Re: [RFC PATCH v2] Workaround for discard on non-conformant nvme devices

[Keith Busch <kbusch@kernel.org>]

On Sun, Nov 10, 2019 at 07:27:08PM +0100, Eduard Hasenleithner wrote:
> +++ linux-5.3.7/drivers/nvme/host/core.c	2019-11-10 19:11:49.419372363 +0100
> @@ -562,8 +562,13 @@ static blk_status_t nvme_setup_discard(s
>  	struct nvme_dsm_range *range;
>  	struct bio *bio;
> 
> -	range = kmalloc_array(segments, sizeof(*range),
> -				GFP_ATOMIC | __GFP_NOWARN);
> +	if (!(ns->ctrl->quirks & NVME_QUIRK_DSM_PAGEALIGN)) {
> +		range = kmalloc_array(segments, sizeof(*range),
> +					GFP_ATOMIC | __GFP_NOWARN);
> +	} else {
> +		/* Device with quirk: use (page aligned) discard_page */
> +		range = NULL;
> +	}

Please initialize 'range' to NULL so we don't need the 'else' case.

> @@ -97,6 +97,11 @@ enum nvme_quirks {
>  	 * Force simple suspend/resume path.
>  	 */
>  	NVME_QUIRK_SIMPLE_SUSPEND		= (1 << 10),
> +
> +	/*
> +	 * Workaround for devices reading DSM beyond end of page.
> +	 */
> +	NVME_QUIRK_DSM_PAGEALIGN		= (1 << 11),

Let's get a better description of this. My recommendation:

	For devices that do not consider the DSM 'Number of Ranges'
	field when determining how much data to DMA. Page aligned and
	sized is always sufficient as that is the largest a DSM range
	list can be.

And let's add the known devices to the nvme pci_driver that require
this.

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