Re: [RFC PATCH 137/162] drm/i915: blt copy between objs using pre-created vma windows

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Quoting Matthew Auld (2020-11-27 12:06:53)
> +int i915_window_blt_copy(struct drm_i915_gem_object *dst,
> +                        struct drm_i915_gem_object *src)
> +{
> +       struct drm_i915_private *i915 = to_i915(src->base.dev);
> +       struct intel_context *ce = i915->gt.engine[BCS0]->blitter_context;
> +       bool src_is_lmem = i915_gem_object_is_lmem(src);
> +       bool dst_is_lmem = i915_gem_object_is_lmem(dst);
> +       struct scatterlist *last_sgl;
> +       struct i915_vma *src_vma, *dst_vma;
> +       struct i915_request *rq;
> +       u64 cur_win_sz, blt_copied, offset;
> +       long timeout;
> +       u32 size;
> +       int err;
> +
> +       src_vma = src_is_lmem ? i915->mm.lmem_window[0] :
> +                               i915->mm.smem_window[0];
> +       dst_vma = dst_is_lmem ? i915->mm.lmem_window[1] :
> +                               i915->mm.smem_window[1];
> +
> +       if (!src_vma || !dst_vma)
> +               return -ENODEV;
> +
> +       blt_copied = 0;
> +
> +       err = i915_window_blt_copy_prepare_obj(src);
> +       if (err)
> +               return err;
> +
> +       err = i915_window_blt_copy_prepare_obj(dst);
> +       if (err) {
> +               i915_gem_object_unpin_pages(src);
> +               return err;
> +       }
> +
> +       mutex_lock(&i915->mm.window_mutex);
> +       src_vma->obj = src;
> +       dst_vma->obj = dst;
> +       do {
> +               cur_win_sz = min_t(u64, BLT_WINDOW_SZ,
> +                                  (src->base.size - blt_copied));
> +               offset = blt_copied >> PAGE_SHIFT;
> +               size = ALIGN(cur_win_sz, src->mm.region->min_page_size) >>
> +                      PAGE_SHIFT;
> +               intel_partial_pages_for_sg_table(src, src_vma->pages, offset,
> +                                                size, &last_sgl);
> +
> +               /*
> +                * Insert pages into vm, expects the pages to the full
> +                * length of VMA. But we may have the pages of <= vma_size.
> +                * Hence altering the vma size to match the total size of
> +                * the pages attached.
> +                */
> +               src_vma->size = size << PAGE_SHIFT;
> +               i915_insert_vma_pages(src_vma, src_is_lmem);
> +               sg_unmark_end(last_sgl);
> +
> +               /*
> +                * Source obj size could be smaller than the dst obj size,
> +                * due to the varying min_page_size of the mem regions the
> +                * obj belongs to. But when we insert the pages into vm,
> +                * the total size of the pages supposed to be multiples of
> +                * the min page size of that mem region.
> +                */
> +               size = ALIGN(cur_win_sz, dst->mm.region->min_page_size) >>
> +                      PAGE_SHIFT;
> +               intel_partial_pages_for_sg_table(dst, dst_vma->pages, offset,
> +                                                size, &last_sgl);
> +
> +               dst_vma->size = size << PAGE_SHIFT;
> +               i915_insert_vma_pages(dst_vma, dst_is_lmem);
> +               sg_unmark_end(last_sgl);
> +
> +               rq = i915_request_create(ce);
> +               if (IS_ERR(rq)) {
> +                       err = PTR_ERR(rq);
> +                       break;
> +               }
> +               if (rq->engine->emit_init_breadcrumb) {
> +                       err = rq->engine->emit_init_breadcrumb(rq);
> +                       if (unlikely(err)) {
> +                               DRM_ERROR("init_breadcrumb failed. %d\n", err);
> +                               break;
> +                       }
> +               }
> +               err = i915_window_blt_copy_batch_prepare(rq, src_vma, dst_vma,
> +                                                        cur_win_sz);
> +               if (err) {
> +                       DRM_ERROR("Batch preparation failed. %d\n", err);
> +                       i915_request_set_error_once(rq, -EIO);
> +               }
> +
> +               i915_request_get(rq);
> +               i915_request_add(rq);
> +
> +               timeout = i915_request_wait(rq, 0, MAX_SCHEDULE_TIMEOUT);

Locked waits.

> +               if (timeout < 0) {
> +                       DRM_ERROR("BLT Request is not completed. %ld\n",
> +                                 timeout);
> +                       err = timeout;
> +                       i915_request_put(rq);
> +                       break;
> +               }
> +
> +               blt_copied += cur_win_sz;
> +               err = 0;
> +               i915_request_put(rq);
> +               flush_work(&i915->gt.engine[BCS0]->retire_work);

Papering (doubtful the paper is successful) over bugs by introducing a
whole load more.

This fails the basic premise that eviction must be pipelined. The PTE
are transient and can be written prior to the copy and kept within the
non-preemptible window of the blt. Thus allowing many evictions to
scheduled in parallel (by either allocating separate contexts, or more
preferably picking a user context).
-Chris
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx



[Index of Archives]     [AMD Graphics]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux