cat: diagnose splice errors by MuntasirSZN · Pull Request #13812 · uutils/coreutils

MuntasirSZN · 2026-08-08T14:44:07Z

Note

The splice() fast path swallowed every error as a silent read/write
fallback, so a splice failure after some bytes were copied (e.g. an
EIO injected by strace in tests/cat/splice.sh) produced no diagnostic.
Once any data has been copied, treat a splice error as fatal and report
it as cat: <input>: … on the read side or cat: write error: … on the
write side; failures before the first byte still
fall back to read/write. Write errors elsewhere now also omit the input
file name.

Now tests/cat/splice.sh passes.

Signed-off-by: MuntasirSZN muntasir.joypurhat@gmail.com

codspeed-hq · 2026-08-08T15:04:54Z

Merging this PR will not alter performance

✅ 342 untouched benchmarks
⏩ 51 skipped benchmarks¹

_{Comparing MuntasirSZN:splice-cat (ecb29d1) with main (8913717)}

51 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

sylvestre

Please add a test to make sure we don't regress in the future

MuntasirSZN · 2026-08-08T15:36:01Z

Done @sylvestre

oech3 · 2026-08-08T15:38:04Z

I don't agree with this approach. This should be shared at uucore for all utils. Also distioction for read/write is not splice specific.

oech3 · 2026-08-08T15:39:07Z

Why cannot we simply use .map_err_context?

oech3 · 2026-08-08T15:52:27Z

Now tests/cat/splice.sh passes.

Note that error distinction is disabled until next releas of GNU at backporting at https://github.com/uutils/coreutils/blob/main/util/fetch-gnu.sh .
The test is failing by different reason (probably due to CI specific reason?, passing locally without PR).

MuntasirSZN · 2026-08-08T16:21:22Z

Now tests/cat/splice.sh passes.

Note that error distinction is disabled until next releas of GNU at backporting at https://github.com/uutils/coreutils/blob/main/util/fetch-gnu.sh .
The test is failing by different reason (probably due to CI specific reason?, passing locally without PR).

Hmm, but not for me? For me locally without pr it doesn't pass.

MuntasirSZN · 2026-08-08T17:02:16Z

@xtqqczze I cannot resolve review that is conflicting, @oech3 is asking for something and you are pushing against it (which he pushed against before) pls read the resolved review comments....

github-actions · 2026-08-08T17:36:16Z

GNU testsuite comparison:

Skip an intermittent issue tests/date/date-locale-hour (fails in this run but passes in the 'main' branch)
Congrats! The gnu test tests/cat/splice is no longer failing!
Note: The gnu test tests/cp/link-heap is now being skipped but was previously passing.
Note: The gnu test tests/csplit/csplit-heap is now being skipped but was previously passing.

oech3 · 2026-08-08T18:15:13Z

I still don't understand this large diff just for appending a new message to existing error message.
New function, enum and macro is unnecessary. No?

MuntasirSZN · 2026-08-09T02:23:44Z

I still don't understand this large diff just for appending a new message to existing error message.
New function, enum and macro is unnecessary. No?

Function, it is (will move to uucore once your prs are done). Enum, no, because it is the idiomatic way as Errors should be in a large enum. Macro, well sylvestre asked for it.

oech3 · 2026-08-09T05:20:15Z

We should remove CatResult and CatError before fixing this to avoid producing specialized types for mostly all utils at least for.

MuntasirSZN · 2026-08-09T05:41:24Z

We should remove CatResult and CatError before fixing this to avoid producing specialized types for mostly all utils at least for.

Isn't that a good thing? In uucore it can have its own error, though CatResult and CatError isn't created by me, so its the maintainers decision.

oech3 · 2026-08-09T05:43:54Z

If we want to share logic for error with file path and distinction, utility specific abstriction is large barrier for sharing code base in my opinion.

xtqqczze

I don’t have any further comments regarding this PR.

oech3 · 2026-08-10T13:08:35Z

I still don't like this PR with too large diff and unshared code base...

oech3 · 2026-08-10T13:12:15Z

I don't understand why is splice.sh(backported from GNU master) is not passing at uutils's main already (without this PR).
Error distinction and catching 2nd+ splice's error should be unnecessary for the test to pass.
I think diff for it should be something tiny.

MuntasirSZN · 2026-08-10T16:23:31Z

I don't understand why is splice.sh(backported from GNU master) is not passing at uutils's main already (without this PR). Error distinction and catching 2nd+ splice's error should be unnecessary for the test to pass. I think diff for it should be something tiny.

I just ran the cat tests at gnu master, and to be honest, splice.sh still fails:

--- exp 2026-08-10 22:21:43.462220266 +0600
+++ err 2026-08-10 22:21:43.462220266 +0600
@@ -1 +0,0 @@
-cat: /dev/zero: Input/output error
--- exp 2026-08-10 22:21:53.500124272 +0600
+++ err 2026-08-10 22:21:53.501124317 +0600
@@ -1 +0,0 @@
-cat: write error: Input/output error
FAIL tests/cat/splice.sh (exit status: 1)

But the new cat-distinct-err.sh passes:

PASS: tests/cat/cat-distinct-err.sh
============================================================================
Testsuite summary for GNU coreutils 9.11
============================================================================
# TOTAL: 1
# PASS:  1
# SKIP:  0
# XFAIL: 0
# FAIL:  0
# XPASS: 0
# ERROR: 0
============================================================================

IDK but only with my patches, even gnu master passes. @oech3

I will investigate in main branch myself to see whats happening....

The splice() fast path swallowed every error as a silent read/write fallback, so a splice failure after some bytes were copied (e.g. an EIO injected by strace in tests/cat/splice.sh) produced no diagnostic. Once any data has been copied, treat a splice error as fatal and report it as "cat: <input>: …" on the read side or "cat: write error: …" on the write side; failures before the first byte still fall back to read/write. Write errors elsewhere now also omit the input file name. Fixes tests/cat/splice.sh (backported in util/fetch-gnu.sh). Signed-off-by: MuntasirSZN <muntasir.joypurhat@gmail.com>

Signed-off-by: MuntasirSZN <muntasir.joypurhat@gmail.com>

oech3 · 2026-08-10T17:01:50Z

+#[inline]
+fn splice_unusable_errno() -> std::io::Error {
+    std::io::Error::from_raw_os_error(rustix::io::Errno::INVAL.raw_os_error())
+}


1 fn is enough. not two.

I think both of them is unnecessary since we can modify/check errno.

oech3 · 2026-08-10T17:06:28Z

-            return Ok(Err(()));
+        match splice(pipe, dest, remaining) {
+            Ok(0) => {
+                // no progress; drain by hand


I believe this match arm is misundertnding.

MuntasirSZN · 2026-08-10T17:10:50Z

@oech3 maybe I found the issue of why splice.sh is failing (as in fetch-gnu of uutils master): So GNU's rule: errors on the first splice cycle → silently fall back to read/write; errors on any later cycle → fatal with diagnosis (input → path, output → write error). uutils currently falls back on every splice error, thats the main issue. Pls correct me if I am wrong... uutils prints cat: /dev/zero: No space left on device for a write error; GNU prints cat: write error: No space left on device. Both stem from the missing Write error variant.

oech3 · 2026-08-10T17:23:47Z

I don't think so since the backported splice.sh's exit code does not depend on splice call completely.

sylvestre reviewed Aug 8, 2026

View reviewed changes

Comment thread src/uu/cat/src/cat.rs Outdated

sylvestre requested changes Aug 8, 2026

View reviewed changes

MuntasirSZN requested a review from sylvestre August 8, 2026 15:35

xtqqczze reviewed Aug 8, 2026

View reviewed changes

Comment thread src/uu/cat/src/cat.rs Outdated