On Fri, Oct 6, 2023 at 8:55 AM Thomas Munro <thomas.munro@gmail.com> wrote:
> ... DELAY_CHKPT_COMPLETE ...
About that... If the lights go out after the truncation and the
delayed logging of the checkpoint, how do we know the truncation has
actually reached the disk? mdtruncate() enqueues fsync() calls, but
if we were already in phase 2 (see proc.h) of a checkpoint at that
moment, they might be processed by the *next* checkpoint, not the one
whose phase 3 we've carefully delayed there, no?