Toward a satisfactory showmanyc

[Paolo Carlini]

Hi,

Nathan Myers wrote:

>Certainly the overwhelming bulk of the implementation work will be in 
>the autoconf machinery.  (Fortunately, "#ifdef FIONREAD" works, which 
>might reduce complexity a little bit.)  
>
>I am sensitive about the performance implications of calling fstat() 
>first.  On any given platform ioctl() or FIONREAD might not exist.
>I'm thinking it should try, in order:
>
>  ioctl
>    If the test is #ifdef'ed out, fine.  If it fails, fine.
>    If it reports a positive number, return it.  If it reports
>    a negative number, it's probably a file bigger than 2G, so
>    skip to the next test.  Otherwise...
>  poll
>    A cheap test -- if it reports "not ready", return 0.
>    Note that it reports "ready" if the next read will say EOF,
>    so "ready" just tells us it's worth looking more closely.
>  fstat
>    If it says the fd is a file, then use st_size as in your 
>    existing patch.  If it says the fd is a pipe or stream, the
>    ioctl above must have been #ifdef'd out; but, if it reports 
>    a non-zero st_size (which works on some targets), use it.  
>    If it fails, report 0.  Probably we should be using fstat64
>    where it exists.
>
Ok, thanks. In the next days I will start implementing your detailed 
outline:
when I posted my first message, today, I didn't know how to check for a 
failure of
ioctl, but in the meanwhile I learned that (from libjava ;) and 
everything seems
rather easily doable exactly as you write.

>For the case of a socket, we could use a funny search after all the 
>above has failed:  set the SO_RCVLOWAT socket option to increasing 
>values until poll() reports "not ready", and then restore 
>SO_RECVLOWAT and report the last "ready" value.  The problem is 
>that since it reports "ready" at EOF, you can't trust it if it 
>never reports "not ready".  So, if you get tired of quadrupling 
>(e.g. you've hit 64K, seven calls) and still get "ready", try 
>a really large value, e.g. 1<<24, -- and if it still reports ready 
>you know it's at EOF, and you report -1; otherwise, just report the 
>64K.
>
>Note that setsockopt SO_RCVLOWAT (like ioctl) is not required by POSIX.
>
>Of course it's up to you to decide how much work you want to put into 
>supporting this; at whatever point you think you and library have
>tried hard enough, return 0.
>
;) Well, in this first round I would probably leave out this refinement 
for sockets and
also the large file support part: we still hope to have this in before 
3.3, and we have not
much time to write and test thoroughly the whole beast.

Paolo.