perfmon.c

优龙2410linux2.6.8内核源代码

		 * task to report completion
		 *
		 * the context is protected by mutex, therefore there
		 * is no risk of being notified of completion before
		 * begin actually on the waitq.
		 */
  		set_current_state(TASK_INTERRUPTIBLE);
  		add_wait_queue(&ctx->ctx_zombieq, &wait);

		UNPROTECT_CTX(ctx, flags);

		/*
		 * XXX: check for signals :
		 * 	- ok of explicit close
		 * 	- not ok when coming from exit_files()
		 */
      		schedule();


		PROTECT_CTX(ctx, flags);


		remove_wait_queue(&ctx->ctx_zombieq, &wait);
  		set_current_state(TASK_RUNNING);

		/*
		 * context is unloaded at this point
		 */
		DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
	}
	else if (task != current) {
#ifdef CONFIG_SMP
		/*
	 	 * switch context to zombie state
	 	 */
		ctx->ctx_state = PFM_CTX_ZOMBIE;

		DPRINT(("zombie ctx for [%d]\n", task->pid));
		/*
		 * cannot free the context on the spot. deferred until
		 * the task notices the ZOMBIE state
		 */
		free_possible = 0;
#else
		pfm_context_unload(ctx, NULL, 0, regs);
#endif
	}

doit:
	/* reload state, may have changed during  opening of critical section */
	state = ctx->ctx_state;

	/*
	 * the context is still attached to a task (possibly current)
	 * we cannot destroy it right now
	 */

	/*
	 * we must free the sampling buffer right here because
	 * we cannot rely on it being cleaned up later by the
	 * monitored task. It is not possible to free vmalloc'ed
	 * memory in pfm_load_regs(). Instead, we remove the buffer
	 * now. should there be subsequent PMU overflow originally
	 * meant for sampling, the will be converted to spurious
	 * and that's fine because the monitoring tools is gone anyway.
	 */
	if (ctx->ctx_smpl_hdr) {
		smpl_buf_addr = ctx->ctx_smpl_hdr;
		smpl_buf_size = ctx->ctx_smpl_size;
		/* no more sampling */
		ctx->ctx_smpl_hdr = NULL;
		ctx->ctx_fl_is_sampling = 0;
	}

	DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
		state,
		free_possible,
		smpl_buf_addr,
		smpl_buf_size));

	if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);

	/*
	 * UNLOADED that the session has already been unreserved.
	 */
	if (state == PFM_CTX_ZOMBIE) {
		pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
	}

	/*
	 * disconnect file descriptor from context must be done
	 * before we unlock.
	 */
	filp->private_data = NULL;

	/*
	 * if we free on the spot, the context is now completely unreacheable
	 * from the callers side. The monitored task side is also cut, so we
	 * can freely cut.
	 *
	 * If we have a deferred free, only the caller side is disconnected.
	 */
	UNPROTECT_CTX(ctx, flags);

	/*
	 * All memory free operations (especially for vmalloc'ed memory)
	 * MUST be done with interrupts ENABLED.
	 */
	if (smpl_buf_addr)  pfm_rvfree(smpl_buf_addr, smpl_buf_size);

	/*
	 * return the memory used by the context
	 */
	if (free_possible) pfm_context_free(ctx);

	return 0;
}

static int
pfm_no_open(struct inode *irrelevant, struct file *dontcare)
{
	DPRINT(("pfm_no_open called\n"));
	return -ENXIO;
}



static struct file_operations pfm_file_ops = {
	.llseek   = no_llseek,
	.read     = pfm_read,
	.write    = pfm_write,
	.poll     = pfm_poll,
	.ioctl    = pfm_ioctl,
	.open     = pfm_no_open,	/* special open code to disallow open via /proc */
	.fasync   = pfm_fasync,
	.release  = pfm_close,
	.flush	  = pfm_flush
};

static int
pfmfs_delete_dentry(struct dentry *dentry)
{
	return 1;
}

static struct dentry_operations pfmfs_dentry_operations = {
	.d_delete = pfmfs_delete_dentry,
};


static int
pfm_alloc_fd(struct file **cfile)
{
	int fd, ret = 0;
	struct file *file = NULL;
	struct inode * inode;
	char name[32];
	struct qstr this;

	fd = get_unused_fd();
	if (fd < 0) return -ENFILE;

	ret = -ENFILE;

	file = get_empty_filp();
	if (!file) goto out;

	/*
	 * allocate a new inode
	 */
	inode = new_inode(pfmfs_mnt->mnt_sb);
	if (!inode) goto out;

	DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));

	inode->i_sb   = pfmfs_mnt->mnt_sb;
	inode->i_mode = S_IFCHR|S_IRUGO;
	inode->i_sock = 0;
	inode->i_uid  = current->fsuid;
	inode->i_gid  = current->fsgid;

	sprintf(name, "[%lu]", inode->i_ino);
	this.name = name;
	this.len  = strlen(name);
	this.hash = inode->i_ino;

	ret = -ENOMEM;

	/*
	 * allocate a new dcache entry
	 */
	file->f_dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
	if (!file->f_dentry) goto out;

	file->f_dentry->d_op = &pfmfs_dentry_operations;

	d_add(file->f_dentry, inode);
	file->f_vfsmnt = mntget(pfmfs_mnt);
	file->f_mapping = inode->i_mapping;

	file->f_op    = &pfm_file_ops;
	file->f_mode  = FMODE_READ;
	file->f_flags = O_RDONLY;
	file->f_pos   = 0;

	/*
	 * may have to delay until context is attached?
	 */
	fd_install(fd, file);

	/*
	 * the file structure we will use
	 */
	*cfile = file;

	return fd;
out:
	if (file) put_filp(file);
	put_unused_fd(fd);
	return ret;
}

static void
pfm_free_fd(int fd, struct file *file)
{
	if (file) put_filp(file);
	put_unused_fd(fd);
}

static int
pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
{
	unsigned long page;

	DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));

	while (size > 0) {
		page = pfm_kvirt_to_pa(buf);

		if (pfm_remap_page_range(vma, addr, page, PAGE_SIZE, PAGE_READONLY)) return -ENOMEM;

		addr  += PAGE_SIZE;
		buf   += PAGE_SIZE;
		size  -= PAGE_SIZE;
	}
	return 0;
}

/*
 * allocate a sampling buffer and remaps it into the user address space of the task
 */
static int
pfm_smpl_buffer_alloc(struct task_struct *task, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
{
	struct mm_struct *mm = task->mm;
	struct vm_area_struct *vma = NULL;
	unsigned long size;
	void *smpl_buf;


	/*
	 * the fixed header + requested size and align to page boundary
	 */
	size = PAGE_ALIGN(rsize);

	DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));

	/*
	 * check requested size to avoid Denial-of-service attacks
	 * XXX: may have to refine this test
	 * Check against address space limit.
	 *
	 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
	 * 	return -ENOMEM;
	 */
	if (size > task->rlim[RLIMIT_MEMLOCK].rlim_cur) return -EAGAIN;

	/*
	 * We do the easy to undo allocations first.
 	 *
	 * pfm_rvmalloc(), clears the buffer, so there is no leak
	 */
	smpl_buf = pfm_rvmalloc(size);
	if (smpl_buf == NULL) {
		DPRINT(("Can't allocate sampling buffer\n"));
		return -ENOMEM;
	}

	DPRINT(("smpl_buf @%p\n", smpl_buf));

	/* allocate vma */
	vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
	if (!vma) {
		DPRINT(("Cannot allocate vma\n"));
		goto error_kmem;
	}
	memset(vma, 0, sizeof(*vma));

	/*
	 * partially initialize the vma for the sampling buffer
	 *
	 * The VM_DONTCOPY flag is very important as it ensures that the mapping
	 * will never be inherited for any child process (via fork()) which is always
	 * what we want.
	 */
	vma->vm_mm	     = mm;
	vma->vm_flags	     = VM_READ| VM_MAYREAD |VM_RESERVED;
	vma->vm_page_prot    = PAGE_READONLY; /* XXX may need to change */

	/*
	 * Now we have everything we need and we can initialize
	 * and connect all the data structures
	 */

	ctx->ctx_smpl_hdr   = smpl_buf;
	ctx->ctx_smpl_size  = size; /* aligned size */

	/*
	 * Let's do the difficult operations next.
	 *
	 * now we atomically find some area in the address space and
	 * remap the buffer in it.
	 */
	down_write(&task->mm->mmap_sem);

	/* find some free area in address space, must have mmap sem held */
	vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
	if (vma->vm_start == 0UL) {
		DPRINT(("Cannot find unmapped area for size %ld\n", size));
		up_write(&task->mm->mmap_sem);
		goto error;
	}
	vma->vm_end = vma->vm_start + size;

	DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));

	/* can only be applied to current task, need to have the mm semaphore held when called */
	if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
		DPRINT(("Can't remap buffer\n"));
		up_write(&task->mm->mmap_sem);
		goto error;
	}

	/*
	 * now insert the vma in the vm list for the process, must be
	 * done with mmap lock held
	 */
	insert_vm_struct(mm, vma);

	mm->total_vm  += size >> PAGE_SHIFT;

	up_write(&task->mm->mmap_sem);

	/*
	 * keep track of user level virtual address
	 */
	ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
	*(unsigned long *)user_vaddr = vma->vm_start;

	return 0;

error:
	kmem_cache_free(vm_area_cachep, vma);
error_kmem:
	pfm_rvfree(smpl_buf, size);

	return -ENOMEM;
}

/*
 * XXX: do something better here
 */
static int
pfm_bad_permissions(struct task_struct *task)
{
	/* inspired by ptrace_attach() */
	DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
		current->uid,
		current->gid,
		task->euid,
		task->suid,
		task->uid,
		task->egid,
		task->sgid));

	return ((current->uid != task->euid)
	    || (current->uid != task->suid)
	    || (current->uid != task->uid)
	    || (current->gid != task->egid)
	    || (current->gid != task->sgid)
	    || (current->gid != task->gid)) && !capable(CAP_SYS_PTRACE);
}

static int
pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
{
	int ctx_flags;

	/* valid signal */

	ctx_flags = pfx->ctx_flags;

	if (ctx_flags & PFM_FL_SYSTEM_WIDE) {

		/*
		 * cannot block in this mode
		 */
		if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
			DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
			return -EINVAL;
		}
	} else {
	}
	/* probably more to add here */

	return 0;
}

static int
pfm_setup_buffer_fmt(struct task_struct *task, pfm_context_t *ctx, unsigned int ctx_flags,
		     unsigned int cpu, pfarg_context_t *arg)
{
	pfm_buffer_fmt_t *fmt = NULL;
	unsigned long size = 0UL;
	void *uaddr = NULL;
	void *fmt_arg = NULL;
	int ret = 0;
#define PFM_CTXARG_BUF_ARG(a)	(pfm_buffer_fmt_t *)(a+1)

	/* invoke and lock buffer format, if found */
	fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
	if (fmt == NULL) {
		DPRINT(("[%d] cannot find buffer format\n", task->pid));
		return -EINVAL;
	}

	/*
	 * buffer argument MUST be contiguous to pfarg_context_t
	 */
	if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);

	ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);

	DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task->pid, ctx_flags, cpu, fmt_arg, ret));

	if (ret) goto error;

	/* link buffer format and context */
	ctx->ctx_buf_fmt = fmt;

	/*
	 * check if buffer format wants to use perfmon buffer allocation/mapping service
	 */
	ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
	if (ret) goto error;

	if (size) {
		/*
		 * buffer is always remapped into the caller's address space
		 */
		ret = pfm_smpl_buffer_alloc(current, ctx, size, &uaddr);
		if (ret) goto error;

		/* keep track of user address of buffer */
		arg->ctx_smpl_vaddr = uaddr;
	}
	ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);

error:
	return ret;
}

static void
pfm_reset_pmu_state(pfm_context_t *ctx)
{
	int i;

	/*
	 * install reset values for PMC.
	 */
	for (i=1; PMC_IS_LAST(i) == 0; i++) {
		if (PMC_IS_IMPL(i) == 0) continue;
		ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
		DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
	}
	/*
	 * PMD registers are set to 0UL when the context in memset()
	 */

	/*
	 * On context switched restore, we must restore ALL pmc and ALL pmd even
	 * when they are not actively used by the task. In UP, the incoming process
	 * may otherwise pick up left over PMC, PMD state from the previous process.
	 * As opposed to PMD, stale PMC can cause harm to the incoming
	 * process because they may change what is being measured.
	 * Therefore, we must systematically reinstall the entire
	 * PMC state. In SMP, the same thing is possible on the
	 * same CPU but also on between