Change Details

The current scheme of calling VOP_GETATTR adds avoidable overhead. On Cascade Lake (no meltdown, mds and similar problems) Linux gets the following on tmpfs: fstat1_processes -t 1 -s 10: average:9299441 fstat1_threads -t 1 -s 10: average:8291153 In contrast, FreeBSD on the same hardware: fstat1_processes -t 1 -s 10: average:7488958 Patched: fstat1_processes -t 1 -s 10: average:7913833 Which mostly catches up to threaded case. Remaining overhead is copying more data and avoidable atomics. I'll patch other filesystems (zfs and ufs) later.

The current scheme of calling VOP_GETATTR adds avoidable overhead. On Cascade Lake (no meltdown, mds and similar problems) Linux gets the following on tmpfs: fstat1_processes -t 1 -s 10: average:9299441 fstat1_threads -t 1 -s 10: average:8291153 In contrast, FreeBSD on the same hardware: fstat1_processes -t 1 -s 10: average:7488958 Patched: fstat1_processes -t 1 -s 10: average:7913833 [there is no difference for threaded case] Which mostly catches up to threaded case. Remaining overhead is copying more data and avoidable atomics. I'll patch other filesystems (zfs and ufs) later.

The current scheme of calling VOP_GETATTR adds avoidable overhead. On Cascade Lake (no meltdown, mds and similar problems) Linux gets the following on tmpfs: fstat1_processes -t 1 -s 10: average:9299441 fstat1_threads -t 1 -s 10: average:8291153 In contrast, FreeBSD on the same hardware: fstat1_processes -t 1 -s 10: average:7488958 Patched: fstat1_processes -t 1 -s 10: average:7913833 [there is no difference for threaded case] Which mostly catches up to threaded case. Remaining overhead is copying more data and avoidable atomics. I'll patch other filesystems (zfs and ufs) later.