[igt-dev] [PATCH i-g-t] tests/i915/gem_exec_gttfill: Reenable test for discrete

[Chris Wilson]

Quoting Zbigniew Kempczyński (2022-04-01 12:38:50)
> Test was turned off unnecessarily for discrete so add softpinning to
> it and enable it again. Parallel execution on different engines and
> same vm is thus limited to 4GiB (in userspace we don't have control
> when same offset would be used for different handles on different
> engines so we don't want to get -ENOSPC for such situation).
> 
> Fixes: https://gitlab.freedesktop.org/drm/intel/-/issues/4086
> 
> Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski at intel.com>
> Cc: Kamil Konieczny <kamil.konieczny at linux.intel.com>
> ---
>  tests/i915/gem_exec_gttfill.c | 40 +++++++++++++++++++++++++++++------
>  1 file changed, 34 insertions(+), 6 deletions(-)
> 
> diff --git a/tests/i915/gem_exec_gttfill.c b/tests/i915/gem_exec_gttfill.c
> index f16428714d..74974d324e 100644
> --- a/tests/i915/gem_exec_gttfill.c
> +++ b/tests/i915/gem_exec_gttfill.c
> @@ -33,6 +33,7 @@ IGT_TEST_DESCRIPTION("Fill the GTT with batches.");
>  struct batch {
>         uint32_t handle;
>         void *ptr;
> +       uint64_t offset;
>  };
>  
>  static void xchg_batch(void *array, unsigned int i, unsigned int j)
> @@ -45,7 +46,7 @@ static void xchg_batch(void *array, unsigned int i, unsigned int j)
>         batches[j] = tmp;
>  }
>  
> -static void submit(int fd, int gen,
> +static void submit(int fd, uint64_t ahnd, int gen,
>                    struct drm_i915_gem_execbuffer2 *eb,
>                    struct drm_i915_gem_relocation_entry *reloc,
>                    struct batch *batches, unsigned int count)
> @@ -56,7 +57,7 @@ static void submit(int fd, int gen,
>  
>         memset(&obj, 0, sizeof(obj));
>         obj.relocs_ptr = to_user_pointer(reloc);
> -       obj.relocation_count = 2;
> +       obj.relocation_count = !ahnd ? 2 : 0;
>  
>         memset(reloc, 0, 2*sizeof(*reloc));
>         reloc[0].offset = eb->batch_start_offset;
> @@ -93,7 +94,17 @@ static void submit(int fd, int gen,
>                 reloc[0].target_handle = obj.handle;
>                 reloc[1].target_handle = obj.handle;
>  
> -               obj.offset = 0;
> +               if (ahnd) {
> +                       uint32_t *delta_ptr = batches[i].ptr + reloc[0].delta;
> +
> +                       *delta_ptr = batches[i].offset;
> +                       batch[1] = batches[i].offset + reloc[0].delta;
> +                       obj.flags = EXEC_OBJECT_PINNED;
> +                       obj.offset = batches[i].offset;
> +                       batch[3] = batches[i].offset;
> +               } else {
> +                       obj.offset = 0;

What is this code?

You are missing out on

commit ab0a539f0c723fd86a4d1440ec82004532b0ba22
Author: Chris Wilson <chris.p.wilson at intel.com>
Date:   Sat Jul 3 10:12:19 2021 +0000

    i915/gem_exec_gttfill: Apply relocations inline

    The goal of gem_exec_gttfill is to watch what happens when we
    oversaturate the available GTT, and force the kernel to perform buffer
    evictions from the virtual address space. We do this by submitting a
    larger working set than we can possibly make resident, but we can still
    provide locations hints to the kernel for placing the objects, as the
    addresses will get reused and eviction occur. The eviction will remain
    randomised on both a single engine, and across the engines, improving
    our chance of spotting issues, but we will no long require as many
    relocations.

    Signed-off-by: Chris Wilson <chris.p.wilson at intel.com>
    Reviewed-by: Zbigniew Kempczynski <zbigniew.kempczynski at intel.com>