Re: [PATCH v4 4/5] async: Add support for queueing on specific node

From: Dan Williams
Date: Fri Sep 21 2018 - 10:57:36 EST


On Thu, Sep 20, 2018 at 3:31 PM Alexander Duyck
<alexander.h.duyck@xxxxxxxxxxxxxxx> wrote:
>
> This patch introduces two new variants of the async_schedule_ functions
> that allow scheduling on a specific node. These functions are
> async_schedule_on and async_schedule_on_domain which end up mapping to
> async_schedule and async_schedule_domain but provide NUMA node specific
> functionality. The original functions were moved to inline function
> definitions that call the new functions while passing NUMA_NO_NODE.
>
> The main motivation behind this is to address the need to be able to
> schedule NVDIMM init work on specific NUMA nodes in order to improve
> performance of memory initialization.
>
> One additional change I made is I dropped the "extern" from the function
> prototypes in the async.h kernel header since they aren't needed.
>
> Signed-off-by: Alexander Duyck <alexander.h.duyck@xxxxxxxxxxxxxxx>
> ---
> include/linux/async.h | 20 +++++++++++++++++---
> kernel/async.c | 36 +++++++++++++++++++++++++-----------
> 2 files changed, 42 insertions(+), 14 deletions(-)
>
> diff --git a/include/linux/async.h b/include/linux/async.h
> index 6b0226bdaadc..9878b99cbb01 100644
> --- a/include/linux/async.h
> +++ b/include/linux/async.h
> @@ -14,6 +14,7 @@
>
> #include <linux/types.h>
> #include <linux/list.h>
> +#include <linux/numa.h>
>
> typedef u64 async_cookie_t;
> typedef void (*async_func_t) (void *data, async_cookie_t cookie);
> @@ -37,9 +38,22 @@ struct async_domain {
> struct async_domain _name = { .pending = LIST_HEAD_INIT(_name.pending), \
> .registered = 0 }
>
> -extern async_cookie_t async_schedule(async_func_t func, void *data);
> -extern async_cookie_t async_schedule_domain(async_func_t func, void *data,
> - struct async_domain *domain);
> +async_cookie_t async_schedule_on(async_func_t func, void *data, int node);
> +async_cookie_t async_schedule_on_domain(async_func_t func, void *data, int node,
> + struct async_domain *domain);

I would expect this to take a cpu instead of a node to not surprise
users coming from queue_work_on() / schedule_work_on()...

> +
> +static inline async_cookie_t async_schedule(async_func_t func, void *data)
> +{
> + return async_schedule_on(func, data, NUMA_NO_NODE);
> +}
> +
> +static inline async_cookie_t
> +async_schedule_domain(async_func_t func, void *data,
> + struct async_domain *domain)
> +{
> + return async_schedule_on_domain(func, data, NUMA_NO_NODE, domain);
> +}
> +
> void async_unregister_domain(struct async_domain *domain);
> extern void async_synchronize_full(void);
> extern void async_synchronize_full_domain(struct async_domain *domain);
> diff --git a/kernel/async.c b/kernel/async.c
> index a893d6170944..1d7ce81c1949 100644
> --- a/kernel/async.c
> +++ b/kernel/async.c
> @@ -56,6 +56,7 @@ synchronization with the async_synchronize_full() function, before returning
> #include <linux/sched.h>
> #include <linux/slab.h>
> #include <linux/workqueue.h>
> +#include <linux/cpu.h>
>
> #include "workqueue_internal.h"
>
> @@ -149,8 +150,11 @@ static void async_run_entry_fn(struct work_struct *work)
> wake_up(&async_done);
> }
>
> -static async_cookie_t __async_schedule(async_func_t func, void *data, struct async_domain *domain)
> +static async_cookie_t __async_schedule(async_func_t func, void *data,
> + struct async_domain *domain,
> + int node)
> {
> + int cpu = WORK_CPU_UNBOUND;
> struct async_entry *entry;
> unsigned long flags;
> async_cookie_t newcookie;
> @@ -194,30 +198,40 @@ static async_cookie_t __async_schedule(async_func_t func, void *data, struct asy
> /* mark that this task has queued an async job, used by module init */
> current->flags |= PF_USED_ASYNC;
>
> + /* guarantee cpu_online_mask doesn't change during scheduling */
> + get_online_cpus();
> +
> + if (node >= 0 && node < MAX_NUMNODES && node_online(node))
> + cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);

...I think this node to cpu helper should be up-leveled for callers. I
suspect using get_online_cpus() may cause lockdep problems to take the
cpu_hotplug_lock() within a "do_something_on()" routine. For example,
I found this when auditing queue_work_on() users:

/*
* Doesn't need any cpu hotplug locking because we do rely on per-cpu
* kworkers being shut down before our page_alloc_cpu_dead callback is
* executed on the offlined cpu.
* Calling this function with cpu hotplug locks held can actually lead
* to obscure indirect dependencies via WQ context.
*/
void lru_add_drain_all(void)

I think it's a gotcha waiting to happen if async_schedule_on() has
more restrictive calling contexts than queue_work_on().