Hi Christoph,
On 4/3/20 3:35 PM, Christoph Hellwig wrote:
[...]
I just stumbled over your above commit, and it really confuses me.
Not the newly added functions, which seems perfectly sane, but why you
left the crazy old functions in place instead of investing a little
bit of extra effort to clean the existing mess up and switch everyone
to the sane new variants?
With crazy old functions I presume you mean the old bpf_probe_read()
which is mapped to BPF_FUNC_probe_read helper or something else entirely?
For the former, basically my main concern was that these would otherwise
break existing tools like bcc/bpftrace/.. unfortunately until they are not
converted over yet to _strict variants.
At least on x86, they would still rely on the broken semantic to probe
kernel and user memory with probe_read where it 'happens to work', but not
on other archs where the address space is not shared.
But once these are fixed, I would love to deprecate these in one way or
another. The warning in 00c42373d397 ("x86-64: add warning for non-canonical
user access address dereferences") should be a good incentive to switch
since people have been hitting it in production as the non-canonical space
is sometimes used in user space to tag pointers, for example.
Thanks,
Daniel
