-
Notifications
You must be signed in to change notification settings - Fork 1.7k
2.1.1-beta crashes with "illegal instruction" #9684
Comments
Running parity with the Other projects around the web seem to have similar issues with that version of Ubuntu on certain hardware; you might want to try building from source and seeing if that works. |
Looks like you tried to run the binary on a machine with old x86_64 CPU, could you share some details of your CPU, e.g. does it support SSE instructions ( |
@joshua-mir thx, i'll try those flags later on. yes, @ordian, the cpu is not the newest (but it supportes multiple versions of SSE)
|
tensorflow/tensorflow#17411 (comment) mentions a similar issue that is caused by AVX instructions that aren't supported by older cpus; whether it's avx or sse or some other newer instruction, the issue should be solved by building from source for your architecture. |
Looking at commits between 2.0.6 and 2.1.1, I'd guess that the suspect is #9414, which implies compiling rocksdb with @iSOcH it would be nice if you could bisect the breaking change down to a single commit. Something like this should work (but this could take ages to compile + require tens of GBs of free storage):
EDIT: I realised that |
Ok, so I guess the issue here is that we distribute binaries on linux with statically linked rocksdb, which was compiled with |
Thanks for the responses. Is it expected, that
Thanks for the update that this was not needed. It struggled with some things it should not have (
This did not create a log file when running the problematic binary. I tried this command: |
@ordian I think compiling with |
I can reproduce this issue on my laptop (i7-3520m):
Note, that it doesn't support AVX2 instructions, while the binary contains them (e.g.
@andresilva can we force |
@ordian Nice find! I still think we should restrict
This is probably nitpicking since I have no idea what kind of performance impact these optimizations have (probably negligible), but at least from your binary it seems GCC is vectorizing some stuff. |
Any fix beside build from source? When a new version will be released? |
@adyshimony the fix (#9725) was merged into master and will probably be released in 1-2 weeks in 2.1.3-beta. |
the fix was included in 2.1.2-beta, please check if it works for you as it does for me |
Thank you. Both release-binaries
work here 👍 |
I just had this crash with
My CPU is not old, it's i9-9900K. |
@gituser can you share the output of |
there you go:
|
Ack sorry, I meant |
There you go:
|
@gituser could you tell us which versions don't crash for you (e.g. from 2.5.x)? (trying to bisect the regression) |
@ordian the last version which worked I think was |
@gituser hm, the error is no reproducible on my CPU, which differs from yours only in a couple of flags ('ept_ad', 'pti', 'epb'), but they shouldn't case that. Does the illegal instruction error occur on startup (e.g. with |
@ordian I'm happy to test, the biggest issue is to reproduce the panic. Not sure why it happens and how to reproduce it. |
Wanted to launch my first ethereum full node for quite some time, but I cannot start the latest version of parity.
Call without arguments (but also with
--help
) crashes with error:Illegal instruction (core dumped)
The failing binary has sha256sum
17ec8d1ae4ed4eeccf1ea142c396a9410dc18575ddeee82cfec1b5347f7b1668
. On another machine with a more recent Ubuntu installation, 2.1.1-beta seems to work so far.dmesg on the affected system contains the following error when the binary is started:
The affected system seems to be able to run 2.0.6 without this problem.
I'm sure you need more information, please tell me how to gather that.
The text was updated successfully, but these errors were encountered: