[PATCH v3] nvme_core: scan namespaces asynchronously

[Sagi Grimberg]

On 15/07/2024 23:34, Stuart Hayes wrote:
> Use async function calls to make namespace scanning happen in parallel.
>
> Without the patch, NVME namespaces are scanned serially, so it can take
> a long time for all of a controller's namespaces to become available,
> especially with a slower (TCP) interface with large number of
> namespaces.
>
> It is not uncommon to have large numbers (hundreds or thousands) of
> namespaces on nvme-of with storage servers.
>
> The time it took for all namespaces to show up after connecting (via
> TCP) to a controller with 1002 namespaces was measured on one system:
>
> network latency   without patch   with patch
>       0                 6s            1s
>      50ms             210s           10s
>     100ms             417s           18s
>
> Measurements taken on another system show the effect of the patch on the
> time nvme_scan_work() took to complete, when connecting to a linux
> nvme-of target with varying numbers of namespaces, on a network of
> 400us.
>
> namespaces    without patch   with patch
>       1            16ms           14ms
>       2            24ms           16ms
>       4            49ms           22ms
>       8           101ms           33ms
>      16           207ms           56ms
>     100           1.4s           0.6s
>    1000          12.9s           2.0s

Not sure how common is the 1000 namespaces use-case, but the dozens of 
namespaces
seems compelling enough.

>
> On the same system, connecting to a local PCIe NVMe drive (a Samsung
> PM1733) instead of a network target:
>
> namespaces    without patch   with patch
>       1            13ms           12ms
>       2            41ms           13ms
>
> Signed-off-by: Stuart Hayes <stuart.w.hayes at gmail.com>
> ---
> changes from V2:
>    * make a separate function nvme_scan_ns_async() that calls
>      nvme_scan_ns(), instead of modifying nvme_scan_ns()
>    * only scan asynchronously from nvme_scan_ns_list(), not from
>      nvme_scan_ns_sequential()
>    * provide more timing data in the commit message
>
> changes from V1:
>    * remove module param to enable/disable async scanning
>    * add scan time measurements to commit message
>
>
>   drivers/nvme/host/core.c | 48 +++++++++++++++++++++++++++++++++-------
>   1 file changed, 40 insertions(+), 8 deletions(-)
>
> diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
> index 782090ce0bc1..dbf05cfea063 100644
> --- a/drivers/nvme/host/core.c
> +++ b/drivers/nvme/host/core.c
> @@ -4,6 +4,7 @@
>    * Copyright (c) 2011-2014, Intel Corporation.
>    */
>   
> +#include <linux/async.h>
>   #include <linux/blkdev.h>
>   #include <linux/blk-mq.h>
>   #include <linux/blk-integrity.h>
> @@ -3952,6 +3953,30 @@ static void nvme_scan_ns(struct nvme_ctrl *ctrl, unsigned nsid)
>   	}
>   }
>   
> +/*
> + * struct async_scan_info - keeps track of controller & NSIDs to scan
> + * @ctrl:	Controller on which namespaces are being scanned
> + * @next_idx:	Index of next NSID to scan in ns_list
> + * @ns_list:	Pointer to list of NSIDs to scan
> + */
> +struct async_scan_info {
> +	struct nvme_ctrl *ctrl;
> +	atomic_t next_idx;

next_nsid ?

> +	__le32 *ns_list;
> +};
> +
> +static void nvme_scan_ns_async(void *data, async_cookie_t cookie)
> +{
> +	struct async_scan_info *scan_info = data;
> +	int idx;
> +	u32 nsid;
> +
> +	idx = (u32)atomic_fetch_add(1, &scan_info->next_idx);
> +	nsid = le32_to_cpu(scan_info->ns_list[idx]);
> +
> +	nvme_scan_ns(scan_info->ctrl, nsid);
> +}
> +
>   static void nvme_remove_invalid_namespaces(struct nvme_ctrl *ctrl,
>   					unsigned nsid)
>   {
> @@ -3975,12 +4000,14 @@ static void nvme_remove_invalid_namespaces(struct nvme_ctrl *ctrl,
>   static int nvme_scan_ns_list(struct nvme_ctrl *ctrl)
>   {
>   	const int nr_entries = NVME_IDENTIFY_DATA_SIZE / sizeof(__le32);
> -	__le32 *ns_list;
> +	struct async_scan_info scan_info;

What initializes next_idx?

>   	u32 prev = 0;
>   	int ret = 0, i;
> +	ASYNC_DOMAIN(domain);
>   
> -	ns_list = kzalloc(NVME_IDENTIFY_DATA_SIZE, GFP_KERNEL);
> -	if (!ns_list)
> +	scan_info.ctrl = ctrl;
> +	scan_info.ns_list = kzalloc(NVME_IDENTIFY_DATA_SIZE, GFP_KERNEL);
> +	if (!scan_info.ns_list)
>   		return -ENOMEM;

I think you can leave the local variable ns_list as is, and just assign 
it to scan_info
after, its common practice to allocate to a local pointer and use it to 
init a struct member.

Plus it will make the patch diff simpler.

>   
>   	for (;;) {
> @@ -3990,28 +4017,33 @@ static int nvme_scan_ns_list(struct nvme_ctrl *ctrl)
>   			.identify.nsid		= cpu_to_le32(prev),
>   		};
>   
> -		ret = nvme_submit_sync_cmd(ctrl->admin_q, &cmd, ns_list,
> -					    NVME_IDENTIFY_DATA_SIZE);
> +		ret = nvme_submit_sync_cmd(ctrl->admin_q, &cmd,
> +					   scan_info.ns_list,
> +					   NVME_IDENTIFY_DATA_SIZE);
>   		if (ret) {
>   			dev_warn(ctrl->device,
>   				"Identify NS List failed (status=0x%x)\n", ret);
>   			goto free;
>   		}
>   
> +		atomic_set(&scan_info.next_idx, 0);
>   		for (i = 0; i < nr_entries; i++) {
> -			u32 nsid = le32_to_cpu(ns_list[i]);
> +			u32 nsid = le32_to_cpu(scan_info.ns_list[i]);
>   
>   			if (!nsid)	/* end of the list? */
>   				goto out;
> -			nvme_scan_ns(ctrl, nsid);
> +			async_schedule_domain(nvme_scan_ns_async, &scan_info,
> +						&domain);
>   			while (++prev < nsid)
>   				nvme_ns_remove_by_nsid(ctrl, prev);
>   		}
> +		async_synchronize_full_domain(&domain);
>   	}
>    out:
>   	nvme_remove_invalid_namespaces(ctrl, prev);
>    free:
> -	kfree(ns_list);
> +	async_synchronize_full_domain(&domain);
> +	kfree(scan_info.ns_list);
>   	return ret;
>   }
>