Simplify UIO_SYSSPACE and UIO_NOCOPY paths in uiomove

[gallatin]

Summary

Uiomove can only block when the segflag is UIO_USERSPACE,
otherwise we end up just doing a bcopy (or nothing) and
moving cursors. So only emit witness warnings and
set deadlock thread flags in the UIO_USERSPACE case.

[kib]

sys/kern/subr_uio.c

218–219

This re-enables witness warning for nofault case. I do not think this is correct.

Simplest fix is to move the warning into else part of the if (nofault) statement below.

219–233

I wonder if we can remove this flag already. Probably not until all buffer cache consumers, like cd9660 and other filesystems, are fixed to use vn_io_fault().

220

nofault is now silently ignored for kernel space. Might be, assert that nofault implies UIO_USERSPACE.

[gallatin]

sys/kern/subr_uio.c

218–219

UIO_USERSPACE path calls maybe_yield() no matter what nofault is set to. I had assumed that required the witness warning. Is this assumption incorrect?

219–233

I was wondering about this. In a quick search, I could not find any callers of uiomove_nofault(), and uiomove_faultflag() is static, I thought I must be missing something.

[kib]

sys/kern/subr_uio.c

218–219

Yield only causes thread to leave CPU, the thread does not become non-schedulable. In other words, yield call does not put thread into sleep or blocked state.

219–233

I am talking about TDP_DEADLKTREAT, not about _nofault() stuff.

[gallatin]

Address kib's feedback

[gallatin]

sys/kern/subr_uio.c

218–219

Thanks. Done.

219–233

Ah, my mistake.

But is there a use to the nofault stuff? It seems like it should be removed if it is not used.

[kib]

sys/kern/subr_uio.c

219–233

The nofault facility AKA TDP_NOFAULTING is actively used.

Perhaps the uiomove_nofault() way of entering the nofault thread state became unused due to circumstances and code changes. I do not think that it should be removed now, because the interface itself is potentially useful and can become utilized again in tree.

[gallatin]

sys/kern/subr_uio.c

219–233

OK. To be clear, I meant the uiomove_nofault() entry point, not TDP_NOFAULTING. If you feel it should be preserved, I'm fine with that. Thanks for the quick feedback!