diff options
| author | Wladimir J. van der Laan <laanwj@gmail.com> | 2013-09-09 12:38:27 +0200 |
|---|---|---|
| committer | Wladimir J. van der Laan <laanwj@gmail.com> | 2013-09-09 13:51:19 +0200 |
| commit | 7d5eafe76d06e6611a6e1e8c175f11a95a4f371f (patch) | |
| tree | 8f5ef01813ada37354040d37227adebf27c72ade /doc/kernel_bugs.md | |
| parent | 199c781c703eb3b4986bd454d29a5049f19290ff (diff) | |
driver: fix cousins of MSAA hang bug
- Change the order in `reset_context` as well
- Clobber PS INPUTS state when MSAA samples changed to make sure it's
always written, as we don't know the current value anymore.
Diffstat (limited to 'doc/kernel_bugs.md')
| -rw-r--r-- | doc/kernel_bugs.md | 22 |
1 files changed, 22 insertions, 0 deletions
diff --git a/doc/kernel_bugs.md b/doc/kernel_bugs.md index 0ccb33a..76e5445 100644 --- a/doc/kernel_bugs.md +++ b/doc/kernel_bugs.md @@ -31,3 +31,25 @@ signal causes the kernel to run out of signals. This causes hangs in rendering. Status: workaround found (see `ETNA_MAX_UNSIGNALED_FLUSHES`) +Command buffer memory management on cubox +------------------------------------------- + + <_rmk_> err, this is not good. + <_rmk_> gckCOMMAND_Start: command queue is at 0xe7846000 + <_rmk_> gckCOMMAND_Start: WL 380000c8 0cf6d19f 40000002 10281000 + <_rmk_> that's what the first 4 words should've been at the beginning (and where when the GPU was started) + <_rmk_> they then conveniently become: 4000002C 19CE0000 AAAAAAAA AAAAAAAA + <_rmk_> the first two change because of the wait being converted to a link + <_rmk_> the second two... + <_rmk_> well, 0xaaaaaaaa is the free page poisoning. + <_rmk_> which suggests that the GPU command page was freed while still in use + <_rmk_> ah ha, that'll be how. get lucky with the cache behaviour and this is what happens... + <_rmk_> page poisoning enabled. When a lowmem page is allocated, it's memset() through its lowmem mapping, which is cacheable. + <_rmk_> this data can sit in the CPU cache. + <_rmk_> it is then ioremap()'d (which I've been dead against for years) which creates a *device* mapping. + <_rmk_> we then write the wait/link through the device mapping. + <_rmk_> at some point later, the cache lines get evicted from the _normal memory_ mapping. + <_rmk_> thereby overwriting the original wait/link commands in the GPU stream with 0xAAAAAAAA + <_rmk_> I wonder what that a command word with a bit pattern of 10101 in the top 5 bits tells the GPU to do... + <_rmk_> the only thing I can say is... if people would damn well listen to me when I say "don't do this" then you wouldn't get these bugs. + <_rmk_> this is one of the reasons why my check in ioremap.c exists to prevent system memory being ioremap'd and therefore this kind of issue cropping up |