Date: Mon, 11 Dec 2000 12:44:55 -0500 (EST) From: Bosko Milekic <bmilekic@technokratis.com> To: Alfred Perlstein <bright@wintelcom.net> Cc: net@freebsd.org Subject: Re: Abusing m_ext for a worthy goal. Message-ID: <Pine.BSF.4.21.0012111223350.21769-100000@jehovah.technokratis.com> In-Reply-To: <20001211014837.W16205@fw.wintelcom.net>
next in thread | previous in thread | raw e-mail | index | archive | help
Hi Alfred,
On Mon, 11 Dec 2000, Alfred Perlstein wrote:
> Ok, part of the mp cleanup got me looking at af_unix sockets.
>
> Here's the deal, descriptors can be called "in-flight" meaning they
> have been passed from one application to another via
> CMG_DATA but not picked up on the other side yet.
>
> The kernel does a bunch of hokey stuff to garbage collect them.
>
> Instead of doing that I'd like to abuse MEXTADD to setup a callback
> to remove the references automagicaly when the mbuf it free'd.
>
> Here's the problem:
>
> I want to use the mbuf clusters, I don't want to use any special
> external storage. I want my callback to free the cluster.
>
> Any hints on how to do this?
>
> here's what I have so far:
>
> MEXTADD(control, mtod(control, char *), control->m_len, unp_mbfree,
> cm, M_RDONLY, EXT_CMSG_DATA);
>
> control is a mbuf with a cluster attached the has fd's written in it.
> cm is a pointer to the mbuf cluster.
Here, the general idea is correct, but you'll have to do things this
way:
#define EXT_CMSG_DATA 666 /* A hack, for now. */
An allocation:
/* Here make sure you're not holding any other mutex. */
mtx_enter(&mclfree.m_mtx, MTX_DEF);
_MCLALLOC(control->m_ext.ext_buf, M_WAIT); /* Or M_DONTWAIT */
mtx_exit(&mclfree.m_mtx, MTX_DEF);
if (control->m_ext.ext_buf != NULL) {
/* Assuming you want to M_RDONLY your data. Otherwise, just
change M_RDONLY to 0. */
MEXTADD(control, control->m_ext.ext_buf, MCLBYTES, unp_mbfree,
cm, M_RDONLY, EXT_CMSG_DATA);
if (control->m_ext.ref_cnt == NULL) {
_MCLFREE(control->m_ext.ext_buf);
/* Allocation of ref. counter failed, so just fail
gracefully instead of panic (this should never happen,
but make sure to check. */
return (ENOBUFS);
}
} else
return (ENOBUFS);
> static void
> unp_mbfree(caddr_t mm, void *claddr)
> {
> struct mbuf *m = (struct mbuf *)mm;
>
> unp_scan(m, unp_discard);
> _MCLFREE(claddr);
> }
Actually, your free routine needs to accept the address of the
cluster as a first argument, and the optional argument as a second
argument. According to your MEXTADD() above, your optional argument will
be something called "cm." In your case, what you need to make your
optional argument be is the actual mbuf "control" (you can cast it to
struct mbuf *). Then, you can call unp_scan() on that mbuf (second
argument) and call _MCLFREE() on the first argument, which will
effectively be the address of the cluster.
> does this look right? would you like to abstract the mbuf
> clusters further before i started using the _ macros?
Abstracting mbuf clusters is deffinately something I'm looking at
doing (as you've probably seen me mention on some posts to some of the
lists), but for now, the above should do. I'd rather not rush into
abstracting the EXT_CLUSTER type further for now "just for the sake of
abstracting it." I want to make sure that we don't lose any performance
whatsoever for what concerns clusters (i.e. I'm trying to decide on
whether the cost of calling a function for clusters as well is worth it
at this point). Briefly, the day will come. :-)
> --
> -Alfred Perlstein - [bright@wintelcom.net|alfred@freebsd.org]
> "I have the heart of a child; I keep it in a jar on my desk."
Cheers,
Bosko Milekic
bmilekic@technokratis.com
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-net" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?Pine.BSF.4.21.0012111223350.21769-100000>