NUMA API for 2.5.12 (1/4)
PATCH 1/1 NUMA API for 2.5.12 (1/4)
Ok all,
I'm going to go ahead and assume (hope) that the no response on the last
posting was because the patch was really large. We'll try this again with 4 smaller patches and see what happens.
This patch implements the NUMA API specified at:
http://lse.sourceforge.net/numa/numa_api.html for the 2.5.12 version of the kernel. The API implements such features as binding processes to CPUs (similar to Robert Love's recent patch), binding to memory blocks, setting launch policies for processes, and rudimentary topology features. The patch is currently used via a prctl() interface with a possible /proc interface in the future. I also have a syscall version if that is preferred.
Again, this is a slightly cleaned-up, and (more) bite-sized version of the
previous patch sent out...
Enjoy, and please send me any comments on the patch, or the API it implements!
-Matt
[email protected]
include/linux/init_task.h +4 -0
@@ -59,6 +59,10 @@ children: LIST_HEAD_INIT(tsk.children), \ sibling: LIST_HEAD_INIT(tsk.sibling), \ thread_group: LIST_HEAD_INIT(tsk.thread_group), \ + numa_restrict: NEW_NUMA_SET, \ + numa_binding: NEW_NUMA_SET, \ + numa_launch_policy: NEW_NUMA_SET, \ + numa_api_lock: RW_LOCK_UNLOCKED, \ wait_chldexit: __WAIT_QUEUE_HEAD_INITIALIZER(tsk.wait_chldexit),\ real_timer: { \ function: it_real_fn \
include/linux/mmzone.h +5 -3
@@ -136,6 +136,7 @@unsigned long node_start_mapnr;unsigned long node_size;int node_id; +int memblk_id;/* A unique ID for each memory block (physical contiguous chunk of memory) */ struct pglist_data *node_next; } pg_data_t;
@@ -163,14 +164,15 @@#define NODE_MEM_MAP (nid) mem_map#define MAX_NR_NODES 1 -#else /* !CONFIG_DISCONTIGMEM */ +#endif /* !CONFIG_DISCONTIGMEM */ -#include <asm/mmzone.h> +#if defined (CONFIG_DISCONTIGMEM ) || defined (CONFIG_NUMA_API ) +#include <asm/mmzone.h> /* page->zone is currently 8 bits ... */ #define MAX_NR_NODES (255 / MAX_NR_ZONES) -#endif /* !CONFIG_DISCONTIGMEM */ +#endif /* CONFIG_DISCONTIGMEM || CONFIG_NUMA_API */ #define MAP_ALIGN (x) ((((x) % sizeof(mem_map_t)) == 0) ? (x) : ((x) + \ sizeof(mem_map_t) - ((x) % sizeof(mem_map_t))))
include/linux/numa.h +76 -0
76 lines not on this page · raw message
include/linux/sched.h +7 -0
@@ -28,6 +28,7 @@#include <linux/securebits.h> #include <linux/fs_struct.h> #include <linux/compiler.h> +#include <linux/numa.h> struct exec_domain ;
@@ -286,6 +287,12 @@struct task_struct *pidhash_next;struct task_struct **pidhash_pprev; +/* additional NUMA stuff */ +numa_set_t numa_restrict; +numa_set_t numa_binding; +numa_set_t numa_launch_policy; +rwlock_t numa_api_lock;/* protects the preceding 3 structs */ +wait_queue_head_t wait_chldexit;/* for wait4() */ struct completion *vfork_done;/* for vfork() */
include/linux/smp.h +1 -0
@@ -90,6 +90,7 @@#define cpu_number_map (cpu) 0#define smp_call_function (func,info,retry,wait) ({ 0; })#define cpu_online_map 1 +#define memblk_online_map 1static inline void smp_send_reschedule (int cpu) { }static inline void smp_send_reschedule_all (void ) { }#define __per_cpu_data
kernel/sched.c +2 -4
@@ -357,7 +357,7 @@runqueue_t *rq;preempt_disable (); - rq =this_rq (); + rq =task_rq (p);spin_lock_irq (&rq->lock); p->state =TASK_RUNNING ;
@@ -371,7 +371,6 @@ p->sleep_avg = p->sleep_avg *CHILD_PENALTY /100 ; p->prio =effective_prio (p); } - p->thread_info->cpu =smp_processor_id ();activate_task (p, rq);spin_unlock_irq (&rq->lock);
@@ -1662,8 +1661,7 @@migration_req_t req;runqueue_t *rq; - new_mask &= cpu_online_map; -if (!new_mask) +if (!(new_mask & cpu_online_map))BUG ();preempt_disable ();