RE: [PATCH RFC] rds: ib: Reduce the contention caused by the asynchronous workers to flush the mr pool

[Praveen Kannoju]

-----Original Message-----
From: Jason Gunthorpe [mailto:jgg@xxxxxxxx] 
Sent: 19 January 2022 06:35 PM
To: Praveen Kannoju <praveen.kannoju@xxxxxxxxxx>
Cc: Leon Romanovsky <leon@xxxxxxxxxx>; Santosh Shilimkar <santosh.shilimkar@xxxxxxxxxx>; David S . Miller <davem@xxxxxxxxxxxxx>; kuba@xxxxxxxxxx; netdev@xxxxxxxxxxxxxxx; linux-rdma@xxxxxxxxxxxxxxx; rds-devel@xxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; Rama Nichanamatlu <rama.nichanamatlu@xxxxxxxxxx>; Rajesh Sivaramasubramaniom <rajesh.sivaramasubramaniom@xxxxxxxxxx>
Subject: Re: [PATCH RFC] rds: ib: Reduce the contention caused by the asynchronous workers to flush the mr pool

On Wed, Jan 19, 2022 at 11:46:16AM +0000, Praveen Kannoju wrote:
 
> 6. Jason, the only function "rds_ib_free_mr" which accesses the 
> introduced bool variable "flush_ongoing" to spawn a flush worker does 
> not crucially impact the availability of MR's, because the flush 
> happens from allocation path as well when necessary.  Hence the 
> Load-store ordering is not essentially needed here, because of which 
> we chose smp_rmb() and smp_wmb() over smp_load_acquire() and 
> smp_store_release().

That seems like a confusing statement, you added barriers which do the same things as acquire/release then say you didn't need acquire release?

I think this is using barriers wrong.

Jason

Jason, 

Yes, we are using the barriers. I was justifying the usage of smp_rmb() and smp_wmb() over smp_load_acquire() and smp_store_release() in the patch.