view power.cpp @ 400:8c7453fb5bd9 api-inversion

Invert audioDB::power_flag / audiodb_power() Here the exciting discovery is that the mmap(), memcpy(), munmap() sequence is in fact not safe. In principle an msync() call should be inserted before unmapping for in-core changes to mmap()ed files to be flushed to disk. In this case we work around the problem entirely, by not mmap()ing anything and doing everything with file descriptors. Amusingly, that's probably not desperately safe either, this time because we have to move the file descriptor position (which is also a shared resource). dup() doesn't save us, as the duplicate file descriptor shares a file position. This applies also to the filling of data_buffer in the query loop, and in fact basically any call to lseek(), which is why I'm not fixing it now. Solution: if you have multiple threads all acting at once on a single database, do one audiodb_open() per thread, for now at least.
author mas01cr
date Thu, 27 Nov 2008 16:22:52 +0000
parents
children a8a5f2ca5380
line wrap: on
line source
#include "audioDB.h"
extern "C" {
#include "audioDB_API.h"
}

/* FIXME: we should not export this symbol to users of the library. */
int audiodb_sync_header(adb_t *adb) {
  off_t pos;
  pos = lseek(adb->fd, (off_t) 0, SEEK_CUR);
  if(pos == (off_t) -1) {
    goto error;
  }
  if(lseek(adb->fd, (off_t) 0, SEEK_SET) == (off_t) -1) {
    goto error;
  }
  if(write(adb->fd, adb->header, O2_HEADERSIZE) != O2_HEADERSIZE) {
    goto error;
  }

  /* can be fsync() if fdatasync() is racily exciting and new */
  fdatasync(adb->fd);
  if(lseek(adb->fd, pos, SEEK_SET) == (off_t) -1) {
    goto error;
  }
  return 0;

 error:
  return 1;
}

int audiodb_power(adb_t *adb) {
  /* FIXME: we should probably include in adb_t information about
   * which mode (O_RDONLY|O_RDWR) the database was opened, so that we
   * can check that it's writeable. */
  if(adb->header->length > 0) {
    return 1;
  }

  adb->header->flags |= O2_FLAG_POWER;
  return audiodb_sync_header(adb);
}