As atomic_inc_not_zero() is implemented with a try_cmpxchg() loop it has
O(N^2) behaviour under contention with N concurrent operations. The
rcuref infrastructure uses atomic_add_negative_relaxed() for the fast
path, which scales better under contention and we get overflow
protection for free.
I've been testing this with will-it-scale using a multi-threaded fstat()
on the same file descriptor on a machine that Jens gave me access (thank
you very much!):
processor : 511
vendor_id : AuthenticAMD
cpu family : 25
model : 160
model name : AMD EPYC 9754 128-Core Processor
and I consistently get a 3-5% improvement on workloads with 256+ and
more threads comparing v6.12-rc1 as base against with these patches
applied.
Note that atomic_inc_not_zero() contained a full memory barrier that we
relied upon. But we only need an acquire barrier and so I replaced the
second load from the file table with a smp_load_acquire(). I'm not
completely sure this is correct or if we could get away with something
else. Linus, maybe you have input here?
Maybe this is all a bad idea but I've wasted enough time on performance
testing this that I at least wanted to have it on list.
Signed-off-by: Christian Brauner <brauner@xxxxxxxxxx>
---
Christian Brauner (4):
fs: protect backing files with rcu
types: add rcuref_long_t
rcuref: add rcuref_long_*() helpers
fs: port files to rcuref_long_t
drivers/gpu/drm/i915/gt/shmem_utils.c | 2 +-
drivers/gpu/drm/vmwgfx/ttm_object.c | 2 +-
fs/eventpoll.c | 2 +-
fs/file.c | 17 ++--
fs/file_table.c | 18 ++--
include/linux/fs.h | 9 +-
include/linux/rcuref_long.h | 166 ++++++++++++++++++++++++++++++++++
include/linux/types.h | 10 ++
lib/rcuref.c | 104 +++++++++++++++++++++
9 files changed, 308 insertions(+), 22 deletions(-)
---
base-commit: 9852d85ec9d492ebef56dc5f229416c925758edc
change-id: 20240927-brauner-file-rcuref-bfa4a4ba915b
