diff mbox series

[6/7] socket: Add recvmsg timestamp test

Message ID 20210705183027.3804093-7-adhemerval.zanella@linaro.org
State Superseded
Headers show
Series Linux 64-bit time_t socket fixes | expand

Checks

Context Check Description
dj/TryBot-apply_patch success Patch applied to master at the time it was sent

Commit Message

Adhemerval Zanella July 5, 2021, 6:30 p.m. UTC
The test check if SO_TIMESTAMP and SO_TIMESTAMPNS is correctly
generated.  The recvmsg() is called with different ancilliary data
buffers, to check if the conversion for 64-bit time on 32-bit
legacy ABIs are handled correctly.

For such archicutures, depending of the remmaning ancially data
space the timestamp can not be appended and MSG_TRUNC is returned
instead.

Checked on x86_64-linux and on i686-linux-gnu on both 5.11 and
4.15 kernel.
---
 sysdeps/unix/sysv/linux/Makefile             |   3 +
 sysdeps/unix/sysv/linux/tst-recvmsg-time64.c |   1 +
 sysdeps/unix/sysv/linux/tst-recvmsg.c        | 171 +++++++++++++++++++
 3 files changed, 175 insertions(+)
 create mode 100644 sysdeps/unix/sysv/linux/tst-recvmsg-time64.c
 create mode 100644 sysdeps/unix/sysv/linux/tst-recvmsg.c

Comments

Florian Weimer July 5, 2021, 7:32 p.m. UTC | #1
* Adhemerval Zanella via Libc-alpha:

> The test check if SO_TIMESTAMP and SO_TIMESTAMPNS is correctly
> generated.  The recvmsg() is called with different ancilliary data
> buffers, to check if the conversion for 64-bit time on 32-bit
> legacy ABIs are handled correctly.
>
> For such archicutures, depending of the remmaning ancially data

“architectures”, “remaining”

> space the timestamp can not be appended and MSG_TRUNC is returned
> instead.

“cannot”

> +static void
> +do_recvmsg (int s, bool support_64_timestamp, void *cmsg, size_t slack,
> +	    size_t tsize, int exp_msg)
> +{
> +  /* A timestamp is expected if 32-bit timestamp are used (support in every
> +     configuration) or if underlying kernel support 64-bit timestamps.

“if [the] underlying kernel support[s]”

> +     Otherwise recvmsg will need extra space do add the 64-bit timestamp.  */

“space [t]o add”

> +  /* If there is not timestamp in the ancilliary data, recvmsg should set

“no[] timestamp”

> +     the flag inidcating it.  */

“indicating”

> +  /* Setup the ancillary data buffer with an extra page with PROT_NONE to
> +     check the possible timestamp conversion on some systems.  */
> +  pagesize = sysconf (_SC_PAGESIZE);
> +  if (pagesize == -1)
> +    FAIL_EXIT1 ("sysconf (_SC_PAGESIZE): %m\n");
> +  void *msgbuf = xmmap (0, 2 * pagesize, PROT_NONE,
> +		       	MAP_ANONYMOUS | MAP_PRIVATE, -1);
> +  xmprotect (msgbuf, pagesize, PROT_READ | PROT_WRITE);

Isn't this support_next_to_fault_allocate?

> +
> +  do_test_send (&srv_addr, 1);
> +  do_recvmsg (srv, false, NULL, 0, 0, 0);
> +
> +  /* If underlying kernel does not support   */
> +  bool support_64_timestamp = support_socket_time64_timestamp (srv);
> +
> +  /* Enable the timestamp using struct timeval precision.  */
> +  xsetsockopt (srv, SOL_SOCKET, SO_TIMESTAMP, &(int){1}, sizeof (int));

Please use setsockopt here, nt the x wrapper, so that the ABI switches
with time64 support.

> +  do_test_send (&srv_addr, array_length (slack));
> +  for (int s = 0; s < array_length (slack); s++)
> +    do_recvmsg (srv, support_64_timestamp, msgbuf, slack[s],
> +		sizeof (struct timeval), s);
> +
> +  /* Now enable timestamp using a higher precision, it overwrites the previous
> +     precision.  */
> +  xsetsockopt (srv, SOL_SOCKET, SO_TIMESTAMPNS, &(int){1}, sizeof (int));

Likewise here.

Thanks,
Florian
Adhemerval Zanella July 5, 2021, 7:51 p.m. UTC | #2
On 05/07/2021 16:32, Florian Weimer wrote:
> * Adhemerval Zanella via Libc-alpha:
> 
>> The test check if SO_TIMESTAMP and SO_TIMESTAMPNS is correctly
>> generated.  The recvmsg() is called with different ancilliary data
>> buffers, to check if the conversion for 64-bit time on 32-bit
>> legacy ABIs are handled correctly.
>>
>> For such archicutures, depending of the remmaning ancially data
> 
> “architectures”, “remaining”
> 
>> space the timestamp can not be appended and MSG_TRUNC is returned
>> instead.
> 
> “cannot”
> 
>> +static void
>> +do_recvmsg (int s, bool support_64_timestamp, void *cmsg, size_t slack,
>> +	    size_t tsize, int exp_msg)
>> +{
>> +  /* A timestamp is expected if 32-bit timestamp are used (support in every
>> +     configuration) or if underlying kernel support 64-bit timestamps.
> 
> “if [the] underlying kernel support[s]”
> 
>> +     Otherwise recvmsg will need extra space do add the 64-bit timestamp.  */
> 
> “space [t]o add”
> 
>> +  /* If there is not timestamp in the ancilliary data, recvmsg should set
> 
> “no[] timestamp”
> 
>> +     the flag inidcating it.  */
> 
> “indicating”

Ack for the typos and semantic fixes.

> 
>> +  /* Setup the ancillary data buffer with an extra page with PROT_NONE to
>> +     check the possible timestamp conversion on some systems.  */
>> +  pagesize = sysconf (_SC_PAGESIZE);
>> +  if (pagesize == -1)
>> +    FAIL_EXIT1 ("sysconf (_SC_PAGESIZE): %m\n");
>> +  void *msgbuf = xmmap (0, 2 * pagesize, PROT_NONE,
>> +		       	MAP_ANONYMOUS | MAP_PRIVATE, -1);
>> +  xmprotect (msgbuf, pagesize, PROT_READ | PROT_WRITE);
> 
> Isn't this support_next_to_fault_allocate?

It is, I will adjust.

> 
>> +
>> +  do_test_send (&srv_addr, 1);
>> +  do_recvmsg (srv, false, NULL, 0, 0, 0);
>> +
>> +  /* If underlying kernel does not support   */
>> +  bool support_64_timestamp = support_socket_time64_timestamp (srv);
>> +
>> +  /* Enable the timestamp using struct timeval precision.  */
>> +  xsetsockopt (srv, SOL_SOCKET, SO_TIMESTAMP, &(int){1}, sizeof (int));
> 
> Please use setsockopt here, nt the x wrapper, so that the ABI switches
> with time64 support.

Currently it does not matter because SO_TIMESTAMP will be handled by
the kernel headers.

> 
>> +  do_test_send (&srv_addr, array_length (slack));
>> +  for (int s = 0; s < array_length (slack); s++)
>> +    do_recvmsg (srv, support_64_timestamp, msgbuf, slack[s],
>> +		sizeof (struct timeval), s);
>> +
>> +  /* Now enable timestamp using a higher precision, it overwrites the previous
>> +     precision.  */
>> +  xsetsockopt (srv, SOL_SOCKET, SO_TIMESTAMPNS, &(int){1}, sizeof (int));
> 
> Likewise here.
> 
> Thanks,
> Florian
>
Florian Weimer July 5, 2021, 8:02 p.m. UTC | #3
* Adhemerval Zanella:

>>> +  do_test_send (&srv_addr, 1);
>>> +  do_recvmsg (srv, false, NULL, 0, 0, 0);
>>> +
>>> +  /* If underlying kernel does not support   */
>>> +  bool support_64_timestamp = support_socket_time64_timestamp (srv);
>>> +
>>> +  /* Enable the timestamp using struct timeval precision.  */
>>> +  xsetsockopt (srv, SOL_SOCKET, SO_TIMESTAMP, &(int){1}, sizeof (int));
>> 
>> Please use setsockopt here, nt the x wrapper, so that the ABI switches
>> with time64 support.
>
> Currently it does not matter because SO_TIMESTAMP will be handled by
> the kernel headers.

With my __setsockopt64 patch, the time64 version of the test will use a
different symbol.  This is what I meant.

Thanks,
Florian
Adhemerval Zanella July 5, 2021, 8:35 p.m. UTC | #4
On 05/07/2021 17:02, Florian Weimer wrote:
> * Adhemerval Zanella:
> 
>>>> +  do_test_send (&srv_addr, 1);
>>>> +  do_recvmsg (srv, false, NULL, 0, 0, 0);
>>>> +
>>>> +  /* If underlying kernel does not support   */
>>>> +  bool support_64_timestamp = support_socket_time64_timestamp (srv);
>>>> +
>>>> +  /* Enable the timestamp using struct timeval precision.  */
>>>> +  xsetsockopt (srv, SOL_SOCKET, SO_TIMESTAMP, &(int){1}, sizeof (int));
>>>
>>> Please use setsockopt here, nt the x wrapper, so that the ABI switches
>>> with time64 support.
>>
>> Currently it does not matter because SO_TIMESTAMP will be handled by
>> the kernel headers.
> 
> With my __setsockopt64 patch, the time64 version of the test will use a
> different symbol.  This is what I meant.

Yeah, I am not really convinced we do need such newer symbols that do
not correlate to 64-bit time_t syscall since I think it is *highly* 
unlikely kernel will 64-bit time_t facilities for multiplexed syscall
(and I would consider this a bug) without any guards or checking
if the program is using 32-bit time_t.

Arnd, which would be kernel take in this regard? Will fcntl/ioctl/etc
exposes constants that might use 64-bit types if userland uses 32-bit
time_t?  For socket constants I see that kernel does take
case of define different SO_* values, but these are for facilities
that are present for old 32-bit syscalls as well.  

This would also require to use kernel types anyway, since the default
glibc provided ones will be 32-bit; and this won't be user-visible
(and not requiring any libc intervention). 

Florian, in any case *currently* this patch does not need to use
xsetsockopt.  I can adapt if your set get merged first.
Florian Weimer July 6, 2021, 9:29 a.m. UTC | #5
* Adhemerval Zanella:

> On 05/07/2021 17:02, Florian Weimer wrote:
>> * Adhemerval Zanella:
>> 
>>>>> +  do_test_send (&srv_addr, 1);
>>>>> +  do_recvmsg (srv, false, NULL, 0, 0, 0);
>>>>> +
>>>>> +  /* If underlying kernel does not support   */
>>>>> +  bool support_64_timestamp = support_socket_time64_timestamp (srv);
>>>>> +
>>>>> +  /* Enable the timestamp using struct timeval precision.  */
>>>>> +  xsetsockopt (srv, SOL_SOCKET, SO_TIMESTAMP, &(int){1}, sizeof (int));
>>>>
>>>> Please use setsockopt here, nt the x wrapper, so that the ABI switches
>>>> with time64 support.
>>>
>>> Currently it does not matter because SO_TIMESTAMP will be handled by
>>> the kernel headers.
>> 
>> With my __setsockopt64 patch, the time64 version of the test will use a
>> different symbol.  This is what I meant.
>
> Yeah, I am not really convinced we do need such newer symbols that do
> not correlate to 64-bit time_t syscall since I think it is *highly* 
> unlikely kernel will 64-bit time_t facilities for multiplexed syscall
> (and I would consider this a bug) without any guards or checking
> if the program is using 32-bit time_t.

What about SO_VM_SOCKETS_CONNECT_TIMEOUT? 

> Florian, in any case *currently* this patch does not need to use
> xsetsockopt.  I can adapt if your set get merged first.

Fine, go with xsetsockopt then.

Thanks,
Florian
Adhemerval Zanella July 6, 2021, 12:55 p.m. UTC | #6
On 06/07/2021 06:29, Florian Weimer wrote:
> * Adhemerval Zanella:
> 
>> On 05/07/2021 17:02, Florian Weimer wrote:
>>> * Adhemerval Zanella:
>>>
>>>>>> +  do_test_send (&srv_addr, 1);
>>>>>> +  do_recvmsg (srv, false, NULL, 0, 0, 0);
>>>>>> +
>>>>>> +  /* If underlying kernel does not support   */
>>>>>> +  bool support_64_timestamp = support_socket_time64_timestamp (srv);
>>>>>> +
>>>>>> +  /* Enable the timestamp using struct timeval precision.  */
>>>>>> +  xsetsockopt (srv, SOL_SOCKET, SO_TIMESTAMP, &(int){1}, sizeof (int));
>>>>>
>>>>> Please use setsockopt here, nt the x wrapper, so that the ABI switches
>>>>> with time64 support.
>>>>
>>>> Currently it does not matter because SO_TIMESTAMP will be handled by
>>>> the kernel headers.
>>>
>>> With my __setsockopt64 patch, the time64 version of the test will use a
>>> different symbol.  This is what I meant.
>>
>> Yeah, I am not really convinced we do need such newer symbols that do
>> not correlate to 64-bit time_t syscall since I think it is *highly* 
>> unlikely kernel will 64-bit time_t facilities for multiplexed syscall
>> (and I would consider this a bug) without any guards or checking
>> if the program is using 32-bit time_t.
> 
> What about SO_VM_SOCKETS_CONNECT_TIMEOUT? 

Should we paper over all drivers or kernel features that do not have
proper 64-bit support?  If so it would add a quite complex multiplex
handling on some syscall (ioctl).  I would expect we will have quite
a few drivers out there which miss proper 64-bit support.

I added support for SO_TIMESTAMP{NS} because it is a features that 
uses libc defined datatype and we include the kernel headers directly
(so user has no opt-out).  I consider most driver as opt-in, where
users usually have to explicit include linux headers and they should 
be aware that if we want to use _TIME_BITS=64 they would need care to
use libc-defined datatypes.

But this is also error prone (and I would consider that drivers using
libc-define datatypes are as well) and I see your point.  I would
prefer if we avoid extra 64-bit symbols, but if you think we would
improve support in the future we can add as a conservative way.

> 
>> Florian, in any case *currently* this patch does not need to use
>> xsetsockopt.  I can adapt if your set get merged first.
> 
> Fine, go with xsetsockopt then.
> 
> Thanks,
> Florian
>
Florian Weimer July 6, 2021, 2:50 p.m. UTC | #7
* Adhemerval Zanella:

> But this is also error prone (and I would consider that drivers using
> libc-define datatypes are as well) and I see your point.  I would
> prefer if we avoid extra 64-bit symbols, but if you think we would
> improve support in the future we can add as a conservative way.

getsockopt/setsockopt is special because there is real length
information.  We don't have that for fcntl, and only sort-of for ioctl.
So the kernel at least has a chance of doing the right thing for
getsockopt/setsockopt.  And we don't need the time32-vs-time64
information in glibc for the present code because it's all encoded in
the option name.  So getsockopt/setsockopt is definitely a corner case.

For recvmsg, sendmsg, sendmmsg, I think we need the time64 alias so that
we can disable the rewriting if it turns out to be too disruptive.

Thanks,
Florian
diff mbox series

Patch

diff --git a/sysdeps/unix/sysv/linux/Makefile b/sysdeps/unix/sysv/linux/Makefile
index faea02bd1b..be6fe08ac2 100644
--- a/sysdeps/unix/sysv/linux/Makefile
+++ b/sysdeps/unix/sysv/linux/Makefile
@@ -237,6 +237,9 @@  sysdep_routines += cmsg_nxthdr
 CFLAGS-recvmmsg.c = -fexceptions -fasynchronous-unwind-tables
 CFLAGS-sendmmsg.c = -fexceptions -fasynchronous-unwind-tables
 
+tests += tst-recvmsg
+tests-time64 += tst-recvmsg-time64
+
 tests-special += $(objpfx)tst-socket-consts.out
 $(objpfx)tst-socket-consts.out: ../sysdeps/unix/sysv/linux/tst-socket-consts.py
 	PYTHONPATH=../scripts \
diff --git a/sysdeps/unix/sysv/linux/tst-recvmsg-time64.c b/sysdeps/unix/sysv/linux/tst-recvmsg-time64.c
new file mode 100644
index 0000000000..6cc8bbff6b
--- /dev/null
+++ b/sysdeps/unix/sysv/linux/tst-recvmsg-time64.c
@@ -0,0 +1 @@ 
+#include "tst-recvmsg.c"
diff --git a/sysdeps/unix/sysv/linux/tst-recvmsg.c b/sysdeps/unix/sysv/linux/tst-recvmsg.c
new file mode 100644
index 0000000000..361dcbb8cc
--- /dev/null
+++ b/sysdeps/unix/sysv/linux/tst-recvmsg.c
@@ -0,0 +1,171 @@ 
+/* Check recvmsg timestamp support.
+   Copyright (C) 2021 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <array_length.h>
+#include <arpa/inet.h>
+#include <errno.h>
+#include <string.h>
+#include <stdio.h>
+#include <support/check.h>
+#include <support/support.h>
+#include <support/xunistd.h>
+#include <support/xsocket.h>
+#include <sys/mman.h>
+
+/* Some extra space added for ancillary data, it might be used to convert
+   32-bit timestamp to 64-bit for _TIME_BITS=64 on older kernels.  */
+static const int slack[] = { 0, 4, 8, 16, 32, 64 };
+
+static size_t pagesize;
+
+static int
+do_test_send (const struct sockaddr_in *addr, int nmsgs)
+{
+  int s = xsocket (AF_INET, SOCK_DGRAM | SOCK_CLOEXEC, 0);
+  xconnect (s, (const struct sockaddr *) addr, sizeof (*addr));
+
+  for (int i = 0; i < nmsgs; i++)
+    xsendto (s, &i, sizeof (i), 0, (const struct sockaddr *) addr,
+	     sizeof (*addr));
+
+  return 0;
+}
+
+static void
+do_recvmsg (int s, bool support_64_timestamp, void *cmsg, size_t slack,
+	    size_t tsize, int exp_msg)
+{
+  /* A timestamp is expected if 32-bit timestamp are used (support in every
+     configuration) or if underlying kernel support 64-bit timestamps.
+     Otherwise recvmsg will need extra space do add the 64-bit timestamp.  */
+  bool exp_timestamp;
+  if (sizeof (time_t) == 4 || support_64_timestamp)
+    exp_timestamp = true;
+  else
+    exp_timestamp = slack >= CMSG_SPACE (tsize);
+
+  int msg;
+  struct iovec iov =
+    {
+      .iov_base = &msg,
+      .iov_len = sizeof (msg)
+    };
+  size_t msgs = CMSG_SPACE (tsize) + slack;
+  struct msghdr msghdr =
+    {
+      .msg_name = NULL,
+      .msg_namelen = 0,
+      .msg_iov = &iov,
+      .msg_iovlen = 1,
+      .msg_controllen = msgs,
+      .msg_control = cmsg != NULL ? cmsg + pagesize - msgs : NULL,
+    };
+  TEST_COMPARE (recvmsg (s, &msghdr, 0), sizeof (int));
+  TEST_COMPARE (msg, exp_msg);
+
+  if (cmsg == NULL)
+    return;
+
+  int timestamp = false;
+  for (struct cmsghdr *cmsg = CMSG_FIRSTHDR (&msghdr);
+       cmsg != NULL;
+       cmsg = CMSG_NXTHDR (&msghdr, cmsg))
+    {
+      if (cmsg->cmsg_level != SOL_SOCKET)
+	continue;
+      if (cmsg->cmsg_type == SCM_TIMESTAMP
+	  && cmsg->cmsg_len == CMSG_LEN (sizeof (struct timeval)))
+	{
+	  struct timeval tv;
+	  memcpy (&tv, CMSG_DATA (cmsg), sizeof (tv));
+	  printf ("SCM_TIMESTAMP:   {%jd, %jd}\n", (intmax_t)tv.tv_sec,
+		  (intmax_t)tv.tv_usec);
+	  timestamp = true;
+	}
+      else if (cmsg->cmsg_type == SCM_TIMESTAMPNS
+	       && cmsg->cmsg_len == CMSG_LEN (sizeof (struct timespec)))
+	{
+	  struct timespec ts;
+	  memcpy (&ts, CMSG_DATA (cmsg), sizeof (ts));
+	  printf ("SCM_TIMESTAMPNS: {%jd, %jd}\n", (intmax_t)ts.tv_sec,
+		  (intmax_t)ts.tv_nsec);
+	  timestamp = true;
+	}
+    }
+
+  /* If there is not timestamp in the ancilliary data, recvmsg should set
+     the flag inidcating it.  */
+  if (exp_timestamp && !timestamp)
+    TEST_VERIFY (msghdr.msg_flags & MSG_TRUNC);
+  else
+    TEST_COMPARE (exp_timestamp, timestamp);
+}
+
+static int
+do_test (void)
+{
+  int srv = xsocket (AF_INET, SOCK_DGRAM, 0);
+  struct sockaddr_in srv_addr =
+    {
+      .sin_family = AF_INET,
+      .sin_addr = {.s_addr = htonl (INADDR_LOOPBACK) },
+    };
+  xbind (srv, (struct sockaddr *) &srv_addr, sizeof (srv_addr));
+  {
+    socklen_t sa_len = sizeof (srv_addr);
+    xgetsockname (srv, (struct sockaddr *) &srv_addr, &sa_len);
+    TEST_VERIFY (sa_len == sizeof (srv_addr));
+  }
+
+  TEST_COMPARE (recvmsg (-1, NULL, 0), -1);
+  TEST_COMPARE (errno, EBADF);
+
+  /* Setup the ancillary data buffer with an extra page with PROT_NONE to
+     check the possible timestamp conversion on some systems.  */
+  pagesize = sysconf (_SC_PAGESIZE);
+  if (pagesize == -1)
+    FAIL_EXIT1 ("sysconf (_SC_PAGESIZE): %m\n");
+  void *msgbuf = xmmap (0, 2 * pagesize, PROT_NONE,
+		       	MAP_ANONYMOUS | MAP_PRIVATE, -1);
+  xmprotect (msgbuf, pagesize, PROT_READ | PROT_WRITE);
+
+  do_test_send (&srv_addr, 1);
+  do_recvmsg (srv, false, NULL, 0, 0, 0);
+
+  /* If underlying kernel does not support   */
+  bool support_64_timestamp = support_socket_time64_timestamp (srv);
+
+  /* Enable the timestamp using struct timeval precision.  */
+  xsetsockopt (srv, SOL_SOCKET, SO_TIMESTAMP, &(int){1}, sizeof (int));
+  do_test_send (&srv_addr, array_length (slack));
+  for (int s = 0; s < array_length (slack); s++)
+    do_recvmsg (srv, support_64_timestamp, msgbuf, slack[s],
+		sizeof (struct timeval), s);
+
+  /* Now enable timestamp using a higher precision, it overwrites the previous
+     precision.  */
+  xsetsockopt (srv, SOL_SOCKET, SO_TIMESTAMPNS, &(int){1}, sizeof (int));
+  do_test_send (&srv_addr, array_length (slack));
+  for (int s = 0; s < array_length (slack); s++)
+    do_recvmsg (srv, support_64_timestamp, msgbuf, slack[s],
+		sizeof (struct timespec), s);
+
+  return 0;
+}
+
+#include <support/test-driver.c>