diff options
author | Dominik Brodowski <linux@dominikbrodowski.net> | 2018-02-11 11:49:47 +0100 |
---|---|---|
committer | Ingo Molnar <mingo@kernel.org> | 2018-02-13 09:04:54 +0100 |
commit | dde3036d62ba3375840b10ab9ec0d568fd773b07 (patch) | |
tree | 4cae26275c84cfe9660540c81b5bca2531c74cb7 /arch/x86/entry/vsyscall | |
parent | 30907fd13bb593202574bb20af58d67c70a1ee14 (diff) | |
download | op-kernel-dev-dde3036d62ba3375840b10ab9ec0d568fd773b07.zip op-kernel-dev-dde3036d62ba3375840b10ab9ec0d568fd773b07.tar.gz |
x86/entry/64: Get rid of the ALLOC_PT_GPREGS_ON_STACK and SAVE_AND_CLEAR_REGS macros
Previously, error_entry() and paranoid_entry() saved the GP registers
onto stack space previously allocated by its callers. Combine these two
steps in the callers, and use the generic PUSH_AND_CLEAR_REGS macro
for that.
This adds a significant amount ot text size. However, Ingo Molnar points
out that:
"these numbers also _very_ significantly over-represent the
extra footprint. The assumptions that resulted in
us compressing the IRQ entry code have changed very
significantly with the new x86 IRQ allocation code we
introduced in the last year:
- IRQ vectors are usually populated in tightly clustered
groups.
With our new vector allocator code the typical per CPU
allocation percentage on x86 systems is ~3 device vectors
and ~10 fixed vectors out of ~220 vectors - i.e. a very
low ~6% utilization (!). [...]
The days where we allocated a lot of vectors on every
CPU and the compression of the IRQ entry code text
mattered are over.
- Another issue is that only a small minority of vectors
is frequent enough to actually matter to cache utilization
in practice: 3-4 key IPIs and 1-2 device IRQs at most - and
those vectors tend to be tightly clustered as well into about
two groups, and are probably already on 2-3 cache lines in
practice.
For the common case of 'cache cold' IRQs it's the depth of
the call chain and the fragmentation of the resulting I$
that should be the main performance limit - not the overall
size of it.
- The CPU side cost of IRQ delivery is still very expensive
even in the best, most cached case, as in 'over a thousand
cycles'. So much stuff is done that maybe contemporary x86
IRQ entry microcode already prefetches the IDT entry and its
expected call target address."[*]
[*] http://lkml.kernel.org/r/20180208094710.qnjixhm6hybebdv7@gmail.com
The "testb $3, CS(%rsp)" instruction in the idtentry macro does not need
modification. Previously, %rsp was manually decreased by 15*8; with
this patch, %rsp is decreased by 15 pushq instructions.
[jpoimboe@redhat.com: unwind hint improvements]
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Denys Vlasenko <dvlasenk@redhat.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: dan.j.williams@intel.com
Link: http://lkml.kernel.org/r/20180211104949.12992-7-linux@dominikbrodowski.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Diffstat (limited to 'arch/x86/entry/vsyscall')
0 files changed, 0 insertions, 0 deletions