perfmon.c

xen虚拟机源代码安装包

}

#ifndef XEN
/*
 * removes virtual mapping of the sampling buffer.
 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
 * a PROTECT_CTX() section.
 */
static int
pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
{
	int r;

	/* sanity checks */
	if (task->mm == NULL || size == 0UL || vaddr == NULL) {
		printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task->pid, task->mm);
		return -EINVAL;
	}

	DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));

	/*
	 * does the actual unmapping
	 */
	down_write(&task->mm->mmap_sem);

	DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));

	r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);

	up_write(&task->mm->mmap_sem);
	if (r !=0) {
		printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task->pid, vaddr, size);
	}

	DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));

	return 0;
}
#endif

/*
 * free actual physical storage used by sampling buffer
 */
#if 0
static int
pfm_free_smpl_buffer(pfm_context_t *ctx)
{
	pfm_buffer_fmt_t *fmt;

	if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;

	/*
	 * we won't use the buffer format anymore
	 */
	fmt = ctx->ctx_buf_fmt;

	DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
		ctx->ctx_smpl_hdr,
		ctx->ctx_smpl_size,
		ctx->ctx_smpl_vaddr));

	pfm_buf_fmt_exit(fmt, current, NULL, NULL);

	/*
	 * free the buffer
	 */
	pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);

	ctx->ctx_smpl_hdr  = NULL;
	ctx->ctx_smpl_size = 0UL;

	return 0;

invalid_free:
	printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", current->pid);
	return -EINVAL;
}
#endif

static inline void
pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
{
	if (fmt == NULL) return;

	pfm_buf_fmt_exit(fmt, current, NULL, NULL);

}

#ifndef XEN
/*
 * pfmfs should _never_ be mounted by userland - too much of security hassle,
 * no real gain from having the whole whorehouse mounted. So we don't need
 * any operations on the root directory. However, we need a non-trivial
 * d_name - pfm: will go nicely and kill the special-casing in procfs.
 */
static struct vfsmount *pfmfs_mnt;

static int __init
init_pfm_fs(void)
{
	int err = register_filesystem(&pfm_fs_type);
	if (!err) {
		pfmfs_mnt = kern_mount(&pfm_fs_type);
		err = PTR_ERR(pfmfs_mnt);
		if (IS_ERR(pfmfs_mnt))
			unregister_filesystem(&pfm_fs_type);
		else
			err = 0;
	}
	return err;
}

static void __exit
exit_pfm_fs(void)
{
	unregister_filesystem(&pfm_fs_type);
	mntput(pfmfs_mnt);
}

static ssize_t
pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
{
	pfm_context_t *ctx;
	pfm_msg_t *msg;
	ssize_t ret;
	unsigned long flags;
  	DECLARE_WAITQUEUE(wait, current);
	if (PFM_IS_FILE(filp) == 0) {
		printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
		return -EINVAL;
	}

	ctx = (pfm_context_t *)filp->private_data;
	if (ctx == NULL) {
		printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", current->pid);
		return -EINVAL;
	}

	/*
	 * check even when there is no message
	 */
	if (size < sizeof(pfm_msg_t)) {
		DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
		return -EINVAL;
	}

	PROTECT_CTX(ctx, flags);

  	/*
	 * put ourselves on the wait queue
	 */
  	add_wait_queue(&ctx->ctx_msgq_wait, &wait);


  	for(;;) {
		/*
		 * check wait queue
		 */

  		set_current_state(TASK_INTERRUPTIBLE);

		DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));

		ret = 0;
		if(PFM_CTXQ_EMPTY(ctx) == 0) break;

		UNPROTECT_CTX(ctx, flags);

		/*
		 * check non-blocking read
		 */
      		ret = -EAGAIN;
		if(filp->f_flags & O_NONBLOCK) break;

		/*
		 * check pending signals
		 */
		if(signal_pending(current)) {
			ret = -EINTR;
			break;
		}
      		/*
		 * no message, so wait
		 */
      		schedule();

		PROTECT_CTX(ctx, flags);
	}
	DPRINT(("[%d] back to running ret=%ld\n", current->pid, ret));
  	set_current_state(TASK_RUNNING);
	remove_wait_queue(&ctx->ctx_msgq_wait, &wait);

	if (ret < 0) goto abort;

	ret = -EINVAL;
	msg = pfm_get_next_msg(ctx);
	if (msg == NULL) {
		printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, current->pid);
		goto abort_locked;
	}

	DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));

	ret = -EFAULT;
  	if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);

abort_locked:
	UNPROTECT_CTX(ctx, flags);
abort:
	return ret;
}

static ssize_t
pfm_write(struct file *file, const char __user *ubuf,
			  size_t size, loff_t *ppos)
{
	DPRINT(("pfm_write called\n"));
	return -EINVAL;
}

static unsigned int
pfm_poll(struct file *filp, poll_table * wait)
{
	pfm_context_t *ctx;
	unsigned long flags;
	unsigned int mask = 0;

	if (PFM_IS_FILE(filp) == 0) {
		printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
		return 0;
	}

	ctx = (pfm_context_t *)filp->private_data;
	if (ctx == NULL) {
		printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", current->pid);
		return 0;
	}


	DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));

	poll_wait(filp, &ctx->ctx_msgq_wait, wait);

	PROTECT_CTX(ctx, flags);

	if (PFM_CTXQ_EMPTY(ctx) == 0)
		mask =  POLLIN | POLLRDNORM;

	UNPROTECT_CTX(ctx, flags);

	DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));

	return mask;
}

static int
pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
{
	DPRINT(("pfm_ioctl called\n"));
	return -EINVAL;
}

/*
 * interrupt cannot be masked when coming here
 */
static inline int
pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
{
	int ret;

	ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);

	DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
		current->pid,
		fd,
		on,
		ctx->ctx_async_queue, ret));

	return ret;
}

static int
pfm_fasync(int fd, struct file *filp, int on)
{
	pfm_context_t *ctx;
	int ret;

	if (PFM_IS_FILE(filp) == 0) {
		printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", current->pid);
		return -EBADF;
	}

	ctx = (pfm_context_t *)filp->private_data;
	if (ctx == NULL) {
		printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", current->pid);
		return -EBADF;
	}
	/*
	 * we cannot mask interrupts during this call because this may
	 * may go to sleep if memory is not readily avalaible.
	 *
	 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
	 * done in caller. Serialization of this function is ensured by caller.
	 */
	ret = pfm_do_fasync(fd, filp, ctx, on);


	DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
		fd,
		on,
		ctx->ctx_async_queue, ret));

	return ret;
}

#ifdef CONFIG_SMP
/*
 * this function is exclusively called from pfm_close().
 * The context is not protected at that time, nor are interrupts
 * on the remote CPU. That's necessary to avoid deadlocks.
 */
static void
pfm_syswide_force_stop(void *info)
{
	pfm_context_t   *ctx = (pfm_context_t *)info;
	struct pt_regs *regs = task_pt_regs(current);
	struct task_struct *owner;
	unsigned long flags;
	int ret;

	if (ctx->ctx_cpu != smp_processor_id()) {
		printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d  but on CPU%d\n",
			ctx->ctx_cpu,
			smp_processor_id());
		return;
	}
	owner = GET_PMU_OWNER();
	if (owner != ctx->ctx_task) {
		printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
			smp_processor_id(),
			owner->pid, ctx->ctx_task->pid);
		return;
	}
	if (GET_PMU_CTX() != ctx) {
		printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
			smp_processor_id(),
			GET_PMU_CTX(), ctx);
		return;
	}

	DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), ctx->ctx_task->pid));	
	/*
	 * the context is already protected in pfm_close(), we simply
	 * need to mask interrupts to avoid a PMU interrupt race on
	 * this CPU
	 */
	local_irq_save(flags);

	ret = pfm_context_unload(ctx, NULL, 0, regs);
	if (ret) {
		DPRINT(("context_unload returned %d\n", ret));
	}

	/*
	 * unmask interrupts, PMU interrupts are now spurious here
	 */
	local_irq_restore(flags);
}

static void
pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
{
	int ret;

	DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
	ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 0, 1);
	DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
}
#endif /* CONFIG_SMP */

/*
 * called for each close(). Partially free resources.
 * When caller is self-monitoring, the context is unloaded.
 */
static int
pfm_flush(struct file *filp)
{
	pfm_context_t *ctx;
	struct task_struct *task;
	struct pt_regs *regs;
	unsigned long flags;
	unsigned long smpl_buf_size = 0UL;
	void *smpl_buf_vaddr = NULL;
	int state, is_system;

	if (PFM_IS_FILE(filp) == 0) {
		DPRINT(("bad magic for\n"));
		return -EBADF;
	}

	ctx = (pfm_context_t *)filp->private_data;
	if (ctx == NULL) {
		printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", current->pid);
		return -EBADF;
	}

	/*
	 * remove our file from the async queue, if we use this mode.
	 * This can be done without the context being protected. We come
	 * here when the context has become unreacheable by other tasks.
	 *
	 * We may still have active monitoring at this point and we may
	 * end up in pfm_overflow_handler(). However, fasync_helper()
	 * operates with interrupts disabled and it cleans up the
	 * queue. If the PMU handler is called prior to entering
	 * fasync_helper() then it will send a signal. If it is
	 * invoked after, it will find an empty queue and no
	 * signal will be sent. In both case, we are safe
	 */
	if (filp->f_flags & FASYNC) {
		DPRINT(("cleaning up async_queue=%p\n", ctx->ctx_async_queue));
		pfm_do_fasync (-1, filp, ctx, 0);
	}

	PROTECT_CTX(ctx, flags);

	state     = ctx->ctx_state;
	is_system = ctx->ctx_fl_system;

	task = PFM_CTX_TASK(ctx);
	regs = task_pt_regs(task);

	DPRINT(("ctx_state=%d is_current=%d\n",
		state,
		task == current ? 1 : 0));

	/*
	 * if state == UNLOADED, then task is NULL
	 */

	/*
	 * we must stop and unload because we are losing access to the context.
	 */
	if (task == current) {
#ifdef CONFIG_SMP
		/*
		 * the task IS the owner but it migrated to another CPU: that's bad
		 * but we must handle this cleanly. Unfortunately, the kernel does
		 * not provide a mechanism to block migration (while the context is loaded).
		 *
		 * We need to release the resource on the ORIGINAL cpu.
		 */
		if (is_system && ctx->ctx_cpu != smp_processor_id()) {

			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
			/*
			 * keep context protected but unmask interrupt for IPI
			 */
			local_irq_restore(flags);

			pfm_syswide_cleanup_other_cpu(ctx);

			/*
			 * restore interrupt masking
			 */
			local_irq_save(flags);

			/*
			 * context is unloaded at this point
			 */
		} else
#endif /* CONFIG_SMP */
		{

			DPRINT(("forcing unload\n"));
			/*
		 	* stop and unload, returning with state UNLOADED
		 	* and session unreserved.
		 	*/
			pfm_context_unload(ctx, NULL, 0, regs);

			DPRINT(("ctx_state=%d\n", ctx->ctx_state));
		}
	}

	/*
	 * remove virtual mapping, if any, for the calling task.
	 * cannot reset ctx field until last user is calling close().
	 *
	 * ctx_smpl_vaddr must never be cleared because it is needed
	 * by every task with access to the context
	 *
	 * When called from do_exit(), the mm context is gone already, therefore
	 * mm is NULL, i.e., the VMA is already gone  and we do not have to
	 * do anything here
	 */
	if (ctx->ctx_smpl_vaddr && current->mm) {
		smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
		smpl_buf_size  = ctx->ctx_smpl_size;
	}

	UNPROTECT_CTX(ctx, flags);

	/*
	 * if there was a mapping, then we systematically remove it
	 * at this point. Cannot be done inside critical section