I failed at my attempt to produce a reliable testcase, so I ended up hacking a debug counter in dtrace_trap. Running dwatch -X read on a idle system with simple ssh activity went from this counter raising by the several thousands to a couple at a time.

I think so. In Illumos, the condition is well checked. Would you like this review to include the a fix for that? If yes, then I'll update D59280 to put a fix for arm32 too.
Also I believe that riscv is affected, but I don't have hw or the will to delve into riscv emulation and asm manuals right now.

(I'd like to make sure the copyout part works. but I need a bit of time to go around it)

In D59341#1361662, @freebsd_dev.thsi.be wrote:

I think so. In Illumos, the condition is well checked. Would you like this review to include the a fix for that? If yes, then I'll update D59280 to put a fix for arm32 too.

It'd be nice to fix both, yeah. Sorry, I already committed the 32-bit arm patch to main, I forgot that it has the same problem.

Also I believe that riscv is affected, but I don't have hw or the will to delve into riscv emulation and asm manuals right now.

(I'd like to make sure the copyout part works. but I need a bit of time to go around it)

Would you be willing to file an issue for that in bugzilla? We can try to get the riscv folks to take a look.

Yes good catch, the flags is technically still there. I agree with you, and I saw it too that the change on the prototype in the comment should have been left for D59449 (where it is formally dropped). I thought let's keep it that way to avoid too much noise, and also because there was already an inconsistency between copy/copystr comment in this regard. In the end, they both need to check that a fault didn't happen. I can update this DR if you prefer having it 100% correct. I'd rather not bother now, but make sure that the Aarch32 copyout case if fixed as well in D59449 (where the IN/OUT are properly split apart, currently for aarch64 only), to explain why we don't need to pass the flags (for aarch32/aarch64 in the IN-direction and for aarch32 in the OUT direction), and align all the prototype definitions at once.

In D59341#1364942, @freebsd_dev.thsi.be wrote:

Yes good catch, the flags is technically still there. I agree with you, and I saw it too that the change on the prototype in the comment should have been left for D59449 (where it is formally dropped). I thought let's keep it that way to avoid too much noise, and also because there was already an inconsistency between copy/copystr comment in this regard. In the end, they both need to check that a fault didn't happen. I can update this DR if you prefer having it 100% correct. I'd rather not bother now, but make sure that the Aarch32 copyout case if fixed as well in D59449 (where the IN/OUT are properly split apart, currently for aarch64 only), to explain why we don't need to pass the flags (for aarch32/aarch64 in the IN-direction and for aarch32 in the OUT direction), and align all the prototype definitions at once.

That's fine with me, thank you.