Enhancement request - support file systems with lower PATH_MAX values

[mykaul]

PATH_MAX seems to be hard-coded everywhere. Unfortunately, when using some file systems (Gluster, for example), the real paths are longer than PATH_MAX.
There is no way (that I found, as opposed to NAME_MAX) that one can query the FS for PATH_MAX - it is just assumed to be 4096 everywhere.

I suspect if we can change dirgen_max and perhaps other functions to be able to get a hard-coded, lower PATH_MAX, we'll be able to overcome this limitation, which currently fails many tests.

[s-m-e]

PATH_MAX is directly hard-coded in most operating systems, for instance in the Linux kernel. I have tried to alter it for testing purposes (in pjdfstest and libfuse2, however not in the kernel itself), but it pretty much always causes issues and inconsistencies in tests. I'd be interested in how this can be overcome, but from experience it's a very deep rabbit hole.

[s-m-e]

Relevant discussions:

• fuse mailing list
• libc-help mailing list

[asomers]

It actually isn't hard-coded. The PATH_MAX in limits.h is considered a minimum, but different file systems can support different maximums. You can retrieve the per-file-system maximum at runtime with pathconf(2) or getconf(1). For example:

> getconf PATH_MAX .
1024

I think this would be a very reasonable enhancement to pjdfstest.

[mykaul]

See https://github.com/bminor/glibc/blob/5f72f9800b250410cad3abfeeb09469ef12b2438/sysdeps/posix/pathconf.c#L89 - it does seem to be hard-coded based on the limits.h that was available for glibc compilation?

[asomers]

Oh, wow, I didn't realize that Linux doesn't implement pathconf(2). That's lame. FreeBSD does, and there is currently one file system (smbfs) that returns a non-default value for _PC_PATH_MAX. It looks like OpenBSD, NetBSD, and OSX also implement pathconf(2).

[mykaul]

I think pathconf is implemented in Linux, just not for PATH_MAX which is assumed to be statically defined.
I could not find the implementation on FreeBSD code, would be happy to take a look.

[asomers]

My grep can't find a pathconf syscall in Linux. Here is the FreeBSD man page for the KPI: https://www.freebsd.org/cgi/man.cgi?query=VOP_PATHCONF&apropos=0&sektion=0&manpath=FreeBSD+12.1-RELEASE+and+Ports&arch=default&format=html . And here is smbfs's implementation, which overrides the default PATH_MAX: https://svnweb.freebsd.org/base/head/sys/fs/smbfs/smbfs_vnops.c?view=markup#l913 .

[gaul]

This is also an issue with s3fs which can only support 1024 byte PATH_MAX due to S3 limitations s3fs-fuse/s3fs-fuse#1602.