perf(rar): switch to unrar-ng batch extract_all_with_callback

[ttys3]

resolve #714

perf test:

RAR Extraction Performance Comparison (Linux kernel v7.0 source)

this is test runs on a physical machine

tested under tmpfs to avoid filesystem I/O impact.
builtin time command comes from fish shell

Linux: kernel 7.0.3
shell: fish
rar/unrar: 7.22
unzip: UnZip 6.00 of 20 April 2009, by Info-ZIP

CPU: 12th Gen Intel(R) Core(TM) i7-12700
RAM: 32GB

Test file: kernel-v7.0.rar (created from Linux kernel v7.0 source tree)

Extraction Performance

Tool	Command	Executed (wall)	User CPU	Sys CPU
unrar (official)	unrar x kernel-v7.0.rar	7.15 s	7.07 s	1.25 s
ouch 0.7.1 (old release)	ouch decompress kernel-v7.0.rar	70.69 s	68.03 s	4.71 s
ouch (this PR)	ouch decompress kernel-v7.0.rar	6.88 s	6.74 s	1.35 s

Reference: Other Formats

Tool	Command	Executed (wall)	User CPU	Sys CPU
unzip	unzip v7.0.zip	7.17 s	6.20 s	0.95 s
rar (compression)	rar a -r kernel-v7.0.rar ./linux-7.0/	20.12 s	102.85 s	10.30 s

Key Takeaways

• This PR (backed by unrar-ng 0.7.6) is ~10.3x faster than the previous release (70.69 s → 6.88 s).
• Performance is on par with the official unrar tool (6.88 s vs 7.15 s).
• Comparable to unzip extracting the same content (6.88 s vs 7.17 s).

Click here to expand and show detailed test steps below

# ensure /tmp is tmpfs, not disk
❯ mount | grep /tmp
tmpfs on /tmp type tmpfs (rw,nosuid,nodev,size=15858944k,nr_inodes=1048576,inode64,usrquota)

❯ cd /tmp
❯ mkdir perf-test
❯ cd perf-test/
❯ curl -LZO https://github.com/ouch-org/ouch/releases/download/0.7.1/ouch-x86_64-unknown-linux-gnu.tar.gz
❯ curl -LZO https://github.com/torvalds/linux/archive/refs/tags/v7.0.zip

❯ tar xvzf ouch-x86_64-unknown-linux-gnu.tar.gz 

❯ time unzip v7.0.zip
Executed in    7.17 secs    fish           external
   usr time    6.20 secs    0.00 millis    6.20 secs
   sys time    0.95 secs    1.99 millis    0.95 secs


❯ time rar a -r kernel-v7.0.rar ./linux-7.0/
________________________________________________________
Executed in   20.12 secs    fish           external
   usr time  102.85 secs    1.40 millis  102.85 secs
   sys time   10.30 secs    1.78 millis   10.29 secs

❯ rm -rf linux-7.0/

❯ time unrar x kernel-v7.0.rar 
________________________________________________________
Executed in    7.15 secs    fish           external
   usr time    7.07 secs    0.00 millis    7.07 secs
   sys time    1.25 secs    2.73 millis    1.25 secs

❯ rm -rf linux-7.0/

❯ time ./ouch-x86_64-unknown-linux-gnu/ouch decompress -d ouch-extracted  kernel-v7.0.rar 
________________________________________________________
Executed in   70.69 secs    fish           external
   usr time   68.03 secs    1.47 millis   68.03 secs
   sys time    4.71 secs    1.74 millis    4.71 secs

❯ time ~/repo/rust/ouch/target/release/ouch decompress -d ouch-pr-extracted  kernel-v7.0.rar 
________________________________________________________
Executed in    6.88 secs    fish           external
   usr time    6.74 secs    0.19 millis    6.73 secs
   sys time    1.35 secs    3.04 millis    1.35 secs

Migrate RAR decompression from upstream unrar 0.5.7 to the maintained fork unrar-ng 0.7.3 via Cargo dep alias (source-level use unrar::* preserved). Replace the per-file read_header + extract_with_base loop with OpenArchive::extract_all_with_callback, which uses the C batch path internally and avoids per-file FFI overhead -- noticeably faster for archives with many small files.

Behavioral notes:

• Directory entries in the archive are now materialized on disk by the C library; the previous loop explicitly skipped non-file headers so empty directories were not created. Most users expect the archive's original directory layout, so this is treated as a fix.
• Per-file errors are captured in the callback and surfaced as Error::Custom with a "failed to extract " title plus a human-readable detail, rather than relying on a bare From conversion (which omitted the filename context).
• From now formats via Display (err.to_string()) instead of the previous Debug-formatted err.code, so messages such as "Wrong password was specified" replace the bare BadPassword enum debug text. Covers LargeDict and the new Unmapped(i32) fallback too.
• Closure handles the new ExtractEvent::LargeDictWarning by emitting an info line with the required vs supported dictionary size, then letting the DLL fail naturally to Err(LargeDict).

[valoq]

Thanks for the patch and picking up the unrar repo.
This replaces the unmaintained https://github.com/muja/unrar.rs with https://github.com/ttys3/unrar.rs as depencency

The PR actually fixes some important issues, including an actively exploited CVE in the backend unrar library

ideally I would like to switch to a pure rust solution, but that does not seem to exist and may be impossible with the unrar library license situation, so this seems to be the best option for ouch at the moment.

Aside from the failed CI tests, there are a few minor things:

src/archive/rar.rs:52 — the closure returns true unconditionally, including from the Err arm. unrar-ng docs states returning false from Start/Ok/Err cancels the rest of the extraction. The previous code bailed on the first error via ?, so extraction now continues past per-file errors and only the first one is recorded (the rest are silently dropped because of the if first_err.is_none() guard at line 38).

"lets the DLL fail naturally to Err(LargeDict)", but per unrar-ng docs, returning true from LargeDictWarning permits the oversized dictionary and proceeds; it's false that rejects it and produces Code::LargeDict. The closure at line 52 returns true, so it's actually permitting the oversized dict. Or did I miss something?

[ttys3]

@valoq updated, PTAL

[ttys3]

updated PR description with real physical machine tests result