TCP partial write

TCP partial write

[Paul Martin]

Is it possible that a (non-blocking) TCP write(2) will write a number
of bytes not multiple of the machine word size? i.e., could a write
request for 4 bytes return 2?

Also is this an OS-dependent behavior or there is a spec for it? (I
could find atomic guarantees for pipes and datagram sockets but not
for TCP)

Thanks,
Paul

Re: TCP partial write

[Rick Jones]

Paul Martin wrote:
> Is it possible that a (non-blocking) TCP write(2) will write a number
> of bytes not multiple of the machine word size? i.e., could a write
> request for 4 bytes return 2?

Yes.

> Also is this an OS-dependent behavior or there is a spec for it? (I
> could find atomic guarantees for pipes and datagram sockets but not
> for TCP)

TCP is a byte-stream. It sends and receives a stream of bytes.  You should/must 
assume that when you do a non-blocking write, it will take any number of the 
bytes you offer from 0 to however many bytes you give it.  And you should/must 
assume that at the other end, your recv/read calls will return with between 0 and 
however many bytes you ask of them, with 0 meaning the remote has said it has 
nothing left to give.

rick jones

Re: TCP partial write

[Stephen Hemminger]

On Mon, 15 Jun 2009 10:09:29 -0700
Rick Jones <rick.jones2@hp.com> wrote:

> Paul Martin wrote:
> > Is it possible that a (non-blocking) TCP write(2) will write a number
> > of bytes not multiple of the machine word size? i.e., could a write
> > request for 4 bytes return 2?
> 
> Yes.
> 
> > Also is this an OS-dependent behavior or there is a spec for it? (I
> > could find atomic guarantees for pipes and datagram sockets but not
> > for TCP)
> 
> TCP is a byte-stream. It sends and receives a stream of bytes.  You should/must 
> assume that when you do a non-blocking write, it will take any number of the 
> bytes you offer from 0 to however many bytes you give it.  And you should/must 
> assume that at the other end, your recv/read calls will return with between 0 and 
> however many bytes you ask of them, with 0 meaning the remote has said it has 
> nothing left to give.
> 

Actually on a blocking socket 0 means other end has closed.

Re: TCP partial write

[Rick Jones]

Stephen Hemminger wrote:
> On Mon, 15 Jun 2009 10:09:29 -0700
> Rick Jones <rick.jones2@hp.com> wrote:

>> And you should/must assume that at the other end, your recv/read calls will
>> return with between 0 and however many bytes you ask of them, with 0 meaning
>> the remote has said it has nothing left to give.
>>
> 
> 
> Actually on a blocking socket 0 means other end has closed.

Overly poetic wording on my part, using "said it has nothing left to give" to 
mean "other end has closed" :)

rick jones