smp.c

是关于linux2.5.1的完全源码

				      ctx, start, end,
				      mm->cpu_vm_mask);

	local_flush_and_out:
		__flush_tlb_range(ctx, start, SECONDARY_CONTEXT, end, PAGE_SIZE, (end-start));
	}
}

void smp_flush_tlb_page(struct mm_struct *mm, unsigned long page)
{
	{
		u32 ctx = CTX_HWBITS(mm->context);
		int cpu = smp_processor_id();

		page &= PAGE_MASK;
		if (mm == current->active_mm && atomic_read(&mm->mm_users) == 1) {
			/* By virtue of being the current address space, and
			 * having the only reference to it, the following operation
			 * is safe.
			 *
			 * It would not be a win to perform the xcall tlb flush in
			 * this case, because even if we switch back to one of the
			 * other processors in cpu_vm_mask it is almost certain that
			 * all TLB entries for this context will be replaced by the
			 * time that happens.
			 */
			mm->cpu_vm_mask = (1UL << cpu);
			goto local_flush_and_out;
		} else {
			/* By virtue of running under the mm->page_table_lock,
			 * and mmu_context.h:switch_mm doing the same, the following
			 * operation is safe.
			 */
			if (mm->cpu_vm_mask == (1UL << cpu))
				goto local_flush_and_out;
		}

		/* OK, we have to actually perform the cross call.  Most likely
		 * this is a cloned mm or kswapd is kicking out pages for a task
		 * which has run recently on another cpu.
		 */
		smp_cross_call_masked(&xcall_flush_tlb_page,
				      ctx, page, 0,
				      mm->cpu_vm_mask);
		if (!(mm->cpu_vm_mask & (1UL << cpu)))
			return;

	local_flush_and_out:
		__flush_tlb_page(ctx, page, SECONDARY_CONTEXT);
	}
}

/* CPU capture. */
/* #define CAPTURE_DEBUG */
extern unsigned long xcall_capture;

static atomic_t smp_capture_depth = ATOMIC_INIT(0);
static atomic_t smp_capture_registry = ATOMIC_INIT(0);
static unsigned long penguins_are_doing_time;

void smp_capture(void)
{
	if (smp_processors_ready) {
		int result = __atomic_add(1, &smp_capture_depth);

		membar("#StoreStore | #LoadStore");
		if (result == 1) {
			int ncpus = smp_num_cpus;

#ifdef CAPTURE_DEBUG
			printk("CPU[%d]: Sending penguins to jail...",
			       smp_processor_id());
#endif
			penguins_are_doing_time = 1;
			membar("#StoreStore | #LoadStore");
			atomic_inc(&smp_capture_registry);
			smp_cross_call(&xcall_capture, 0, 0, 0);
			while (atomic_read(&smp_capture_registry) != ncpus)
				membar("#LoadLoad");
#ifdef CAPTURE_DEBUG
			printk("done\n");
#endif
		}
	}
}

void smp_release(void)
{
	if (smp_processors_ready) {
		if (atomic_dec_and_test(&smp_capture_depth)) {
#ifdef CAPTURE_DEBUG
			printk("CPU[%d]: Giving pardon to imprisoned penguins\n",
			       smp_processor_id());
#endif
			penguins_are_doing_time = 0;
			membar("#StoreStore | #StoreLoad");
			atomic_dec(&smp_capture_registry);
		}
	}
}

/* Imprisoned penguins run with %pil == 15, but PSTATE_IE set, so they
 * can service tlb flush xcalls...
 */
extern void prom_world(int);
extern void save_alternate_globals(unsigned long *);
extern void restore_alternate_globals(unsigned long *);
void smp_penguin_jailcell(int irq, struct pt_regs *regs)
{
	unsigned long global_save[24];

	clear_softint(1 << irq);

	preempt_disable();

	__asm__ __volatile__("flushw");
	save_alternate_globals(global_save);
	prom_world(1);
	atomic_inc(&smp_capture_registry);
	membar("#StoreLoad | #StoreStore");
	while (penguins_are_doing_time)
		membar("#LoadLoad");
	restore_alternate_globals(global_save);
	atomic_dec(&smp_capture_registry);
	prom_world(0);

	preempt_enable();
}

extern unsigned long xcall_promstop;

void smp_promstop_others(void)
{
	if (smp_processors_ready)
		smp_cross_call(&xcall_promstop, 0, 0, 0);
}

extern void sparc64_do_profile(unsigned long pc, unsigned long o7);

static unsigned long current_tick_offset;

#define prof_multiplier(__cpu)		cpu_data[(__cpu)].multiplier
#define prof_counter(__cpu)		cpu_data[(__cpu)].counter

void smp_percpu_timer_interrupt(struct pt_regs *regs)
{
	unsigned long compare, tick, pstate;
	int cpu = smp_processor_id();
	int user = user_mode(regs);

	/*
	 * Check for level 14 softint.
	 */
	{
		unsigned long tick_mask;

		if (SPARC64_USE_STICK)
			tick_mask = (1UL << 16);
		else
			tick_mask = (1UL << 0);

		if (!(get_softint() & tick_mask)) {
			extern void handler_irq(int, struct pt_regs *);

			handler_irq(14, regs);
			return;
		}
		clear_softint(tick_mask);
	}

	do {
		if (!user)
			sparc64_do_profile(regs->tpc, regs->u_regs[UREG_RETPC]);
		if (!--prof_counter(cpu)) {
			if (cpu == boot_cpu_id) {
				irq_enter(cpu, 0);

				kstat.irqs[cpu][0]++;
				timer_tick_interrupt(regs);

				irq_exit(cpu, 0);
			}

			update_process_times(user);

			prof_counter(cpu) = prof_multiplier(cpu);
		}

		/* Guarentee that the following sequences execute
		 * uninterrupted.
		 */
		__asm__ __volatile__("rdpr	%%pstate, %0\n\t"
				     "wrpr	%0, %1, %%pstate"
				     : "=r" (pstate)
				     : "i" (PSTATE_IE));

		/* Workaround for Spitfire Errata (#54 I think??), I discovered
		 * this via Sun BugID 4008234, mentioned in Solaris-2.5.1 patch
		 * number 103640.
		 *
		 * On Blackbird writes to %tick_cmpr can fail, the
		 * workaround seems to be to execute the wr instruction
		 * at the start of an I-cache line, and perform a dummy
		 * read back from %tick_cmpr right after writing to it. -DaveM
		 *
		 * Just to be anal we add a workaround for Spitfire
		 * Errata 50 by preventing pipeline bypasses on the
		 * final read of the %tick register into a compare
		 * instruction.  The Errata 50 description states
		 * that %tick is not prone to this bug, but I am not
		 * taking any chances.
		 */
		if (!SPARC64_USE_STICK) {
		__asm__ __volatile__("rd	%%tick_cmpr, %0\n\t"
				     "ba,pt	%%xcc, 1f\n\t"
				     " add	%0, %2, %0\n\t"
				     ".align	64\n"
				  "1: wr	%0, 0x0, %%tick_cmpr\n\t"
				     "rd	%%tick_cmpr, %%g0\n\t"
				     "rd	%%tick, %1\n\t"
				     "mov	%1, %1"
				     : "=&r" (compare), "=r" (tick)
				     : "r" (current_tick_offset));
		} else {
		__asm__ __volatile__("rd	%%asr25, %0\n\t"
				     "add	%0, %2, %0\n\t"
				     "wr	%0, 0x0, %%asr25\n\t"
				     "rd	%%asr24, %1\n\t"
				     : "=&r" (compare), "=r" (tick)
				     : "r" (current_tick_offset));
		}

		/* Restore PSTATE_IE. */
		__asm__ __volatile__("wrpr	%0, 0x0, %%pstate"
				     : /* no outputs */
				     : "r" (pstate));
	} while (tick >= compare);
}

static void __init smp_setup_percpu_timer(void)
{
	int cpu = smp_processor_id();
	unsigned long pstate;

	prof_counter(cpu) = prof_multiplier(cpu) = 1;

	/* Guarentee that the following sequences execute
	 * uninterrupted.
	 */
	__asm__ __volatile__("rdpr	%%pstate, %0\n\t"
			     "wrpr	%0, %1, %%pstate"
			     : "=r" (pstate)
			     : "i" (PSTATE_IE));

	/* Workaround for Spitfire Errata (#54 I think??), I discovered
	 * this via Sun BugID 4008234, mentioned in Solaris-2.5.1 patch
	 * number 103640.
	 *
	 * On Blackbird writes to %tick_cmpr can fail, the
	 * workaround seems to be to execute the wr instruction
	 * at the start of an I-cache line, and perform a dummy
	 * read back from %tick_cmpr right after writing to it. -DaveM
	 */
	if (!SPARC64_USE_STICK) {
	__asm__ __volatile__("
		rd	%%tick, %%g1
		ba,pt	%%xcc, 1f
		 add	%%g1, %0, %%g1
		.align	64
	1:	wr	%%g1, 0x0, %%tick_cmpr
		rd	%%tick_cmpr, %%g0"
	: /* no outputs */
	: "r" (current_tick_offset)
	: "g1");
	} else {
	__asm__ __volatile__("
		rd	%%asr24, %%g1
		add	%%g1, %0, %%g1
		wr	%%g1, 0x0, %%asr25"
	: /* no outputs */
	: "r" (current_tick_offset)
	: "g1");
	}

	/* Restore PSTATE_IE. */
	__asm__ __volatile__("wrpr	%0, 0x0, %%pstate"
			     : /* no outputs */
			     : "r" (pstate));
}

void __init smp_tick_init(void)
{
	int i;
	
	boot_cpu_id = hard_smp_processor_id();
	current_tick_offset = timer_tick_offset;
	cpu_present_map = 0;
	for (i = 0; i < linux_num_cpus; i++)
		cpu_present_map |= (1UL << linux_cpus[i].mid);
	for (i = 0; i < NR_CPUS; i++) {
		__cpu_number_map[i] = -1;
		__cpu_logical_map[i] = -1;
	}
	__cpu_number_map[boot_cpu_id] = 0;
	prom_cpu_nodes[boot_cpu_id] = linux_cpus[0].prom_node;
	__cpu_logical_map[0] = boot_cpu_id;
	prof_counter(boot_cpu_id) = prof_multiplier(boot_cpu_id) = 1;
}

cycles_t cacheflush_time;
unsigned long cache_decay_ticks;

extern unsigned long cheetah_tune_scheduling(void);
extern unsigned long timer_ticks_per_usec_quotient;

static void __init smp_tune_scheduling(void)
{
	unsigned long orig_flush_base, flush_base, flags, *p;
	unsigned int ecache_size, order;
	cycles_t tick1, tick2, raw;

	/* Approximate heuristic for SMP scheduling.  It is an
	 * estimation of the time it takes to flush the L2 cache
	 * on the local processor.
	 *
	 * The ia32 chooses to use the L1 cache flush time instead,
	 * and I consider this complete nonsense.  The Ultra can service
	 * a miss to the L1 with a hit to the L2 in 7 or 8 cycles, and
	 * L2 misses are what create extra bus traffic (ie. the "cost"
	 * of moving a process from one cpu to another).
	 */
	printk("SMP: Calibrating ecache flush... ");
	if (tlb_type == cheetah) {
		cacheflush_time = cheetah_tune_scheduling();
		goto report;
	}

	ecache_size = prom_getintdefault(linux_cpus[0].prom_node,
					 "ecache-size", (512 * 1024));
	if (ecache_size > (4 * 1024 * 1024))
		ecache_size = (4 * 1024 * 1024);
	orig_flush_base = flush_base =
		__get_free_pages(GFP_KERNEL, order = get_order(ecache_size));

	if (flush_base != 0UL) {
		__save_and_cli(flags);

		/* Scan twice the size once just to get the TLB entries
		 * loaded and make sure the second scan measures pure misses.
		 */
		for (p = (unsigned long *)flush_base;
		     ((unsigned long)p) < (flush_base + (ecache_size<<1));
		     p += (64 / sizeof(unsigned long)))
			*((volatile unsigned long *)p);

		/* Now the real measurement. */
		__asm__ __volatile__("
		b,pt	%%xcc, 1f
		 rd	%%tick, %0

		.align	64
1:		ldx	[%2 + 0x000], %%g1
		ldx	[%2 + 0x040], %%g2
		ldx	[%2 + 0x080], %%g3
		ldx	[%2 + 0x0c0], %%g5
		add	%2, 0x100, %2
		cmp	%2, %4
		bne,pt	%%xcc, 1b
		 nop
	
		rd	%%tick, %1"
		: "=&r" (tick1), "=&r" (tick2), "=&r" (flush_base)
		: "2" (flush_base), "r" (flush_base + ecache_size)
		: "g1", "g2", "g3", "g5");

		__restore_flags(flags);

		raw = (tick2 - tick1);

		/* Dampen it a little, considering two processes
		 * sharing the cache and fitting.
		 */
		cacheflush_time = (raw - (raw >> 2));

		free_pages(orig_flush_base, order);
	} else {
		cacheflush_time = ((ecache_size << 2) +
				   (ecache_size << 1));
	}
report:
	/* Convert cpu ticks to jiffie ticks. */
	cache_decay_ticks = ((long)cacheflush_time * timer_ticks_per_usec_quotient);
	cache_decay_ticks >>= 32UL;
	cache_decay_ticks = (cache_decay_ticks * HZ) / 1000;

	printk("Using heuristic of %ld cycles, %ld ticks.\n",
	       cacheflush_time, cache_decay_ticks);
}

/* /proc/profile writes can call this, don't __init it please. */
int setup_profiling_timer(unsigned int multiplier)
{
	unsigned long flags;
	int i;

	if ((!multiplier) || (timer_tick_offset / multiplier) < 1000)
		return -EINVAL;

	save_and_cli(flags);
	for (i = 0; i < NR_CPUS; i++) {
		if (cpu_present_map & (1UL << i))
			prof_multiplier(i) = multiplier;
	}
	current_tick_offset = (timer_tick_offset / multiplier);
	restore_flags(flags);

	return 0;
}