Bug 8414

Created attachment 7143 [details] [review]
dmablit - DDX

Created attachment 7144 [details] [review]
dmablit - DRM

Created attachment 7145 [details] [review]
EXA/DRI mem "sharing"

This is a hack for "sharing" framebuffer memory between EXA and DRI. It allows
to load the DRM module without statically taking offscreen memory from EXA for
DRI back and depth buffers.

These are my current patches which have received a couple of hours testing, they
seem ok sofar.

They copy extensively from via_dmablit which is:
Copyright (C) 2005 Thomas Hellstrom, All Rights Reserved.

Functionality wise, the current mach64 dmablit implementation is tailored for
EXA UTS/DFS. Among other things, it hijacks the DMA descriptor ring, assumes a
single call at a time, VT switch does not work, ...

Performance wise, the memcpy-based UTS/DFS have an average throughput 155 MB/s
and 6.5 MB/s here. The dmablit-based UTS/DFS get 105 MB/s and 52 MB/s for
typical desktop usage.

As expected, the big win is dmablit DFS which gets a speedup of 8x and greatly
offloads the CPU when opening more than 3 windows (on an 8 MB card) or switching
workspaces.

Created attachment 7228 [details] [review]
dmablit - DDX - try 1

Created attachment 7229 [details] [review]
dmablit - DRM - try 1

These patches should run ok with DRI and VT switches. They reserve space for
the SG list similar to the DMA ring and the two are forced to run in sync (this
was done for AGP, but not PCI, cards for now). They also add parameter
checking, comments, etc.

I got two lockups in 3 days with the previous patch. This patch handles the
case where VBLANK and BUSMASTER interrupts coincide, I haven't had any lockups
in the rest 4 days, but this maybe because of luck, interrupt handling needs to
become less crude.

Created attachment 7286 [details] [review]
dmablit - DDX - try 2

Created attachment 7287 [details] [review]
dmablit - DRM - try 2

- drop the separate allocation for the SG list, increase the space for the DMA
  ring to 32 KB and use it for the SG list also

- profile for software fallbacks in the DDX

- add option to enable/disable DFS

These patches also cause lockups. This and the irq handling part are the two
remaining isssues.

Comment on attachment 7286 [details] [review]
dmablit - DDX - try 2

Of course, this is not obsolete yet.

Comment on attachment 7145 [details] [review]
EXA/DRI mem "sharing"

committed (23cec754158c94b43225b2070fc0a0b003e7bd25)

Sorry about the phenomenal bug spam, guys.  Adding xorg-team@ to the QA contact so bugs don't get lost in future.

Please send this patch to xorg-driver-ati@lists.x.org  and xorg-devel@lists.x.org. Posting it in a bug report makes it quite easy to ignore. :\

Actually, the DRM bits need a ton of work, since mach64 DRM isn't even upstream anymore.

Mass closure: This bug has been untouched for more than six years, and is not
obviously still valid. Please reopen this bug or file a new report if you continue to experience issues with current releases.