Skip to contentSkip to conversation

O(1) MQ scheduler J9 patch for IA64

Linux-ia64 David Mosberger 3 participants 4 messages Feb 8, 2002

lore
Share & tools
lore
#0 lore raw

Erich> I had to add a function smp_call_function_nowait which sends Erich> an IPI and returns immediately in order to get rid of a Erich> IA64-specific race condition when the mmu_context wraps Erich> around. Now it makes a very stable and fast impression, Erich> interactive work is even possible while doing a "hackbench Erich> 55" and ping flooding the machine... Tested on 2 CPU BigSur, Erich> 4 CPU LION, 16 CPU AzusA.

Hmmh, I'm a bit worried that this is a fragile solution. It would break if there ever was a scenario where a task migrated from one CPU to another while the target CPU has interrupts disabled. I do not think this can happen with Ingo's current scheduler but I'd prefer a more robust solution.

A somewhat cheesy solution might be to have a per-CPU flag that indicates whether a CPU should flush its TLB before switching to the next task. We could check for this in activate_mm(). Do you want to try this?

	--david
#1 lore raw
Quoted 4 lines · David Mosberger

A somewhat cheesy solution might be to have a per-CPU flag that indicates whether a CPU should flush its TLB before switching to the next task. We could check for this in activate_mm(). Do you want to try this?

Would this be a good thing to put in the mmu_context? It seems like that could isolate it within arch-specific code.

Cheers,
Bill

#2 lore raw
In reply to David Mosberger
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · William Lee Irwin III

Bill> On Thu, Feb 07, 2002 at 10:11:28PM -0800, David Mosberger

Quoted 4 lines · Bill> wrote:
Quoted 4 lines

A somewhat cheesy solution might be to have a per-CPU flag that indicates whether a CPU should flush its TLB before switching to the next task. We could check for this in activate_mm(). Do you want to try this?

Bill> Would this be a good thing to put in the mmu_context? It seems Bill> like that could isolate it within arch-specific code.

It's arch-specific either way. The flag needs to be per CPU, not per task.

	--david
#3 lore raw

[Linux-ia64] Re: O(1) MQ scheduler J9 patch for IA64 -> K3 patch

In reply to David Mosberger
Quoted 14 lines · David Mosberger

Erich> I had to add a function smp_call_function_nowait which sends Erich> an IPI and returns immediately in order to get rid of a Erich> IA64-specific race condition when the mmu_context wraps Erich> around. Now it makes a very stable and fast impression, Erich> interactive work is even possible while doing a "hackbench Erich> 55" and ping flooding the machine... Tested on 2 CPU BigSur, Erich> 4 CPU LION, 16 CPU AzusA.

Hmmh, I'm a bit worried that this is a fragile solution. It would break if there ever was a scenario where a task migrated from one CPU to another while the target CPU has interrupts disabled. I do not think this can happen with Ingo's current scheduler but I'd prefer a more robust solution.

Quoted 4 lines

A somewhat cheesy solution might be to have a per-CPU flag that indicates whether a CPU should flush its TLB before switching to the next task. We could check for this in activate_mm(). Do you want to try this?

OK, I also feel better with the new solution. Didn't want to put it into the runqueues because it really is an IA-64 problem. It's architecture-specific and the CPU flags are in ia64_ctx. The context wrapping is so seldom that the single lock protecting the CPU mask shouldn't matter. But if somebody decides to use this mechanism generally for flush_tlb_all() we'd have to change it...

Besides this: the attached patch is for the latest version of the O(1) scheduler (K3). It has a different sys_sched_yield() from Ingo's because that one has a problem with more than 8 CPUs. Also the set_cpus_allowed() function has the limitation that it just works if the current process tries to change its mask.

Best regards,
Erich

arch/ia64/kdb/kdba_bt.c +1 -1
diff -urN 2.4.17-ia64-kdbv2.1-K3/arch/ia64/kdb/kdba_bt.c 2.4.17-ia64-kdbv2.1-K3ia64/arch/ia64/kdb/kdba_bt.c
@@ -197,7 +197,7 @@
 	}
 #ifdef CONFIG_SMP
 	else if (task_has_cpu(p)) {
-		sw = kdb_sw[p->processor];
+		sw = kdb_sw[p->cpu];
 	}
 #endif
 	else {
arch/ia64/kernel/entry.S +4 -2
diff -urN 2.4.17-ia64-kdbv2.1-K3/arch/ia64/kernel/entry.S 2.4.17-ia64-kdbv2.1-K3ia64/arch/ia64/kernel/entry.S
@@ -161,7 +161,8 @@
 	mov r8=r13			// return pointer to previously running task
 	mov r13=in0			// set "current" pointer
 	;;
-(p6)	ssm psr.i			// renable psr.i AFTER the ic bit is serialized
+//(p6)	ssm psr.i			// interrupt delivery should not be enabled
+					// with the new O(1) MQ scheduler
 	DO_LOAD_SWITCH_STACK
 
 #ifdef CONFIG_SMP
@@ -170,7 +171,8 @@
 	br.ret.sptk.many rp		// boogie on out in new context
 
 .map:
-	rsm psr.i | psr.ic
+	//rsm psr.i | psr.ic
+	rsm psr.ic
 	movl r25=PAGE_KERNEL
 	;;
 	srlz.d
arch/ia64/kernel/irq_ia64.c +9 -0
diff -urN 2.4.17-ia64-kdbv2.1-K3/arch/ia64/kernel/irq_ia64.c 2.4.17-ia64-kdbv2.1-K3ia64/arch/ia64/kernel/irq_ia64.c
@@ -148,6 +148,14 @@
 	flags:		SA_INTERRUPT,
 	name:		"IPI"
 };
+
+extern void smp_task_migration_interrupt(int irq, void *dev_id, struct pt_regs *regs);
+
+static struct irqaction task_migration_irqaction = {
+	handler:	smp_task_migration_interrupt,
+	flags:		SA_INTERRUPT,
+	name:		"Task migration"
+};
 #endif
 
 void
@@ -172,6 +180,7 @@
 	register_percpu_irq(IA64_SPURIOUS_INT_VECTOR, NULL);
 #ifdef CONFIG_SMP
 	register_percpu_irq(IA64_IPI_VECTOR, &ipi_irqaction);
+	register_percpu_irq(IA64_TASK_MIGRATION, &task_migration_irqaction);
 #endif
 	platform_irq_init();
 }
arch/ia64/kernel/process.c +1 -5
diff -urN 2.4.17-ia64-kdbv2.1-K3/arch/ia64/kernel/process.c 2.4.17-ia64-kdbv2.1-K3ia64/arch/ia64/kernel/process.c
@@ -125,9 +125,6 @@
 cpu_idle (void *unused)
 {
 	/* endless idle loop with no priority at all */
-	init_idle();
-	current->nice = 20;
-	current->counter = -100;
 
 
 	while (1) {
@@ -136,11 +133,10 @@
 			min_xtp();
 #endif
 
-		while (!current->need_resched) {
+		if (!current->need_resched) {
 #ifdef CONFIG_IA64_SGI_SN
 			snidle();
 #endif
-			continue;
 		}
 
 #ifdef CONFIG_IA64_SGI_SN
arch/ia64/kernel/setup.c +3 -3
diff -urN 2.4.17-ia64-kdbv2.1-K3/arch/ia64/kernel/setup.c 2.4.17-ia64-kdbv2.1-K3ia64/arch/ia64/kernel/setup.c
@@ -375,10 +375,10 @@
 {
 #ifdef CONFIG_SMP
 #	define lpj	c->loops_per_jiffy
-#	define cpu	c->processor
+#	define cpum	c->processor
 #else
 #	define lpj	loops_per_jiffy
-#	define cpu	0
+#	define cpum	0
 #endif
 	char family[32], features[128], *cp;
 	struct cpuinfo_ia64 *c = v;
@@ -417,7 +417,7 @@
 		   "cpu MHz    : %lu.%06lu\n"
 		   "itc MHz    : %lu.%06lu\n"
 		   "BogoMIPS   : %lu.%02lu\n\n",
-		   cpu, c->vendor, family, c->model, c->revision, c->archrev,
+		   cpum, c->vendor, family, c->model, c->revision, c->archrev,
 		   features, c->ppn, c->number,
 		   c->proc_freq / 1000000, c->proc_freq % 1000000,
 		   c->itc_freq / 1000000, c->itc_freq % 1000000,
arch/ia64/kernel/smp.c +36 -0
diff -urN 2.4.17-ia64-kdbv2.1-K3/arch/ia64/kernel/smp.c 2.4.17-ia64-kdbv2.1-K3ia64/arch/ia64/kernel/smp.c
@@ -200,6 +200,42 @@
 }
 
 void
+smp_send_reschedule_all(void)
+{
+	send_IPI_all(IA64_IPI_RESCHEDULE);
+}
+
+static spinlock_t migration_lock = SPIN_LOCK_UNLOCKED;
+static task_t *new_task;
+
+
+/*
+ * This function sends a 'task migration' IPI to another CPU.
+ * Must be called from syscall contexts, with interrupts *enabled*.
+ */
+void smp_migrate_task(int cpu, task_t *p)
+{
+	/*
+	 * The target CPU will unlock the migration spinlock:
+	 */
+	spin_lock(&migration_lock);
+	new_task = p;
+	platform_send_ipi(cpu, IA64_TASK_MIGRATION, IA64_IPI_DM_INT, 0);
+}
+
+/*
+ * Task migration callback.
+ */
+asmlinkage void smp_task_migration_interrupt(int irq, void *dev_id, struct pt_regs *regs)
+{
+	task_t *p;
+
+	p = new_task;
+	spin_unlock(&migration_lock);
+	sched_task_migrated(p);
+}
+
+void
 smp_flush_tlb_all (void)
 {
 	smp_call_function ((void (*)(void *))__flush_tlb_all,0,1,1);
arch/ia64/kernel/smpboot.c +11 -5
diff -urN 2.4.17-ia64-kdbv2.1-K3/arch/ia64/kernel/smpboot.c 2.4.17-ia64-kdbv2.1-K3ia64/arch/ia64/kernel/smpboot.c
@@ -356,6 +356,7 @@
 	local_irq_enable();
 	calibrate_delay();
 	local_cpu_data->loops_per_jiffy = loops_per_jiffy;
+	ia64_disable_timer();
 	/*
 	 * Allow the master to continue.
 	 */
@@ -379,7 +380,8 @@
 	Dprintk("CPU %d is set to go.\n", smp_processor_id());
 	while (!atomic_read(&smp_commenced))
 		;
-
+	/* reenable timer interrupts */
+	ia64_cpu_local_tick();
 	Dprintk("CPU %d is starting idle.\n", smp_processor_id());
 	return cpu_idle();
 }
@@ -416,11 +418,10 @@
 	if (!idle)
 		panic("No idle process for CPU %d", cpu);
 
-	task_set_cpu(idle, cpu);	/* we schedule the first task manually */
+	init_idle(idle, cpu);
 
 	ia64_cpu_to_sapicid[cpu] = sapicid;
 
-	del_from_runqueue(idle);
 	unhash_process(idle);
 	init_tasks[cpu] = idle;
 
@@ -481,8 +482,7 @@
 	printk("Boot processor id 0x%x/0x%x\n", 0, boot_cpu_id);
 
 	global_irq_holder = 0;
-	current->processor = 0;
-	init_idle();
+	current->cpu = 0;
 
 	/*
 	 * If SMP should be disabled, then really disable it!
@@ -569,3 +569,9 @@
 		smp_num_cpus = 1;
 	}
 }
+
+/* Number of ticks we consider an idle tasks still cache-hot.
+ * For Itanium: with 1GB/s bandwidth we need 4ms to fill up 4MB L3 cache...
+ * So let's try 10 ticks.
+ */
+unsigned long cache_decay_ticks=10;
arch/ia64/kernel/time.c +7 -1
diff -urN 2.4.17-ia64-kdbv2.1-K3/arch/ia64/kernel/time.c 2.4.17-ia64-kdbv2.1-K3ia64/arch/ia64/kernel/time.c
@@ -209,7 +209,7 @@
 /*
  * Encapsulate access to the itm structure for SMP.
  */
-void __init
+void
 ia64_cpu_local_tick (void)
 {
 	int cpu = smp_processor_id();
@@ -298,3 +298,9 @@
 	efi_gettimeofday((struct timeval *) &xtime);
 	ia64_init_itm();
 }
+
+void __init ia64_disable_timer(void)
+{
+	ia64_set_itv(IA64_TIMER_VECTOR | IA64_TIMER_MASK);
+}
+
arch/ia64/mm/fault.c +1 -2
diff -urN 2.4.17-ia64-kdbv2.1-K3/arch/ia64/mm/fault.c 2.4.17-ia64-kdbv2.1-K3ia64/arch/ia64/mm/fault.c
@@ -194,8 +194,7 @@
   out_of_memory:
 	up_read(&mm->mmap_sem);
 	if (current->pid == 1) {
-		current->policy |= SCHED_YIELD;
-		schedule();
+		yield();
 		down_read(&mm->mmap_sem);
 		goto survive;
 	}
arch/ia64/mm/tlb.c +5 -2
diff -urN 2.4.17-ia64-kdbv2.1-K3/arch/ia64/mm/tlb.c 2.4.17-ia64-kdbv2.1-K3ia64/arch/ia64/mm/tlb.c
@@ -37,7 +37,8 @@
 	lock:	SPIN_LOCK_UNLOCKED,
 	next:	1,
 	limit:	(1 << 15) - 1,		/* start out with the safe (architected) limit */
-	max_ctx: ~0U
+	max_ctx: ~0U,
+	flush: 0UL
 };
 
 /*
@@ -76,7 +77,9 @@
 			ia64_ctx.limit = tsk_context;
 	}
 	read_unlock(&tasklist_lock);
-	flush_tlb_all();
+	//flush_tlb_all(); /* potential race condition with O(1) scheduler [EF] */
+	ia64_ctx.flush = ((1<<smp_num_cpus) - 1) & ~(1UL<<smp_processor_id());
+	__flush_tlb_all();
 }
 
 static inline void
arch/ia64/tools/print_offsets.c +1 -1
diff -urN 2.4.17-ia64-kdbv2.1-K3/arch/ia64/tools/print_offsets.c 2.4.17-ia64-kdbv2.1-K3ia64/arch/ia64/tools/print_offsets.c
@@ -54,7 +54,7 @@
     { "IA64_TASK_PTRACE_OFFSET",	offsetof (struct task_struct, ptrace) },
     { "IA64_TASK_SIGPENDING_OFFSET",	offsetof (struct task_struct, sigpending) },
     { "IA64_TASK_NEED_RESCHED_OFFSET",	offsetof (struct task_struct, need_resched) },
-    { "IA64_TASK_PROCESSOR_OFFSET",	offsetof (struct task_struct, processor) },
+    { "IA64_TASK_PROCESSOR_OFFSET",	offsetof (struct task_struct, cpu) },
     { "IA64_TASK_THREAD_OFFSET",	offsetof (struct task_struct, thread) },
     { "IA64_TASK_THREAD_KSP_OFFSET",	offsetof (struct task_struct, thread.ksp) },
 #ifdef CONFIG_PERFMON
include/asm-ia64/bitops.h +60 -0

80 lines not on this page · raw message

include/asm-ia64/hw_irq.h +4 -0
diff -urN 2.4.17-ia64-kdbv2.1-K3/include/asm-ia64/hw_irq.h 2.4.17-ia64-kdbv2.1-K3ia64/include/asm-ia64/hw_irq.h
@@ -49,6 +49,7 @@
 #define IA64_PERFMON_VECTOR		0xee	/* performanc monitor interrupt vector */
 #define IA64_TIMER_VECTOR		0xef	/* use highest-prio group 15 interrupt for timer */
 #define	IA64_MCA_WAKEUP_VECTOR		0xf0	/* MCA wakeup (must be >MCA_RENDEZ_VECTOR) */
+#define IA64_TASK_MIGRATION		0xfb	/* task migration interrupt vector */
 #define IA64_IPI_RESCHEDULE		0xfd	/* SMP reschedule */
 #define IA64_IPI_VECTOR			0xfe	/* inter-processor interrupt vector */
 
@@ -65,6 +66,9 @@
         IA64_IPI_DM_EXTINT =    0x7,    /* pend an 8259-compatible interrupt. */
 };
 
+/* bit for masking and discarding timer interrupts on IA64 */
+#define IA64_TIMER_MASK         (1<<16)
+
 extern __u8 isa_irq_to_vector_map[16];
 #define isa_irq_to_vector(x)	isa_irq_to_vector_map[(x)]
 
include/asm-ia64/mmu_context.h +39 -0
diff -urN 2.4.17-ia64-kdbv2.1-K3/include/asm-ia64/mmu_context.h 2.4.17-ia64-kdbv2.1-K3ia64/include/asm-ia64/mmu_context.h
@@ -33,6 +33,9 @@
 	unsigned int next;	/* next context number to use */
 	unsigned int limit;	/* next >= limit => must call wrap_mmu_context() */
 	unsigned int max_ctx;	/* max. context value supported by all CPUs */
+	unsigned long flush;    /* CPU mask for those which have to perform a
+				 * local TLB flush before switching the context.
+				 */
 };
 
 extern struct ia64_ctx ia64_ctx;
@@ -44,6 +47,25 @@
 {
 }
 
+/* 
+ * When the context counter wraps around all TLBs need to be flushed because
+ * an old context number might have been reused. This is signalled by a bit
+ * set in ia64_ctx.flush, which is checked in the routine below. Called by
+ * activate_mm().                                        <[email protected]>
+ */
+static inline void
+delayed_tlb_flush (void)
+{
+	extern void __flush_tlb_all (void);
+
+	if (unlikely(ia64_ctx.flush & (1<<smp_processor_id()))) {
+		spin_lock(&ia64_ctx.lock);
+		__flush_tlb_all();
+		ia64_ctx.flush &= ~(1<<smp_processor_id());
+		spin_unlock(&ia64_ctx.lock);
+	}
+}
+
 static inline void
 get_new_mmu_context (struct mm_struct *mm)
 {
@@ -113,11 +135,28 @@
 	 * We may get interrupts here, but that's OK because interrupt
 	 * handlers cannot touch user-space.
 	 */
+	delayed_tlb_flush();
 	ia64_set_kr(IA64_KR_PT_BASE, __pa(next->pgd));
 	get_mmu_context(next);
 	reload_context(next);
 }
 
+/*
+ * Needed for the O(1) MQ scheduler.
+ */
+#if MAX_PRIO >= 192
+# error update this function. */
+#endif
+
+static inline int sched_find_first_bit(unsigned long *b)
+{
+	if (unlikely(b[0]))
+		return __ffs(b[0]);
+	if (b[1])
+		return 64 + __ffs(b[1]);
+	return __ffs(b[2]) + 128;
+}
+
 #define switch_mm(prev_mm,next_mm,next_task,cpu)	activate_mm(prev_mm, next_mm)
 
 # endif /* ! __ASSEMBLY__ */
include/asm-ia64/pgalloc.h +3 -0
diff -urN 2.4.17-ia64-kdbv2.1-K3/include/asm-ia64/pgalloc.h 2.4.17-ia64-kdbv2.1-K3ia64/include/asm-ia64/pgalloc.h
@@ -160,9 +160,12 @@
 
 #ifdef CONFIG_SMP
   extern void smp_flush_tlb_all (void);
+  extern void smp_flush_tlb_all_nowait (void);
 # define flush_tlb_all()	smp_flush_tlb_all()
+# define flush_tlb_all_nowait()	smp_flush_tlb_all_nowait()
 #else
 # define flush_tlb_all()	__flush_tlb_all()
+# define flush_tlb_all_nowait()	__flush_tlb_all()
 #endif
 
 /*
include/asm-ia64/smp.h +4 -7
diff -urN 2.4.17-ia64-kdbv2.1-K3/include/asm-ia64/smp.h 2.4.17-ia64-kdbv2.1-K3ia64/include/asm-ia64/smp.h
@@ -27,7 +27,7 @@
 #define SMP_IRQ_REDIRECTION	(1 << 0)
 #define SMP_IPI_REDIRECTION	(1 << 1)
 
-#define smp_processor_id()	(current->processor)
+#define smp_processor_id()	(current->cpu)
 
 extern struct smp_boot_data {
 	int cpu_count;
@@ -48,6 +48,9 @@
 
 extern unsigned long ap_wakeup_vector;
 
+extern void smp_send_reschedule(int cpu);
+extern void smp_send_reschedule_all(void);
+
 /*
  * Function to map hard smp processor id to logical id.  Slow, so
  * don't use this in performance-critical code.
@@ -110,12 +113,6 @@
 
 #define NO_PROC_ID		0xffffffff	/* no processor magic marker */
 
-/*
- * Extra overhead to move a task from one cpu to another (due to TLB and cache misses).
- * Expressed in "negative nice value" units (larger number means higher priority/penalty).
- */
-#define PROC_CHANGE_PENALTY	20
-
 extern void __init init_smp_config (void);
 extern void smp_do_timer (struct pt_regs *regs);
 
include/asm-ia64/spinlock.h +1 -1
diff -urN 2.4.17-ia64-kdbv2.1-K3/include/asm-ia64/spinlock.h 2.4.17-ia64-kdbv2.1-K3ia64/include/asm-ia64/spinlock.h
@@ -84,7 +84,7 @@
 	"mov r29 = 1\n"						\
 	";;\n"							\
 	"1:\n"							\
-	"ld4.bias r2 = [%0]\n"					\
+	"ld4 r2 = [%0]\n"					\
 	";;\n"							\
 	"cmp4.eq p0,p7 = r0,r2\n"				\
 	"(p7) br.cond.spnt.few 1b \n"				\
include/linux/smp.h +0 -6
diff -urN 2.4.17-ia64-kdbv2.1-K3/include/linux/smp.h 2.4.17-ia64-kdbv2.1-K3ia64/include/linux/smp.h
@@ -24,12 +24,6 @@
 extern void smp_send_stop(void);
 
 /*
- * sends a 'reschedule' event to another CPU:
- */
-extern void FASTCALL(smp_send_reschedule(int cpu));
-
-
-/*
  * Boot processor call to load the other CPU's
  */
 extern void smp_boot_cpus(void);
kdb/kdbmain.c +1 -1
diff -urN 2.4.17-ia64-kdbv2.1-K3/kdb/kdbmain.c 2.4.17-ia64-kdbv2.1-K3ia64/kdb/kdbmain.c
@@ -2344,7 +2344,7 @@
 	for_each_task(p) {
 		kdb_printf("0x%p %08d %08d  %1.1d  %3.3d  %s  0x%p%c%s\n",
 			   (void *)p, p->pid, p->p_pptr->pid,
-			   task_has_cpu(p), p->processor,
+			   task_has_cpu(p), p->cpu,
 			   (p->state == 0)?"run ":(p->state>0)?"stop":"unrn",
 			   (void *)(&p->thread),
 			   (p == current) ? '*': ' ',
kernel/printk.c +2 -0
diff -urN 2.4.17-ia64-kdbv2.1-K3/kernel/printk.c 2.4.17-ia64-kdbv2.1-K3ia64/kernel/printk.c
@@ -25,6 +25,8 @@
 #include <linux/init.h>
 #include <linux/module.h>
 #include <linux/interrupt.h>			/* For in_interrupt() */
+#include <linux/config.h>
+#include <linux/delay.h>
 
 #include <asm/uaccess.h>
 
kernel/sched.c +31 -60

124 lines not on this page · raw message

kernel/timer.c +1 -12
diff -urN 2.4.17-ia64-kdbv2.1-K3/kernel/timer.c 2.4.17-ia64-kdbv2.1-K3ia64/kernel/timer.c
@@ -584,18 +584,7 @@
 	int cpu = smp_processor_id(), system = user_tick ^ 1;
 
 	update_one_process(p, user_tick, system, cpu);
-	if (p->pid) {
-		if (--p->counter <= 0) {
-			p->counter = 0;
-			p->need_resched = 1;
-		}
-		if (p->nice > 0)
-			kstat.per_cpu_nice[cpu] += user_tick;
-		else
-			kstat.per_cpu_user[cpu] += user_tick;
-		kstat.per_cpu_system[cpu] += system;
-	} else if (really_local_bh_count() || really_local_irq_count() > 1)
-		kstat.per_cpu_system[cpu] += system;
+	scheduler_tick(user_tick, system);
 }
 
 /*