Socket recv() hang on large message with MSG_WAITA

问题:

I have an application that reads large files from a server and hangs frequently on a particular machine. It has worked successfully under RHEL5.2 for a long time. We have recently upgraded to RHEL6.1 and it now hangs regularly.

I have created a test app that reproduces the problem. It hangs approx 98 times out of 100.

#include <errno.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/param.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <unistd.h>
#include <netdb.h>
#include <sys/socket.h>
#include <sys/time.h>

int mFD = 0;

void open_socket()
{
  struct addrinfo hints, *res;
  memset(&hints, 0, sizeof(hints));
  hints.ai_socktype = SOCK_STREAM;
  hints.ai_family = AF_INET;

  if (getaddrinfo("localhost", "60000", &hints, &res) != 0)
  {
    fprintf(stderr, "Exit %d\n", __LINE__);
    exit(1);
  }

  mFD = socket(res->ai_family, res->ai_socktype, res->ai_protocol);

  if (mFD == -1)
  {
    fprintf(stderr, "Exit %d\n", __LINE__);
    exit(1);
  }

  if (connect(mFD, res->ai_addr, res->ai_addrlen) < 0)
  {
    fprintf(stderr, "Exit %d\n", __LINE__);
    exit(1);
  }

  freeaddrinfo(res);
}

void read_message(int size, void* data)
{
  int bytesLeft = size;
  int numRd = 0;

  while (bytesLeft != 0)
  {
    fprintf(stderr, "reading %d bytes\n", bytesLeft);

    /* Replacing MSG_WAITALL with 0 works fine */
    int num = recv(mFD, data, bytesLeft, MSG_WAITALL);

    if (num == 0)
    {
      break;
    }
    else if (num < 0 && errno != EINTR)
    {
      fprintf(stderr, "Exit %d\n", __LINE__);
      exit(1);
    }
    else if (num > 0)
    {
      numRd += num;
      data += num;
      bytesLeft -= num;
      fprintf(stderr, "read %d bytes - remaining = %d\n", num, bytesLeft);
    }
  }

  fprintf(stderr, "read total of %d bytes\n", numRd);
}

int main(int argc, char **argv)
{
  open_socket();

  uint32_t raw_len = atoi(argv[1]);
  char raw[raw_len];

  read_message(raw_len, raw);

  return 0;
}

Some notes from my testing:

• If "localhost" maps to the loopback address 127.0.0.1, the app hangs on the call to recv() and NEVER returns.
• If "localhost" maps to the ip of the machine, thus routing the packets via the ethernet interface, the app completes successfully.
• When I experience a hang, the server sends a "TCP Window Full" message, and the client responds with a "TCP ZeroWindow" message (see image and attached tcpdump capture). From this point, it hangs forever with the server sending keep-alives and the client sending ZeroWindow messages. The client never seems to expand its window, allowing the transfer to complete.
• During the hang, if I examine the output of "netstat -a", there is data in the servers send queue but the clients receive queue is empty.
• If I remove the MSG_WAITALL flag from the recv() call, the app completes successfully.
• The hanging issue only arises using the loopback interface on 1 particular machine. I suspect this may all be related to timing dependencies.
• As I drop the size of the 'file', the likelihood of the hang occurring is reduced

The source for the test app can be found here:

Socket test source

The tcpdump capture from the loopback interface can be found here:

tcpdump capture

I reproduce the issue by issuing the following commands:

>  gcc socket_test.c -o socket_test
>  perl -e 'for (1..6000000){ print "a" }' | nc -l 60000
>  ./socket_test 6000000

This sees 6000000 bytes sent to the test app which tries to read the data using a single call to recv().

I would love to hear any suggestions on what I might be doing wrong or any further ways to debug the issue.

回答1:

MSG_WAITALL should block until all data has been received. From the manual page on recv:

This flag requests that the operation block until the full request is satisfied.

However, the buffers in the network stack probably are not large enough to contain everything, which is the reason for the error messages on the server. The client network stack simply can't hold that much data.

The solution is either to increase the buffer sizes (SO_RCVBUF option to setsockopt), split the message into smaller pieces, or receiving smaller chunks putting it into your own buffer. The last is what I would recommend.

Edit: I see in your code that you already do what I suggested (read smaller chunks with own buffering,) so just remove the MSG_WAITALL flag and it should work.

Oh, and when recv returns zero, that means the other end have closed the connection, and that you should do it too.

回答2:

Consider these two possible rules:

1. The receiver may wait for the sender to send more before receiving what has already been sent.

2. The sender may wait for the receiver to receive what has already been sent before sending more.

We can have either of these rules, but we cannot have both of these rules.

Why? Because if the receiver is permitted to wait for the sender, that means the sender cannot wait for the receiver to receive before sending more, otherwise we deadlock. And if the sender is permitted to wait for the receiver, that means the receiver cannot wait for the sender to send before receiving more, otherwise we deadlock.

If both of these things happen at the same time, we deadlock. The sender will not send more until the receiver receives what has already been sent, and the receiver will not receive what has already been sent unless the sender send more. Boom.

TCP chooses rule 2 (for reasons that should be obvious). Thus it cannot support rule 1. But in your code, you are the receiver, and you are waiting for the sender to send more before you receive what has already been sent. So this will deadlock.