On Wed, Nov 8, 2017 at 6:05 AM, Lars Schotte <[email protected]> wrote: > Wait a moment. I have been seeing one issue on this realtek cards, > reproducible on Linux and FreeBSD.
Ours are different from Linux and FreeBSD one. And our way of interrupt throttle in re(4) is unique (openbsd adopted our way). > > The problem was that when you hit a realtek card with gigabit speeds, > both ways at the same time, for example with NFS, having a pf firewall > in place, you will get a watchdog timeout. It works for me on Dfly. Full line rate, dual side. > > Problems like this do not happen on Intel NICs. Well, never compare Realtek's stuff w/ Intel's :) > Some report that problems go away with Realtek official driver (from > their website), but havent tried that (I solved the problem by buying a > intel nic). Dfly's re(4) chip initialization is taken from Realtek's official site. They don't have datasheet for new chips, and the initialization consists large amount of magic values. > > Also surprising was that I tested some old PCI RTL8169SC and they max > at 150 Mbit/s even though they should (in theory) transfer gigabit. I won't be surprised. This 8169SC is really old, and is PCI. I'd suggest you to try their PCIe chips (though I am not recommending it), latest ones have no issue to TX (w/o TSO)/RX 941Mbps for TCP payload; relative older PCIe one normally can TX (w/o TSO) @800Mbps and RX @941Mbps for TCP stream. Given they are targeting desktop/laptop, the performance is reasonable. > > To me it blows my mind how can someone sell such hw. And ppl like me > not even noticing until having replaced all switches with gigabit and > noticing a bottleneck somewhere in the net I did not expect. > > On cheap mainboards sometime they do put in Atheros, you do not see > that happening there. For me, every realtek nic is a 100Mbit/s nic no > matter if it is a 100Mbit/s nic or gigabit. See my test result. Newer PCIe chips are much better. > > I d like to know your opinions on it. And second I am interested if > dragonfly has a driver that is somewhat different from FreeBSD's or > Linux's (they behave the same way). As I said, it's different. Thanks, sephe > > Thanks. > > On Tue, 7 Nov 2017 20:37:10 +0100 > Daniel Bilik <[email protected]> wrote: > >> On Mon, 6 Nov 2017 22:11:02 +0800 >> Sepherosa Ziehau <[email protected]> wrote: >> >> > For users have re(4): >> > https://leaf.dragonflybsd.org/~sephe/re194.diff >> > ... >> > Please test it, and let me know the result. >> >> Thank you. Tested with... >> >> re0@pci0:3:0:0: class=0x020000 card=0x81681849 chip=0x816810ec >> rev=0x11 hdr=0x00 vendor = 'Realtek Semiconductor Co., Ltd.' >> device = 'RTL8111/8168/8411 PCI Express Gigabit Ethernet >> Controller' >> >> ... works fine, no problems observed. >> >> -- >> Dan > > > > -- > Lars Schotte > Mudroňova 13 > 92101 Piešťany -- Tomorrow Will Never Die
