perfmon.c

xen虚拟机源代码安装包

pfm_restore_monitoring(struct task_struct *task)
{
	pfm_context_t *ctx = PFM_GET_CTX(task);
	struct thread_struct *th = &task->thread;
	unsigned long mask, ovfl_mask;
	unsigned long psr, val;
	int i, is_system;

	is_system = ctx->ctx_fl_system;
	ovfl_mask = pmu_conf->ovfl_val;

	if (task != current) {
		printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task->pid, current->pid);
		return;
	}
	if (ctx->ctx_state != PFM_CTX_MASKED) {
		printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
			task->pid, current->pid, ctx->ctx_state);
		return;
	}
	psr = pfm_get_psr();
	/*
	 * monitoring is masked via the PMC.
	 * As we restore their value, we do not want each counter to
	 * restart right away. We stop monitoring using the PSR,
	 * restore the PMC (and PMD) and then re-establish the psr
	 * as it was. Note that there can be no pending overflow at
	 * this point, because monitoring was MASKED.
	 *
	 * system-wide session are pinned and self-monitoring
	 */
	if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
		/* disable dcr pp */
		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
		pfm_clear_psr_pp();
	} else {
		pfm_clear_psr_up();
	}
	/*
	 * first, we restore the PMD
	 */
	mask = ctx->ctx_used_pmds[0];
	for (i = 0; mask; i++, mask>>=1) {
		/* skip non used pmds */
		if ((mask & 0x1) == 0) continue;

		if (PMD_IS_COUNTING(i)) {
			/*
			 * we split the 64bit value according to
			 * counter width
			 */
			val = ctx->ctx_pmds[i].val & ovfl_mask;
			ctx->ctx_pmds[i].val &= ~ovfl_mask;
		} else {
			val = ctx->ctx_pmds[i].val;
		}
		ia64_set_pmd(i, val);

		DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
			i,
			ctx->ctx_pmds[i].val,
			val));
	}
	/*
	 * restore the PMCs
	 */
	mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
	for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
		if ((mask & 0x1) == 0UL) continue;
		th->pmcs[i] = ctx->ctx_pmcs[i];
		ia64_set_pmc(i, th->pmcs[i]);
		DPRINT(("[%d] pmc[%d]=0x%lx\n", task->pid, i, th->pmcs[i]));
	}
	ia64_srlz_d();

	/*
	 * must restore DBR/IBR because could be modified while masked
	 * XXX: need to optimize 
	 */
	if (ctx->ctx_fl_using_dbreg) {
		pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
		pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
	}

	/*
	 * now restore PSR
	 */
	if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
		/* enable dcr pp */
		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
		ia64_srlz_i();
	}
	pfm_set_psr_l(psr);
}
#endif

static inline void
pfm_save_pmds(unsigned long *pmds, unsigned long mask)
{
	int i;

	ia64_srlz_d();

	for (i=0; mask; i++, mask>>=1) {
		if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
	}
}

#ifndef XEN
/*
 * reload from thread state (used for ctxw only)
 */
static inline void
pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
{
	int i;
	unsigned long val, ovfl_val = pmu_conf->ovfl_val;

	for (i=0; mask; i++, mask>>=1) {
		if ((mask & 0x1) == 0) continue;
		val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
		ia64_set_pmd(i, val);
	}
	ia64_srlz_d();
}

/*
 * propagate PMD from context to thread-state
 */
static inline void
pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
{
	struct thread_struct *thread = &task->thread;
	unsigned long ovfl_val = pmu_conf->ovfl_val;
	unsigned long mask = ctx->ctx_all_pmds[0];
	unsigned long val;
	int i;

	DPRINT(("mask=0x%lx\n", mask));

	for (i=0; mask; i++, mask>>=1) {

		val = ctx->ctx_pmds[i].val;

		/*
		 * We break up the 64 bit value into 2 pieces
		 * the lower bits go to the machine state in the
		 * thread (will be reloaded on ctxsw in).
		 * The upper part stays in the soft-counter.
		 */
		if (PMD_IS_COUNTING(i)) {
			ctx->ctx_pmds[i].val = val & ~ovfl_val;
			 val &= ovfl_val;
		}
		thread->pmds[i] = val;

		DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
			i,
			thread->pmds[i],
			ctx->ctx_pmds[i].val));
	}
}
#else
static inline void
xenpfm_restore_pmds(pfm_context_t* ctx)
{
	int i;
	unsigned long ovfl_val = pmu_conf->ovfl_val;
	unsigned long mask = ctx->ctx_all_pmds[0];
	unsigned long val;

	for (i = 0; mask; i++, mask >>= 1) {
		if ((mask & 0x1) == 0)
			continue;

		val = ctx->ctx_pmds[i].val;
		/*
		 * We break up the 64 bit value into 2 pieces
		 * the lower bits go to the machine state in the
		 * thread (will be reloaded on ctxsw in).
		 * The upper part stays in the soft-counter.
		 */
		if (PMD_IS_COUNTING(i)) {
			ctx->ctx_pmds[i].val = val & ~ovfl_val;
			val &= ovfl_val;
		}
		ia64_set_pmd(i, val);		
	}
	ia64_srlz_d();
}
#endif

#ifndef XEN
/*
 * propagate PMC from context to thread-state
 */
static inline void
pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
{
	struct thread_struct *thread = &task->thread;
	unsigned long mask = ctx->ctx_all_pmcs[0];
	int i;

	DPRINT(("mask=0x%lx\n", mask));

	for (i=0; mask; i++, mask>>=1) {
		/* masking 0 with ovfl_val yields 0 */
		thread->pmcs[i] = ctx->ctx_pmcs[i];
		DPRINT(("pmc[%d]=0x%lx\n", i, thread->pmcs[i]));
	}
}



static inline void
pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
{
	int i;

	for (i=0; mask; i++, mask>>=1) {
		if ((mask & 0x1) == 0) continue;
		ia64_set_pmc(i, pmcs[i]);
	}
	ia64_srlz_d();
}
#else
static inline void
xenpfm_restore_pmcs(pfm_context_t* ctx)
{
	int i;
	unsigned long mask = ctx->ctx_all_pmcs[0];

	for (i = 0; mask; i++, mask >>= 1) {
		if ((mask & 0x1) == 0)
			continue;
		ia64_set_pmc(i, ctx->ctx_pmcs[i]);
		DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
	}
	ia64_srlz_d();
	
}
#endif

static inline int
pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
{
	return memcmp(a, b, sizeof(pfm_uuid_t));
}

static inline int
pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
{
	int ret = 0;
	if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
	return ret;
}

static inline int
pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
{
	int ret = 0;
	if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
	return ret;
}


static inline int
pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
		     int cpu, void *arg)
{
	int ret = 0;
	if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
	return ret;
}

static inline int
pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
		     int cpu, void *arg)
{
	int ret = 0;
	if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
	return ret;
}

static inline int
pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
{
	int ret = 0;
	if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
	return ret;
}

static inline int
pfm_buf_fmt_restart_active(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
{
	int ret = 0;
	if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
	return ret;
}

static pfm_buffer_fmt_t *
__pfm_find_buffer_fmt(pfm_uuid_t uuid)
{
	struct list_head * pos;
	pfm_buffer_fmt_t * entry;

	list_for_each(pos, &pfm_buffer_fmt_list) {
		entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
		if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
			return entry;
	}
	return NULL;
}
 
/*
 * find a buffer format based on its uuid
 */
static pfm_buffer_fmt_t *
pfm_find_buffer_fmt(pfm_uuid_t uuid)
{
	pfm_buffer_fmt_t * fmt;
	spin_lock(&pfm_buffer_fmt_lock);
	fmt = __pfm_find_buffer_fmt(uuid);
	spin_unlock(&pfm_buffer_fmt_lock);
	return fmt;
}
 
int
pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
{
	int ret = 0;

	/* some sanity checks */
	if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;

	/* we need at least a handler */
	if (fmt->fmt_handler == NULL) return -EINVAL;

	/*
	 * XXX: need check validity of fmt_arg_size
	 */

	spin_lock(&pfm_buffer_fmt_lock);

	if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
		printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
		ret = -EBUSY;
		goto out;
	} 
	list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
	printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);

out:
	spin_unlock(&pfm_buffer_fmt_lock);
 	return ret;
}
EXPORT_SYMBOL(pfm_register_buffer_fmt);

int
pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
{
	pfm_buffer_fmt_t *fmt;
	int ret = 0;

	spin_lock(&pfm_buffer_fmt_lock);

	fmt = __pfm_find_buffer_fmt(uuid);
	if (!fmt) {
		printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
		ret = -EINVAL;
		goto out;
	}
	list_del_init(&fmt->fmt_list);
	printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);

out:
	spin_unlock(&pfm_buffer_fmt_lock);
	return ret;

}
EXPORT_SYMBOL(pfm_unregister_buffer_fmt);

extern void update_pal_halt_status(int);

static int
pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
{
	unsigned long flags;
	/*
	 * validy checks on cpu_mask have been done upstream
	 */
	LOCK_PFS(flags);

	DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
		pfm_sessions.pfs_sys_sessions,
		pfm_sessions.pfs_task_sessions,
		pfm_sessions.pfs_sys_use_dbregs,
		is_syswide,
		cpu));

	if (is_syswide) {
		/*
		 * cannot mix system wide and per-task sessions
		 */
		if (pfm_sessions.pfs_task_sessions > 0UL) {
			DPRINT(("system wide not possible, %u conflicting task_sessions\n",
			  	pfm_sessions.pfs_task_sessions));
			goto abort;
		}

		if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;

		DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));

#ifndef XEN
		pfm_sessions.pfs_sys_session[cpu] = task;
#else
		pfm_sessions.pfs_sys_session[cpu] = XENOPROF_TASK;
#endif

		pfm_sessions.pfs_sys_sessions++ ;

	} else {
		if (pfm_sessions.pfs_sys_sessions) goto abort;
		pfm_sessions.pfs_task_sessions++;
	}

	DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
		pfm_sessions.pfs_sys_sessions,
		pfm_sessions.pfs_task_sessions,
		pfm_sessions.pfs_sys_use_dbregs,
		is_syswide,
		cpu));

	/*
	 * disable default_idle() to go to PAL_HALT
	 */
	update_pal_halt_status(0);

	UNLOCK_PFS(flags);

	return 0;

error_conflict:
	DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
#ifndef XEN
  		pfm_sessions.pfs_sys_session[cpu]->pid,
#else
		-1,
#endif
		cpu));
abort:
	UNLOCK_PFS(flags);

	return -EBUSY;

}

static int
pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
{
	unsigned long flags;
	/*
	 * validy checks on cpu_mask have been done upstream
	 */
	LOCK_PFS(flags);

	DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
		pfm_sessions.pfs_sys_sessions,
		pfm_sessions.pfs_task_sessions,
		pfm_sessions.pfs_sys_use_dbregs,
		is_syswide,
		cpu));


	if (is_syswide) {
		pfm_sessions.pfs_sys_session[cpu] = NULL;
		/*
		 * would not work with perfmon+more than one bit in cpu_mask
		 */
		if (ctx && ctx->ctx_fl_using_dbreg) {
			if (pfm_sessions.pfs_sys_use_dbregs == 0) {
				printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
			} else {
				pfm_sessions.pfs_sys_use_dbregs--;
			}
		}
		pfm_sessions.pfs_sys_sessions--;
	} else {
		pfm_sessions.pfs_task_sessions--;
	}
	DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
		pfm_sessions.pfs_sys_sessions,
		pfm_sessions.pfs_task_sessions,
		pfm_sessions.pfs_sys_use_dbregs,
		is_syswide,
		cpu));

	/*
	 * if possible, enable default_idle() to go into PAL_HALT
	 */
	if (pfm_sessions.pfs_task_sessions == 0 && pfm_sessions.pfs_sys_sessions == 0)
		update_pal_halt_status(1);

	UNLOCK_PFS(flags);

	return 0;