Hi Corinna,
sorry for the blast from the past, but I am renewing my efforts to
upstream Git for Windows' patches that can be upstreamed.
On Thu, 17 Dec 2015, Corinna Vinschen wrote:
> On Dec 17 19:05, Johannes Schindelin wrote:
> > [...]
> > diff --git a/winsup/cygwin/uinfo.cc b/winsup/cygwin/uinfo.cc
> > index c9b3e09..a5d6270 100644
> > --- a/winsup/cygwin/uinfo.cc
> > +++ b/winsup/cygwin/uinfo.cc
> > [...]
> > +static size_t
> > +fetch_env(LPCWSTR key, char *buf, size_t size)
> ^^^
> space
>
> > +{
> > + WCHAR wbuf[32767];
>
> Ok, there are a couple of problems here. First, since this buffer
> is a filename buffer, use NT_MAX_PATH from winsup.h as buffer size.
>
> But then again, please avoid allocating 64K buffers on the stack.
> That's what tmp_pathbuf:w_get () is for.
Excellent. I did it exactly as you suggested.
> > + DWORD max = sizeof wbuf / sizeof *wbuf;
> > + DWORD len = GetEnvironmentVariableW (key, wbuf, max);
>
> This call to GetEnvironmentVariableW looks gratuitous to me. Why don't
> you simply call getenv? It did the entire job already, it avoids the
> requirement for a local buffer, and in case of $HOME it even did the
> Win32->POSIX path conversion. If there's a really good reason for using
> GetEnvironmentVariableW it begs at least for a longish comment.
My only worry is that `getenv("HOME")` might receive a "Cygwin-ified"
version of the value. That is, `getenv("HOME")` might return something
like `/cygdrive/c/Users/corinna` when we expect it to return
`C:\Users\corinna` instead.
I do not think that the current iteration is resilient against that.
This problem might not be a big issue with Cygwin (I don't think it
automatically converts environment variables that look like paths from
Windows to Unix-style), but it will most likely cause issues with MSYS2
(where we do precisely that with environment variables that look like
paths). Meaning: it will probably take some follow-up work to make this
work correctly, even if it is just to verify that things work when `HOME`
is in Unix-style already while calling into the runtime.
> > +
> > + if (!len || len >= max)
> > + return 0;
> > +
> > + len = sys_wcstombs (buf, size, wbuf, len);
> > + return len && len < size ? len : 0;
> > +}
> > +
> > +static char *
> > +fetch_home_env (void)
> > +{
> > + char home[32767];
> > + size_t max = sizeof home / sizeof *home, len;
> > +
> > + if (fetch_env (L"HOME", home, max)
> > + || ((len = fetch_env (L"HOMEDRIVE", home, max))
> > + && fetch_env (L"HOMEPATH", home + len, max - len))
> > + || fetch_env (L"USERPROFILE", home, max))
> > + {
> > + tmp_pathbuf tp;
> > + cygwin_conv_path (CCP_WIN_A_TO_POSIX | CCP_ABSOLUTE,
> > + home, tp.c_get(), NT_MAX_PATH);
> ^^^
> space
> > + return strdup(tp.c_get());
> ^^^ ^^^
> space......s
>
> Whoa, tp.c_get() twice to access the same space? That's a dirty trick
> which may puzzle later readers of the code and heavily depends on
> knowing the internals of tmp_pathbuf. Please use a variable and only
> assign tp.c_get () once.
>
> OTOH, the above's a case for a cygwin_create_path call, rather than
> cygwin_conv_path+strdup. Also, if there's *really* a good reason to use
> GetEnvironmentVariableW, you should collapse sys_wcstombs+cygwin_conv_path+
> strdup into a single cygwin_create_path (CCP_WIN_W_TO_POSIX, ...).
Right, that `cygwin_create_path()` call nicely avoids all the problems of
my original code.
>
> > [...]
> > @@ -1079,6 +1123,7 @@ cygheap_pwdgrp::get_shell (cyg_ldap *pldap, cygpsid
> > &sid, PCWSTR dom,
> > case NSS_SCHEME_FALLBACK:
> > return NULL;
> > case NSS_SCHEME_WINDOWS:
> > + case NSS_SCHEME_ENV:
> > break;
> > case NSS_SCHEME_CYGWIN:
> > if (pldap->fetch_ad_account (sid, false, dnsdomain))
>
> You know that I don't exactly like the "env" idea, but if we implement
> it anyway, wouldn't it make sense to add some kind of $SHELL handling as
> well, for symmetry?
I have decided against that.
The reason: the home directory is a very different thing from the `SHELL`
variable because Windows users _do_ have a home directory even if it is
called differently in Windows speak, while they do not have any POSIX
shell available. There is `COMSPEC`, of course, but it is _not_ a POSIX
shell and cannot be used in place of `SHELL`.
For that reason, I do not believe that we need to do anything about
`SHELL`.
> > [...]
> > @@ -1487,6 +1497,16 @@ of each schema when used with
> > <literal>db_home:</literal>
> > for a detailed description.</listitem>
> > </varlistentry>
> > <varlistentry>
> > + <term><literal>env</literal></term>
> > + <listitem>Derives the home directory of the current user from the
> > + environment variable <literal>HOME</literal> (falling back to
> > + <literal>HOMEDRIVE\HOMEPATH</literal> and
> > + <literal>USERPROFILE</literal>, in that order). This is faster
> > + than the <term><literal>windows</literal></term> schema at the
> > + expense of determining only the current user's home directory
> > + correctly.</listitem>
>
> In both case of the documentation it might make sense to add a few words
> along the lines of "This schema is skipped for any other account",
> wouldn't it?
Yes!
(Belated) thank you very much for your review!
Dscho
