fuse_file_info struct fixes needed

[ThomasWaldmann]

The new struct layout test from #38 has some TODO in it - remove the commented fuse_file_info member offset prints and run the test to see it.

I couldn't make it work, so this is for somebody else to solve.

The struct test output indicates that these changes are needed:

u2204 x86_64+arm64 fuse 3.10 fuse_file_info: fh offset 16 -> 24
u2204 x86_64+arm64 fuse 3.10 fuse_file_info: lock_owner offset 24 -> 32
u2404 x86_64+arm64 fuse 3.14 fuse_file_info: fh offset 16 -> 24
u2404 x86_64+arm64 fuse 3.14 fuse_file_info: lock_owner offset 24 -> 32

netbsd x86_64 fuse 2.9 fuse_file_info: fh offset 24 -> 16
netbsd x86_64 fuse 2.9 fuse_file_info: lock_owner offset 32 -> 24

[mxmlnkn]

I can reproduce this locally. You did add all possible libfuse include paths, which led to the libfuse 2 headers being used even when the test is for libfuse 3.

The issue on NetBSD is more frustrating. It reports a version of 2.9 and the struct seems to look like this:

struct fuse_file_info {
 int32_t flags;
 uint32_t fh_old;
 int32_t writepage:1;
 uint32_t direct_io:1;
 uint32_t keep_cache:1;
 uint32_t flush:1;
 uint32_t nonseekable:1;
 uint32_t flock_release:1;
 uint32_t cache_readdir:1;
 uint32_t padding:26;
 uint64_t fh;
 uint64_t lock_owner;
 uint32_t poll_events;
};

This is some weird chimera interface between libFUSE 3.x, which changed int writepage (in 2.9) to unsigned int writepage : 1; and libFUSE 2.x because it still has that unsigned long fh_old;, which was removed in libFUSE 3.x ...

[mxmlnkn]

Did I get banned from Github or is it down? I was getting 503 error for the test file downloads from other repos and now I am getting 503 errors during the VM setup...

Edit: Seems down to some extent.

[ThomasWaldmann]

Oh, crap, I didn't consider that it would pick the wrong header. Thanks for finding that!

Guess NetBSD has to be considered separately due to its librefuse fork (IIRC).

[mxmlnkn]

The FUSE version reporting on NetBSD seems to be broken... It reports 2.9 in the CI even though NetBSD 10.1 should print 3.10 according to the source.

Furthermore, the writepage member changes from a simple int32 in NetBSD 9 to a bit field in NetBSD 10... I need to have the correct version to determine the correct struct layout, but I can't get it...

[mxmlnkn]

After trying to fix the structs and all the struct tests passing, I suddenly get this issue: tests/test_examples.py::test_read_write_file_system[cli0] perfused: perfuse_node_inactive: perfuse_node_fsync failed error = 22: Resource temporarily unavailable ... I think NetBSD is majorly non-working. I don't know what happens in the current master version, but it is probably not correct. It is a mystery to me why the example test run through, but after fixing the structs, they don't run anymore. The fuse_operations struct is also very vaguely defined ad-hoc in the header. I doubt that it works correctly for all compiled FUSE_USE_VERSION vs. included version ... the ABI seems ill-defined.

I see that you are also giving this a try. I am giving up for now. I have put my struct layout testing changes and my try for fixing the structs on the master branch guarded with a if _system == 'NetBSD_False'.