Re: [RFC PATCH 4/4 v0.3] sched/umcg: RFC: implement UMCG syscalls

[Peter Oskolkov]

On Mon, Jul 19, 2021 at 9:07 AM Thierry Delisle <tdelisle@xxxxxxxxxxxx> wrote:
>
>  > /**
>  >  * @idle_servers_ptr: a single-linked list pointing to the list
>  >  *                    of idle servers. Can be NULL.
>  >  *
>  >  * Readable/writable by both the kernel and the userspace: the
>  >  * userspace adds items to the list, the kernel removes them.
>  >  *
>  >  * This is a single-linked list (stack): head->next->next->next->NULL.
>  >  * "next" nodes are idle_servers_ptr fields in struct umcg_task.
>  >  *
>  >  * Example:
>  >  *
>  >  *  a running worker             idle server 1        idle server 2
>  >  *
>  >  * struct umct_task:             struct umcg_task:    struct umcg_task:
>  >  *    state                         state state
>  >  *    api_version                   api_version api_version
>  >  *    ...                           ...                  ...
>  >  *    idle_servers_ptr --> head --> idle_servers_ptr -->
> idle_servers_ptr --> NULL
>  >  *    ...                           ...                  ...
>  >  *
>  >  *
>  >  * Due to the way struct umcg_task is aligned, idle_servers_ptr
>  >  * is aligned at 8 byte boundary, and so has its first byte as zero
>  >  * when it holds a valid pointer.
>  >  *
>  >  * When pulling idle servers from the list, the kernel marks nodes as
>  >  * "deleted" by ORing the node value (the pointer) with 1UL atomically.
>  >  * If a node is "deleted" (i.e. its value AND 1UL is not zero),
>  >  * the kernel proceeds to the next node.
>  >  *
>  >  * The kernel checks at most [nr_cpu_ids * 2] first nodes in the list.
>  >  *
>  >  * It is NOT considered an error if the kernel cannot find an idle
>  >  * server.
>  >  *
>  >  * The userspace is responsible for cleanup/gc (i.e. for actually
>  >  * removing nodes marked as "deleted" from the list).
>  >  */
>  > uint64_t    idle_servers_ptr;    /* r/w */
>
> I don't understand the reason for using this ad-hoc scheme, over using a
> simple
> eventfd to do the job. As I understand it, the goal here is to let
> servers that
> cannot find workers to run, block instead of spinning. Isn't that
> exactly what
> the eventfd interface is for?

Latency/efficiency: on worker wakeup an idle server can be picked from
the list and context-switched into synchronously, on the same CPU.
Using FDs and select/poll/epoll will add extra layers of abstractions;
synchronous context-switches (not yet fully implemented in UMCG) will
most likely be impossible. This patchset seems much more efficient and
lightweight than whatever can be built on top of FDs.

>
> Have you considered an idle_fd field, the kernel writes 1 to the fd when a
> worker is appended to the idle_workers_ptr? Servers that don't find work can
> read the fd or alternatively use select/poll/epoll. Multiple workers are
> expected to share fds, either a single global fd, one fd per server, or any
> other combination the scheduler may fancy.
>