Linux non-blocking fifo (on demand logging)

This is a (very) old thread, but I've run into a similar problem of late. In fact, what I needed is a cloning of stdin to stdout with a copy to a pipe that is non blocking. the proposed ftee in the first answer really helped there, but was (for my use case) too volatile. Meaning I lost data I could have processed if I had gotten to it in time.

The scenario I was faced with is that I have a process (some_process) that aggregates some data and writes its results every three seconds to stdout. The (simplified) setup looked like this (in the real setup I am using a named pipe):

some_process | ftee >(onlineAnalysis.pl > results) | gzip > raw_data.gz

Now, raw_data.gz has to be compressed and has to be complete. ftee does this job very well. But the pipe I am using in the middle was too slow to grab the data flushed out - but it was fast enough to process everything if it could get to it, which was tested with a normal tee. However, a normal tee blocks if anything happens to the unnamed pipe, and as I want to be able to hook in on demand, tee is not an option. Back to the topic: It got better when I put a buffer in between, resulting in:

some_process | ftee >(mbuffer -m 32M| onlineAnalysis.pl > results) | gzip > raw_data.gz

But that was still losing data I could have processed. So I went ahead and extended the ftee proposed before to a buffered version (bftee). It still has all the same properties, but uses an (inefficient ?) internal buffer in case a write fails. It still loses data if the buffer runs full, but it works beautifully for my case. As always there is a lot of room for improvement, but as I copied the code off of here I'd like to share it back to people that might have a use for it.

/* bftee - clone stdin to stdout and to a buffered, non-blocking pipe 
    (c) racic@stackoverflow
    (c) fabraxias@stackoverflow
    WTFPL Licence */

    #include <stdio.h>
    #include <stdlib.h>
    #include <string.h>
    #include <sys/types.h>
    #include <sys/stat.h>
    #include <fcntl.h>
    #include <errno.h>
    #include <signal.h>
    #include <unistd.h>

    // the number of sBuffers that are being held at a maximum
    #define BUFFER_SIZE 4096
    #define BLOCK_SIZE 2048

    typedef struct {
      char data[BLOCK_SIZE];
      int bytes;
    } sBuffer;

    typedef struct {
      sBuffer *data;  //array of buffers
      int bufferSize; // number of buffer in data
      int start;      // index of the current start buffer
      int end;        // index of the current end buffer
      int active;     // number of active buffer (currently in use)
      int maxUse;     // maximum number of buffers ever used
      int drops;      // number of discarded buffer due to overflow
      int sWrites;    // number of buffer written to stdout
      int pWrites;    // number of buffers written to pipe
    } sQueue;

    void InitQueue(sQueue*, int);              // initialized the Queue
    void PushToQueue(sQueue*, sBuffer*, int);  // pushes a buffer into Queue at the end 
    sBuffer *RetrieveFromQueue(sQueue*);       // returns the first entry of the buffer and removes it or NULL is buffer is empty
    sBuffer *PeakAtQueue(sQueue*);             // returns the first entry of the buffer but does not remove it. Returns NULL on an empty buffer
    void ShrinkInQueue(sQueue *queue, int);    // shrinks the first entry of the buffer by n-bytes. Buffer is removed if it is empty
    void DelFromQueue(sQueue *queue);          // removes the first entry of the queue

    static void sigUSR1(int);                  // signal handled for SUGUSR1 - used for stats output to stderr
    static void sigINT(int);                   // signla handler for SIGKILL/SIGTERM - allows for a graceful stop ?

    sQueue queue;                              // Buffer storing the overflow
    volatile int quit;                         // for quiting the main loop

    int main(int argc, char *argv[])
    {   
        int readfd, writefd;
        struct stat status;
        char *fifonam;
        sBuffer buffer;
        ssize_t bytes;
        int bufferSize = BUFFER_SIZE;

        signal(SIGPIPE, SIG_IGN);
        signal(SIGUSR1, sigUSR1);
        signal(SIGTERM, sigINT);
        signal(SIGINT,  sigINT);

        /** Handle commandline args and open the pipe for non blocking writing **/

        if(argc < 2 || argc > 3)
        {   
            printf("Usage:\n someprog 2>&1 | %s FIFO [BufferSize]\n"
                   "FIFO - path to a named pipe, required argument\n"
                   "BufferSize - temporary Internal buffer size in case write to FIFO fails\n", argv[0]);
            exit(EXIT_FAILURE);
        }

        fifonam = argv[1];
        if (argc == 3) {
          bufferSize = atoi(argv[2]);
          if (bufferSize == 0) bufferSize = BUFFER_SIZE;
        }

        readfd = open(fifonam, O_RDONLY | O_NONBLOCK);
        if(-1==readfd)
        {   
            perror("bftee: readfd: open()");
            exit(EXIT_FAILURE);
        }

        if(-1==fstat(readfd, &status))
        {
            perror("bftee: fstat");
            close(readfd);
            exit(EXIT_FAILURE);
        }

        if(!S_ISFIFO(status.st_mode))
        {
            printf("bftee: %s in not a fifo!\n", fifonam);
            close(readfd);
            exit(EXIT_FAILURE);
        }

        writefd = open(fifonam, O_WRONLY | O_NONBLOCK);
        if(-1==writefd)
        {
            perror("bftee: writefd: open()");
            close(readfd);
            exit(EXIT_FAILURE);
        }

        close(readfd);


        InitQueue(&queue, bufferSize);
        quit = 0;

        while(!quit)
        {
            // read from STDIN
            bytes = read(STDIN_FILENO, buffer.data, sizeof(buffer.data));

            // if read failed due to interrupt, then retry, otherwise STDIN has closed and we should stop reading
            if (bytes < 0 && errno == EINTR) continue;
            if (bytes <= 0) break;

            // save the number if read bytes in the current buffer to be processed
            buffer.bytes = bytes;

            // this is a blocking write. As long as buffer is smaller than 4096 Bytes, the write is atomic to a pipe in Linux
            // thus, this cannot be interrupted. however, to be save this should handle the error cases of partial or interrupted write none the less.
            bytes = write(STDOUT_FILENO, buffer.data, buffer.bytes);
            queue.sWrites++;

            if(-1==bytes) {
                perror("ftee: writing to stdout");
                break;
            }

            sBuffer *tmpBuffer = NULL;

            // if the queue is empty (tmpBuffer gets set to NULL) the this does nothing - otherwise it tries to write
            // the buffered data to the pipe. This continues until the Buffer is empty or the write fails.
            // NOTE: bytes cannot be -1  (that would have failed just before) when the loop is entered. 
            while ((bytes != -1) && (tmpBuffer = PeakAtQueue(&queue)) != NULL) {
               // write the oldest buffer to the pipe
               bytes = write(writefd, tmpBuffer->data, tmpBuffer->bytes);

               // the  written bytes are equal to the buffer size, the write is successful - remove the buffer and continue
               if (bytes == tmpBuffer->bytes) {
                 DelFromQueue(&queue);
                 queue.pWrites++;
               } else if (bytes > 0) {
                 // on a positive bytes value there was a partial write. we shrink the current buffer
                 //  and handle this as a write failure
                 ShrinkInQueue(&queue, bytes);
                 bytes = -1;
               }
            }
            // There are several cases here:
            // 1.) The Queue is empty -> bytes is still set from the write to STDOUT. in this case, we try to write the read data directly to the pipe
            // 2.) The Queue was not empty but is now -> bytes is set from the last write (which was successful) and is bigger 0. also try to write the data
            // 3.) The Queue was not empty and still is not -> there was a write error before (even partial), and bytes is -1. Thus this line is skipped.
            if (bytes != -1) bytes = write(writefd, buffer.data, buffer.bytes);

            // again, there are several cases what can happen here
            // 1.) the write before was successful -> in this case bytes is equal to buffer.bytes and nothing happens
            // 2.) the write just before is partial or failed all together - bytes is either -1 or smaller than buffer.bytes -> add the remaining data to the queue
            // 3.) the write before did not happen as the buffer flush already had an error. In this case bytes is -1 -> add the remaining data to the queue
            if (bytes != buffer.bytes)
              PushToQueue(&queue, &buffer, bytes);
            else 
              queue.pWrites++;
        }

        // once we are done with STDIN, try to flush the buffer to the named pipe
        if (queue.active > 0) {
           //set output buffer to block - here we wait until we can write everything to the named pipe
           // --> this does not seem to work - just in case there is a busy loop that waits for buffer flush aswell. 
           int saved_flags = fcntl(writefd, F_GETFL);
           int new_flags = saved_flags & ~O_NONBLOCK;
           int res = fcntl(writefd, F_SETFL, new_flags);

           sBuffer *tmpBuffer = NULL;
           //TODO: this does not handle partial writes yet
           while ((tmpBuffer = PeakAtQueue(&queue)) != NULL) {
             int bytes = write(writefd, tmpBuffer->data, tmpBuffer->bytes);
             if (bytes != -1) DelFromQueue(&queue);
           }
        }

        close(writefd);

    }


    /** init a given Queue **/
    void InitQueue (sQueue *queue, int bufferSize) {
      queue->data = calloc(bufferSize, sizeof(sBuffer));
      queue->bufferSize = bufferSize;
      queue->start = 0;
      queue->end = 0;
      queue->active = 0;
      queue->maxUse = 0;
      queue->drops = 0;
      queue->sWrites = 0;
      queue->pWrites = 0;
    }

    /** push a buffer into the Queue**/
    void PushToQueue(sQueue *queue, sBuffer *p, int offset)
    {

        if (offset < 0) offset = 0;      // offset cannot be smaller than 0 - if that is the case, we were given an error code. Set it to 0 instead
        if (offset == p->bytes) return;  // in this case there are 0 bytes to add to the queue. Nothing to write

        // this should never happen - offset cannot be bigger than the buffer itself. Panic action
        if (offset > p->bytes) {perror("got more bytes to buffer than we read\n"); exit(EXIT_FAILURE);}

        // debug output on a partial write. TODO: remove this line
        // if (offset > 0 ) fprintf(stderr, "partial write to buffer\n");

        // copy the data from the buffer into the queue and remember its size
        memcpy(queue->data[queue->end].data, p->data + offset , p->bytes-offset);
        queue->data[queue->end].bytes = p->bytes - offset;

        // move the buffer forward
        queue->end = (queue->end + 1) % queue->bufferSize;

        // there is still space in the buffer
        if (queue->active < queue->bufferSize)
        {
            queue->active++;
            if (queue->active > queue->maxUse) queue->maxUse = queue->active;
        } else {
            // Overwriting the oldest. Move start to next-oldest
            queue->start = (queue->start + 1) % queue->bufferSize;
            queue->drops++;
        }
    }

    /** return the oldest entry in the Queue and remove it or return NULL in case the Queue is empty **/
    sBuffer *RetrieveFromQueue(sQueue *queue)
    {
        if (!queue->active) { return NULL; }

        queue->start = (queue->start + 1) % queue->bufferSize;
        queue->active--;
        return &(queue->data[queue->start]);
    }

    /** return the oldest entry in the Queue or NULL if the Queue is empty. Does not remove the entry **/
    sBuffer *PeakAtQueue(sQueue *queue)
    {
        if (!queue->active) { return NULL; }
        return &(queue->data[queue->start]);
    }

    /*** Shrinks the oldest entry i the Queue by bytes. Removes the entry if buffer of the oldest entry runs empty*/
    void ShrinkInQueue(sQueue *queue, int bytes) {

      // cannot remove negative amount of bytes - this is an error case. Ignore it
      if (bytes <= 0) return;

      // remove the entry if the offset is equal to the buffer size
      if (queue->data[queue->start].bytes == bytes) {
        DelFromQueue(queue);
        return;
      };

      // this is a partial delete
      if (queue->data[queue->start].bytes > bytes) {
        //shift the memory by the offset
        memmove(queue->data[queue->start].data, queue->data[queue->start].data + bytes, queue->data[queue->start].bytes - bytes);
        queue->data[queue->start].bytes = queue->data[queue->start].bytes - bytes;
        return;
      }

      // panic is the are to remove more than we have the buffer
      if (queue->data[queue->start].bytes < bytes) {
        perror("we wrote more than we had - this should never happen\n");
        exit(EXIT_FAILURE);
        return;
      }
    }

    /** delete the oldest entry from the queue. Do nothing if the Queue is empty **/
    void DelFromQueue(sQueue *queue)
    {
        if (queue->active > 0) {
          queue->start = (queue->start + 1) % queue->bufferSize;
          queue->active--;
        }
    }

    /** Stats output on SIGUSR1 **/
    static void sigUSR1(int signo) {
      fprintf(stderr, "Buffer use: %i (%i/%i), STDOUT: %i PIPE: %i:%i\n", queue.active, queue.maxUse, queue.bufferSize, queue.sWrites, queue.pWrites, queue.drops);
    }

    /** handle signal for terminating **/
    static void sigINT(int signo) {
      quit++;
      if (quit > 1) exit(EXIT_FAILURE);
    }

This version takes one more (optional) argument which specifies the number of the blocks that are to buffered for the pipe. My sample call now looks like this:

some_process | bftee >(onlineAnalysis.pl > results) 16384 | gzip > raw_data.gz

resulting in 16384 blocks to be buffered before discards happen. this uses about 32 Mbyte more memory, but... who cares ?

Of course, in the real environment I am using a named pipe so that I can attach and detach as needed. There is looks like this:

mkfifo named_pipe
some_process | bftee named_pipe 16384 | gzip > raw_data.gz &
cat named_pipe | onlineAnalysis.pl > results

Also, the process reacts on signals as follows: SIGUSR1 -> print counters to STDERR SIGTERM, SIGINT -> first exits the main loop and flushed the buffer to the pipe, the second terminated the program immediatly.

Maybe this helps someone in the future... Enjoy


Inspired by your question I've written a simple program that will let you do this:

$ myprogram 2>&1 | ftee /tmp/mylog

It behaves similarly to tee but clones the stdin to stdout and to a named pipe (a requirement for now) without blocking. This means that if you want to log this way it may happen that you're gonna lose your log data, but I guess it's acceptable in your scenario. The trick is to block SIGPIPE signal and to ignore error on writing to a broken fifo. This sample may be optimized in various ways of course, but so far, it does the job I guess.

/* ftee - clone stdin to stdout and to a named pipe 
(c) racic@stackoverflow
WTFPL Licence */

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <errno.h>
#include <signal.h>
#include <unistd.h>

int main(int argc, char *argv[])
{
    int readfd, writefd;
    struct stat status;
    char *fifonam;
    char buffer[BUFSIZ];
    ssize_t bytes;
    
    signal(SIGPIPE, SIG_IGN);

    if(2!=argc)
    {
        printf("Usage:\n someprog 2>&1 | %s FIFO\n FIFO - path to a"
            " named pipe, required argument\n", argv[0]);
        exit(EXIT_FAILURE);
    }
    fifonam = argv[1];

    readfd = open(fifonam, O_RDONLY | O_NONBLOCK);
    if(-1==readfd)
    {
        perror("ftee: readfd: open()");
        exit(EXIT_FAILURE);
    }

    if(-1==fstat(readfd, &status))
    {
        perror("ftee: fstat");
        close(readfd);
        exit(EXIT_FAILURE);
    }

    if(!S_ISFIFO(status.st_mode))
    {
        printf("ftee: %s in not a fifo!\n", fifonam);
        close(readfd);
        exit(EXIT_FAILURE);
    }

    writefd = open(fifonam, O_WRONLY | O_NONBLOCK);
    if(-1==writefd)
    {
        perror("ftee: writefd: open()");
        close(readfd);
        exit(EXIT_FAILURE);
    }

    close(readfd);

    while(1)
    {
        bytes = read(STDIN_FILENO, buffer, sizeof(buffer));
        if (bytes < 0 && errno == EINTR)
            continue;
        if (bytes <= 0)
            break;

        bytes = write(STDOUT_FILENO, buffer, bytes);
        if(-1==bytes)
            perror("ftee: writing to stdout");
        bytes = write(writefd, buffer, bytes);
        if(-1==bytes);//Ignoring the errors
    }
    close(writefd); 
    return(0);
}

You can compile it with this standard command:

$ gcc ftee.c -o ftee

You can quickly verify it by running e.g.:

$ ping www.google.com | ftee /tmp/mylog

$ cat /tmp/mylog

Also note - this is no multiplexer. You can only have one process doing $ cat /tmp/mylog at a time.


It seems like bash <> redirection operator (3.6.10 Opening File Descriptors for Reading and WritingSee) makes writing to file/fifo opened with it non-blocking. This should work:

$ mkfifo /tmp/mylog
$ exec 4<>/tmp/mylog
$ myprogram 2>&1 | tee >&4
$ cat /tmp/mylog # on demend

Solution given by gniourf_gniourf on #bash IRC channel.


However, this would create a ever growing log file even if not used until the drive runs out of space.

Why not periodically rotate the logs? There's even a program to do it for you logrotate.

There's also a system for generating log messages and doing different things with them according to type. It's called syslog.

You could even combine the two. Have your program generate syslog messages, configure syslog to place them in a file and use logrotate to ensure they don't fill the disk.


If it turned out that you were writing for a small embedded system and the program's output is heavy there are a variety of techniques you might consider.

  • Remote syslog: send the syslog messages to a syslog server on the network.
  • Use the severity levels availble in syslog to do different things with the messages. E.g. discard "INFO" but log and forward "ERR" or greater. E.g. to console
  • Use a signal handler in your program to reread configuration on HUP and vary log generation "on demand" this way.
  • Have your program listen on a unix socket and write messages down it when open. You could even implement and interactive console into your program this way.
  • Using a configuration file, provide granular control of logging output.