On Tue, Apr 16, 2024 at 4:10 AM Jody Bruchon <[email protected]> wrote:
>
> This revision reduces code size in concat_path_file_fast() by 19 bytes. It
> also applies the new get_d_namlen() optimization to a readdir() loop in
> runit/svlogd.c.
>
> If there are any changes that need to be made for inclusion then please let
> me know as soon as possible. I'd like to finish this project up. If the
> old/slow and new/fast code should be chosen with a compile-time config option
> then I'm happy to do that as well.
+#elif defined _DIRENT_HAVE_D_RECLEN
+ const size_t base = (sizeof(struct dirent) - sizeof(((struct
dirent *)0)->d_name)) - offsetof(struct dirent, d_name) - 1;
+ size_t skip;
+
+ skip = dirent->d_reclen - (sizeof(struct dirent) -
sizeof(((struct dirent *)0)->d_name));
+ if (skip > 0) skip -= base;
+ return skip + strlen(dirent->d_name + skip);
What is it doing? (Probably needs a comment).
Shouldn't this be just
+ const size_t prefix_sz = offsetof(struct dirent, d_name);
+ return de->d_reclen - prefix_sz;
?
+ const char *filename = dirp->d_name;
...
+ while (*filename == '/') {
+ filename++;
+ namelen--;
+ }
d_name's never contain slashes. It's not allowed in unix filenames.
+ buf = (char *)malloc(end_offset + 1);
+ if (!buf) return NULL;
Use xmalloc.
+ if (dirp->d_name && DOT_OR_DOTDOT(dirp->d_name))
+ return NULL;
dirp->d_name is never NULL, don't check for that.
But the biggest problem is this: I don't see much improvement
in the time it takes to copy data.
Here I copy a Linux kernel tree, 1450M mbytes, 89679 files, on tmpfs
("ramdisk"):
$ time ./busybox_old cp -a linux linux1; rm -rf linux1; time ./busybox
cp -a linux linux1; rm -rf linux1
real 0m10.464s
user 0m1.067s <<< almost the same
sys 0m8.558s
real 0m11.070s
user 0m0.959s <<< almost the same
sys 0m9.204s
and the time, even on tmpfs, is heavily dominated by actual copying,
not constructing filenames.
_______________________________________________
busybox mailing list
[email protected]
https://lists.busybox.net/mailman/listinfo/busybox
