i965/drm: Rewrite relocation handling.

The execbuf2 kernel API requires us to construct two kinds of lists.
First is a "validation list" (struct drm_i915_gem_exec_object2[])
containing each BO referenced by the batch.  (The batch buffer itself
must be the last entry in this list.)  Each validation list entry
contains a pointer to the second kind of list: a relocation list.
The relocation list contains information about pointers to BOs that
the kernel may need to patch up if it relocates objects within the VMA.

This is a very general mechanism, allowing every BO to contain pointers
to other BOs.  libdrm_intel models this by giving each drm_intel_bo a
list of relocations to other BOs.  Together, these form "reloc trees".

Processing relocations involves a depth-first-search of the relocation
trees, starting from the batch buffer.  Care has to be taken not to
double-visit buffers.  Creating the validation list has to be deferred
until the last minute, after all relocations are emitted, so we have the
full tree present.  Calculating the amount of aperture space required to
pin those BOs also involves tree walking, which is expensive, so libdrm
has hacks to try and perform less expensive estimates.

For some reason, it also stored the validation list in the global
(per-screen) bufmgr structure, rather than as an local variable in the
execbuffer function, requiring locking for no good reason.

It also assumed that the batch would probably contain a relocation
every 2 DWords - which is absurdly high - and simply aborted if there
were more relocations than the max.  This meant the first relocation
from a BO would allocate 180kB of data structures!

This is way too complicated for our needs.  i965 only emits relocations
from the batchbuffer - all GPU commands and state such as SURFACE_STATE
live in the batch BO.  No other buffer uses relocations.  This means we
can have a single relocation list for the batchbuffer.  We can add a BO
to the validation list (set) the first time we emit a relocation to it.
We can easily keep a running tally of the aperture space required for
that list by adding the BO size when we add it to the validation list.

This patch overhauls the relocation system to do exactly that.  There
are many nice benefits:

- We have a flat relocation list instead of trees.
- We can produce the validation list up front.
- We can allocate smaller arrays and dynamically grow them.
- Aperture space checks are now (a + b <= c) instead of a tree walk.
- brw_batch_references() is a trivial validation list walk.
  It should be straightforward to make it O(1) in the future.
- We don't need to bloat each drm_bacon_bo with 32B of reloc data.
- We don't need to lock in execbuffer, as the data structures are
  context-local, and not per-screen.
- Significantly less code and a better match for what we're doing.
- The simpler system should make it easier to take advantage of
  I915_EXEC_NO_RELOC in a future patch.

Improves performance in Synmark 7.0's OglBatch7:

    - Skylake GT4e: 12.1499% +/- 2.29531%  (n=130)
    - Apollolake:   3.89245% +/- 0.598945% (n=35)

Improves performance in GFXBench4's gl_driver2 test:

    - Skylake GT4e: 3.18616% +/- 0.867791% (n=229)
    - Apollolake:   4.1776%  +/- 0.240847% (n=120)

v2: Feedback from Chris Wilson:
    - Omit explicit zero initializers for garbage execbuf fields.
    - Use .rsvd1 = ctx_id rather than i915_execbuffer2_set_context_id
    - Drop unnecessary fencing assertions.
    - Only use _WR variant of execbuf ioctl when necessary.
    - Shrink the arrays to be smaller by default.

Reviewed-by: Jason Ekstrand <[email protected]>

1 files changed, 1 insertions, 1 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_draw.c b/src/mesa/drivers/dri/i965/brw_draw.c
index a6e229f2210..bf09915d0c3 100644
--- a/src/mesa/drivers/dri/i965/brw_draw.c
+++ b/src/mesa/drivers/dri/i965/brw_draw.c
@@ -601,7 +601,7 @@ retry:
 
       brw->no_batch_wrap = false;
 
-      if (drm_bacon_bufmgr_check_aperture_space(&brw->batch.bo, 1)) {
+      if (!brw_batch_has_aperture_space(brw, 0)) {
          if (!fail_next) {
             intel_batchbuffer_reset_to_saved(brw);
             intel_batchbuffer_flush(brw);