Hi all- I think there's general agreement that the current entry trampoline sucks, mainly because it's dog-slow. Thanks, Spectre.
There are three possible fixes I know of: a) Linus' hack: use R11 for scratch space. This doesn't actually speed it up, but it improves the addressing situation a bit. I don't like it, though: it causes the SYSCALL64 path to forget the arithmetic flags and all of the MSR_SYCALL_MASK flags. The latter may be a showstopper, given that we've seen some evidence of nasty Wine use cases that expect setting EFLAGS.NT and doing a system call to actually do something intelligent. Similarly, there could easily be user programs out there that set AC because they want alignment checking and expect AC to remain set across system calls. b) Move the trampoline within 2G of the entry text and copy it for each CPU. This is certainly possible, but it's a bit gross, and it uses num_possible_cpus() * 64 bytes of memory (rounded up to a page). It will also result in more complicated code. c) This series. Just make %gs work in the entry trampoline. It's actually a net code deletion. I suspect that (b) would be faster in code that does a lot of system calls and doesn't totally blow away the cache or the TLB between system calls. I suspect that (c) is faster in code that does cache-cold system calls. Andy Lutomirski (2): x86/entry/64: Use the TSS sp2 slot for rsp_scratch x86/pti/64: Remove the SYSCALL64 entry trampoline arch/x86/entry/entry_64.S | 66 +----------------------------- arch/x86/include/asm/processor.h | 5 +++ arch/x86/include/asm/thread_info.h | 1 + arch/x86/kernel/asm-offsets_64.c | 1 + arch/x86/kernel/cpu/common.c | 11 +---- arch/x86/kernel/kprobes/core.c | 10 +---- arch/x86/kernel/process_64.c | 2 - arch/x86/kernel/vmlinux.lds.S | 10 ----- arch/x86/mm/cpu_entry_area.c | 5 --- arch/x86/mm/pti.c | 24 ++++++++++- 10 files changed, 33 insertions(+), 102 deletions(-) -- 2.17.1
