[hugojosealmeida]

Summary

When +D/+d (or other warning-guarded soft errors) targets a missing or invalid directory, enter_dir() returns failure and sets err, which unconditionally drives usage() and prints the version/usage/help block. The -w/-t flags only set Fwarn, suppressing the warning text but not the error-driven help text. A SoftErr flag now marks these warning-level failures; when warnings are suppressed (Fwarn set), main() exits silently with an error status instead of calling usage(), while normal (warnings-enabled) runs retain the existing warning + help output. Hard errors are unaffected and still show help text.

Changes

• src/arg.c: declare extern int SoftErr; and set SoftErr = 1 on warning-guarded soft failures (enter_dir, enter_efsys, enter_fd_lst), with braces fixed on the Readlink branches.
• src/main.c: define int SoftErr = 0; and, before calling usage(), call Error(ctx) when err && Fwarn && SoftErr so suppressed soft errors exit silently.

Verification

• make -> build succeeds
• ./lsof -w +D /test -> no output, exit 1
• ./lsof -t +D /test -> no output, exit 1
• ./lsof +D /test -> warning + help text preserved, exit 1
• ./lsof -w -z -> hard error still shows help text, exit 1
• bash tests/case-20-exit-status.bash ... -> pass
• bash tests/case-20-handle-missing-files.bash ... -> pass

Risks / Follow-ups

See CI failure description at #374 (comment)

---

This PR is AI Assisted.

Closes #312

[hugojosealmeida]

The following CI steps are failing at Build and test using auto toolset on ci/circleci: linux-archlinux, ci/circleci: linux-fedora38 and ci/circleci: linux-opensuse15

FAIL: tests/LTlock

I am not able to see details of ./test-suite.log at https://app.circleci.com/pipelines/github/lsof-org/lsof/884/details?workflowId=8c9fec36-da21-4ac9-9914-68b5ef27c635.

Nevertheless it passes in my local environment (opensuse/leap:15):

NAME="openSUSE Leap"
VERSION="15.6"

(...)
SKIP: lib/dialects/linux/tests/case-20-ux-socket-endpoint.bash
PASS: lib/dialects/linux/tests/case-20-ux-socket-endpoint-unaccepted.bash
PASS: lib/dialects/linux/tests/case-20-unbal-parens.bash
PASS: tests/LTbasic
PASS: tests/LTbigf
PASS: tests/LTdnlc
PASS: tests/LTlock
SKIP: tests/LTnfs
PASS: tests/LTnlink
PASS: tests/LTsock
PASS: tests/LTszoff
PASS: tests/LTunix
PASS: tests/LTbasic2
============================================================================
Testsuite summary for lsof 4.99.7
============================================================================
# TOTAL: 42
# PASS:  31
# SKIP:  11
# XFAIL: 0
# FAIL:  0
# XPASS: 0
# ERROR: 0
============================================================================
make[2]: Leaving directory '/src'
make[1]: Leaving directory '/src'
4cf32bdea806:/src # cat ./test
test-driver     test-suite.log  tests/          
4cf32bdea806:/src # cat ./test
test-driver     test-suite.log  tests/          
4cf32bdea806:/src # cat ./test-suite.log 
===================================
   lsof 4.99.7: ./test-suite.log
===================================

# TOTAL: 42
# PASS:  31
# SKIP:  11
# XFAIL: 0
# FAIL:  0
# XPASS: 0
# ERROR: 0

Maybe someone could support me to understand what is really failing at circleci (x86_64 vs. my aarch64 container??).

[jiegec]

Please do not commit temp files

[hugojosealmeida]

accidentally committed

[hugojosealmeida]

seems the repo's .gitignore ignores *.o, *.lo, etc., but not *.o.tmp, so libtool temp files can slip through. Maybe this also could be added to .gitignore?!

[jiegec]

I prefer to change the return value, instead of using another global variable to express this difference, e.g. return (2).

[hugojosealmeida]

I see, nevertheless, 0 and 1 are the conventional success/failure returns that are used throughout lsof codebase, like read it as boolean. A bare 2  has no such convention, it is kind we need to guess what it means. I will take a look at this to make the solution more elegant, maybe using a 2  with a comment in the function header...

[jiegec]

Maybe an enum for this?

[hugojosealmeida]

In this case a full three-value enum ( OK ,  ERR ,  SOFT_ERR) would be cleaner, but the impact is huge. But 0 and 1 shall remain untouched and so I do agree that having the anonymous enum (e.g. SOFT_ERR = 2) is the compromise.

[hugojosealmeida]

I have replaced global SoftErr with enum { SOFT_ERR = 2 }, preserving existing 0/1 returns. Soft errors propagate through helpers without masking hard errors. Added regression tests.