[PATCH v4 00/17] netdev: Eliminate duplicate barriers on weakly-ordered archs

[Sinan Kaya]

On 3/21/2018 10:56 AM, David Miller wrote:
> From: Sinan Kaya <okaya at codeaurora.org>
> Date: Mon, 19 Mar 2018 22:42:15 -0400
> 
>> Code includes wmb() followed by writel() in multiple places. writel()
>> already has a barrier on some architectures like arm64.
>>
>> This ends up CPU observing two barriers back to back before executing the
>> register write.
>>
>> Since code already has an explicit barrier call, changing writel() to
>> writel_relaxed().
>>
>> I did a regex search for wmb() followed by writel() in each drivers
>> directory.
>> I scrubbed the ones I care about in this series.
>>
>> I considered "ease of change", "popular usage" and "performance critical
>> path" as the determining criteria for my filtering.
> 
> I agree that for performance sensitive operations, specifically writing
> doorbell registers in the hot paths or RX and TX packet processing, this
> is a good change.
> 
> However, in configuration paths and whatnot, it is much less urgent and
> useful.
> 
> Therefore I think it would work better if you concentrated solely on
> hot code path cases.
> 
> You can, on a driver by driver basis, submit the other transformations
> in the slow paths, and let the driver maintainers decide whether to
> take those on or not.
> 
> Also, please stick exactly to the case where we have:
> 
> 	wmb/mb/etc.
> 	writel()
> 

OK

> Because I see some changes where we have:
> 
> 	writel()
> 
> 	barrier()
> 
> 	writel()
> 

barrier() on ARM is a write barrier. Apparently, it is a compiler barrier
on Intel. I briefly discussed the barrier() behavior in rdma mailing list [1].

Our conclusion is that code should have used wmb() if it really needed
to synchronize memory contents to the device and barrier() is already
wrong. It just guarantees that code doesn't move. writel() already has
a compiler barrier inside. It won't move to begin with.

Like you suggested, we decided to leave these changes alone and even
skip those drivers.

I'll take another look at the patches.

> for exmaple, and you are turning that second writel() into a relaxed
> on as well.  The above is using a compile barrier, not a memory
> barrier, so effectively it is two writel()'s in sequence which is
> not what this patch set is about.
> 
> If anything, that compile barrier() is superfluous and could be
> removed.  But that is also a separate change from what this patch
> series is doing here.
> 

agreed, I'll remove such changes.

> Finally, it makes it that much easier if we can see the preceeding
> memory barrier in the context of the patch that adjusts the writel
> into a writel_relaxed.
> 
> In one case, a macro DOORBELL() is changed to use writel().  This
> makes it so that the patch reviewer has to scan over the entire
> driver in question to see exactly how DOORBELL() is used and whether
> it fits the criteria for the writel_relaxed() transformation.
> 
> I would suggest that you adjust the name of the macro in a situation
> like this, f.e. to DOORBELL_RELAXED().

makes sense.

> 
> Thank you.
> 

[1] https://patchwork.kernel.org/project/LKML/list/?submitter=145491

-- 
Sinan Kaya
Qualcomm Datacenter Technologies, Inc. as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project.