Clear PG_writeback after zil_commit() for sync I/O #912
Conversation
When writing via ->writepage() the writeback bit was always cleared as part of the txg commit callback. However, when the I/O is being done synchronsously to a slog device we can safely run the callback immediately after zil_commit(). This will significantly reduce the delay for calls such as fsync() and msync() which block until the data is safe on disk. Issue openzfs#700 Issue openzfs#907
|
One thing worries me: are you sure there's no side effects to clearing the writeback bit early? Sure, the data has been written to the ZIL, but if it has been written in immediate mode, then the buffer is still dirty and it's still waiting to be written to the pool in the next TXG sync. I'm not familiar with how buffers and writeback work in Linux memory management, but if clearing the writeback bit means "you can reclaim this", then maybe it would be wrong to do it at this point. Just wondering. |
Yes. I'm sure, and here is why. The Obviously we're safe against a crash here too since the ZIL will get replayed. |
|
2b28613 Clear PG_writeback after zil_commit() for sync I/O |
`zcache hits` displays the hits-by-size histogram, showing the predicted cache hit rate with varying cache size. The data is based on cache accesses since the last `zcache hits --clear`. The problem is that the relevant time period for evaluating cache effectiveness may not match the time since the histogram was cleared. This commit makes the zettacache store a history of hits-by-size histograms, by default one each day. The `zcache hits` command will now sum up the histograms for the requested time/date range, which by default is the entire history (since the cache size was last changed by `zcache add` or `zcache remove`). The new `--begin`, `--end`, and `--duration` flags can be used to select a more specific time range. The `zcache hits --clear` command will create a new histogram, saving the current one in the history. The internal representation of the history can be dumped with `zcache hits --raw`. This can be fed into `zcache hits --from-raw` to query and display the histogram from a different system. The implementation required a new kind of block-based log to store the large, variable-size histograms. The JsonBlockBasedLog stores Rust structs that are serializable by Serde (as opposed to a normal BlockBasedLog which stores fixed-size C structs, anotated with `#[repr(C)]`). There is no mechanism to discard the historical hits-by-size histograms. Performance evaluation found good performance even with enough entries to cover many decades of history. Therefore, reducing the number of entries is left to future work.
When writing via ->writepage() the writeback bit was always cleared
as part of the txg commit callback. However, when the I/O is being
done synchronsously to a slog device we can safely run the callback
immediately after zil_commit().
This will significantly reduce the delay for calls such as fsync()
and msync() which block until the data is safe on disk.
Issue #700
Issue #907