OSDN Git Service

locking: remove spin_lock_prefetch
authorMateusz Guzik <mjguzik@gmail.com>
Sat, 12 Aug 2023 16:15:54 +0000 (18:15 +0200)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sat, 12 Aug 2023 16:18:47 +0000 (09:18 -0700)
commitc8afaa1b0f8bc93d013ab2ea6b9649958af3f1d3
treecdf2d4c2f22ab40d7d31ba99ffdce18e3dfcc6cd
parent3feecb1b848359b152dd66b26c24c2454a64fc15
locking: remove spin_lock_prefetch

The only remaining consumer is new_inode, where it showed up in 2001 as
commit c37fa164f793 ("v2.4.9.9 -> v2.4.9.10") in a historical repo [1]
with a changelog which does not mention it.

Since then the line got only touched up to keep compiling.

While it may have been of benefit back in the day, it is guaranteed to
at best not get in the way in the multicore setting -- as the code
performs *a lot* of work between the prefetch and actual lock acquire,
any contention means the cacheline is already invalid by the time the
routine calls spin_lock().  It adds spurious traffic, for short.

On top of it prefetch is notoriously tricky to use for single-threaded
purposes, making it questionable from the get go.

As such, remove it.

I admit upfront I did not see value in benchmarking this change, but I
can do it if that is deemed appropriate.

Removal from new_inode and of the entire thing are in the same patch as
requested by Linus, so whatever weird looks can be directed at that guy.

Link: https://git.kernel.org/pub/scm/linux/kernel/git/tglx/history.git/commit/fs/inode.c?id=c37fa164f793735b32aa3f53154ff1a7659e6442
Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
arch/alpha/include/asm/processor.h
arch/arm64/include/asm/processor.h
arch/ia64/include/asm/processor.h
arch/mips/include/asm/mach-cavium-octeon/cpu-feature-overrides.h
arch/powerpc/include/asm/processor.h
arch/sparc/include/asm/processor_64.h
arch/x86/include/asm/processor.h
fs/inode.c
include/linux/prefetch.h