Re: data modified by second NRT write with same memory location

[Philippe Gerum <rpm@xenomai.org>]

yo sang <stlfatboy@hotmail.com> writes:

> class RTMem
> {
> public:
> 	RTMem() {}
>
> 	RTMem(const char* data, int size, RTMemType type) {
> 		Data = Buf;
> 		memcpy(Data, data, size);
> 		Size = size;
> 		Type = type;
> 	}
>
> 	int			Size{};
> 	RTMemType	Type{};
> 	char*		Data{};	
> 	char		Buf[256]{};
>
> };

<snip>

> static void *realtime_thread(void *arg)
> {
>    struct timespec ts;
>    int ret;
>    RTMem buf;
>    for (;;) {
>        ret = recvfrom(s, &buf, sizeof(buf), 0, NULL, 0);
>        if (ret <= 0)
>			fail("recvfrom");
>        printf("   => %s \n", buf.Data);

<snip>

>
>     RTMem data("data1", 6, RTMEM_1);
>     ret = write(fd, &data, sizeof(data));
>     if (ret != sizeof(data))
>         fail("write");
>     memcpy(data.Data, "datadata", 9);
>     data.Size = 9;
>     data.Type = RTMEM_2;
>     ret = write(fd, &data, sizeof(data));
>     if (ret != sizeof(data))
>         fail("write");

This example code is broken, badly. Please consider what the reader
thread actually reads, which is definitely not a value, but a descriptor
containing a pointer to some value, which the writer thread will
overwrite before realtime_thread() runs. Which explains the outcome.

At the very least, you should have a distinct RTMem instance to feed
each write() with. This said, since XDDP is by design an inter-stage,
inter-process IPC, passing process-private pointers in messages looks
wrong in the first place.

-- 
Philippe.