Richard Gooch writes:
> Hm. This is a problem with your patch, right? What about the problem I
> reported back on 16-FEB? It's still happening with 2.3.48. The problem
> does not occur with 2.2.14.
>
> > Hi, all. I've been noticing odd behaviour with named pipes under
> > recent 2.3.x kernels (at least since 2.3.36 and possibly before).
> >
> > If you open a FIFO with O_RDONLY and then call read(2), and then
> > another process writes to the FIFO, the read(2) call doesn't return. A
> > subsequent writer process does wake up the reader, however.
> >
> > Has anybody else noticed this behaviour?
>
> I've noticed this problem on UP and SMP systems. It doesn't always
> happen, but it *does* happen :-(
>
> Regards,
>
> Richard....
I have attached a mail that I send to the lkml on 26-DEC-1999. It explains
the exact behaviour you have reported and contains a fix for it.
I have been applying this patch since kernel 2.3.32 it did still cleanly
apply for 2.3.49. I have also send the patch to Linus some time ago but
it probably got lost in the noise.
-- Markus----- oldmail -----
Hi,
I think I've found a bug in the FIFO (named pipes) implementation of Linux 2.3.3[234] (probably older kernels too). When doing the following
Terminal 1:
$ mkfifo /tmp/f $ cat /tmp/f
Terminal 2:
$ echo hello >/tmp/f
I would have expected that the cat does now produce `hello' instead it just sits there (most of the time). After several echo invocations the cat will output everything at once and then exit.
I don't have the POSIX specs but this behaviour looks broken.
The reason for this behaviour is a race condition in fifo_open which the following patch attempts to fix. Note that this race condition is probably much more likely to trigger on an UP system.
What happens is that between the interruptible_sleep_on call of the reader (line 63 in fs/fifo.c) and the down for the pipe semaphore (immediately afterwards), the writer does open the pipe writes to it and closes it. The reader then checks the writer count, finds it to be zero and continues to sleep.
This is my first (official) patch to the Linux kernel, so be careful it is possible that I got it all wrong (it works on my system but that's just a system for personal use that doesn't get much stress testing).
The patch is against 2.3.33 but should apply to 2.3.34 too.
Bye
-- Markus
--- /tmp/linux/fs/fifo.c Wed Dec 15 13:18:28 1999 +++ linux/fs/fifo.c Thu Dec 23 03:07:46 1999 @@ -41,6 +41,7 @@ PIPE_BASE(*inode) = (char *) page; PIPE_START(*inode) = PIPE_LEN(*inode) = 0; PIPE_READERS(*inode) = PIPE_WRITERS(*inode) = 0; + PIPE_CONNECT(*inode) = 0; } switch (filp->f_mode) { @@ -51,11 +52,13 @@ * opened, even when there is no process writing the FIFO. */ filp->f_op = &connecting_fifo_fops; - if (PIPE_READERS(*inode)++ == 0) + if (PIPE_READERS(*inode)++ == 0 && PIPE_WRITERS(*inode)) { + PIPE_CONNECT(*inode) = 1; wake_up_interruptible(PIPE_WAIT(*inode)); + } if (!(filp->f_flags & O_NONBLOCK)) { - while (!PIPE_WRITERS(*inode)) { + while (!PIPE_CONNECT(*inode)) { if (signal_pending(current)) goto err_rd; up(PIPE_SEM(*inode)); @@ -69,7 +72,7 @@ } } - if (PIPE_WRITERS(*inode)) + if (PIPE_CONNECT(*inode)) filp->f_op = &read_fifo_fops; break; @@ -80,14 +83,17 @@ * errno=ENXIO when there is no process reading the FIFO. */ ret = -ENXIO; - if ((filp->f_flags & O_NONBLOCK) && !PIPE_READERS(*inode)) + if ((filp->f_flags & O_NONBLOCK) && !PIPE_READERS(*inode) && + !PIPE_CONNECT(*inode)) goto err; filp->f_op = &write_fifo_fops; - if (!PIPE_WRITERS(*inode)++) + if (!PIPE_WRITERS(*inode)++ && PIPE_READERS(*inode)) { + PIPE_CONNECT(*inode) = 1; wake_up_interruptible(PIPE_WAIT(*inode)); + } - while (!PIPE_READERS(*inode)) { + while (!PIPE_CONNECT(*inode)) { if (signal_pending(current)) goto err_wr; up(PIPE_SEM(*inode)); @@ -107,8 +113,10 @@ PIPE_READERS(*inode)++; PIPE_WRITERS(*inode)++; - if (PIPE_READERS(*inode) == 1 || PIPE_WRITERS(*inode) == 1) + if (PIPE_READERS(*inode) == 1 || PIPE_WRITERS(*inode) == 1) { + PIPE_CONNECT(*inode) = 1; wake_up_interruptible(PIPE_WAIT(*inode)); + } break; default: --- /tmp/linux/include/linux/pipe_fs_i.h Wed Dec 15 13:18:18 1999 +++ linux/include/linux/pipe_fs_i.h Thu Dec 23 01:43:02 1999 @@ -9,6 +9,7 @@ unsigned int writers; unsigned int waiting_readers; unsigned int waiting_writers; + unsigned int connect; }; /* Differs from PIPE_BUF in that PIPE_SIZE is the length of the actual @@ -24,6 +25,7 @@ #define PIPE_WRITERS(inode) ((inode).i_pipe->writers) #define PIPE_WAITING_READERS(inode) ((inode).i_pipe->waiting_readers) #define PIPE_WAITING_WRITERS(inode) ((inode).i_pipe->waiting_writers) +#define PIPE_CONNECT(inode) ((inode).i_pipe->connect) #define PIPE_EMPTY(inode) (PIPE_LEN(inode) == 0) #define PIPE_FULL(inode) (PIPE_LEN(inode) == PIPE_SIZE)
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Tue Mar 07 2000 - 21:00:17 EST