From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001 From: Quentin Perret Date: Tue, 24 Nov 2020 15:33:49 +0000 Subject: NOUPSTREAM: ANDROID: sched: Track wake_q length Some partners have value-adds based on aosp/540066, which cannot be carried in ACK in its entirety as it no longer makes sense as-is (the select_idle_capacity() rework upstream solved the issue differently). It seems that those partners do not actually need the wake-wide tweaks, they only need to access the wake_q length for wake-up balance. To support this, add minimal tracking to the wake_q infrastructure in the core kernel, but do that by adding a pointer to the wake_q_head to task_struct directly to not litter all sched classes with an additional sibling_count_hint argument to the select_task_rq callbacks. Modules needing to access the wake_q length can do so by dereferencing p->wake_q_head in the wake-up path when it is non-NULL. [CPNOTE: 30/06/21] Lee: Vendor specific reference counting Bug: 173981591 Signed-off-by: Quentin Perret Change-Id: I9a98167face92e70aba847d9f04d0c216065478c Signed-off-by: Lee Jones --- include/linux/sched.h | 1 + include/linux/sched/wake_q.h | 2 ++ 2 files changed, 3 insertions(+) diff --git a/include/linux/sched.h b/include/linux/sched.h --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -1139,6 +1139,7 @@ struct task_struct { raw_spinlock_t pi_lock; struct wake_q_node wake_q; + int wake_q_count; #ifdef CONFIG_RT_MUTEXES /* PI waiters blocked on a rt_mutex held by this task: */ diff --git a/include/linux/sched/wake_q.h b/include/linux/sched/wake_q.h --- a/include/linux/sched/wake_q.h +++ b/include/linux/sched/wake_q.h @@ -38,6 +38,7 @@ struct wake_q_head { struct wake_q_node *first; struct wake_q_node **lastp; + int count; }; #define WAKE_Q_TAIL ((struct wake_q_node *) 0x01) @@ -52,6 +53,7 @@ static inline void wake_q_init(struct wake_q_head *head) { head->first = WAKE_Q_TAIL; head->lastp = &head->first; + head->count = 0; } static inline bool wake_q_empty(struct wake_q_head *head)