[PATCH v6 0/1] linux-user: add openat2 support in linux-user

[PATCH v6 0/1] linux-user: add openat2 support in linux-user

[Michael Vogt]

Hi,

This is v6 of the openat2 support in linux-user. Thanks agan for the
excellent feedback from Laurent Vivier on v5.

This version is very close to v5 and only fixes the small details
that I overlooked in my previous patch (sorry for that). This time
I hopefully also fixed my git send email setup.

Thanks again,
 Michael

v5 -> v6
- do not use get_errno(fd) in do_guest_openat()
- do not put declarations in the middle of the code
- do not return early in do_openat2() when we get a faked file

v4 -> v5
- drop "*use_returned_fd" from maybe_do_fake_open() and use return value
  -2 to signal to the caller to continue
- keep "pathname" in parameter to do_guest_openat() for a cleaner diff
- fix two missing get_errno(fd)

v3 -> v4:
- fix typos in the commit message

v2 -> v3:
- fix coding style (braches)
- improve argument args/naming in do_openat2()
- merge do_openat2/do_guest_openat2
- do size checks first in do_openat2
- add "copy_struct_from_user" and use in "do_openat2()"
- drop using openat2.h and create "struct open_how_v0"
- log if open_how guest struct is bigger than our supported struct

v1 -> v2:
- do not include <sys/syscall.h>
- drop do_guest_openat2 from qemu.h and make static
- drop "safe" from do_guest_openat2
- ensure maybe_do_fake_open() is correct about when the result should
  be used or not
- Extract do_openat2() helper from do_syscall1()
- Call user_unlock* if a lock call fails
- Fix silly incorrect use of "target_open_how" when "open_how" is required
- Fix coding style comments
- Fix validation of arg4 in openat2
- Fix missing zero initialization of open_how
- Define target_open_how with abi_* types
- Warn about unimplemented size if "size" of openat2 is bigger than
  target_open_how


Michael Vogt (1):
  linux-user: add openat2 support in linux-user

 linux-user/syscall.c      | 108 +++++++++++++++++++++++++++++++++++++-
 linux-user/syscall_defs.h |   7 +++
 2 files changed, 113 insertions(+), 2 deletions(-)

-- 
2.45.2

[PATCH v6 1/1] linux-user: add openat2 support in linux-user

[Michael Vogt]

This commit adds support for the `openat2()` syscall in the
`linux-user` userspace emulator.

It is implemented by extracting a new helper `maybe_do_fake_open()`
out of the exiting `do_guest_openat()` and share that with the
new `do_guest_openat2()`. Unfortunately we cannot just make
do_guest_openat2() a superset of do_guest_openat() because the
openat2() syscall is stricter with the argument checking and
will return an error for invalid flags or mode combinations (which
open()/openat() will ignore).

The implementation is similar to SYSCALL_DEFINE(openat2), i.e.
a new `copy_struct_from_user()` is used that works the same
as the kernels version to support backwards-compatibility
for struct syscall argument.

Instead of including openat2.h we create a copy of `open_how`
as `open_how_ver0` to ensure that if the structure grows we
can log a LOG_UNIMP warning.

Note that in this commit using openat2() for a "faked" file in
/proc will ignore the "resolve" flags. This is not great but it
seems similar to the exiting behavior when openat() is called
with a dirfd to "/proc". Here too the fake file lookup may
not catch the special file because "realpath()" is used to
determine if the path is in /proc. Alternatively to ignoring
we could simply fail with `-TARGET_ENOSYS` (or similar) if
`resolve` flags are passed and we found something that looks
like a file in /proc that needs faking.

Signed-off-by: Michael Vogt <mvogt@redhat.com>
Buglink: https://github.com/osbuild/bootc-image-builder/issues/619
---
 linux-user/syscall.c      | 108 +++++++++++++++++++++++++++++++++++++-
 linux-user/syscall_defs.h |   7 +++
 2 files changed, 113 insertions(+), 2 deletions(-)

diff --git a/linux-user/syscall.c b/linux-user/syscall.c
index b693aeff5b..68ebda4ec8 100644
--- a/linux-user/syscall.c
+++ b/linux-user/syscall.c
@@ -602,6 +602,34 @@ static int check_zeroed_user(abi_long addr, size_t ksize, size_t usize)
     return 1;
 }
 
+/*
+ * Copies a target struct to a host struct, in a way that guarantees
+ * backwards-compatibility for struct syscall arguments.
+ *
+ * Similar to kernels uaccess.h:copy_struct_from_user()
+ */
+static int
+copy_struct_from_user(void *dst, size_t ksize, abi_ptr src, size_t usize)
+{
+    size_t size = MIN(ksize, usize);
+    size_t rest = MAX(ksize, usize) - size;
+
+    /* Deal with trailing bytes. */
+    if (usize < ksize) {
+        memset(dst + size, 0, rest);
+    } else if (usize > ksize) {
+        int ret = check_zeroed_user(src, ksize, usize);
+        if (ret <= 0) {
+            return ret ?: -TARGET_E2BIG;
+        }
+    }
+    /* Copy the interoperable parts of the struct. */
+    if (copy_from_user(dst, src, size)) {
+        return -TARGET_EFAULT;
+    }
+    return 0;
+}
+
 #define safe_syscall0(type, name) \
 static type safe_##name(void) \
 { \
@@ -653,6 +681,15 @@ safe_syscall3(ssize_t, read, int, fd, void *, buff, size_t, count)
 safe_syscall3(ssize_t, write, int, fd, const void *, buff, size_t, count)
 safe_syscall4(int, openat, int, dirfd, const char *, pathname, \
               int, flags, mode_t, mode)
+
+struct open_how_ver0 {
+    __u64 flags;
+    __u64 mode;
+    __u64 resolve;
+};
+safe_syscall4(int, openat2, int, dirfd, const char *, pathname, \
+              const struct open_how_ver0 *, how, size_t, size)
+
 #if defined(TARGET_NR_wait4) || defined(TARGET_NR_waitpid)
 safe_syscall4(pid_t, wait4, pid_t, pid, int *, status, int, options, \
               struct rusage *, rusage)
@@ -8334,8 +8371,9 @@ static int open_net_route(CPUArchState *cpu_env, int fd)
 }
 #endif
 
-int do_guest_openat(CPUArchState *cpu_env, int dirfd, const char *fname,
-                    int flags, mode_t mode, bool safe)
+static int maybe_do_fake_open(CPUArchState *cpu_env, int dirfd,
+                              const char *fname, int flags, mode_t mode,
+                              bool safe)
 {
     g_autofree char *proc_name = NULL;
     const char *pathname;
@@ -8418,6 +8456,17 @@ int do_guest_openat(CPUArchState *cpu_env, int dirfd, const char *fname,
         return fd;
     }
 
+    return -2;
+}
+
+int do_guest_openat(CPUArchState *cpu_env, int dirfd, const char *pathname,
+                    int flags, mode_t mode, bool safe)
+{
+    int fd = maybe_do_fake_open(cpu_env, dirfd, pathname, flags, mode, safe);
+    if (fd > -2) {
+        return fd;
+    }
+
     if (safe) {
         return safe_openat(dirfd, path(pathname), flags, mode);
     } else {
@@ -8425,6 +8474,56 @@ int do_guest_openat(CPUArchState *cpu_env, int dirfd, const char *fname,
     }
 }
 
+
+static int do_openat2(CPUArchState *cpu_env, abi_long dirfd,
+                      abi_ptr guest_pathname, abi_ptr guest_open_how,
+                      abi_long guest_size)
+{
+    struct open_how_ver0 how = {0};
+    char *pathname;
+    int ret;
+
+    if (guest_size < sizeof(struct target_open_how_ver0)) {
+        return -TARGET_EINVAL;
+    }
+    ret = copy_struct_from_user(&how, sizeof(how), guest_open_how, guest_size);
+    if (ret) {
+        if (ret == -TARGET_E2BIG) {
+            qemu_log_mask(LOG_UNIMP,
+                          "Unimplemented openat2 open_how size: %lu\n",
+                          guest_size);
+        }
+        return ret;
+    }
+    pathname = lock_user_string(guest_pathname);
+    if (!pathname) {
+        return -TARGET_EFAULT;
+    }
+
+    how.flags = target_to_host_bitmask(how.flags, fcntl_flags_tbl);
+    how.mode = tswap64(how.mode);
+    how.resolve = tswap64(how.resolve);
+
+    /*
+     * Ideally we would pass "how->resolve" flags into this helper too but
+     * the lookup for files that need faking is based on "realpath()" so
+     * neither a dirfd for "proc" nor restrictions via "resolve" flags can
+     * be honored right now.
+     */
+    int fd = maybe_do_fake_open(cpu_env, dirfd, pathname, how.flags, how.mode,
+                                true);
+    if (fd > -2) {
+        ret = get_errno(fd);
+    } else {
+        ret = get_errno(safe_openat2(dirfd, pathname, &how,
+                                     sizeof(struct open_how_ver0)));
+    }
+
+    fd_trans_unregister(ret);
+    unlock_user(pathname, guest_pathname, 0);
+    return ret;
+}
+
 ssize_t do_guest_readlink(const char *pathname, char *buf, size_t bufsiz)
 {
     ssize_t ret;
@@ -9197,6 +9296,11 @@ static abi_long do_syscall1(CPUArchState *cpu_env, int num, abi_long arg1,
         fd_trans_unregister(ret);
         unlock_user(p, arg2, 0);
         return ret;
+#if defined(TARGET_NR_openat2)
+    case TARGET_NR_openat2:
+        ret = do_openat2(cpu_env, arg1, arg2, arg3, arg4);
+        return ret;
+#endif
 #if defined(TARGET_NR_name_to_handle_at) && defined(CONFIG_OPEN_BY_HANDLE)
     case TARGET_NR_name_to_handle_at:
         ret = do_name_to_handle_at(arg1, arg2, arg3, arg4, arg5);
diff --git a/linux-user/syscall_defs.h b/linux-user/syscall_defs.h
index 8ed53904ed..b83fa34663 100644
--- a/linux-user/syscall_defs.h
+++ b/linux-user/syscall_defs.h
@@ -2753,4 +2753,11 @@ struct target_sched_param {
     abi_int sched_priority;
 };
 
+/* from kernel's include/uapi/linux/openat2.h */
+struct target_open_how_ver0 {
+    abi_ullong flags;
+    abi_ullong mode;
+    abi_ullong resolve;
+};
+
 #endif
-- 
2.45.2

Re: [PATCH v6 1/1] linux-user: add openat2 support in linux-user

[Laurent Vivier]

Le 20/09/2024 à 11:22, Michael Vogt a écrit :
> This commit adds support for the `openat2()` syscall in the
> `linux-user` userspace emulator.
> 
> It is implemented by extracting a new helper `maybe_do_fake_open()`
> out of the exiting `do_guest_openat()` and share that with the
> new `do_guest_openat2()`. Unfortunately we cannot just make
> do_guest_openat2() a superset of do_guest_openat() because the
> openat2() syscall is stricter with the argument checking and
> will return an error for invalid flags or mode combinations (which
> open()/openat() will ignore).
> 
> The implementation is similar to SYSCALL_DEFINE(openat2), i.e.
> a new `copy_struct_from_user()` is used that works the same
> as the kernels version to support backwards-compatibility
> for struct syscall argument.
> 
> Instead of including openat2.h we create a copy of `open_how`
> as `open_how_ver0` to ensure that if the structure grows we
> can log a LOG_UNIMP warning.
> 
> Note that in this commit using openat2() for a "faked" file in
> /proc will ignore the "resolve" flags. This is not great but it
> seems similar to the exiting behavior when openat() is called
> with a dirfd to "/proc". Here too the fake file lookup may
> not catch the special file because "realpath()" is used to
> determine if the path is in /proc. Alternatively to ignoring
> we could simply fail with `-TARGET_ENOSYS` (or similar) if
> `resolve` flags are passed and we found something that looks
> like a file in /proc that needs faking.
> 
> Signed-off-by: Michael Vogt <mvogt@redhat.com>
> Buglink: https://github.com/osbuild/bootc-image-builder/issues/619
> ---
>   linux-user/syscall.c      | 108 +++++++++++++++++++++++++++++++++++++-
>   linux-user/syscall_defs.h |   7 +++
>   2 files changed, 113 insertions(+), 2 deletions(-)
> 

Reviewed-by: Laurent Vivier <laurent@vivier.eu>

Re: [PATCH v6 1/1] linux-user: add openat2 support in linux-user

[Richard Henderson]

On 9/20/24 11:22, Michael Vogt wrote:
> +static int do_openat2(CPUArchState *cpu_env, abi_long dirfd,
> +                      abi_ptr guest_pathname, abi_ptr guest_open_how,
> +                      abi_long guest_size)

abi_ulong guest_size, as the "real" type is the unsigned size_t.

> +            qemu_log_mask(LOG_UNIMP,
> +                          "Unimplemented openat2 open_how size: %lu\n",
> +                          guest_size);

Use of %lu, and host "long" in general, is always wrong in qemu.

You did not build test everything: this errors out for 32-bit guests.
We must use TARGET_ABI_LONG_lu here.

> @@ -9197,6 +9296,11 @@ static abi_long do_syscall1(CPUArchState *cpu_env, int num, abi_long arg1,
>           fd_trans_unregister(ret);
>           unlock_user(p, arg2, 0);
>           return ret;
> +#if defined(TARGET_NR_openat2)
> +    case TARGET_NR_openat2:
> +        ret = do_openat2(cpu_env, arg1, arg2, arg3, arg4);
> +        return ret;
> +#endif

TARGET_NR_openat2 is universally defined.
This ifdef is not required.

An update to strace.list is missing.

Finally, this does not pass LTP.  In particular,

src/testcases/kernel/syscalls/openat2/openat202.c:62: TFAIL: resolve-no-magiclinks: 
openat2() passed unexpectedly

With RESOLVE_NO_MAGICLINKS, the open of /proc/self/exe handled by maybe_do_fake_open 
should fail with ELOOP.  I *think* that's the only magic link that we handle there, but 
please double check.


r~