Re: Question on socket non block read

[Scott Klement <rpg400-l@xxxxxxxxxxxxxxxx>]

I've found that it's more robust to call read/recv first, and only call 
select() if read/recv returns that it "would block".

There are three reasons why this approach is better:

1) The same pattern can be used with SSL calls. Calling select() first 
is not valid in SSL because select waits on the raw (still encrypted) 
socket. It's possible for select() to wait for the network even though 
there's data in the SSL buffer available to return immediately. Data 
that has already been removed from the TCP channel, and decrypted and 
SSL has it in a buffer to be returned -- but since it's out of the TCP 
buffer, select does not know about it.  I realize that this example did 
not use SSL, but I prefer to use the same pattern for both,

2) When data is available on the socket already, calling recv/read first 
is a bit more efficient, since there's already data and so no need for 
the call to select().

3) There used to be sporadic situations where select() would  get stuck 
waiting when there's data in the TCP read buffer.  I think all of these 
were fixed, they were bugs in the OS ages ago...  but using this pattern 
worked around the problem.

What was the reason for calling select() first?  As long as you are 
using non-blocking sockets, I have never seen a disadvantage to calling 
read/recv first...

-SK


On 9/17/2015 2:03 PM, Bradley Stone wrote:
> select() should be called before the read()/recv() to know if the socket is
> ready for the read (or write).
>
> The timeout of 60 seconds seems quite big.  This isn't a delay if that's
> what you're thinking.
>
> If you specify 60 seconds that means it will wait on the select() for up to
> 60 seconds for "readiness" and then error if not.  If it's found to be
> ready right away, it will pass through without error.