Defining PATH_MAX for a filesystem?

2019-03-12 03:40发布

I'm presently writing a filesystem. The statvfs (and even the statfs) structs contain a field specifying the maximum length of a name in that path. As PATH_MAX is defined in the pathconf manpage (getconf), this means it is defined on a per-directory basis (and thus, determined by the underlying filesystem). How does one specify this value?

6条回答
虎瘦雄心在
2楼-- · 2019-03-12 04:23

PATH_MAX is a system wide setting and is usually defined in pathmax.h as:

define PATH_MAX (pathconf ("/", _PC_PATH_MAX) < 1 ? 1024 \
            : pathconf ("/", _PC_PATH_MAX))
查看更多
Viruses.
3楼-- · 2019-03-12 04:24

PATH_MAX mostly behaves as a property of the file system function call interface, so I don't think it makes much sense to have it vary across directories.

For example, renaming or moving a directory with large directory trees in it may make the longest absolute pathname longer and it would be complicated and inefficient to limit that.

Instead, PATH_MAX serves to allow the kernel to copy passed pathnames to temporary unpaged memory, which can then be processed without needing to allow for a page fault at each access. Allocating huge amounts of such memory may block most other things the kernel is doing or even cause kernel panics.

查看更多
干净又极端
4楼-- · 2019-03-12 04:32

POSIX allows _PC_PATH_MAX to vary based on the current directory, but that doesn't mean that systems which don't vary it aren't compliant.

The real reason for PATH_MAX existing is that the kernel copies the pathname into kernelspace before doing any actual work with it.

Your assertion that there is a PATH_MAX-related field in statvfs is just wrong. That's related to NAME_MAX, which is a different thing.

查看更多
老娘就宠你
5楼-- · 2019-03-12 04:33

I do not enough about other OSes but imho this is a system-wide setting in at least FreeBSD 5.2.1

PATH_MAX is found in #62 sys/syslimits.h


Because static int ufs_pathconf() which returns the PATHCONF information for UFS FS, uses this variable in the manner you specified.

/*
 * Return POSIX pathconf information applicable to ufs filesystems.
 */
int
ufs_pathconf(ap)
    struct vop_pathconf_args /* {
        struct vnode *a_vp;
        int a_name;
        int *a_retval;
    } */ *ap;
{

    switch (ap->a_name) {
    .
    .
    .
    .
    case _PC_PATH_MAX:
        *ap->a_retval = PATH_MAX;
        return (0);
    .
    .
    .
    .

    default:
        return (EINVAL);
    }
    /* NOTREACHED */
}
查看更多
冷血范
6楼-- · 2019-03-12 04:39

On Linux, glibc's implementation of pathconf returns a compile-time constant value of PATH_MAX so there is no runtime magic FUSE or anyone else can perform to adjust it. (See sysdeps/unix/sysv/linux/pathconf.c which falls through to sysdeps/posix/pathconf.c.) The answer to your question "How do I specify my filesystem's PATH_MAX?" is "You can't. glibc doesn't let you and FUSE is just the messenger."

The end result is a sticky situation. Here's a blog post that discusses the code that does and does not care about PATH_MAX. Software that relies on paths no longer than PATH_MAX was broken long ago by other filesystems so it's safe for you to ignore PATH_MAX.

On MacOS X (and probably other BSDs): The implementation of pathconf is entirely in the kernel and can be swapped out per filesystem. OSXFUSE includes a NOOP version of pathconf which should return the usual compile-time constants. However, in my tests it seems to be catching another NOOP function along the way which returns an ENXIO and I can't get pathconf to work.

Bonus: for NAME_MAX, implement statfs and set f_namemax.

查看更多
再贱就再见
7楼-- · 2019-03-12 04:43

Since this question is tagged "FUSE" ...

I just ran into this issue while working on a FUSE filesystem. I wrote an e-mail to the FUSE developers, seeking clarification. Reply from the current libfuse maintainer (January 2018): There is not a way to specify the maximum path length in a FUSE filesystem [driver].

Is there a way for a FUSE filesystem to inform software running on top of it about the correct maximum path length?

Not at the moment, no.

If not, should there be?

Probably yes. Patches welcome :-)

For reference: Full e-mail thread

查看更多
登录 后发表回答