Hi Zack,
On 10/29/21 16:53, Zack Weinberg via Libc-alpha wrote:
On Fri, Oct 29, 2021, at 9:55 AM, Theo de Raadt wrote:
<rsbecker@xxxxxxxxxxxxx> wrote:
On October 29, 2021 7:29 AM, Alejandro Colomar wrote:
On 10/29/21 13:15, Alejandro Colomar wrote:
Hi,
As the manual pages says, SUSv2 marked it as LEGACY, and POSIX doesn't
have it at all. The manual page goes further and says "This function
is obsolete. Do not use it." in its first lines.
...
The community finally had the balls to get rid of gets(3).
getpass(3) shares the same flaw, that the buffer size isn't passed.
This has been an issue in the past
I was about to post exactly the same thing. getpass(3) is not deprecated because there's a better replacement, it's deprecated because it's _unsafe_. The glibc implementation wraps getline(3) and therefore doesn't truncate the passphrase or overflow a fixed-size buffer, no matter how long the input is, but portable code cannot rely on that. And come to think of it, using getline(3) means that prefixes of the passphrase may be left lying around in malloc's free lists.
(getpass also cannot be made thread safe, due to recycling of a static buffer, but a program in which multiple threads are racing to prompt the user for passwords would be a UX disaster anyway, so I don't think that's a critical flaw the way it is for e.g. strtok(3).)
The Linux manpage project's documentation is, as I understand it, for Linux with glibc _first_, but not _only_; it should not describe this function as not-deprecated just because glibc has patched its worst problems and doesn't offer any better API.
readpassphrase(3) has a few too many features/extensions for my taste, but
at least it is harder to abuse.
I am inclined to agree that readpassphrase has too many knobs, and I can't think of any legitimate present-day use for several of them, which is not a good property for an API handling security-critical data. Also, it relies on the caller to size the buffer for the passphrase, and therefore risks truncating people's passphrases.
With my libxcrypt hat on I've thought a bit about replacements for getpass. The conclusion I came to is that the easy changes are all putting lipstick on a pig, and if I was going to work on this at all I was going to design a privilege-separated authentication service that could be asked to take over a tty, read a passphrase, check it, and return just success or failure to the caller. Neither the passphrase itself, nor any strings derived from it, would ever be in the caller's address space. But this is obviously well out of scope for the C library.
zw
I happen to be working on replacing getpass(3) in shadow-utils. As
there is no replacement in glibc, I'm making the code depend on libbsd
on GNU systems.
I developed a function similar to getpass(3), but which allocates a
buffer (similar to asprintf(3)). I only allocate once, and bail out if
the password exceeds PASS_MAX, so no leaks in allocated memory (modulo
bugs that I may have not noticed).
I also enforce both clearing and freeing the memory, by requiring a
specific clean-up function.
The prototypes for the function and the clean-up are:
```
void erase_pass(char *p);
[[gnu::malloc(erase_pass)]] char *shdw_getpass(const char *prompt);
```
And the implementation is:
```
#include "prototypes.h"
#include <limits.h>
#include <readpassphrase.h>
#include <stdio.h>
#include <stdlib.h>
#if !defined(PASS_MAX)
#define PASS_MAX BUFSIZ
#endif
char *
agetpass(const char *prompt)
{
char *p;
size_t len;
p = malloc(PASS_MAX);
if (p == NULL)
return NULL;
if (readpassphrase(prompt, p, PASS_MAX, 0) == NULL)
goto fail;
len = strlen(p);
if (len == 0)
return p;
if (p[len - 1] != '\n')
goto truncated;
p[len - 1] = '\0';
return p;
truncated:
memzero(p, PASS_MAX);
fail:
free(p);
return NULL;
}
void
erase_pass(char *p)
{
if (p != NULL)
memzero(p, PASS_MAX);
free(p);
}
```
Would you mind implementing readpassphrase(3) in glibc so that it's
easier to use something safe and portable without resorting to
compatibility libraries? Also, I'd like some review of this function,
if you think the API could be improved. Maybe agetpass() would be a
simple almost-drop-in replacement for getpass(3), so if you like it for
glibc, let's discuss it.
I chose a predefined buffer size to not have to pass a buffer size all
the time, which could be error-prone. I also allocated the buffer
internally, to make it easier to replace getpass(3). It may be
desirable to use existing buffers, and pass them through a pointer, but
for shadow-utils, it was simpler to keep the getpass(3) API.
I don't know what was the practice with PASS_MAX regarding the NUL byte,
but to avoid creating a buffer of a power of two plus one, I decided
that the NUL byte would be within PASS_MAX. Another solution would be
to declare PASS_MAX to be something like BUFSIZ-1, and then use
PASS_MAX+1, but I opted for simplicity.
What are your thoughts?
Cheers,
Alex
