Re: [PATCH] error-injection: Add prompt for function error injection

[Benjamin Tissoires]

On Dec 02 2022, Alexei Starovoitov wrote:
> On Fri, Dec 02, 2022 at 03:55:38PM +0100, Benjamin Tissoires wrote:
> > 
> > 
> > On 12/1/22 22:13, Linus Torvalds wrote:
> > > On Thu, Dec 1, 2022 at 8:59 AM Alexei Starovoitov
> > > <alexei.starovoitov@xxxxxxxxx> wrote:
> > > > 
> > > > The hid-bpf framework depends on it.
> > > 
> > > Ok, this is completely unacceptably disgusting hack.
> > 
> > [foreword: I have read the other replies, just replying to this one
> > because it is the explicit ask for a fix]
> > 
> > > 
> > > That needs fixing.
> > > 
> > > > Either hid-bpf or bpf core can add
> > > > "depends on FUNCTION_ERROR_INJECTION"
> > > 
> > > No, it needs to be narrowed down a lot. Nobody sane wants error
> > > injection just because they want some random HID thing.
> > > 
> > > And no, BPF shouldn't need it either.
> > > 
> > > This needs to be narrowed down to the point where HID can say "I want
> > > *this* particular call to be able to be a bpf call.
> > 
> > So, would the following be OK? I still have a few concerns I'll explain
> > after the patch.
> > 
[...]
> > 
> > While this patch removes the need for ALLOW_ERROR_INJECTION it has a
> > couple of drawbacks:
> > - suddenly we lose the nice separation of concerns between bpf core and
> > its users (HID in my case)
> > - it would need to be changed in 6.3 simply because of the previous
> > point, so it is just a temporary fix.

And third, it was bogus because the check_attach_modify_return() test was
inverted.

> 
> Agree, but it works short term.
> A silver lining is BTF_SET approach consumes 4 bytes per mark
> while ALLOW_ERROR_INJECTION consumes 16 bytes for struct error_injection_entry
> and then another 48 bytes per mark for struct ei_entry.
> 
> An alternative would be to define a known prefix like "bpf_modret_"
> or "bpf_hook_" and rename these three functions.

Not a big fan of that idea, sorry. It would work, for sure, but I don't
want to prefix my subsystem API by "bpf_modret_" which would make it
look like it is not part of my subsystem.

> 
> Then there will be no need for #include <linux/hid_bpf.h> in bpf core.

So I took your other advice to look into register_btf_kfunc_id_set().
And given that the internals of kfuncs are no more than a binary list of
ids, we can easily adapt it to have a better API to declare which
functions are fmodret. See [1] for the new series.

The net benefit are that now we can declare those functions outside of
any ALLOW_ERROR_INJECTION, outside of bpf-core, and also we can tag
sleepable ones which was not the case previously.

So I am rather happy with that v2. I'm sure there will be bikeshedding,
but this one looks better than my previous patch here.

> 
> > So I am not sure if this would qualify HID-BPF for 6.2. Please speak up.
> 
> Since that was the only thing I think it's fine to stay in the queue.

My plan is to see if we can get this validated in the next 2 days and if
not, drop it from 6.2 and reintroduce it in 6.3. After the weekend I
realized that delaying HID_BPF for one more release was not too much of an
issue in the end.

Cheers,
Benjamin


[1] https://lore.kernel.org/all/20221205164856.705656-2-benjamin.tissoires@xxxxxxxxxx/