Codestin Search App

kleisauke · 2026-03-28T17:00:06Z

Opened as a draft since ThreadSanitizer is not officially supported by CIFuzz.

Context: #4713 (comment).
Depends on: google/oss-fuzz#15235.

Use `libraries` instead of `requires` when calling `pkg.generate()` to make the configuration succeed with internal dependencies (i.e. subprojects). Non-functional change.

Helps: libvips/libvips#4981.

kleisauke · 2026-03-29T14:08:00Z

This is starting to take shape. I've opened PR #4985 and google/oss-fuzz#15235, which this PR depends on.

TSan is also far more reliable when GLib is built with it, which allows me to remove most of the previously needed suppressions. :)

Instead, exit the loop when `progress()` returns an error.

Do this only when building with `-Ddeprecated=true`.

kleisauke · 2026-03-31T08:15:48Z

The changeset at kleisauke/libvips@8ed96ff...6967b16 (which I'll split out later) fixes all known data races found in this PR. The TSan suppressions now apply only to GLib:

libvips/suppressions/tsan.supp

Lines 3 to 16 in 6967b16

    
           ## The suppressions below match only the top frame of report stacks 
        
           # GClosure mixes atomic compare and exchange with non atomic reads 
        
           # (read of cl->n_guards and cl->derivative_flag is non atomic) 
        
           # https://gitlab.gnome.org/GNOME/glib/-/issues/1672 
        
           race_top:closure_invoke_notifiers 
        
           race_top:g_cclosure_marshal_VOID__POINTER 
        
           race_top:vips_INT__VOID 
        
           ## The suppressions below match all frames of report stacks 
        
           # Unsynchronized read of once->status from g_once 
        
           # https://gitlab.gnome.org/GNOME/glib/-/issues/1672 
        
           race:g_once_impl

Note that accessing a shared variable from multiple threads without synchronization (where at least one access is a write) is always a data race and causes UB.

dloebl · 2026-03-31T08:32:57Z

That's awesome! Thanks Kleis!
How severe were the TSan findings so far? I'm asking because we still are on libvips v8.16.0 because of #4622

kleisauke · 2026-03-31T09:23:11Z

I think(?) the only severe issue was the one fixed in commit 6967b16 (the code was introduced in v8.16.1), where the eval callback could race with the one in vips_sink_base_progress(). However, this would likely only affect cases where vips_image_set_progress() is enabled and the eval signal is connected via g_signal_connect().

Other than that, the remaining data races involved only flags, which are likely not severe.

jcupitt · 2026-03-31T09:25:45Z

Note that accessing a shared variable from multiple threads without synchronization (where at least one access is a write) is always a data race and causes UB.

libvips does that quite often for events which are not time-critical. Do we need to lock all these, or can we just suppress them?

kleisauke · 2026-03-31T10:26:59Z

Note that accessing a shared variable from multiple threads without synchronization (where at least one access is a write) is always a data race and causes UB.

libvips does that quite often for events which are not time-critical. Do we need to lock all these, or can we just suppress them?

I think this falls under what are referred to as "Benign" data races. This blog post also discusses why such races can still be problematic:
https://bartoszmilewski.com/2020/08/11/benign-data-races-considered-harmful/

I'm aware that GLib uses __ATOMIC_SEQ_CST semantics for its g_atomic API, so it's not using weak atomics. Let me run benchmarks on the previously mentioned changeset using https://github.com/kleisauke/vips-microbench, to see whether it has a measurable impact on performance.

jcupitt · 2026-03-31T10:44:58Z

https://bartoszmilewski.com/2020/08/11/benign-data-races-considered-harmful/

Ah interesting, I guess that's true. Then I suppose we should use glib atomics for these cross-thread "signal" variables.

It'd be good to avoid mutexes or cond vars if possible, they can limit SMP scaling.

Ensure it corresponds to the image passed to `vips_sink_disc()`. This fixes a regression introduced in bbd9207 where eval callbacks failed to propagate to the image that was originally marked with `vips_image_set_progress()`.

This reverts commit e13d378, no longer needed after 8154442.

This reverts commit 6967b16, no longer needed after 8154442.

kleisauke force-pushed the cifuzz-add-tsan branch 2 times, most recently from 8a948f2 to 58c1cfc Compare March 29, 2026 13:40

meson: fix pkg-config generation with internal dependencies

56f0d01

Use `libraries` instead of `requires` when calling `pkg.generate()` to make the configuration succeed with internal dependencies (i.e. subprojects). Non-functional change.

kleisauke mentioned this pull request Mar 29, 2026

meson: fix pkg-config generation with internal dependencies #4985

Open

kleisauke added a commit to kleisauke/oss-fuzz that referenced this pull request Mar 29, 2026

CIFuzz: support ThreadSanitizer

d2d26d7

Helps: libvips/libvips#4981.

kleisauke mentioned this pull request Mar 29, 2026

CIFuzz: support ThreadSanitizer google/oss-fuzz#15235

Open

kleisauke added 2 commits March 30, 2026 14:43

CI: disable TSan workflow

763fc6d

CIFuzz: integrate ThreadSanitizer

e7cd7b7

kleisauke force-pushed the cifuzz-add-tsan branch 2 times, most recently from 3313b54 to 6aa50c5 Compare March 30, 2026 17:03

kleisauke added 16 commits March 31, 2026 09:52

CIFuzz: built GLib with TSan instrumentation

f746a3c

fuzz: ensure TSan exits after first reported error

3aed0e9

fuzz: ensure TSan suppressions are applied

15f3224

fuzz: disable TSan errors from non-instrumented code

8ed96ff

Prefer use of g_once_init*

d59908d

threadpool: make pool->stop and pool->error atomic

1e3d57b

Simplify previous commit

af66c63

threadpool: avoid double check

5fd1b69

Instead, exit the loop when `progress()` returns an error.

image: make ->kill atomic

284d5cd

image: avoid attaching/updating the time struct to ->progress_signal

bbd9207

Do this only when building with `-Ddeprecated=true`.

region: skip FALSE assignment if region is already valid

5e822b4

region: hoist VIPS_IMAGE_SIZEOF_PEL calculation to avoid race

6a9fdb9

sink: fix ->processed data race on 64-bit platforms

e13d378

sinkdisc: skip error check if operation already failed

dac8deb

sinkdisc: make ->kill atomic

610f507

cgifsave: avoid triggering eval callbacks

6967b16

kleisauke force-pushed the cifuzz-add-tsan branch from 6aa50c5 to 6967b16 Compare March 31, 2026 07:57

threadpool: remove need of worker->stop

22f1e2b

kleisauke added 7 commits April 1, 2026 13:18

webpsave: trigger eval callbacks on save->ready

6fa9d66

Ensure it corresponds to the image passed to `vips_sink_disc()`. This fixes a regression introduced in bbd9207 where eval callbacks failed to propagate to the image that was originally marked with `vips_image_set_progress()`.

Merge branch 'master' into cifuzz-add-tsan

64ff522

dzsave: check cancellation on save->ready

3a992e7

threadpool: guard progress reporting with allocate_lock mutex

8154442

Revert "sink: fix ->processed data race on 64-bit platforms"

14e2839

This reverts commit e13d378, no longer needed after 8154442.

Revert "cgifsave: avoid triggering eval callbacks"

be55e48

This reverts commit 6967b16, no longer needed after 8154442.

Simplify 8154442

44cd859

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

CIFuzz: integrate ThreadSanitizer#4981

CIFuzz: integrate ThreadSanitizer#4981
kleisauke wants to merge 27 commits intolibvips:masterfrom
kleisauke:cifuzz-add-tsan

kleisauke commented Mar 28, 2026 •

edited

Loading

Uh oh!

kleisauke commented Mar 29, 2026

Uh oh!

kleisauke commented Mar 31, 2026

Uh oh!

dloebl commented Mar 31, 2026

Uh oh!

kleisauke commented Mar 31, 2026

Uh oh!

jcupitt commented Mar 31, 2026

Uh oh!

kleisauke commented Mar 31, 2026

Uh oh!

jcupitt commented Mar 31, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Conversation

kleisauke commented Mar 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

kleisauke commented Mar 29, 2026

Uh oh!

kleisauke commented Mar 31, 2026

Uh oh!

dloebl commented Mar 31, 2026

Uh oh!

kleisauke commented Mar 31, 2026

Uh oh!

jcupitt commented Mar 31, 2026

Uh oh!

kleisauke commented Mar 31, 2026

Uh oh!

jcupitt commented Mar 31, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

kleisauke commented Mar 28, 2026 •

edited

Loading