Hallo,

I tried to connect my ASUS 24T1E monitor (a device which is thought to be used as TV as well as as Computer Monitor) via HDMI (and DVI-to-HDMI-Converter, but this makes no difference to the problem I experience) to my Radeon HD4650. I have a setup that uses the open source radeon driver together with KMS, the kernel used is 2.6.39. 
The problem that ocurrs is, that, as soon as X.org is started, the picture on the screen suffers from massive overscan (I loose about 40px on each side), while in console-only mode (with the high-resolution console that KMS features) there is no overscan until X.org is started. After X.org is started, the console also suffers from the overscan problem.
In several forums you can read that this problem may be corrected by supplying a custom EDID to the graphics driver that contains no extensions sections so that the connected monitor is treated as a common digital monitor, and this solution also works for me, but of course only in KMS-disabled mode as only this mode allows one to use the CustomEDID option in xorg.conf. The difference between the two EDIDs of course lies beyond ModeLines: even with the exact similar mode line, the picture is overscanning in KMS-enabled mode (without custom EDID) where it's not with KMS disabled and a custom EDID. So there must be something else in this EDID that makes X.org (and not the framebuffer console) trigger this weird overscanning behaviour. 
If you execute xrandr --verbose, one of the differences is, that without a custom EDID, xrandr offers the underscan properties (which, btw, are able to correct the overscan problem, but for the price of loosing resolution) while with a custom EDID with the extensions section stripped, it doesn't, but instead offers e.g. a property "dvi_monitor_type" and "scaler". So X.org seems to evaluate the EDID file for more than only valid ModeLines (e.g. for the extensions section, which contains the information that there is a TV set connected and audio channel may be used) and trigger, in this case incorrectly, the overscan behaviour. 
So if there isn't a way yet to tell X.org via xrandr to use the monitor not as TV but as common DVI monitor and ignore this part of EDID and so stop overscanning, there should be, as many users suffer from this problem and not every HDMI-capable TV features the setting to correct buggy computer output and show the image 1:1. Putting a Custom EDID without extensions in place (which is also not possible in KMS mode) can solve the problem simply because X.org doesn't switch to this TV mode then, but is not the proper solution as it sacrifices the possibility of passing audio through the HDMI cable because this part of the extension also misses then. 
That there is no problem with the mode lines that the EDID from the monitor contains in general is - imho - shown by the fact that the framebuffer console displays without overscanning until X.org is started.