On Tue, Mar 19, 2013 at 07:53:45AM +0200, Oleksii Shevchuk wrote: > > > + if (! (f->writable && f->fd >= 0)) > > > + return -1; > > -1 means EPERM. Something different is needed. > > EINVAL/EBUSY/EFAULT ? At least if !f->writable then the return code should be 0.
> > > + fdatasync(f->fd); > > Apparently the only error code which could happen here is EIO. > > Are you sure that we want to ignore it? > > I think we don't care at that point > > > What is the effect of fdatasync every 10 msg during high load? Wouldn't > > this decrease the throughput significantly? > > In case of high load and idle timer strategy should be the way to have > guaranted maximum count of not saved messages. But I have doubt about > it. I can set this as optional parameter with high defaults, and > introduce additional timer with maximal commit timeout. What do you > think about it? I think that this is the same thing as with fixed times: if you set the interval to 100000 msgs, you get a random sync every few days, which is useless. If you set it low, than you limit the throughput. I think that the msg-based limit doesn't make much sense. Zbyszek _______________________________________________ systemd-devel mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/systemd-devel
