Re: [PATCH] nvmet: introduce use_vfs ns-attr

[Mark Ruijter <MRuijter@onestopsystems.com>]

I will install the patch tomorrow on the systems that I used before and run some tests so that I can compare the performance numbers.

Mark

> Op 27 okt. 2019 om 16:03 heeft "hch@lst.de" <hch@lst.de> het volgende geschreven:
> 
> On Fri, Oct 25, 2019 at 08:44:00AM +0000, Mark Ruijter wrote:
>> 
>> Hi Keith,
>> 
>> I am indeed not using buffered io.
>> Using the VFS increases my 4k random write performance from 200K to 650K when using raid1. 
>> So the difference is huge and becomes more significant when the underlying drives or raid0 can handle more iops.
> 
> Can you try the patch below to use block layer plugging in nvmet?  That
> should be the only major difference in how we do I/O.
> 
>> 1. Currently a controller id collision can occur when using a clustered HA setup. See this message:
>>>>> [1122789.054677] nvme nvme1: Duplicate cntlid 4 with nvme0, rejecting.
>> 
>> The controller ID is currently hard wired.
>> 
>>       ret = ida_simple_get(&cntlid_ida,
>>                             NVME_CNTLID_MIN, NVME_CNTLID_MAX,
>>                             GFP_KERNEL);
>> 
>> So two nodes exporting the exact same volume using the same port configuration can easily come up with the same controller id.
>> I would like to propose to make it configurable, but with the current logic setting a default.
>> SCST for example allows manual target id selection for this reason.
> 
> We can allow some control there using a new configfs file.  But what
> would be even better is an actually integrated cluster manager, which
> we'd need to support features such as persistent reservations.
> 
>> 2. The Model of the drives has been hard wired to Linux. As I see it this should be configurable with 'Linux' as default value.
>> I'll provide code that makes that work.
> 
> Yes, please send a patch.
> 
>> 3. A NVMEoF connected disk on the initiator seems to queue forever when the target dies.
>> It would be nice if we had the ability to select either 'queue foreever' or 'failfast'.
> 
> Making this configurable has been a long time todo list item.  At some
> point in the past Hannes (added to Cc) signed up for it, but it seems
> to have dropped off his priority list.
> 
> ---
> From 87ab0d6f9e092cde04775452131f90e8b4c46a66 Mon Sep 17 00:00:00 2001
> From: Christoph Hellwig <hch@lst.de>
> Date: Sun, 27 Oct 2019 15:59:08 +0100
> Subject: nvmet: use block layer plugging in nvmet_bdev_execute_rw
> 
> Signed-off-by: Christoph Hellwig <hch@lst.de>
> ---
> drivers/nvme/target/io-cmd-bdev.c | 3 +++
> 1 file changed, 3 insertions(+)
> 
> diff --git a/drivers/nvme/target/io-cmd-bdev.c b/drivers/nvme/target/io-cmd-bdev.c
> index 04a9cd2a2604..ed1a8d0ab30e 100644
> --- a/drivers/nvme/target/io-cmd-bdev.c
> +++ b/drivers/nvme/target/io-cmd-bdev.c
> @@ -147,6 +147,7 @@ static void nvmet_bdev_execute_rw(struct nvmet_req *req)
>    int sg_cnt = req->sg_cnt;
>    struct bio *bio;
>    struct scatterlist *sg;
> +    struct blk_plug plug;
>    sector_t sector;
>    int op, op_flags = 0, i;
> 
> @@ -185,6 +186,7 @@ static void nvmet_bdev_execute_rw(struct nvmet_req *req)
>    bio->bi_end_io = nvmet_bio_done;
>    bio_set_op_attrs(bio, op, op_flags);
> 
> +    blk_start_plug(&plug);
>    for_each_sg(req->sg, sg, req->sg_cnt, i) {
>        while (bio_add_page(bio, sg_page(sg), sg->length, sg->offset)
>                != sg->length) {
> @@ -202,6 +204,7 @@ static void nvmet_bdev_execute_rw(struct nvmet_req *req)
>        sector += sg->length >> 9;
>        sg_cnt--;
>    }
> +    blk_finish_plug(&plug);
> 
>    submit_bio(bio);
> }
> -- 
> 2.20.1
> 
_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme