On 05/03/18 12:57 PM, Sagi Grimberg wrote:
Keith, while we're on this, regardless of cmb, is SQE memcopy and
update ordering always guaranteed?
If you look at mlx4 (rdma device driver) that works exactly the same as
nvme you will find:
qp->sq.head += nreq;
* Make sure that descriptors are written before
* doorbell record.
* Make sure doorbells don't leak out of SQ spinlock
* and reach the HCA out of order.
To me, it looks like the wmb() is redundant as writel should guarantee
the order. (Indeed, per Sinan's comment, writel on arm64 starts with a
wmb() which means, on that platform, there are two wmb() calls in a row.)
The mmiowb() call, on the other hand, looks correct per my understanding
of it's purpose with respect to the spinlock.