x86_64 allyesconfig build hang

[nickdesaulniers]

forking from #319 ; it seems that D53765 diff 183714 the build seems to be hanging.

I'm not sure how to provide feedback to help debug; I feel like I'd need to dump the cycle of optimizations we keep bouncing between? Maybe a preprocessed source and command line flags are a good starting point?

[nickdesaulniers]

export.zip let me know if that works, @topperc

[nickdesaulniers]

seems to be KASAN related (cc @dvyukov @ramosian-glider ).

$ make CC=clang allyesconfig
$ ./scripts/config -d KASAN
$ make CC=clang arch/x86/mm/fault.o
$ file arch/x86/mm/fault.o 
arch/x86/mm/fault.o: ELF 64-bit LSB relocatable, x86-64, version 1 (SYSV), not stripped

(I used menuconfig to turn off KASAN properly)

[nickdesaulniers]

seems to be KASAN related

maybe not, even with KASAN disabled, I'm seeing different hangs:

➜  ~ ps -ef | grep clang | rev | cut -d " " -f 1 | rev | grep c$ | sort | uniq | wc -l
17
➜  ~ ps -ef | grep clang | rev | cut -d " " -f 1 | rev | grep c$ | sort | uniq        
drivers/ata/sata_dwc_460ex.c
drivers/char/pcmcia/cm4000_cs.c
drivers/extcon/extcon-max77843.c
drivers/i2c/busses/i2c-taos-evm.c
drivers/infiniband/hw/cxgb4/qp.c
drivers/pinctrl/intel/pinctrl-cherryview.c
drivers/spi/spi-topcliff-pch.c
drivers/tty/serial/sh-sci.c
drivers/usb/mtu3/mtu3_gadget.c
net/ipv6/netfilter/ip6t_rt.c
sound/pci/echoaudio/darla20.c
sound/pci/echoaudio/darla24.c
sound/pci/echoaudio/gina20.c
sound/pci/echoaudio/layla20.c
sound/pci/hda/patch_realtek.c
sound/soc/codecs/wm_adsp.c
sound/x86/intel_hdmi_audio.c

[nickdesaulniers]

Another reproducer

[nickdesaulniers]

removing -fsanitize-coverage=trace-pc or -fsanitize-coverage=trace-cmp seems to help. Maybe these can't be used together?

[topperc]

export.zip is hanging in BranchFolding.cpp operating on the function do_user_addr_fault. Waiting on a build with debug symbols so I can run lldb.

[nickdesaulniers]

-fsanitize-coverage=trace-pc is added by CC_HAS_SANCOV_TRACE_PC which is depended on by CONFIG_KCOV. Turning that off, we can proceed. So allyesconfig minus CONFIG_KCOV I can proceed with both the second reproducer and the first. (KASAN not related)

Going to try to do a full build minus kcov, but it seems the diagnosis is that -fsanitize-coverage=trace-pc or -fsanitize-coverage=trace-cmp is currently broken with the current asm goto patch implementation.

[nickdesaulniers]

but it seems the diagnosis is that -fsanitize-coverage=trace-pc or -fsanitize-coverage=trace-cmp is currently broken with the current asm goto patch implementation.

Spoke too soon. Still seeing hangs... let me see if I can diagnose further.

[nickdesaulniers]

so turning off KASAN helped one of the problematic translation units build. Maybe it's BOTH KCOV AND KASAN that are broken (for different translations). Let me try that.

[topperc]

Just put up a new patch that let me get through export.zip

[nickdesaulniers]

w/ D53765 diff 183738 I can no longer repro hangs with the two above cases. Running an allyesconfig to triple check, but this is likely fixed. Will then revisit #319 .

[dileks]

w/ D53765 diff 183738 I have also no hanging. I can boot into QEMU, but on bare metal I get...

Kernel panic - not syncing: stack-protector: Kernel stack is corrupted in intel_idle_probe

UPDATE: My kernel-config is attached.

config-5.0.0-rc3-3-amd64-cbl.txt

[nickdesaulniers]

That's not good, but I wouldn't bother trying to boot on metal until #319 is solved.

[dileks]

$ make CC=clang arch/x86/mm/fault.o

I had seen such breakage before without CONFIG_KASAN and CONFIG_KCOV set. My latest build was fine with D53765 diff 183738 (no breakage in arch/x86/mm/fault.o).

Can you comment why I can boot in QEMU (just curious)?