Messages in this thread | | | Date | Mon, 30 Jun 2008 18:16:50 +0200 | From | Andi Kleen <> | Subject | Re: [PATCH 3/3] Fix copy_user on x86_64 |
| |
One optimization (I haven't explored) is that the startup overhead of rep ; movs might make it worthwhile to branch to a special version for copies larger than 8 bytes (which are optimized in uaccess for constant counts) and smaller than the threshold.
Best heuristic unfortunately varies per CPU core (and possible even per stepping)
> > and we historically haven't cared much, because the above _naturally_ > happens for the bulk of the important cases (copy whole pages, which > happens not just in the VM for COW, but also when a user reads a regular > file in aligned chunks). But again, for networking buffers, it _might_ > make sense to try to help trigger this case explicitly.
The problem (I did some measurements on that a long time ago) is that often the source and destination alignments are different and you can't really fix that up in copy_*_user.
That is why I dropped the original "fix up alignments" code.
One way to do that would be use a alignment hint from the user address while allocating the destination buffer (or rather adding an offset), so that source and destination are (mis)aligned in the same way and that can be fixed up then.
But that is not trivial for networking. Not sure it would be really worth the effort.
For the file system is not feasible at all because the page cache offsets are fixed.
-Andi
| |