RE: [PATCH 1/1] nvme-fabrics: Generate uevent on namespace AENs

From: "Rao, Vinay" <Vinay.Rao@dell.com>
To: Christoph Hellwig <hch@lst.de>, Keith Busch <kbusch@kernel.org>
Cc: Chaitanya Kulkarni <Chaitanya.Kulkarni@wdc.com>,
	"Belanger, Martin" <Martin.Belanger@dell.com>,
	"Hayes, Stuart" <Stuart.Hayes@dell.com>,
	"Tarikere, Madhu" <Madhu.Tarikere@dell.com>,
	Martin Belanger <nitram_67@hotmail.com>,
	"linux-nvme@lists.infradead.org" <linux-nvme@lists.infradead.org>,
	"axboe@fb.com" <axboe@fb.com>,
	"sagi@grimberg.me" <sagi@grimberg.me>
Subject: RE: [PATCH 1/1] nvme-fabrics: Generate uevent on namespace AENs
Date: Mon, 3 May 2021 07:20:57 +0000	[thread overview]
Message-ID: <BN8PR19MB29290320440FE9740E699BAC8F5B9@BN8PR19MB2929.namprd19.prod.outlook.com> (raw)
In-Reply-To: <20210503065235.GB32627@lst.de>

Forcing RAE bit on all the log pages that kernel cares about is one of the option.

Can we also think also think about an another mechanism where the kernel would generate uevent after kernel does a get log page. This would make kernel always to get the get log data first from the controller. 

-----Original Message-----
From: Christoph Hellwig <hch@lst.de> 
Sent: Monday, May 3, 2021 12:23 PM
To: Keith Busch
Cc: Rao, Vinay; Chaitanya Kulkarni; Belanger, Martin; Hayes, Stuart; Tarikere, Madhu; Christoph Hellwig; Martin Belanger; linux-nvme@lists.infradead.org; axboe@fb.com; sagi@grimberg.me
Subject: Re: [PATCH 1/1] nvme-fabrics: Generate uevent on namespace AENs

[EXTERNAL EMAIL] 

On Fri, Apr 30, 2021 at 09:21:04AM -0700, Keith Busch wrote:
> On Fri, Apr 30, 2021 at 05:14:55AM +0000, Rao, Vinay wrote:
> > + Madhu
> > 
> > Hi Chaitanya,
> > 
> > Some of the AEN's like ANA state change and Name Space attribute change events are of interest outside the NVME driver. 
> > 
> > It would be good to consider these events to be propagated up to user space. I am not yet convinced on how propagating this above to user space is problematic.
> 
> The namespace attribute change event pairs with the Changed Namespace 
> List log, and reading that log will change the result for a subsequent 
> reader. User space racing with the kernel on a log access when there 
> are read side effects creates non-deterministic behavior, and that is 
> problematic.

The only way I could think of making this work is by:

forcing the RAE bit on for all log pages that the kernel cares about, and only delivering the uevent on controllers that actually do support the RAE bit (IIRC it got added in 1.3, but I'd have to check).

_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme