If it's taking 10 seconds to do that then I'm guessing you are using a really inefficient algorithm. I've done a realtime light-propagation for voxels (26-directions), it could handle 20x20x20 fills well within a single frame on a single core (and rebuilding and uploading the mesh) after tuning it.

Something that you may have fallen for is to put all adjacent voxels onto the queue when the current voxel is hollow... rather than first check if each adjacent voxel is hollow and then put it in the queue. Depending on how you traverse, this can considerably shorten the number of iterations. Also, make sure that your voxel checks are inlined and fast.

You could do it completly on the GPU.
256^3 could be placed on a single 4096x4096 texture (16x16 patches of 256x256 texel). You need two of these textures and switch them as render source/target between render frames. The shader will check per texel all 26 neighbor texel(when diagonal filling is allowed) or 6 neighbor texel and decides if the target texel should be filled. [s]The worst case should be 512 render frames, which should be no problem on a current hardware.[/s]

Edit: the worst case could be higher....

well the fill itself takes 1.5 seconds per chunk, all added together everything it comes to 10 seconds (which would be more than one chunk) so you think you could fill a 256x256x256 chunk in less than 1.5 seconds? I mean the closer it gets to 10 milliseconds the more id be happy...

Well, filling a 256x256x256 chunk it's hard to say, I mean, we are talking 16.7m voxels... all I have to go on is the performance of my minecraft-like lighting engine, I can create spherical holes 50x50x50 and propagate light into them in a single frame, if even that, I can easily do it 10 times per second and the FPS drops from ~500 to somewhere around ~350... and like I mentioned, that also includes mesh rebuild and upload. Some basic math on those rather sketchy numbers would indicate that I could repeat it 1000+ times a second if I avoid rendering. And since mine is more than 5x5x5 times (125+) smaller than yours, logic would indicate that it indeed it is possible to bring it down to somewhere around 0.2-0.3s if you flood fill something really complicated spanning the entire volume. But those numbers may also be way off, I can't really benchmark it in any meaningful way.

And light-propagation is a tad bit more complicated as I actually propagate light values and not just a single bit... and more importantly, I also do this in all 26 directions (!) which I know for a fact significantly slows it down, even though I've written some really tricky code to speed it up. But it depends on what you are interfacing with, if it's some OO class linked from a DLL with virtual methods and such, there might not be all that much you can do unless you can are able to have direct access into memory.

But really, if you want to speed it up, you need direct access and some serious fine-tuning, even switching an if-else-statement around can significantly increase performance. And I should also add that, my code is also significantly slowed down because the world is divded into blocks and not a single continous chunk of data... meaning, there is "significant" overhead in determining boundaries of blocks and computing block local coordinates for indexing.

You could do it completly on the GPU.
256^3 could be placed on a single 4096x4096 texture (16x16 patches of 256x256 texel). You need two of these textures and switch them as render source/target between render frames. The shader will check per texel all 26 neighbor texel(when diagonal filling is allowed) or 6 neighbor texel and decides if the target texel should be filled. [s]The worst case should be 512 render frames, which should be no problem on a current hardware.[/s]

Edit: the worst case could be higher....

Indeed, worst case would be a single tile wide "snake" going back and forth through the entire cube, that would not be pretty... but then again not really something that would occur either.