percpu area using gs segment

This speeds up smp again, by storing the struct processor
in a percpu area and avoiding an expensive cpu_number every call
of current_processor(), as well as getting the cpu_number by
an offset into the percpu area.  Untested on 64 bit
and work remains to use other percpu arrays.

TESTED: (NCPUS=8) -smp 1 boots to login shell ~2x slower than uniprocessor
TESTED: (NCPUS=8) -smp 2 boots to INIT but hangs there
TESTED: (NCPUS=8) -smp 4 gets stuck seemingly within rumpdisk and hangs
TESTED: (NCPUS=1) uniprocessor is a bit faster than normal

Message-Id: <20230924103428.455966-3-damien@zamaudio.com>

4 files changed, 11 insertions, 19 deletions

diff --git a/kern/cpu_number.h b/kern/cpu_number.h
index 0be2d338..1abe3dbb 100644
--- a/kern/cpu_number.h
+++ b/kern/cpu_number.h
@@ -37,7 +37,8 @@ extern int	master_cpu;	/* 'master' processor - keeps time */
 
 #if	(NCPUS == 1)
 	/* cpu number is always 0 on a single processor system */
-#define	cpu_number()	(0)
+#define	cpu_number()		(0)
+#define	cpu_number_slow()	(0)
 
 #endif	/* NCPUS == 1 */
 
diff --git a/kern/processor.c b/kern/processor.c
index 2cd6d46c..76735381 100644
--- a/kern/processor.c
+++ b/kern/processor.c
@@ -60,14 +60,12 @@ struct kmem_cache pset_cache;
 int	master_cpu;
 
 struct processor_set default_pset;
-struct processor processor_array[NCPUS];
 
 queue_head_t		all_psets;
 int			all_psets_count;
 def_simple_lock_data(, all_psets_lock);
 
 processor_t	master_processor;
-processor_t	processor_ptr[NCPUS];
 
 /*
  *	Bootstrap the processor/pset system so the scheduler can run.
@@ -81,10 +79,9 @@ void pset_sys_bootstrap(void)
 	for (i = 0; i < NCPUS; i++) {
 		/*
 		 *	Initialize processor data structures.
-		 *	Note that cpu_to_processor(i) is processor_ptr[i].
+		 *	Note that cpu_to_processor is processor_ptr.
 		 */
-		processor_ptr[i] = &processor_array[i];
-		processor_init(processor_ptr[i], i);
+		processor_init(processor_ptr(i), i);
 	}
 	master_processor = cpu_to_processor(master_cpu);
 	queue_init(&all_psets);
diff --git a/kern/processor.h b/kern/processor.h
index 79386627..fc204ffa 100644
--- a/kern/processor.h
+++ b/kern/processor.h
@@ -112,6 +112,7 @@ typedef struct processor Processor;
 extern struct processor	processor_array[NCPUS];
 
 #include <kern/cpu_number.h>
+#include <machine/percpu.h>
 
 /*
  *	Chain of all processor sets.
@@ -196,23 +197,15 @@ extern processor_t	master_processor;
 #define	PROCESSOR_ASSIGN	4	/* Assignment is changing */
 #define PROCESSOR_SHUTDOWN	5	/* Being shutdown */
 
-/*
- *	Use processor ptr array to find current processor's data structure.
- *	This replaces a multiplication (index into processor_array) with
- *	an array lookup and a memory reference.  It also allows us to save
- *	space if processor numbering gets too sparse.
- */
-
-extern processor_t	processor_ptr[NCPUS];
-
-#define cpu_to_processor(i)	(processor_ptr[i])
+#define processor_ptr(i)	(&percpu_array[i].processor)
+#define cpu_to_processor	processor_ptr
 
-#define current_processor()	(processor_ptr[cpu_number()])
+#define current_processor()	(percpu_ptr(struct processor, processor))
 #define current_processor_set()	(current_processor()->processor_set)
 
 /* Compatibility -- will go away */
 
-#define cpu_state(slot_num)	(processor_ptr[slot_num]->state)
+#define cpu_state(slot_num)	(processor_ptr(slot_num)->state)
 #define cpu_idle(slot_num)	(cpu_state(slot_num) == PROCESSOR_IDLE)
 
 /* Useful lock macros */
diff --git a/kern/startup.c b/kern/startup.c
index 2eb3a739..88608c7d 100644
--- a/kern/startup.c
+++ b/kern/startup.c
@@ -74,6 +74,7 @@ boolean_t reboot_on_panic = TRUE;
 
 #if	NCPUS > 1
 #include <machine/mp_desc.h>
+#include <kern/smp.h>
 #include <kern/machine.h>
 #endif	/* NCPUS > 1 */
 
@@ -281,7 +282,7 @@ void cpu_launch_first_thread(thread_t th)
 {
 	int	mycpu;
 
-	mycpu = cpu_number();
+	mycpu = cpu_number_slow();
 
 	cpu_up(mycpu);