Date: Tue, 5 May 2015 00:03:04 +0300 From: Slawa Olhovchenkov <slw@zxy.spb.ru> To: John Baldwin <jhb@freebsd.org> Cc: Gleb Smirnoff <glebius@freebsd.org>, svn-src-head@freebsd.org, svn-src-all@freebsd.org, src-committers@freebsd.org Subject: Re: svn commit: r282280 - in head/sys/dev: e1000 ixgbe ixl Message-ID: <20150504210304.GA1394@zxy.spb.ru> In-Reply-To: <2463555.FfYUgqxYi8@ralph.baldwin.cx> References: <201504301823.t3UINd74073186@svn.freebsd.org> <26088556.xbkUe5VAyp@ralph.baldwin.cx> <20150504185125.GL34544@FreeBSD.org> <2463555.FfYUgqxYi8@ralph.baldwin.cx>
next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, May 04, 2015 at 04:01:28PM -0400, John Baldwin wrote: > On Monday, May 04, 2015 09:51:25 PM Gleb Smirnoff wrote: > > John, > > > > On Mon, May 04, 2015 at 09:58:49AM -0400, John Baldwin wrote: > > J> > J> > J> So the importaing thing here is that if_get_counter() is still doing > > J> > J> > J> per-ifnet stats. The stat you underlined above is per-queue instead. > > J> > J> > J> We well need more explicitly knowledge of queues outside of drivers > > J> > J> > J> and in the stack itself to support a generic framework for per-queue > > J> > J> > J> stats. > > J> > J> > > > J> > J> > This depends on how generic we want the API to be. Of course, we can add > > J> > J> > an extra argument to if_get_counter(). > > J> > J> > > > J> > J> > However, if we don't expect the number of queues to exceed a reasonable > > J> > J> > number of 255 :), we can fit the functionality into existing API. > > J> > J> > We can keep the queue number in the highest 8 bits of the ift_counter > > J> > J> > parameter. > > J> > J> > > > J> > J> > #define IFCOUNTER_MASK 0x00ffffff > > J> > J> > #define IFCOUNTER_QUEUE(c) ((c) >> 24) > > J> > J> > > J> > J> I'd prefer that expose queues more directly and figure out how to handle > > J> > J> per-queue stats then (e.g. do we have some sort of driver-independent > > J> > J> structure that each ifnet has 1 or more of that maps to a queue and has > > J> > J> driver provided methods, etc. If so you could have a driver method for > > J> > J> queue stats). Note that I did use if_get_counter to report the > > J> > J> per-interface stats instead of adding a new sysctl. > > J> > > > J> > What do you prefer: an extra argument to the if_get_counter() or a extra > > J> > ifop? > > J> > > J> As I said, I'd prefer we expose queues to the stack more directly complete > > J> with per-queue ops (e.g. I'd like a per-queue if_transmit thing, though > > J> probably more like the old if_start). > > > > Your answer seems quite orthogonal to my question. I reread it couple of times, > > but still can't figure out how exactly do you prefet to fetch per-queue stats. > > Can you please explain in more detail? > > struct if_queue { > struct ifnet *ifq_parent; > void (*ifq_get_counter)(struct if_queue *, ift_counter); > ... > }; > > (Pretend that if_queue is a new object type and that each RX or TX queue on a > NIC has one.) What about bulk get? One request for 1.5K queues stats.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20150504210304.GA1394>