Question

I’m writing a high-loaded daemon that should be run on the FreeBSD 8.0 and on Linux as well. The main purpose of daemon is to pass files that are requested by their identifier. Identifier is converted into local filename/file size via request to db. And then I use sequential mmap() calls to pass file blocks with send().

However sometimes there are mismatch of filesize in db and filesize on filesystem (realsize < size in db). In this situation I’ve sent all real data blocks and when next data block is mapped — mmap returns no errors, just usual address (I’ve checked errno variable also, it’s equal to zero after mmap). And when daemon tries to send this block it gets Segmentation Fault. (This behaviour is guarantedly issued on FreeBSD 8.0 amd64)

I was using safe check before open to ensure size with stat() call. However real life shows to me that segfault still can be raised in rare situtaions.

So, my question is there a way to check whether pointer is accessible before dereferencing it? When I’ve opened core in gdb, gdb says that given address is out of bound.
Probably there is another solution somebody can propose.

#include <fcntl.h>
#include <stdio.h>
#include <stdlib.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <time.h>
#include <unistd.h>

#define FILENAME        "./datafile"

int main()
{
    unsigned long i, j;

    srand(time(NULL));
    unsigned long pagesize = sysconf(_SC_PAGESIZE);

    unsigned long basesize = 4 * pagesize;
    unsigned long cropsize = 2 * pagesize;

    // create 4*pagesize sized file
    int f = creat(FILENAME, 0644);
    for (i = 0; i < basesize; i++) {
        unsigned char c = (unsigned char)rand();
        if (write(f, &c, 1) < 1) { perror("write"); break; }
    }
    close(f);

    f = open(FILENAME, O_RDONLY);

    // walk trough file
    unsigned char xor = 0;
    unsigned long offset = 0;
    for (j = 0; j < 4; j++) {
        // trunc file to 2*pagesize
        if (j == 2) truncate(FILENAME, cropsize);

        char *data = mmap(NULL, pagesize, PROT_READ, MAP_PRIVATE, f, offset);
        if (data == MAP_FAILED) { perror("mmap"); break; }
        printf("mmap: %lu@%lu for %i\n", pagesize, offset, f);

        for (i = 0; i < pagesize; i++) xor ^= data[i];

        offset += pagesize;
    }

    close(f);

    return 0;
}

Answer 1

Of course I can’t prove it from here, but I strongly suspect that you just have a book-keeping bug in your code. If you call mmap and pass in a size, and it succeeds, you shouldn’t get SIGSEGV.

I recommend that you apply valgrind to your investigation.

On many linux systems /proc/PID/maps will show you what regions are mapped with what access permissions.