Commit b8002058 strengthened openat()'s /proc detection by calling
realpath(3) on the given path, which allows various paths and symlinks
that points to the /proc file system to be intercepted correctly.

Using realpath(3), though, has a side effect that it reads the symlinks
along the way, and thus changes their atime. The results in the
following code snippet already get ~now instead of the real atime:

  int fd = open("/path/to/a/symlink", O_PATH | O_NOFOLLOW);
  struct stat st;
  fstat(fd, st);
  return st.st_atime;

This change opens a path that doesn't appear to be part of /proc
directly and checks the destination of /proc/self/fd/n to determine if
it actually refers to a file in /proc.

Neither this nor the existing code works with symlinks or indirect paths
(e.g.  /tmp/../proc/self/exe) that points to /proc/self/exe because it
is itself a symlink, and both realpath(3) and /proc/self/fd/n will
resolve into the location of QEMU.

Resolves: https://gitlab.com/qemu-project/qemu/-/issues/2004
Signed-off-by: Shu-Chun Weng <scw@google.com>
---
 linux-user/syscall.c | 47 +++++++++++++++++++++++++++++++++++---------
 1 file changed, 38 insertions(+), 9 deletions(-)

diff --git a/linux-user/syscall.c b/linux-user/syscall.c
index e384e14248..7c3772301f 100644
--- a/linux-user/syscall.c
+++ b/linux-user/syscall.c
@@ -8308,8 +8308,7 @@ static int open_net_route(CPUArchState *cpu_env, int fd)
 int do_guest_openat(CPUArchState *cpu_env, int dirfd, const char *fname,
                     int flags, mode_t mode, bool safe)
 {
-    g_autofree char *proc_name = NULL;
-    const char *pathname;
+    g_autofree char *pathname = NULL;
     struct fake_open {
         const char *filename;
         int (*fill)(CPUArchState *cpu_env, int fd);
@@ -8334,12 +8333,42 @@ int do_guest_openat(CPUArchState *cpu_env, int dirfd, const char *fname,
         { NULL, NULL, NULL }
     };
 
-    /* if this is a file from /proc/ filesystem, expand full name */
-    proc_name = realpath(fname, NULL);
-    if (proc_name && strncmp(proc_name, "/proc/", 6) == 0) {
-        pathname = proc_name;
+    if (strncmp(fname, "/proc/", 6) == 0) {
+        pathname = g_strdup(fname);
     } else {
-        pathname = fname;
+        g_autofree char *proc_name = NULL;
+        struct stat proc_stat;
+        int fd;
+
+        if (safe) {
+            fd = safe_openat(dirfd, path(fname), flags, mode);
+        } else {
+            fd = openat(dirfd, path(fname), flags, mode);
+        }
+        if (fd < 0) {
+            return fd;
+        }
+
+        /*
+         * Try to get the real path of the file we just opened. We avoid calling
+         * `realpath(3)` because it calls `readlink(2)` on symlinks which
+         * changes their atime. Note that since `/proc/self/exe` is a symlink,
+         * `pathname` will never resolve to it (neither will `realpath(3)`).
+         * That's why we check `fname` against the "/proc/" prefix first.
+         */
+        proc_name = g_strdup_printf("/proc/self/fd/%d", fd);
+        if (lstat(proc_name, &proc_stat) < 0 || !S_ISLNK(proc_stat.st_mode)) {
+            /* No procfs or something weird. Not going to dig further. */
+            return fd;
+        }
+        pathname = g_new(char, proc_stat.st_size + 1);
+        readlink(proc_name, pathname, proc_stat.st_size + 1);
+
+        /* if this is not a file from /proc/ filesystem, the fd is good as-is */
+        if (strncmp(pathname, "/proc/", 6) != 0) {
+            return fd;
+        }
+        close(fd);
     }
 
     if (is_proc_myself(pathname, "exe")) {
@@ -8390,9 +8419,9 @@ int do_guest_openat(CPUArchState *cpu_env, int dirfd, const char *fname,
     }
 
     if (safe) {
-        return safe_openat(dirfd, path(pathname), flags, mode);
+        return safe_openat(dirfd, pathname, flags, mode);
     } else {
-        return openat(dirfd, path(pathname), flags, mode);
+        return openat(dirfd, pathname, flags, mode);
     }
 }

On 12/8/23 23:42, Shu-Chun Weng wrote:
> Commit b8002058 strengthened openat()'s /proc detection by calling
> realpath(3) on the given path, which allows various paths and symlinks
> that points to the /proc file system to be intercepted correctly.
>
> Using realpath(3), though, has a side effect that it reads the symlinks
> along the way, and thus changes their atime. The results in the
> following code snippet already get ~now instead of the real atime:
>
>    int fd = open("/path/to/a/symlink", O_PATH | O_NOFOLLOW);
>    struct stat st;
>    fstat(fd, st);
>    return st.st_atime;
>
> This change opens a path that doesn't appear to be part of /proc
> directly and checks the destination of /proc/self/fd/n to determine if
> it actually refers to a file in /proc.
>
> Neither this nor the existing code works with symlinks or indirect paths
> (e.g.  /tmp/../proc/self/exe) that points to /proc/self/exe because it
> is itself a symlink, and both realpath(3) and /proc/self/fd/n will
> resolve into the location of QEMU.
>
> Resolves: https://gitlab.com/qemu-project/qemu/-/issues/2004
> Signed-off-by: Shu-Chun Weng <scw@google.com>
> ---
>   linux-user/syscall.c | 47 +++++++++++++++++++++++++++++++++++---------
>   1 file changed, 38 insertions(+), 9 deletions(-)
>
> diff --git a/linux-user/syscall.c b/linux-user/syscall.c
> index e384e14248..7c3772301f 100644
> --- a/linux-user/syscall.c
> +++ b/linux-user/syscall.c
> @@ -8308,8 +8308,7 @@ static int open_net_route(CPUArchState *cpu_env, int fd)
>   int do_guest_openat(CPUArchState *cpu_env, int dirfd, const char *fname,
>                       int flags, mode_t mode, bool safe)
>   {
> -    g_autofree char *proc_name = NULL;
> -    const char *pathname;
> +    g_autofree char *pathname = NULL;
>       struct fake_open {
>           const char *filename;
>           int (*fill)(CPUArchState *cpu_env, int fd);
> @@ -8334,12 +8333,42 @@ int do_guest_openat(CPUArchState *cpu_env, int dirfd, const char *fname,
>           { NULL, NULL, NULL }
>       };
>
> -    /* if this is a file from /proc/ filesystem, expand full name */
> -    proc_name = realpath(fname, NULL);
> -    if (proc_name && strncmp(proc_name, "/proc/", 6) == 0) {
> -        pathname = proc_name;
> +    if (strncmp(fname, "/proc/", 6) == 0) {
> +        pathname = g_strdup(fname);
>       } else {
> -        pathname = fname;
> +        g_autofree char *proc_name = NULL;
> +        struct stat proc_stat;
> +        int fd;
> +
> +        if (safe) {
> +            fd = safe_openat(dirfd, path(fname), flags, mode);
> +        } else {
> +            fd = openat(dirfd, path(fname), flags, mode);
> +        }
> +        if (fd < 0) {
> +            return fd;
> +        }
> +
> +        /*
> +         * Try to get the real path of the file we just opened. We avoid calling
> +         * `realpath(3)` because it calls `readlink(2)` on symlinks which
> +         * changes their atime. Note that since `/proc/self/exe` is a symlink,
> +         * `pathname` will never resolve to it (neither will `realpath(3)`).
> +         * That's why we check `fname` against the "/proc/" prefix first.
> +         */
> +        proc_name = g_strdup_printf("/proc/self/fd/%d", fd);
> +        if (lstat(proc_name, &proc_stat) < 0 || !S_ISLNK(proc_stat.st_mode)) {
> +            /* No procfs or something weird. Not going to dig further. */
> +            return fd;
> +        }
> +        pathname = g_new(char, proc_stat.st_size + 1);
> +        readlink(proc_name, pathname, proc_stat.st_size + 1);

this gives a build error for me:
./qemu/linux-user/syscall.c:8365:9: error: ignoring return value of ‘readlink’ declared with attribute ‘warn_unused_result’ [-Werror=unused-result]

Other than that, this patch indeed fixes the issue #2004

Reviewed-by: Helge Deller <deller@gmx.de>


Helge

> +
> +        /* if this is not a file from /proc/ filesystem, the fd is good as-is */
> +        if (strncmp(pathname, "/proc/", 6) != 0) {
> +            return fd;
> +        }
> +        close(fd);
>       }
>
>       if (is_proc_myself(pathname, "exe")) {
> @@ -8390,9 +8419,9 @@ int do_guest_openat(CPUArchState *cpu_env, int dirfd, const char *fname,
>       }
>
>       if (safe) {
> -        return safe_openat(dirfd, path(pathname), flags, mode);
> +        return safe_openat(dirfd, pathname, flags, mode);
>       } else {
> -        return openat(dirfd, path(pathname), flags, mode);
> +        return openat(dirfd, pathname, flags, mode);
>       }
>   }
>
>

Yup, reproduced on a system with glibc built with fortified source. Sending
out version 4.

On Thu, Dec 28, 2023 at 7:49 AM Helge Deller <deller@gmx.de> wrote:

> On 12/8/23 23:42, Shu-Chun Weng wrote:
> > Commit b8002058 strengthened openat()'s /proc detection by calling
> > realpath(3) on the given path, which allows various paths and symlinks
> > that points to the /proc file system to be intercepted correctly.
> >
> > Using realpath(3), though, has a side effect that it reads the symlinks
> > along the way, and thus changes their atime. The results in the
> > following code snippet already get ~now instead of the real atime:
> >
> >    int fd = open("/path/to/a/symlink", O_PATH | O_NOFOLLOW);
> >    struct stat st;
> >    fstat(fd, st);
> >    return st.st_atime;
> >
> > This change opens a path that doesn't appear to be part of /proc
> > directly and checks the destination of /proc/self/fd/n to determine if
> > it actually refers to a file in /proc.
> >
> > Neither this nor the existing code works with symlinks or indirect paths
> > (e.g.  /tmp/../proc/self/exe) that points to /proc/self/exe because it
> > is itself a symlink, and both realpath(3) and /proc/self/fd/n will
> > resolve into the location of QEMU.
> >
> > Resolves: https://gitlab.com/qemu-project/qemu/-/issues/2004
> > Signed-off-by: Shu-Chun Weng <scw@google.com>
> > ---
> >   linux-user/syscall.c | 47 +++++++++++++++++++++++++++++++++++---------
> >   1 file changed, 38 insertions(+), 9 deletions(-)
> >
> > diff --git a/linux-user/syscall.c b/linux-user/syscall.c
> > index e384e14248..7c3772301f 100644
> > --- a/linux-user/syscall.c
> > +++ b/linux-user/syscall.c
> > @@ -8308,8 +8308,7 @@ static int open_net_route(CPUArchState *cpu_env,
> int fd)
> >   int do_guest_openat(CPUArchState *cpu_env, int dirfd, const char
> *fname,
> >                       int flags, mode_t mode, bool safe)
> >   {
> > -    g_autofree char *proc_name = NULL;
> > -    const char *pathname;
> > +    g_autofree char *pathname = NULL;
> >       struct fake_open {
> >           const char *filename;
> >           int (*fill)(CPUArchState *cpu_env, int fd);
> > @@ -8334,12 +8333,42 @@ int do_guest_openat(CPUArchState *cpu_env, int
> dirfd, const char *fname,
> >           { NULL, NULL, NULL }
> >       };
> >
> > -    /* if this is a file from /proc/ filesystem, expand full name */
> > -    proc_name = realpath(fname, NULL);
> > -    if (proc_name && strncmp(proc_name, "/proc/", 6) == 0) {
> > -        pathname = proc_name;
> > +    if (strncmp(fname, "/proc/", 6) == 0) {
> > +        pathname = g_strdup(fname);
> >       } else {
> > -        pathname = fname;
> > +        g_autofree char *proc_name = NULL;
> > +        struct stat proc_stat;
> > +        int fd;
> > +
> > +        if (safe) {
> > +            fd = safe_openat(dirfd, path(fname), flags, mode);
> > +        } else {
> > +            fd = openat(dirfd, path(fname), flags, mode);
> > +        }
> > +        if (fd < 0) {
> > +            return fd;
> > +        }
> > +
> > +        /*
> > +         * Try to get the real path of the file we just opened. We
> avoid calling
> > +         * `realpath(3)` because it calls `readlink(2)` on symlinks
> which
> > +         * changes their atime. Note that since `/proc/self/exe` is a
> symlink,
> > +         * `pathname` will never resolve to it (neither will
> `realpath(3)`).
> > +         * That's why we check `fname` against the "/proc/" prefix
> first.
> > +         */
> > +        proc_name = g_strdup_printf("/proc/self/fd/%d", fd);
> > +        if (lstat(proc_name, &proc_stat) < 0 ||
> !S_ISLNK(proc_stat.st_mode)) {
> > +            /* No procfs or something weird. Not going to dig further.
> */
> > +            return fd;
> > +        }
> > +        pathname = g_new(char, proc_stat.st_size + 1);
> > +        readlink(proc_name, pathname, proc_stat.st_size + 1);
>
> this gives a build error for me:
> ./qemu/linux-user/syscall.c:8365:9: error: ignoring return value of
> ‘readlink’ declared with attribute ‘warn_unused_result’
> [-Werror=unused-result]
>
> Other than that, this patch indeed fixes the issue #2004
>
> Reviewed-by: Helge Deller <deller@gmx.de>
>
>
> Helge
>
> > +
> > +        /* if this is not a file from /proc/ filesystem, the fd is good
> as-is */
> > +        if (strncmp(pathname, "/proc/", 6) != 0) {
> > +            return fd;
> > +        }
> > +        close(fd);
> >       }
> >
> >       if (is_proc_myself(pathname, "exe")) {
> > @@ -8390,9 +8419,9 @@ int do_guest_openat(CPUArchState *cpu_env, int
> dirfd, const char *fname,
> >       }
> >
> >       if (safe) {
> > -        return safe_openat(dirfd, path(pathname), flags, mode);
> > +        return safe_openat(dirfd, pathname, flags, mode);
> >       } else {
> > -        return openat(dirfd, path(pathname), flags, mode);
> > +        return openat(dirfd, pathname, flags, mode);
> >       }
> >   }
> >
> >
>
>

ping~

On Fri, Dec 8, 2023 at 2:42 PM Shu-Chun Weng <scw@google.com> wrote:

> Commit b8002058 strengthened openat()'s /proc detection by calling
> realpath(3) on the given path, which allows various paths and symlinks
> that points to the /proc file system to be intercepted correctly.
>
> Using realpath(3), though, has a side effect that it reads the symlinks
> along the way, and thus changes their atime. The results in the
> following code snippet already get ~now instead of the real atime:
>
>   int fd = open("/path/to/a/symlink", O_PATH | O_NOFOLLOW);
>   struct stat st;
>   fstat(fd, st);
>   return st.st_atime;
>
> This change opens a path that doesn't appear to be part of /proc
> directly and checks the destination of /proc/self/fd/n to determine if
> it actually refers to a file in /proc.
>
> Neither this nor the existing code works with symlinks or indirect paths
> (e.g.  /tmp/../proc/self/exe) that points to /proc/self/exe because it
> is itself a symlink, and both realpath(3) and /proc/self/fd/n will
> resolve into the location of QEMU.
>
> Resolves: https://gitlab.com/qemu-project/qemu/-/issues/2004
> Signed-off-by: Shu-Chun Weng <scw@google.com>
> ---
>  linux-user/syscall.c | 47 +++++++++++++++++++++++++++++++++++---------
>  1 file changed, 38 insertions(+), 9 deletions(-)
>
> diff --git a/linux-user/syscall.c b/linux-user/syscall.c
> index e384e14248..7c3772301f 100644
> --- a/linux-user/syscall.c
> +++ b/linux-user/syscall.c
> @@ -8308,8 +8308,7 @@ static int open_net_route(CPUArchState *cpu_env, int
> fd)
>  int do_guest_openat(CPUArchState *cpu_env, int dirfd, const char *fname,
>                      int flags, mode_t mode, bool safe)
>  {
> -    g_autofree char *proc_name = NULL;
> -    const char *pathname;
> +    g_autofree char *pathname = NULL;
>      struct fake_open {
>          const char *filename;
>          int (*fill)(CPUArchState *cpu_env, int fd);
> @@ -8334,12 +8333,42 @@ int do_guest_openat(CPUArchState *cpu_env, int
> dirfd, const char *fname,
>          { NULL, NULL, NULL }
>      };
>
> -    /* if this is a file from /proc/ filesystem, expand full name */
> -    proc_name = realpath(fname, NULL);
> -    if (proc_name && strncmp(proc_name, "/proc/", 6) == 0) {
> -        pathname = proc_name;
> +    if (strncmp(fname, "/proc/", 6) == 0) {
> +        pathname = g_strdup(fname);
>      } else {
> -        pathname = fname;
> +        g_autofree char *proc_name = NULL;
> +        struct stat proc_stat;
> +        int fd;
> +
> +        if (safe) {
> +            fd = safe_openat(dirfd, path(fname), flags, mode);
> +        } else {
> +            fd = openat(dirfd, path(fname), flags, mode);
> +        }
> +        if (fd < 0) {
> +            return fd;
> +        }
> +
> +        /*
> +         * Try to get the real path of the file we just opened. We avoid
> calling
> +         * `realpath(3)` because it calls `readlink(2)` on symlinks which
> +         * changes their atime. Note that since `/proc/self/exe` is a
> symlink,
> +         * `pathname` will never resolve to it (neither will
> `realpath(3)`).
> +         * That's why we check `fname` against the "/proc/" prefix first.
> +         */
> +        proc_name = g_strdup_printf("/proc/self/fd/%d", fd);
> +        if (lstat(proc_name, &proc_stat) < 0 ||
> !S_ISLNK(proc_stat.st_mode)) {
> +            /* No procfs or something weird. Not going to dig further. */
> +            return fd;
> +        }
> +        pathname = g_new(char, proc_stat.st_size + 1);
> +        readlink(proc_name, pathname, proc_stat.st_size + 1);
> +
> +        /* if this is not a file from /proc/ filesystem, the fd is good
> as-is */
> +        if (strncmp(pathname, "/proc/", 6) != 0) {
> +            return fd;
> +        }
> +        close(fd);
>      }
>
>      if (is_proc_myself(pathname, "exe")) {
> @@ -8390,9 +8419,9 @@ int do_guest_openat(CPUArchState *cpu_env, int
> dirfd, const char *fname,
>      }
>
>      if (safe) {
> -        return safe_openat(dirfd, path(pathname), flags, mode);
> +        return safe_openat(dirfd, pathname, flags, mode);
>      } else {
> -        return openat(dirfd, path(pathname), flags, mode);
> +        return openat(dirfd, pathname, flags, mode);
>      }
>  }
>
>