Re: [PATCH] smb: Work around Clang __bdos() type confusion

From: Ard Biesheuvel
Date: Thu Jan 25 2024 - 13:32:06 EST


On Thu, 25 Jan 2024 at 19:00, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>
> On Thu, Jan 25, 2024 at 01:19:19PM +0100, Ard Biesheuvel wrote:
> > On Wed, 24 Jan 2024 at 00:47, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> > >
> > > Recent versions of Clang gets confused about the possible size of the
> > > "user" allocation, and CONFIG_FORTIFY_SOURCE ends up emitting a
> > > warning[1]:
> > >
> > > repro.c:126:4: warning: call to '__write_overflow_field' declared with 'warning' attribute: detected write beyond size of field (1st parameter); maybe use struct_group()? [-Wattribute-warning]
> > > 126 | __write_overflow_field(p_size_field, size);
> > > | ^
> > >
> > > for this memset():
> > >
> > > int len;
> > > __le16 *user;
> > > ...
> > > len = ses->user_name ? strlen(ses->user_name) : 0;
> > > user = kmalloc(2 + (len * 2), GFP_KERNEL);
> > > ...
> > > if (len) {
> > > ...
> > > } else {
> > > memset(user, '\0', 2);
> > > }
> > >
> > > While Clang works on this bug[2], switch to using a direct assignment,
> > > which avoids memset() entirely which both simplifies the code and silences
> > > the false positive warning. (Making "len" size_t also silences the
> > > warning, but the direct assignment seems better.)
> > >
> > > Reported-by: Nathan Chancellor <nathan@xxxxxxxxxx>
> > > Closes: https://github.com/ClangBuiltLinux/linux/issues/1966 [1]
> > > Link: https://github.com/llvm/llvm-project/issues/77813 [2]
> > > Cc: Steve French <sfrench@xxxxxxxxx>
> > > Cc: Paulo Alcantara <pc@xxxxxxxxxxxxx>
> > > Cc: Ronnie Sahlberg <ronniesahlberg@xxxxxxxxx>
> > > Cc: Shyam Prasad N <sprasad@xxxxxxxxxxxxx>
> > > Cc: Tom Talpey <tom@xxxxxxxxxx>
> > > Cc: linux-cifs@xxxxxxxxxxxxxxx
> > > Cc: llvm@xxxxxxxxxxxxxxx
> > > Signed-off-by: Kees Cook <keescook@xxxxxxxxxxxx>
> > > ---
> > > fs/smb/client/cifsencrypt.c | 2 +-
> > > 1 file changed, 1 insertion(+), 1 deletion(-)
> > >
> > > diff --git a/fs/smb/client/cifsencrypt.c b/fs/smb/client/cifsencrypt.c
> > > index ef4c2e3c9fa6..6322f0f68a17 100644
> > > --- a/fs/smb/client/cifsencrypt.c
> > > +++ b/fs/smb/client/cifsencrypt.c
> > > @@ -572,7 +572,7 @@ static int calc_ntlmv2_hash(struct cifs_ses *ses, char *ntlmv2_hash,
> > > len = cifs_strtoUTF16(user, ses->user_name, len, nls_cp);
> > > UniStrupr(user);
> > > } else {
> > > - memset(user, '\0', 2);
> > > + *(u16 *)user = 0;
> >
> > Is 'user' guaranteed to be 16-bit aligned?
>
> It's the first two bytes of a kmalloced address range, which I'm nearly
> certain will be sanely aligned, as those allocs are commonly used for
> holding structs, etc.
>

Ah yes, this kmalloc() was carefully hidden in the commit log :-)