Re: [PATCH 0/8] pxafb cleanup

On Wed, Feb 27, 2008 at 5:25 PM, Alexandre Rusev [off-list ref] wrote:

eric miao wrote:
 > The following series of patches try to address several issues of the
 > current PXA framebuffer driver. Some of these patches are already
 > submitted for review, here are more:
 >
 > pxa: un-nest pxafb_parse_options() to cleanup the coding style issue
 > pxa: fix various coding style issues for pxafb
 > pxa: purge unnecessary pr_debug and comments from pxafb
 > pxa: sanitize the usage of #ifdef .. processing pxafb parameters
 > pxa: convert fb driver to use ioremap() and __raw_{readl, writel}
 > pxa: introduce "struct pxafb_dma_buff" for palette and dma descriptors
 > pxa: introduce register independent LCD connection type for pxafb
 > pxa: make lubbock/mainstone/zylonite/littleton to use new LCD connection type
 >
 > Board maintainers are encouraged to use the new way for LCD connection
 > types.
 >
 > To further clean-up the driver, I suggest to make the following changes:
 > 1. removal of set_ctrlr_state(), the original code is written so because
 > the fb_blank() is called within interrupt context in 2.4 kernel, but this is
 > no longer the case in 2.6. And states can be simplified.
 >
 > 2. introduce a "vram" module parameter or boot option to indicate the
 > size of the video memory, so that frame buffer offset can be specified
 > within this video memory range, and PAN_DISPLAY can be done.
 > And the start offset of each overlay can also be specified.
 >
 > 3. a "pxafb_layer" structure describing each possible layers,
 > base, overlay1 and overlay2 for pxa27x can be described. From this
 > layer information, frame buffer device can be dynamically allocated
 > and registered. Palette manipulation, parameter checking and
 > layer activation code can be generalized.
 >
 Some considerations on dynamic allocation of framebuffer/overlys memory
 (if applicable).
 Early implementation of framebuffer/overlays for pxa270 contained the
 following bug:
 when resolution of framebuffer/overlay or bpp  was changed and that
 framebuffer/overlay
 (or some part of it) was mapped by userspace application the driver just
 invoke kfree/dma_free_writecombine,
 with the memory area still leaving mapped to userspace!
 This could cause the corruption of kernel memory.
 (for example DirectFB just tried to draw at the previously mapped area
 so no more new picture appeared at screen;) )

 When we do NOTallocate maximal possible sized piece of memory for
 baseplane and overlay in advance
 this problem seems still to be an issue :(
 I suppose that driver-specific vm_area_ops to be attached by driver to
 each mapping did by userland
 due to avoid kernel memory corruption at unmap.

Userland app should really unmap and dereference the previously
mapped area for the overlay(s), but a fix in the driver is also necessary.

 Yet the problem is that if userland applicatein frequently use/unuse and
 change resolution/bpp of overlay
 and forget to unmap it each time, then all consistent/writecombine
 memory would be locked by this application.


 Allocating maximal possible size (and not using it) is quite adversary
 against writecombine/consistent memory pool.
 With current definition of CONSISTENT_SIZE for architecture it would be
 even impossible to allocate
 all the needed memory for all overlays! :(

Not necessary true. We don't have to allocate the video memory from
consistent pool, which is mainly for things like DMA descriptor and
small FIFO areas. Normal memory can be allocated and reserved
for video usage.

-- 
Cheers
- eric

-------------------------------------------------------------------------
This SF.net email is sponsored by: Microsoft
Defy all challenges. Microsoft(R) Visual Studio 2008.
http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/