Webserver logs would only be somewhat accurate if you got the logs from all servers, and only for desktop systems that are used to browse the Web -- and even then, only if nobody used a web proxy, connected from a network behind a NAT device, and was individually identifiable within the logs to account for overlap.

Statistics for one subdomain of one domain of mine are as follows:

1. Windows 693 71.37%
2. Linux 124 12.77%
3. Macintosh 84 8.65%
4. FreeBSD 68 7.00%
5. (not set) 2 0.21%

I'm sure some of those numbers are inaccurate.