OK, here is an alternative solution for the kernel-dcache / user-icache flush issue in uprobes which I think follows Dave Miller's suggested approach. As a reminder: the goal is to make sure the user-space icache does not have stale data after the kernel rewrite of an instruction in the user's uprobe "execute out of line" (xol) page. It seems only ARM currently finds the flush_dcache_page() call insufficient, but then apparently only two architectures (other than ARM) support uprobes.
I've modified events/uprobes.c to simply call the copy_to_user_page() function instead of doing a memcpy() followed by a flush_dcache_page() call. This results in a net reduction of one line of code in that file. Then I modified copy_to_user_page() and/or flushing function(s) it calls to treat a NULL vma pointer to mean: "assume the user icache address range is now invalid". In the majority of cases this is pretty basic and should be safe as nothing could have been doing this previously. In some cases this now results in flushing more icache than is necessary. For the mips, sh, sparc, and alpha architectures something more complicated is necessary and I have not currently done that. I am not certain this approach can be made to work cleanly for those architectures, although there is probably always the last resort of flushing all icache. On the other hand, it appears only x86, powerpc, and (god-willing) ARM currently support uprobes.
I have only tested this on ARM (arndale) at this point.
The preliminary patch follows in my next email.
(BTW, is depending on the C compiler short-circuiting conditonals acceptable style?)
-dl