[PATCH 3/3] linux: Optimize realpath stack usage

[Paul Eggert]

On 8/10/20 1:48 PM, Adhemerval Zanella wrote:

> Regarding syscalls usage, for a sucessful path without symlinks it
> trades 2 syscalls (getcwd/lstat) for 5 (openat, readlink, fstat, stat,
>   and close). 

Thanks for looking into this.

There's no need for the new __resolve_path function to call __lstat64. It does 
that only to determine whether the file is a symlink or a directory, and it can 
do the former with __readlink (which it's gonna do already) and the latter by 
going into the next iteration and seeing whether it gets NOTDIR. Avoiding 
__lstat64 means we don't have to worry about irrelevant EOVERFLOW hassles.

In the Linux __realpath_system, the code should be prepared for the /proc 
syscall to fail because /proc is not mounted. It can fall back on __resolve_path 
in that case. The code should be prepared for the fstat64 to fail due to 
EOVERFLOW. And I'm not quite getting why the code is comparing fstat64's result 
to stat64's result -- perhaps explain this in a comment and how EOVERFLOW is 
dealt with?