Re: [PATCHv2 12/17] nvme: add Clang context annotations for nvme_queue::cq_poll_lock

[Christoph Hellwig]

On Sun, Jun 14, 2026 at 06:45:27PM +0530, Nilay Shroff wrote:
> nvme_queue::cqes, nvme_queue::cq_head, and nvme_queue::cq_phase are
> protected by nvme_queue::cq_poll_lock. Annotate these fields with
> __guarded_by(&cq_poll_lock) and annotate helpers accessing them with
> __must_hold(&cq_poll_lock) so that Clang's context analysis can
> validate the locking requirements.
> 
> IRQ-based queues do not use cq_poll_lock and instead rely on interrupt
> serialization. Annotate nvme_irq() and nvme_irq_check() with
> __context_unsafe() to suppress the corresponding context analysis
> warnings.
> 
> nvme_poll() invokes nvme_cqe_pending() as a lockless fast-path check
> before acquiring cq_poll_lock. This check is intentionally kept outside
> the lock because nvme_poll() may be called repeatedly in a tight polling
> loop until completions are found. The result is only advisory, as the
> completion queue is subsequently revalidated under cq_poll_lock by
> nvme_poll_cq(). Suppress the corresponding context analysis warning by
> annotating the lockless invocation of nvme_cqe_pending() with
> context_unsafe().

I'm not sure the annotation really buys us anything in this form,
as it makes bold claims about a guard and then ignores it for the
most common case.