +               ret = ww_mutex_lock_slow_interruptible(
> +                       &bos[contended]->gem.resv->lock, ctx);
> +               if (!ret) {
> +                       slow_locked = contended;
> +                       goto retry;
> +               }
> +       }
> +       ww_acquire_fini(ctx);
> +
> +       return ret;
> +}
> +
> +static void lima_gem_unlock_bos(struct lima_bo **bos, u32 nr_bos,
> +                               struct ww_acquire_ctx *ctx)
> +{
> +       int i;
> +
> +       for (i = 0; i < nr_bos; i++)
> +               ww_mutex_unlock(&bos[i]->gem.resv->lock);
> +       ww_acquire_fini(ctx);
> +}

Not to make you keep shooting for a moving target, Eric just posted a
patch[1] a few days ago that can replace these 2 functions. Would be
good to use if you respin, but otherwise can be a follow-on patch.

Rob

[1] https://patchwork.freedesktop.org/patch/291016/