Valgrind fails in CI

[phischu]

No description provided.

[mattisboeckle]

I did some investigating on these valgrind errors and found that it was complaining about "uninitialized value was created by a heap allocation" when using bytearrays.

If we just initialize the allocated memory to 0 the valgrind errors disappear (at least locally)

[jiribenes]

I'm all for using calloc in bytearrays :)

[b-studios]

That would be great! Is there any performance implications of doing that?

[marvinborner]

I'm confused, this shouldn't be a problem, right? Why are we reading from the bytearray when it's not initialized yet (especially in all IO tests)? If I remember correctly, we've discussed calloc-ing before and decided against it(?)

I think the performance implications could be significant, since bytearrays could be used for loading multiple gigabytes of data etc.

[phischu]

It is good to know that initializing the array would fix it, but we should not actually initialize them to zero every time. Rather, we should find the underlying out-of-bounds access. I can not reproduce those errors locally, but @mattisboeckle since you can, please take a look.

[mattisboeckle]

I did another pass at this, but as far as I can tell we never actually access anything out of bounds.

I'm starting to believe valgrind can't tell what happens in uv_fs_read, where our buffer is filled.
According to their docs https://docs.libuv.org/en/v1.x/fs.html#c.uv_fs_read the function should be equivalent to preadv, so I tried rewriting c_fs_read with that and valgrind stops complaining.

Probably not a bug on our end, not sure what to do

[phischu]

Try to find a way to silence those warnings specifically, perhaps telling valgrind about uv_fs_read. Also please remove the line that initializes bytearrays again so we see what CI is complaining about.

[jiribenes]

FWIW, for debugging, it might be worth it to memset the bytes of the bytearray to something recognisable like 0xEF and then trace in gdb/lldb + rwatch buf/watchpoint set variable --watch read buf to see when it gets read 🤔

[mattisboeckle]

I'm pretty sure there is no real problem here.
Without valgrind the programs work exactly as I would expect them to. Correct results, no errors. Definitely no out of bounds accesses, because that would mean segfaults.

I'll try and see if valgrind can handle libuv better somehow, but otherwise just suppressing the errors should be fine.