On Wed, 3 Apr 2024 10:04:36 GMT, Alan Bateman <al...@openjdk.org> wrote:
> This change drops the adjustments to the virtual thread scheduler's target
> parallelism when virtual threads do file operations on files that are opened
> for buffered I/O. These operations are usually just too short to have any
> benefit and may have a negative benefit when reading/writing a small number
> of bytes. There is no change for read/write operations on files opened for
> direct I/O or when writing to files that are opened with options for
> synchronized I/O file integrity (O_SYNC/O_DSYNC and equivalents). Sergery
> Kuksenko is polishing benchmarks that includes this area, this is for a
> future PR.
>
> In addition, the blocker mechanism is updated to handle reentrancy as can
> happen if debugging code is added to ForkJoinPool, if there is preemption
> when attempting to compensate, or potentially forced preemption in the
> future. This part is a pre-requisite to the changes to better support object
> monitor there are more places where preemption is possible and this quickly
> leads to unbalanced begin/end.
>
> The changes have been baking in the loom repo for several months.
src/java.base/share/classes/jdk/internal/misc/CarrierThread.java line 104:
> 102: if (compensating == COMPENSATING) {
> 103: ForkJoinPools.endCompensatedBlock(getPool(),
> compensateValue);
> 104: compensating = NOT_COMPENSATING;
For the sake for consistency between the state and the `compensateValue`, would
it be good to reset `compensateValue` to `0` when we switch to
`NOT_COMPENSATING`?
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/18598#discussion_r1557714608
