[Intel-gfx] [RFC 0/7] Save megabytes of wasted sg entries

Tvrtko Ursulin tursulin at ursulin.net
Thu Oct 13 09:03:57 UTC 2016


From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>

We can decrease the i915 slab usage by 1-6 MiB* easily by doing
more sg list coalescing and having the tables at the correct size
at the same time.

This series extracts the sg table building code into helper so
that all places in i915 can benefit.

Another problem is that it is hard to know in advance how many
entries we will need (due coalescing).

Rather than try to pre-scan the pages before creating the table,
it looks easier to keep our pessimistic table sizing at creation,
and trim the waste afterwards.

*** There is a bug here somewhere so please don't run it.   ***
*** I am sending it as RFC only since my debug capabilities ***
*** are severely limited at the moment.			    ***

*) I measured the memory savings by looking at the unused number
   of sg entries at some points:

    * Logging into KDE there is 1-2 MiB of unused sg entries.
    * Running the T-Rex benchamrk aroun 3 Mib.
    * Similarly for Manhattan 5-6 MiB.

Tvrtko Ursulin (7):
  drm/i915: Extract sg creation into a helper
  drm/i915: Use i915_sg_create for userptr
  drm/i915: Use i915_sg_create for partial views
  drm/i915: Use i915_sg_create for rotated pages view
  drm/i915: Use i915_sg_create for dmabuf
  lib/scatterlist: Add sg_trim_table
  drm/i915: Trim sg table after creating it

 drivers/gpu/drm/i915/i915_drv.h         |  34 +++++
 drivers/gpu/drm/i915/i915_gem.c         | 218 +++++++++++++++++++++++++-------
 drivers/gpu/drm/i915/i915_gem_dmabuf.c  |  45 +++----
 drivers/gpu/drm/i915/i915_gem_gtt.c     |  76 ++++-------
 drivers/gpu/drm/i915/i915_gem_userptr.c |  51 ++------
 include/linux/scatterlist.h             |   2 +
 lib/scatterlist.c                       |  53 ++++++++
 7 files changed, 314 insertions(+), 165 deletions(-)

-- 
2.7.4



More information about the Intel-gfx mailing list