WIP: Add a mechanism to abort a blend file reading on critical error. #105085

Bastien Montagne · 2023-02-22T18:01:19+01:00

Bastien Montagne commented

2023-02-22 18:01:19 +01:00

This commit introduces a C++-exception based mechanism to abort a blend
file reading.

The generated BMain is then marked as invalid, and calling code (like
e.g. #BKE_blendfile_read etc.) will not proceed in such case.

This is an implementation of the #105083 design task.

TODOs:

Check further about usages of BLO_read_from_file, this patch should
cover 'regular' blendfile reading and linking/appending code, but
another check pass is needed here.

Write tests to validate at least the basics of this new mechanism.

Some notes:

This commit also has a fix to Main freeing function in case it is in
'split' state. This should be committed separately.

There is a memleak from read_libblock, not clear whether this is caused
by the forged .blend file or this patch yet.

The only actual change currently is adding a check on bScreen->scene
pointer being non-NULL (related to #99836), but we likely want way more
usages of this 'abort fileread'.

Given the extense of the change, I do not think this should be
considered for 3.5 and previous LTS releases. For these a simple 'check
and assign first available scene' "fix" would probably be enough to
address #99836 ?

This commit introduces a C++-exception based mechanism to abort a blend file reading. The generated BMain is then marked as invalid, and calling code (like e.g. #BKE_blendfile_read etc.) will not proceed in such case. This is an implementation of the #105083 design task. ------------------------------ TODOs: Check further about usages of `BLO_read_from_file`, this patch should cover 'regular' blendfile reading and linking/appending code, but another check pass is needed here. Write tests to validate at least the basics of this new mechanism. ------------------------------ Some notes: This commit also has a fix to Main freeing function in case it is in 'split' state. This should be committed separately. There is a memleak from `read_libblock`, not clear whether this is caused by the forged .blend file or this patch yet. The only actual change currently is adding a check on bScreen->scene pointer being non-NULL (related to #99836), but we likely want way more usages of this 'abort fileread'. Given the extense of the change, I do not think this should be considered for 3.5 and previous LTS releases. For these a simple 'check and assign first available scene' "fix" would probably be enough to address #99836 ?

Bastien Montagne added this to the Core project 2023-02-22 18:01:44 +01:00

Bastien Montagne added the

Module

Core

label 2023-02-22 18:02:17 +01:00

Bastien Montagne referenced this pull request

2023-02-23 11:08:14 +01:00

Add a mechanism to abort blend file reading on critical error. #105083

Brecht Van Lommel commented

2023-02-23 17:47:51 +01:00

I'm unsure about using exceptions in code that is still written in C style. This can lead to memory leaks and hanging mutex locks, and you can't tell from looking at any function by itself that it needs to using RAII.

Is it possible to have more explicit C style error checking? I don't know how deep these call stacks go and how many functions this affects.

I'm unsure about using exceptions in code that is still written in C style. This can lead to memory leaks and hanging mutex locks, and you can't tell from looking at any function by itself that it needs to using RAII. Is it possible to have more explicit C style error checking? I don't know how deep these call stacks go and how many functions this affects.

Bastien Montagne commented

2023-02-23 18:24:32 +01:00

@brecht this is certainly possible, but it will then become way more verbose and intrusive, since basically each level of the call stack would need to periodically check the error status/flag to abort the reading... And even then, it would not give us the same 'instant abort' behavior as the one we get with exceptions...

Brecht Van Lommel commented

2023-02-23 18:35:45 +01:00

I'm fine with catching exceptions as a fallback, but don't think we should be throwing them.

It would be more verbose to use error checking, but as far as I can tell this patch adds just one BLO_read_assert_message, and all we have to do is go up the call stack and add some tests for that case, meanwhile checking there are no memory leaks or hanging mutex locks in those functions?

I didn't understand why it would not be as instant.

I'm fine with catching exceptions as a fallback, but don't think we should be throwing them. It would be more verbose to use error checking, but as far as I can tell this patch adds just one `BLO_read_assert_message`, and all we have to do is go up the call stack and add some tests for that case, meanwhile checking there are no memory leaks or hanging mutex locks in those functions? I didn't understand why it would not be as instant.

Bastien Montagne force-pushed F-abort-file-reading from b9a626007a to bfdcd8de3a

2023-02-27 18:17:27 +01:00

Compare

Bastien Montagne commented

2023-02-27 18:22:02 +01:00

Updated to use the new main invalid flag and check it in a few 'critical' places, instead of using exceptions... This will get way more verbose if we keep adding these in the future.

Brecht Van Lommel requested changes 2023-02-28 10:27:04 +01:00

source/blender/blenloader/intern/readfile.cc

						
				@ -3604,1 +3635,4 @@

				    main->is_locked_for_linking = false;

				    return;

				  }

				  do_versions_after_linking_cycles(main);

Brecht Van Lommel commented

2023-02-28 10:24:35 +01:00

I suggest to write it like this instead of returning early and having main->is_locked_for_linking = false duplicated.

if (!main->is_read_invalid) {
  do_versions_after_linking_280(fd, main);
}
if (!main->is_read_invalid) {
 do_versions_after_linking_290(fd, main);
}
...

And to do this for every do_versions, even if not needed now I think there is some chance of this getting overlooked in the future.

I suggest to write it like this instead of returning early and having `main->is_locked_for_linking = false` duplicated. ``` if (!main->is_read_invalid) { do_versions_after_linking_280(fd, main); } if (!main->is_read_invalid) { do_versions_after_linking_290(fd, main); } ... ``` And to do this for every `do_versions`, even if not needed now I think there is some chance of this getting overlooked in the future.

mont29 marked this conversation as resolved

Bastien Montagne force-pushed F-abort-file-reading from bfdcd8de3a to 3dc8035e1f

2023-02-28 18:29:58 +01:00

Compare

Brecht Van Lommel approved these changes 2023-02-28 18:45:00 +01:00

Bastien Montagne commented

2023-03-01 11:52:38 +01:00

Committed as 3e5ce23c99, b3f42d8e98 and 1ed783eed3.

Committed as 3e5ce23c99, b3f42d8e98 and 1ed783eed3.

Bastien Montagne closed this pull request

2023-03-01 11:52:39 +01:00

Bastien Montagne deleted branch F-abort-file-reading