On Jan 20 17:10, Ben Wijen wrote:
> Not having to query file information improves unlink speed.
> ---
> winsup/cygwin/syscalls.cc | 78 ++++++++++++++++++++++++++-------------
> 1 file changed, 52 insertions(+), 26 deletions(-)
>
> diff --git a/winsup/cygwin/syscalls.cc b/winsup/cygwin/syscalls.cc
> index ab0c4c2d6..b5ab6ac5e 100644
> --- a/winsup/cygwin/syscalls.cc
> +++ b/winsup/cygwin/syscalls.cc
> @@ -1272,6 +1272,28 @@ _unlink_ntpc_ (path_conv& pc, bool shareable)
> return status;
> }
>
> +NTSTATUS
> +unlink_nt (const char *ourname, ULONG eflags)
> +{
> + uint32_t opt = PC_SYM_NOFOLLOW | PC_SKIP_SYM_CHECK | PC_SKIP_FS_CHECK;
> + if (!(eflags & FILE_NON_DIRECTORY_FILE))
> + opt &= ~PC_SKIP_FS_CHECK;
> +
> + path_conv pc (ourname, opt, NULL);
> + if (pc.error || pc.isspecial ())
> + return STATUS_CANNOT_DELETE;
> +
> + OBJECT_ATTRIBUTES attr;
> + PUNICODE_STRING ntpath = pc.get_nt_native_path ();
> + InitializeObjectAttributes(&attr, ntpath, 0, NULL, NULL);
> + NTSTATUS status = _unlink_nt (&attr, eflags);
Sorry again, but I don't see the advantage of not using the intelligence
already collected in path_conv by neglecting to call the unlink
variation not using them. It's also unclear to me, why the new code
doesn't try_to_bin right away, rather than stomping ahead and falling
back to the current code for each such file. In an rm -rf on a file
hirarchy used by other users, you could end up with a much slower
operation.
Wouldn't it make more sense to streamline the existing _unlink_nt? I
don't claim it's the most streamlined way to delete files, but at least
it has documented workarounds for problems encountered on the way,
already.
Corinna
