[Intel-gfx] [PATCH]drm_i915: remove unnecessary object pin for mode set

[Jesse Barnes]

On Fri, 15 May 2009 10:49:33 +0800
Shaohua Li <shaohua.li at intel.com> wrote:

> I get this log when using xrandr set display resolution in kms:
> [   52.795810] [drm:i915_gem_object_unbind] *ERROR* Attempting to
> unbind pinned buffer [   52.795824] ------------[ cut
> here ]------------ [   52.795829] WARNING:
> at /workshop/kernel/git/my/linux/drivers/gpu/drm/i915/i915_gem_tiling.c:320
> i915_gem_set_tiling+0x3a9/0x412() [   52.795836] Hardware name:
> Calistoga & ICH7M Chipset [   52.795840] failed to unbind object for
> tiling switchModules linked in: [   52.795851] Pid: 1991, comm: X Not
> tainted 2.6.30-rc5 #318 [   52.795855] Call Trace: [   52.795868]
> [<c02475f8>] warn_slowpath_fmt+0x77/0xa6 [   52.795879]
> [<c0617e60>] ? _spin_unlock_irqrestore+0x1c/0x31 [   52.795887]
> [<c0247b9a>] ? release_console_sem+0x180/0x1ad [   52.795894]
> [<c0248078>] ? vprintk+0x2ab/0x2e0 [   52.795905]  [<c0292f64>] ?
> __generic_file_aio_write_nolock+0x3d4/0x40c [   52.795913]
> [<c0615650>] ? printk+0x14/0x16 [   52.795922]  [<c043fbbf>] ?
> i915_gem_object_unbind+0x43/0x200 [   52.795931]  [<c0443e1f>]
> i915_gem_set_tiling+0x3a9/0x412 [   52.795940]  [<c042cd43>]
> drm_ioctl+0x215/0x29b [   52.795947]  [<c0443a76>] ?
> i915_gem_set_tiling+0x0/0x412 [   52.795957]  [<c02fffb7>] ?
> ext3_file_write+0x1f/0x90 [   52.795966]  [<c02b738a>] ?
> do_sync_write+0xb0/0xee [   52.795976]  [<c02c2744>]
> vfs_ioctl+0x55/0x6e [   52.795985]  [<c02c2bf3>]
> do_vfs_ioctl+0x496/0x4cf [   52.795993]  [<c02b7cee>] ?
> vfs_write+0xf5/0x109 [   52.796002]  [<c02c2c5d>] sys_ioctl+0x31/0x4a
> [   52.796011]  [<c021de70>] sysenter_do_call+0x12/0x22
> [   52.796017] ---[ end trace 8cb90508f72e0a4d ]---
> 
> dri_bo_unreference() try to free the front buffer and will set the
> tiling to NONE, but the bo is pined in intel_pipe_set_base(), so the
> warning is reported. It appears the bo isn't required to be pinned in
> intel_pipe_set_base(), as either 2d driver or intel fb will pin the
> front buffer, and the pin will confuse 2d driver.
> 
> Signed-off-by: Shaohua Li <shaohua.li at intel.com>
> 

Yes the 2D driver is confused here, but I'm pretty sure we want to keep
the pin in pipe_set_base.  With KMS we shouldn't have to do userspace
pinning at all.  I think this is really a 2D teardown problem, not a
kernel problem.

-- 
Jesse Barnes, Intel Open Source Technology Center