| Message ID | 20231208224207.613445-3-scw@google.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | linux-user: openat() fixes | expand |
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); > } > } > >
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); > > } > > } > > > > > >
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); } }
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(-)