perfmon.c

这个linux源代码是很全面的~基本完整了~使用c编译的~由于时间问题我没有亲自测试~但就算用来做参考资料也是非常好的

	UNLOCK_PFS();

	return -EBUSY;

}

static int
pfm_unreserve_session(struct task_struct *task, int is_syswide, unsigned long cpu_mask)
{
	pfm_context_t *ctx;
	unsigned long m;
	unsigned int n, i;

	ctx = task ? task->thread.pfm_context : NULL;

	/*
	 * validy checks on cpu_mask have been done upstream
	 */
	LOCK_PFS();

	DBprintk(("[%d] sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu_mask=0x%lx\n",
		task->pid,
		pfm_sessions.pfs_sys_sessions,
		pfm_sessions.pfs_task_sessions,
		pfm_sessions.pfs_sys_use_dbregs,
		is_syswide,
		cpu_mask));
		

	if (is_syswide) {
		m = cpu_mask; n = 0;
		for(i=0; m; i++, m>>=1) {
			if ((m & 0x1) == 0UL) continue;
			pfm_sessions.pfs_sys_session[i] = NULL;
			n++;
		}
		/* 
		 * 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("perfmon: invalid release for [%d] sys_use_dbregs=0\n", task->pid);
			} else {
				pfm_sessions.pfs_sys_use_dbregs--;
			}
		}
		pfm_sessions.pfs_sys_sessions -= n;

		DBprintk(("CPU%d sys_sessions=%u\n", 
			smp_processor_id(), pfm_sessions.pfs_sys_sessions));
	} else {
		pfm_sessions.pfs_task_sessions--;
		DBprintk(("[%d] task_sessions=%u\n", 
			task->pid, pfm_sessions.pfs_task_sessions));
	}

	UNLOCK_PFS();

	return 0;
}



/*
 * XXX: do something better here
 */
static int
pfm_bad_permissions(struct task_struct *task)
{
	/* stolen from bad_signal() */
	return (current->session != task->session)
	    && (current->euid ^ task->suid) && (current->euid ^ task->uid)
	    && (current->uid ^ task->suid) && (current->uid ^ task->uid);
}

static int
pfx_is_sane(struct task_struct *task, pfarg_context_t *pfx)
{
	unsigned long smpl_pmds = pfx->ctx_smpl_regs[0];
	int ctx_flags;
	int cpu;

	/* valid signal */

	/* cannot send to process 1, 0 means do not notify */
	if (pfx->ctx_notify_pid == 1) {
		DBprintk(("invalid notify_pid %d\n", pfx->ctx_notify_pid));
		return -EINVAL;
	}
	ctx_flags = pfx->ctx_flags;

	if ((ctx_flags & PFM_FL_INHERIT_MASK) == (PFM_FL_INHERIT_ONCE|PFM_FL_INHERIT_ALL)) {
		DBprintk(("invalid inherit mask 0x%x\n",ctx_flags & PFM_FL_INHERIT_MASK));
		return -EINVAL;
	}

	if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
		DBprintk(("cpu_mask=0x%lx\n", pfx->ctx_cpu_mask));
		/*
		 * cannot block in this mode 
		 */
		if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
			DBprintk(("cannot use blocking mode when in system wide monitoring\n"));
			return -EINVAL;
		}
		/*
		 * must only have one bit set in the CPU mask
		 */
		if (hweight64(pfx->ctx_cpu_mask) != 1UL) {
			DBprintk(("invalid CPU mask specified\n"));
			return -EINVAL;
		}
		/*
		 * and it must be a valid CPU
		 */
		cpu = ffz(~pfx->ctx_cpu_mask);
		if (cpu_is_online(cpu) == 0) {
			DBprintk(("CPU%d is not online\n", cpu));
			return -EINVAL;
		}
		/*
		 * check for pre-existing pinning, if conflicting reject
		 */
		if (task->cpus_allowed != ~0UL && (task->cpus_allowed & (1UL<<cpu)) == 0) {
			DBprintk(("[%d] pinned on 0x%lx, mask for CPU%d \n", task->pid, 
				task->cpus_allowed, cpu));
			return -EINVAL;
		}

	} else {
		/*
		 * must provide a target for the signal in blocking mode even when
		 * no counter is configured with PFM_FL_REG_OVFL_NOTIFY
		 */
		if ((ctx_flags & PFM_FL_NOTIFY_BLOCK) && pfx->ctx_notify_pid == 0) {
			DBprintk(("must have notify_pid when blocking for [%d]\n", task->pid));
			return -EINVAL;
		}
#if 0
		if ((ctx_flags & PFM_FL_NOTIFY_BLOCK) && pfx->ctx_notify_pid == task->pid) {
			DBprintk(("cannot notify self when blocking for [%d]\n", task->pid));
			return -EINVAL;
		}
#endif
	}
	/* verify validity of smpl_regs */
	if ((smpl_pmds & pmu_conf.impl_pmds[0]) != smpl_pmds) {
		DBprintk(("invalid smpl_regs 0x%lx\n", smpl_pmds));
		return -EINVAL;
	}
	/* probably more to add here */

	return 0;
}

static int
pfm_context_create(struct task_struct *task, pfm_context_t *ctx, void *req, int count, 
		   struct pt_regs *regs)
{
	pfarg_context_t tmp;
	void *uaddr = NULL;
	int ret, cpu = 0;
	int ctx_flags;
	pid_t notify_pid;

	/* a context has already been defined */
	if (ctx) return -EBUSY;

	/*
	 * not yet supported
	 */
	if (task != current) return -EINVAL;

	if (__copy_from_user(&tmp, req, sizeof(tmp))) return -EFAULT;

	ret = pfx_is_sane(task, &tmp);
	if (ret < 0) return ret;

	ctx_flags = tmp.ctx_flags;

	ret =  -EBUSY;

	ret = pfm_reserve_session(task, ctx_flags & PFM_FL_SYSTEM_WIDE, tmp.ctx_cpu_mask);
	if (ret) goto abort;

	ret = -ENOMEM;

	ctx = pfm_context_alloc();
	if (!ctx) goto error;

	/* record the creator (important for inheritance) */
	ctx->ctx_owner = current;

	notify_pid = tmp.ctx_notify_pid;

	spin_lock_init(&ctx->ctx_lock);

	if (notify_pid == current->pid) {

		ctx->ctx_notify_task = current;
		task->thread.pfm_context = ctx;

	} else if (notify_pid!=0) {
		struct task_struct *notify_task;

		read_lock(&tasklist_lock);

		notify_task = find_task_by_pid(notify_pid);

		if (notify_task) {

			ret = -EPERM;

			/*
			 * check if we can send this task a signal
			 */
			if (pfm_bad_permissions(notify_task)) {
				read_unlock(&tasklist_lock);
				goto buffer_error;
			}

			/* 
		 	 * make visible
		 	 * must be done inside critical section
		 	 *
		 	 * if the initialization does not go through it is still
		 	 * okay because child will do the scan for nothing which
		 	 * won't hurt.
		 	 */
			task->thread.pfm_context = ctx;

			/*
			 * will cause task to check on exit for monitored
			 * processes that would notify it. see release_thread()
			 * Note: the scan MUST be done in release thread, once the
			 * task has been detached from the tasklist otherwise you are
			 * exposed to race conditions.
			 */
			atomic_add(1, &ctx->ctx_notify_task->thread.pfm_notifiers_check);

			ctx->ctx_notify_task = notify_task;
		}
		read_unlock(&tasklist_lock);
	}

	/*
	 * notification process does not exist
	 */
	if (notify_pid != 0 && ctx->ctx_notify_task == NULL) {
		ret = -EINVAL;
		goto buffer_error;
	}

	if (tmp.ctx_smpl_entries) {
		DBprintk(("sampling entries=%lu\n",tmp.ctx_smpl_entries));

		ret = pfm_smpl_buffer_alloc(ctx, tmp.ctx_smpl_regs, 
						 tmp.ctx_smpl_entries, &uaddr);
		if (ret<0) goto buffer_error;

		tmp.ctx_smpl_vaddr = uaddr;
	}
	/* initialization of context's flags */
	ctx->ctx_fl_inherit   = ctx_flags & PFM_FL_INHERIT_MASK;
	ctx->ctx_fl_block     = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
	ctx->ctx_fl_system    = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
	ctx->ctx_fl_frozen    = 0;
	/*
	 * setting this flag to 0 here means, that the creator or the task that the
	 * context is being attached are granted access. Given that a context can only
	 * be created for the calling process this, in effect only allows the creator
	 * to access the context. See pfm_protect() for more.
	 */
	ctx->ctx_fl_protected = 0;

	/* for system wide mode only (only 1 bit set) */
	ctx->ctx_cpu         = cpu;

	atomic_set(&ctx->ctx_last_cpu,-1); /* SMP only, means no CPU */

	/* may be redudant with memset() but at least it's easier to remember */
	atomic_set(&ctx->ctx_saving_in_progress, 0); 
	atomic_set(&ctx->ctx_is_busy, 0); 

	sema_init(&ctx->ctx_restart_sem, 0); /* init this semaphore to locked */

	if (__copy_to_user(req, &tmp, sizeof(tmp))) {
		ret = -EFAULT;
		goto buffer_error;
	}

	DBprintk(("context=%p, pid=%d notify_task=%p\n",
			(void *)ctx, task->pid, ctx->ctx_notify_task));

	DBprintk(("context=%p, pid=%d flags=0x%x inherit=%d block=%d system=%d\n", 
			(void *)ctx, task->pid, ctx_flags, ctx->ctx_fl_inherit, 
			ctx->ctx_fl_block, ctx->ctx_fl_system));

	/*
	 * when no notification is required, we can make this visible at the last moment
	 */
	if (notify_pid == 0) task->thread.pfm_context = ctx;
	/*
	 * pin task to CPU and force reschedule on exit to ensure
	 * that when back to user level the task runs on the designated
	 * CPU.
	 */
	if (ctx->ctx_fl_system) {
		ctx->ctx_saved_cpus_allowed = task->cpus_allowed;
		task->cpus_allowed = 1UL << cpu;
		task->need_resched = 1;
		DBprintk(("[%d] rescheduled allowed=0x%lx\n", task->pid,task->cpus_allowed));
	}

	return 0;

buffer_error:
	pfm_context_free(ctx);
error:
	pfm_unreserve_session(task, ctx_flags & PFM_FL_SYSTEM_WIDE , tmp.ctx_cpu_mask);
abort:
	/* make sure we don't leave anything behind */
	task->thread.pfm_context = NULL;

	return ret;
}

static inline unsigned long
pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
{
	unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
	unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
	extern unsigned long carta_random32 (unsigned long seed);

	if (reg->flags & PFM_REGFL_RANDOM) {
		new_seed = carta_random32(old_seed);
		val -= (old_seed & mask);	/* counter values are negative numbers! */
		if ((mask >> 32) != 0)
			/* construct a full 64-bit random value: */
			new_seed |= carta_random32(old_seed >> 32) << 32;
		reg->seed = new_seed;
	}
	reg->lval = val;
	return val;
}

static void
pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int flag)
{
	unsigned long mask = ovfl_regs[0];
	unsigned long reset_others = 0UL;
	unsigned long val;
	int i, is_long_reset = (flag == PFM_PMD_LONG_RESET);

	/*
	 * now restore reset value on sampling overflowed counters
	 */
	mask >>= PMU_FIRST_COUNTER;
	for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
		if (mask & 0x1) {
			val = pfm_new_counter_value(ctx->ctx_soft_pmds + i, is_long_reset);
			reset_others |= ctx->ctx_soft_pmds[i].reset_pmds[0];

			DBprintk_ovfl(("[%d] %s reset soft_pmd[%d]=%lx\n", current->pid,
				  is_long_reset ? "long" : "short", i, val));

			/* upper part is ignored on rval */
			pfm_write_soft_counter(ctx, i, val);
		}
	}

	/*
	 * Now take care of resetting the other registers
	 */
	for(i = 0; reset_others; i++, reset_others >>= 1) {

		if ((reset_others & 0x1) == 0) continue;

		val = pfm_new_counter_value(ctx->ctx_soft_pmds + i, is_long_reset);

		if (PMD_IS_COUNTING(i)) {
			pfm_write_soft_counter(ctx, i, val);
		} else {
			ia64_set_pmd(i, val);
		}
		DBprintk_ovfl(("[%d] %s reset_others pmd[%d]=%lx\n", current->pid,
			  is_long_reset ? "long" : "short", i, val));
	}
	ia64_srlz_d();
}

static int
pfm_write_pmcs(struct task_struct *task, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
{
	struct thread_struct *th = &task->thread;
	pfarg_reg_t tmp, *req = (pfarg_reg_t *)arg;
	unsigned long value, reset_pmds;
	unsigned int cnum, reg_flags, flags;
	int i;
	int ret = -EINVAL;

	if (!CTX_IS_ENABLED(ctx)) return -EINVAL;

	/* XXX: ctx locking may be required here */

	for (i = 0; i < count; i++, req++) {

		if (__copy_from_user(&tmp, req, sizeof(tmp))) return -EFAULT;

		cnum       = tmp.reg_num;
		reg_flags  = tmp.reg_flags;
		value      = tmp.reg_value;
		reset_pmds = tmp.reg_reset_pmds[0];
		flags      = 0;

		/* 
		 * we reject all non implemented PMC as well
		 * as attempts to modify PMC[0-3] which are used
		 * as status registers by the PMU
		 */
		if (!PMC_IS_IMPL(cnum) || cnum < 4) {
			DBprintk(("pmc[%u] is unimplemented or invalid\n", cnum));
			goto error;
		}
		/*
		 * A PMC used to configure monitors must be:
		 * 	- system-wide session: privileged monitor
		 * 	- per-task : user monitor
		 * any other configuration is rejected.
		 */
		if (PMC_IS_MONITOR(cnum) || PMC_IS_COUNTING(cnum)) {
			DBprintk(("pmc[%u].pm=%ld\n", cnum, PMC_PM(cnum, value))); 

			if (ctx->ctx_fl_system ^ PMC_PM(cnum, value)) {
				DBprintk(("pmc_pm=%ld fl_system=%d\n", PMC_PM(cnum, value), ctx->ctx_fl_system));
				goto error;
			}
		}

		if (PMC_IS_COUNTING(cnum)) {
			pfm_monitor_t *p = (pfm_monitor_t *)&value;
			/*
		 	 * enforce generation of overflow interrupt. Necessary on all
		 	 * CPUs.
		 	 */
			p->pmc_oi = 1;

			if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
				/*
			 	 * must have a target for the signal
			 	 */
				if (ctx->ctx_notify_task == NULL) {
					DBprintk(("cannot set ovfl_notify: no notify_task\n"));
					goto error;
				}
				flags |= PFM_REGFL_OVFL_NOTIFY;
			}

			if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;

			/* verify validity of reset_pmds */
			if ((reset_pmds & pmu_conf.impl_pmds[0]) != reset_pmds) {
				DBprintk(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
				goto error;
			}