Re: playing with qemu's 8139 nic and FreeBSD (loopback mode missing?

On Mon, May 08, 2006 at 12:44:43AM +0400, Igor Kovalenko wrote:
> Hi Juergen,
> 
> Juergen Lock wrote:
> > On Sun, May 07, 2006 at 12:38:06AM +0400, Igor Kovalenko wrote:
> >> Hi Juergen,
> >>
> >> Juergen Lock wrote:
> >>> [Cc'ing glebius_at_ because he did most of the recent re(4) commits, and
> >>> -current in case possible other driver developers didn't see this thread
> >>> in -emulation]
> >>>
> >>> On Wed, May 03, 2006 at 07:46:07AM +0400, Igor Kovalenko wrote:
> >>>> Juergen Lock wrote:
> >>>>> On Sat, Apr 29, 2006 at 10:48:48AM +0400, Igor Kovalenko wrote:
> >>>>>> Juergen Lock wrote:
> >>>>>>> On Fri, Apr 28, 2006 at 08:25:02PM +0400, Igor Kovalenko wrote:
> >>>>>>>> Juergen Lock wrote:
> >>>>>>>>> I played with
> >>>>>>>>> 	qemu -monitor stdio -m 256 -cdrom 6.1-RC1-i386-disc1.iso 
> >>>>>>>>> 	-usb -soundhw es1370 -kernel-kqemu -net nic,model=rtl8139 
> >>>>>>>>> 	-net user
> >>>>>>>>> and got it as far as
> >>>>>>>>> 	re0: diagnostic failed, failed to receive packet in loopback 
> >>>>>>>>> 	mode
> >>>>>>>>> (followed by a panic :) with the (experimental) patches below.
> >>>>>>>>>
> >>>>>>>>> Anyone in the mood to implement loopback mode for this nic?
> >>>>>>>>>
> >>>>>>>>> Hmm actually...  I just found the original posting in the archive,
> >>>>>>>>> is C+ mode implemented now?  If not re is probably not what I want,
> >>>>>>>> The rtl8139 is set up with PCI rev ID 0x20 which should be enough for 
> >>>>>>>> OS driver
> >>>>>>>> to detect C+ mode features. C+ mode is OK, tested with Linux driver.
> >>>>>>> Cool, so I want FreeBSD's re driver.  That one checks TxConfig
> >>>>>>> tho, as changed in my patch (inside #if 0).  And when changed,
> >>>>>>> it still doesn't work as mentioned above because the driver expects
> >>>>>>> loopback mode to be working.
> >>>>>>>>> but the rl driver that it attaches without that #if 0'd (now) hunk
> >>>>>>>>> below doesnt seem to be able to get data thru either and I get
> >>>>>>>>> 	rl0: watchdog timeout
> >>>>>>>>> in dmesg, which usually means the driver doesnt receive interrupts.
> >>>>>>>>>
> >>>>>>>>> What the heck, I'll append a log of a run just doing in 
> >>>>>>>>> fixit->cdrom:
> >>>>>>>>> 	ifconfig rl0 10.0.2.15
> >>>>>>>>> and then exiting (which is enough to trigger the watchdog timeout...)
> >>>>>>>>>
> >>>>>>>> I'm too lasy to test with fresh freebsd installation :)
> >>>>>>> No need to install FreeBSD, you can get away by just using
> >>>>>>> fixit mode of an install iso, i.e. disc1.  (which actually is
> >>>>>>> what I did above. :)
> >>>>>>>
> >>>>>>> You can look at 6.1RC's re driver here:
> >>>>>>> http://www.freebsd.org/cgi/cvsweb.cgi/src/sys/dev/re/if_re.c?annotate=1.46.2.14
> >>>>>>> which includes:
> >>>>>>> http://www.freebsd.org/cgi/cvsweb.cgi/src/sys/pci/if_rlreg.h?annotate=1.51.2.3
> >>>>>>>
> >>>>>>> And 6.1RC disc1 iso is e.g. here:
> >>>>>>> ftp://ftp.ru.freebsd.org:/pub/FreeBSD/ISO-IMAGES-i386/6.1/6.1-RC1-i386-disc1.iso
> >>>>>>>
> >>>>>>>
> >>>>>> Thanks, that iso pointer made it.
> >>>>>>
> >>>>> :)
> >>>>>
> >>>>>> Please try the following on top of your patch, at least ping should now 
> >>>>>> work:
> >>>>> Thanks, that seems to get the rl driver going.  Now to fix C+ mode
> >>>>> (re driver) change the #if 0 in my patch to #if 1...
> >>>>>
> >>>>>
> >>>> I believe freebsd re driver is somewhat broken, e.g. it does not follow 
> >>>> documented
> >>>> procedure to detect hardware features (e.g. 8139 c+ mode)
> >>> Hmm, a bit of googling didn't reveal docs about this, do you have
> >>> a pointer?  (I only found the data sheet at
> >>> 	http://people.freebsd.org/~wpaul/RealTek/spec-8139cp(150).pdf
> >>> which doesnt seem to mention detecting c+ hardware specifically)
> >> Well, I might have misread some docs; please do not consider this as an 
> >> assault :)
> >> I remember PCI rev id >= 0x20 is C+ mode for realtek 8139.
> > 
> > I don't.  It may well be the doc you saw this in just is no longer
> > online, or I didn't came up with the right search terms...
> >>>> and in tries to use 8169
> >>>> registers (e.g. 0xda) on 8139 hardware etc.
> >>> Oh, then it must have mis-detected the nic as a 8169 because
> >>> the code in question reads:
> >>>
> >>>        if (sc->rl_type == RL_8169)
> >>>                CSR_WRITE_2(sc, RL_MAXRXPKTLEN, 16383);
> > 
> >  And with your patch i get
> > 	RTL8139: not implemented write(b) addr=0xda val=0x00
> > but that can't be this line anyway because it seems to be a byte access...
> > Any of the re(4) experts have an idea where this comes from?
> > (log generated by defining DEBUG_RTL8139 in the patched hw/rtl8139.c)
> > 
> 
> I see the re driver actually is doing 2 byte write at RL_TXSTART so the second
> byte goes to RL_MAXRXPKTLEN. Comment in if_rlreg.h says RL_TXSTART is 8bit
> so this is a bug in if_re.c line 2063
> 
> This is not affecting qemu's card emulation (the extra byte write is ignored)
> so I think the problem is somewhere else.

I finally got around to playing with this a bit again, the following
patch gets ping and small fetch (http) going but is not enough for scp:

Index: hw/rtl8139.c
_at__at_ -1008,15 +1038,22 _at__at_
         val = cpu_to_le32(rxdw1);
         cpu_physical_memory_write(cplus_rx_ring_desc+4,  (uint8_t *)&val, 4);
 
-        /* seek to next Rx descriptor */
-        if (rxdw0 & CP_RX_EOR)
-        {
-            s->currCPlusRxDesc = 0;
-        }
-        else
-        {
-            ++s->currCPlusRxDesc;
-        }
+#ifdef DEBUG_RTL8139
+        printf("RTL8139: +++ written C+ mode RX descriptor %d %08x %08x %08x %08x\n",
+               descriptor,
+               rxdw0, rxdw1, rxbufLO, rxbufHI);
+#endif
+	/* seek to next Rx descriptor unless in loopback mode */
+	if (!(TxLoopBack == (s->TxConfig & TxLoopBack))) {
+		if (rxdw0 & CP_RX_EOR)
+		{
+		    s->currCPlusRxDesc = 0;
+		}
+		else
+		{
+		    ++s->currCPlusRxDesc;
+		}
+	}
 
 #if defined(DEBUG_RTL8139)
         printf("RTL8139: done C+ Rx mode ----------------\n");