Re: [PATCH V2 28/45] vfio: return mr from vfio_get_xlat_addr

[Steven Sistare <steven.sistare@oracle.com>]

On 2/24/2025 2:20 PM, Peter Xu wrote:
> On Mon, Feb 24, 2025 at 11:50:50AM -0500, Steven Sistare wrote:
>>>> I can do that, but it would add cycles.  Is this considered a high performance
>>>> path that may be called frequently?
>>>
>>> AFAICT, any vIOMMU mapping isn't high perf path.  In this specific path,
>>> the refcount op should be buried in any dma map operations..
>>
>> memory_region_ref contains a comment that implies we should avoid taking a
>> ref if possible:
>>       * Memory regions without an owner are supposed to never go away;
>>       * we do not ref/unref them because it slows down DMA sensibly.
> 
> That's for internal / permanent MRs that don't have an owner to speed
> things up, and that should be orthogonal to this, as that'll automatically
> take effect even if we use memory_region_[un]ref() here.

Yes, I understand that is a different type of mr, but my point is that the
author claims that memory_region_[un]ref() can noticeably slow DMA.

> AFAIU we should always suggest using memory_region_[un]ref() in callers
> when there's MR references.
> 
> I'm also ok without the boosted refcount, but then please document the RCU
> implications on referencing the MR, and the MR reference must not be used
> after rcu_read_unlock(), in that case.

I'll do that, to avoid any possibility of performance regression.

- Steve