does the request function block

does the request function block

[Christian Widmer]

does the request function of a bock device driver to be attomic or may it 
block? in the linux device driver 2 it says that it must be attomic and 
may not block. that makes sense to me since it also sais that the bottom-
half of the driver will call the request function too. but this is not realy
nessesery when i remove all request before releasing the io_request_lock.
on the other side block devices like nbd or brbd do send date using a socket
in there request function.

who is right and why?

Re: does the request function block

[Peter T. Breuer]

"Christian Widmer wrote:"
> does the request function of a bock device driver to be attomic or may it 
> block? in the linux device driver 2 it says that it must be attomic and 
> may not block. that makes sense to me since it also sais that the bottom-

It may not block - it's executed by the kernel in order to empty the
queue of requests for the device, and the kernel executes it with the
io lock held. So that's that. It can return without emptying the
queue, but that's another story. It must return immediately.

> half of the driver will call the request function too. but this is not realy
> nessesery when i remove all request before releasing the io_request_lock.

?? This seems to me to be orthogonal to to your question.

> on the other side block devices like nbd or brbd do send date using a socket
> in there request function.

No they don't. NBD moves the requests to an internal queue when the
request function is run. The function does not block. The internal
queue is later emptied by another means, in another context.

> who is right and why?

Well, since the hypothesis on which the question is based is mistaken,
it's hard to say!

Peter

Re: does the request function block

[Christian Widmer]

On Tuesday 28 August 2001 20:28, you wrote:
> No they don't. NBD moves the requests to an internal queue when the
> request function is run. The function does not block. The internal
> queue is later emptied by another means, in another context.

ok - does 
	result = sock_sendmsg(sock, &msg, size);
block? something in the back of my brains sais yes, but i might be
wrong since i'm new to linux kernel programing. if it does block 
nbd blocks. it realeases the io_request_lock lock and calls 
nbd_send_req which calls nbd_xmit and that results in a call to 
sock_sendmst.

in the documention ob the brbd it sais: that the request function 
is allways called in the context of a process doing I/O the kflushd
or the kupdate kernel thread.