While rebuilding a superblock, the transaction_id, data_block_size, and
nr_data_blocks could be inherited from the current superblock if available
(assumed it's a multiple-activated copy, so supposed partially valid),
or derived from the probed data mappings and device details, that saves
the hassle of manual input.
Note that either the current superblock or user-overrides must be compatible
with the probed values, thus the filters are applied.
A pre-built node could be under populated (less than half-full) due to
following reasons:
- A single shared leaf generated by dm-thin key removal. The residency
could drop below 50%, until it reaches the merge threshold (33% or 44%,
depends on its context).
- A shared root, which could have any possible nr_entries.
- Underfull shared nodes (less than 33% residency) caused by kernel issues.
To avoid producing under populated nodes, those kinds of pre-built nodes,
except the roots, will be merged into their siblings.
The ref counts now are handled in a straightforward method:
The pre-built subtrees serve as snapshots of potentially shared nodes.
They hold an initial ref count on the pre-built nodes and their children.
The temporary ref counts are later dropped at the end of device building.
This way fixes the ref counts of unshifted nodes, thus the 'reserve'
operation introduced by commit 6d16c58 is reverted.
- Do not use shared internals as Defs
- Build Defs with common leaf sequences only, that could forms
more common Defs between partially shared subtrees.
- Update crc32c to 0.6, which allows it to be built on AArch64.
- Update base64 to 0.13
- Update byteorder to 0.14
- Update io-uring to 0.4
- Update libc to 0.2.83
- Update nix to 0.19
- Update nom to 6.0.1
- Update quick-xml to 0.20
- Update rand to 0.8
- Update tempfile to 3.2
- Update tui to 0.14
Signed-off-by: Kay Lin <i@v2bv.net>
Multithreaded sync-io has performance similar to async-io. Also,
sync-io saves the hassle of setting ulimits to get io_uring working
on some systems (commit ba7fd7b). Now we default to sync-io, and
leave async-io as a hidden option for testing and benchmarking.
The blocks storing metadata itself are located continuously within
a certain reserved range, hence there's no need to use a block set
as the representation.
- Allow packing unaligned number of bitmap entries that could happen
with the last bitmap. Unused trailing entries are zeroed by memset.
- Fix none_free_before in index_entry
A leaf node should be counted only if it is referenced by some internal
nodes, since the leaves generated by <def> sections might be unshifted
and merged with exclusive mappings or other shared defs, or they might
not even be used by any of the devices if the xml was tampered. The
internal nodes should be handled in the same manner.
The new 'reserve' operation is designed for this purpose. Applications
could reserve a block for later use, without setting its ref count
immediately. This way saves the hassles of tracking unused leaves.
- Fix out-of-bounds index
- Automatically flush queued writes before function return
- Track allocated blocks in write_batcher (might be space consuming)
