Commit Graph

3 Commits

Author SHA1 Message Date
ispyisail f8c1b5206a Address review on #905: offered-list semantics, dump ordering, FreeBSD, legacy dumps
Five things raised in review, plus tests for the parts that were only
described in prose.

clearPendingCrashDump() did not do what its comment said. It called
pendingCrashDumpFiles() again at clear time, so it deleted whatever was
in the directory then, not what had been offered. The offer sits inside a
modal dialog that stays open as long as the user reads it, and
SingleApplication keys its socket on the binary path, so a second
QElectroTech build running alongside is a separate process that can crash
and write a dump in that window. Re-listing deleted that dump unseen --
the exact failure this change exists to fix. The list is now taken once
in QETApp::checkCrashDump() and passed to both
pendingCrashDumpContents() and clearPendingCrashDump().

The ring is now written before the backtrace. backtrace() unwinds through
libgcc, which calls dl_iterate_phdr and takes the loader lock; warming it
in install() removes the allocation but not the lock. Crashing inside
dlopen() (Qt plugin loading), or on a corrupted stack, could therefore
hang or re-fault the handler at the backtrace and lose the ring with it.
Order is now header, signal, ring, backtrace, so the cheapest and most
valuable part is already on disk before anything that can block. The
class comment claimed the handler takes no locks; that was not strictly
true and now says so.

QET_CRASH_BACKTRACE comes from find_package(Backtrace) rather than
__has_include(<execinfo.h>). The header exists on FreeBSD but backtrace()
lives in libexecinfo there, so the probe compiled and the link failed.

A crash_dump.log left by a pre-#905 version is migrated into crashes/ at
startup, named from its own mtime. Otherwise upgrading stranded it: the
new code never looks at that path, so the dump from the crash that
prompted the upgrade would sit there unoffered forever.

Also from the review: dumps are capped at the 10 newest, so a crash loop
cannot fill the log directory before any dialog is shown; crashDumpDir()
no longer creates the directory as a side effect of a const getter
(ensureCrashDumpDir() does that for the callers that write); and redact()
now masks an AppImage's per-run /tmp/.mount_XXXXXX prefix, which
backtrace_symbols_fd() writes into every frame.

Two test executables, both of which were checked to fail against the
behaviour they replace:

- tst_crashhandler covers CrashHandler::formatInt(), which had no
  coverage at all despite running only inside a signal handler, where
  nothing can assert: zero, negatives, INT_MIN (negated through unsigned,
  since -INT_MIN is UB), INT_MAX, truncation and a zero-sized buffer,
  each checked against a sentinel-filled buffer so a write past the
  reported length fails.
- tst_crashdumps covers the bookkeeping: ordering, empty dumps, the
  exclusion of this run's own path, the cap, concatenation of every
  offered dump, that clearing deletes only what was offered, and what
  redact() masks. qetlogger.cpp needs exactly one symbol from the
  application, QETApp::dataDir(), which the test supplies itself.

Not addressed here: the timestamp in crash_<timestamp>_<pid> is the
launch time, not the crash time -- correct as observed, and the commit
message that implied otherwise was the thing that was wrong. Resolvable
QET frames for AppImage/Flatpak/Snap/Debian need -rdynamic and archived
debug symbols, which is a packaging discussion, not this change.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-09-18 14:31:49 +12:00
ispyisail 0646f9ca4f Harden the crash reporter: keep every dump, and say what crashed
Three weaknesses, all visible in ChuckNr11's report on #898 -- "the report
appeared only once despite there being 10 or more crashes".

One dump per run instead of one per install
-------------------------------------------
crashDumpPath() was a single fixed crash_dump.log, and the handler opens it
O_TRUNC, so each crash destroyed the evidence from the one before. Ten
crashes left one dump. Dumps now go to a crashes/ directory named
crash_<timestamp>_<pid>.log, and every pending one is offered together,
newest first, with a banner saying how many there are. A crash that repeats
is exactly the case where the earlier dumps matter, because the difference
between them is the evidence.

The name is built in normal context and handed to CrashHandler::install(),
which copies it into a preallocated buffer as before -- the handler still
writes to one fixed path, so its no-allocation invariant is untouched.

The dump now says which signal fired
------------------------------------
The header is built once at install(), so every dump looked identical no
matter what killed the process -- and SIGSEGV and SIGABRT point at very
different bugs. Written with an async-signal-safe integer formatter into a
stack buffer, since snprintf is not on the POSIX safe list.

...and where it was
-------------------
The ring said what the program was doing; nothing said where it died. The
dump now carries a backtrace. backtrace() is warmed once in install() so
its first-call lazy resolution cannot allocate inside the handler, and
backtrace_symbols_fd() writes straight to the fd -- unlike
backtrace_symbols(), which mallocs and must never be used here. Guarded on
__has_include(<execinfo.h>) so platforms without it are unaffected.

QET's own frames currently resolve as offsets rather than names, since the
binary does not export its dynamic symbols. They are still resolvable
offline: the header records the exact git SHA. Building with -rdynamic
would give names directly, but that is a build-flag decision for its own
change.

Deliberately unchanged: the four invariants in crashhandler.h. Nothing
added here allocates, blocks, takes a lock, or swallows the crash.

Verified: three consecutive SIGSEGVs now leave three separate dumps, each
carrying "Signal: 11" and a backtrace with resolved Qt frames; launching
afterwards offers all three in one dialog, newest first, and clears them
once shown. ctest 8/8.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-09-17 22:44:59 +12:00
ispyisail 5dec36cb29 Add crash-time ring flush and a diagnostics export UI (discussion #644, steps 4-5)
Stacked on the steps 1-3 branch (feature-diagnostic-logging, PR #646).
Kept as its own PR rather than folded into that one, matching the
discussion's own framing: step 4 is explicitly "the highest-risk piece
... lands last, behind its own switch."

## Step 4 -- crash-time ring flush (CrashHandler)

Installs a handler for SIGSEGV/SIGABRT/SIGBUS/SIGFPE/SIGILL (POSIX) /
SetUnhandledExceptionFilter (Windows) that flushes the in-memory ring to
a fixed crash_dump.log before the process dies.

This required reworking LogRing (step 3) to be genuinely lock-free, not
just mutex-protected: a signal handler that blocks on a lock the
crashing thread (or another thread) already holds turns a clean crash
into a hang -- no ring dump *and* no core dump, worse than doing
nothing. append() now claims a slot with a single atomic fetch-add;
dumpToFd() reads the preallocated entries directly and writes them with
write(2) only, looping on EINTR/short writes. Accepted tradeoff: at most
one entry can be read torn if a crash lands mid-append into that exact
slot -- documented in logring.h, and the alternative (a seqlock to
detect and retry) wasn't judged worth the complexity for that window.

Other invariants implemented per the discussion:
- sigaltstack with a static 64 KiB buffer, SA_ONSTACK -- a stack-
  overflow SIGSEGV has no usable stack for a handler without one.
- Nothing under the actual handler touches Qt, QString or the
  allocator: the dump path and a small header (version/git/OS/Qt) are
  precomputed into fixed char buffers by install(), which runs once at
  startup in normal context.
- Atomic test-and-set so only the first crash writes a dump; a second
  concurrent/nested fault goes straight to restore-and-re-raise.
- After writing, the handler restores SIG_DFL and re-raises (POSIX) /
  returns EXCEPTION_CONTINUE_SEARCH (Windows) so the OS's own crash
  path -- core dump, Windows Error Reporting -- still runs. A handler
  that "fixed" the crash by swallowing the signal would destroy exactly
  the post-mortem evidence this whole design exists to preserve.

Tested in this environment: POSIX/Linux only, all five signals. Sent
each directly to a running process and confirmed (a) crash_dump.log is
written with the correct header and ring contents, mode 0600, and (b)
the process still terminates via the signal with the kernel's own
"core dumped" flag set (exit code 128+signal, confirmed for all five).
The Windows path is implemented per the discussion's guidance but is
untested -- no Windows build available in this sandbox.

## Step 5 -- getting the data back out

- QETApp::checkCrashDump(), called from checkBackupFiles() only when
  there's no stale project file to recover this run (so the two
  prompts never both show, per the discussion), offers an unretrieved
  crash dump via DiagnosticsReportDialog and then deletes it regardless
  of the user's choice -- offered exactly once.
- A new "Aide > Enregistrer un rapport de diagnostic..." action
  (QETMainWindow) builds the same kind of report from the *current*
  session (QetLogger::buildDiagnosticsReport(): header + this session's
  log file) for a manual "attach this to a bug report" flow, not tied
  to a crash.
- Both go through QetLogger::redact() before ever reaching the user:
  the one redaction implemented is a literal replace of the home
  directory with "~", since an absolute path under it leaks the
  account name. The discussion's fancier "optionally redact project
  filenames too" isn't attempted -- reliably telling a project path
  apart from arbitrary log text is a much fuzzier problem than a
  literal prefix match.
- DiagnosticsReportDialog shows the full (already-redacted) content
  before saving, per the discussion: "the user is about to attach this
  to a public tracker."

Verified in a real GUI session (Xvfb): triggered a SIGSEGV, relaunched,
confirmed the crash-report dialog appears with the right header/content,
confirmed it does not reappear on a second relaunch, and confirmed the
manual "Save report" action produces a correctly-formatted report and
saves it to a chosen path.

Built clean, no new warnings.

## Build systems

Registered in both: cmake/qet_compilation_vars.cmake, and
qelectrotech.pro. The .pro needed explicit globs for the new
sources/logging/ui/ subfolder -- sources/logging/*.{h,cpp} was already
globbed, but unlike the other ui/ subfolders that one had no entry of
its own, so diagnosticsreportdialog.{h,cpp} would not have been built
under qmake.
2026-08-06 23:04:27 +12:00