From: renevant@xxxxxxxxxxxxxxxx > Messing with the Realtek nic driver didn't work my pc crashed soon after. > > > It looks like i've hit on a stable combination at the moment. It's looking > like either... > > * Passing "nomsi" as a kernel parameter has worked somehow, when doing > /proc/interrupts it looks like everything that used to be using msi is still > using msi, except possibly the hdmi audio on the video card. I've not tried that, but I know that MSI interrupts can be a bag of worms. > * This patch: > https://git.kernel.org/cgit/linux/kernel/git/tip/tip.git/commit/?h=x86/build&id=b399fe355b30d0102e7690 > c99e6f764ddfd32ec3 Ah yes gcc developers: when I say -fno-sse I want that to disable sse2, sse3, avx, avx2 and all the other extra instructions that might appear in the next version. In a local driver makefile I have: # With gcc 4.7 (and maybe some earlier ones) we have to explicitly # disable sse in order to stop the xmm registers being saved to stack # in every varargs function. # I think we have to explicitly disable all the newer instructions, so set # core2 (the first sane Intel 64bit cpu) and disable all its simd instructions. CFLAGS += -mno-mmx -mno-sse -mno-sse2 -mno-sse3 # The compiler on kapit doesn't support these option... XOPTS := -march=core2 -mtune=generic XFLAGS := $(shell $(CC) $(XOPTS) -E - </dev/null >/dev/null 2>&1 && echo $(XOPTS)) CFLAGS += $(XFLAGS) I must have a look at the NetBSD kernel build options. i386 still targets 486 so is ok, but amd64 will need the instructions disabled. > I'm currently running a bit of a frankenstien kernel because I have tried a > bunch of stuff. > > Everything is working at the moment stabily for me, i'm only seeing this when > doing transfers through the wireless VPN.... > > [ 839.482251] xhci_hcd 0000:02:00.0: WARN Successful completion on short TX: > needs XHCI_TRUST_TX_LENGTH quirk? > [ 844.480798] handle_tx_event: 2158 callbacks suppressed If you are seeing that, you need to add the quirk. (Or make the code in xhci-ring.c assume the quirk is set.) Without the quirk the xhci code believes that the transfer filled the RX buffer, even though the controller indicated a short length. Given that one of the VHDL definitions used to build xhci controllers has this bug. It would make sense for this to be enabled without a quirk. The driver is already unconditionally doing the check. > * I set the TRBS per segment back to 64 and the sg list back to "~0u" 256 is probably a better value for TRBS per segment. It will save some ring expansion requests happening later. > * readq/writeq reversion Definitely, that one wasn't tested on a range of hosts at all. > * LPM reversion That didn't make any difference to the problems I'm seeing with the ax88179 card using the ASMedia controller. > * merged usb3-port-power-v4 from > https://git.kernel.org/cgit/linux/kernel/git/djbw/usb.git/ > > * merged x86/build and x86/urgent from > https://git.kernel.org/cgit/linux/kernel/git/tip/tip.git/ I wonder if there is anything in there that has a real effect? David -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
