Messages in this thread | | | From | Andy Lutomirski <> | Date | Thu, 13 Dec 2012 18:11:55 -0800 | Subject | Re: [PATCH] Add VDSO time function support for x86 32-bit kernel |
| |
On Thu, Dec 13, 2012 at 5:49 PM, H. Peter Anvin <hpa@zytor.com> wrote: > On 12/13/2012 05:42 PM, Andy Lutomirski wrote: >> >> The 64-bit/x32 case is currently very simple and fast because it uses >> absolute addressing. Admittedly, pcrel references are free, so >> changing this wouldn't cost much. For native, it'll be slower, but >> maybe no one cares. I seem to care about this more than anyone else, >> and I don't use 32 bit code. :) >> > > pcrel is actually cheaper than absolute addressing in 64-bit mode. > >> The benefit of switching is that the vdso code could be the same in >> all three cases. (Actually, it's even better than that. All of the >> VVAR magic could be the same in the vdso and the kernel -- the kernel >> linker script would just have to have an appropriate symbol to see the >> appropriate mapping.) >> >> >> This: >> >> __attribute__((visibility("hidden"))) int foo; >> >> int get_foo(void) >> { >> return foo; >> } >> >> generates a rip-relative access on 64 bits and GOTOFF on 32 bits. >> >> The only reason I didn't use a real symbol in the first place is >> because I couldn't figure out how to get gcc to emit an absolute >> relocation in pic code. > > Well, then, we wouldn't need to do that... this is starting to sound > like a significant win.
How will this avoid breaking checkpoint/restore in userspace? If the vdso is not just plain old code, criu presumably needs to know about it. Should there be an arch_prctl(ARCH_MAP_VDSO, addr) to create a vdso mapping somewhere?
--Andy
| |