linux-kernel.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: Oleg Nesterov <oleg@redhat.com>
To: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>,
	Ingo Molnar <mingo@elte.hu>,
	Andrew Morton <akpm@linux-foundation.org>,
	Linus Torvalds <torvalds@linux-foundation.org>,
	Masami Hiramatsu <mhiramat@redhat.com>,
	Randy Dunlap <rdunlap@xenotime.net>,
	Ananth N Mavinakayanahalli <ananth@in.ibm.com>,
	Jim Keniston <jkenisto@linux.vnet.ibm.com>,
	Frederic Weisbecker <fweisbec@gmail.com>,
	"Frank Ch. Eigler" <fche@redhat.com>,
	LKML <linux-kernel@vger.kernel.org>,
	Roland McGrath <roland@redhat.com>
Subject: Re: [PATCH v2 7/11] Uprobes Implementation
Date: Tue, 13 Apr 2010 20:35:37 +0200	[thread overview]
Message-ID: <20100413183537.GA17538@redhat.com> (raw)
In-Reply-To: <20100331155228.4181.61294.sendpatchset@localhost6.localdomain6>

On 03/31, Srikar Dronamraju wrote:
>
> --- /dev/null
> +++ b/kernel/uprobes.c
> ...
> +static struct uprobe_process *find_uprocess(struct pid *tg_leader)
> +{
> +	struct uprobe_process *uproc;
> +	struct task_struct *tsk = get_pid_task(tg_leader, PIDTYPE_PID);
> +
> +	if (!tsk)
> +		return NULL;
> +
> +	if (!tsk->utask || !tsk->utask->uproc) {
> +		put_task_struct(tsk);
> +		return NULL;
> +	}
> +
> +	uproc = tsk->utask->uproc;
> +	BUG_ON(uproc->tg_leader != tg_leader);
> +	atomic_inc(&uproc->refcount);
> +	put_task_struct(tsk);
> +	return uproc;

Looks like, this doesn't need get/put task_struct, you could just
use pid_task() under rcu_read_lock().

> +static void cleanup_uprocess(struct uprobe_process *uproc,
> +					struct task_struct *start)
> +{
> +	struct task_struct *tsk = start;
> +
> +	if (!start)
> +		return;
> +
> +	rcu_read_lock();
> +	do {
> +		if (tsk->utask) {
> +			kfree(tsk->utask);
> +			tsk->utask = NULL;
> +		}
> +		tsk = next_thread(tsk);

This doesn't look right. We can't trust ->thread_group list even under
rcu_read_lock(). The task can exit and __exit_signal() can remove it
from ->thread_group list before we take rcu_read_lock().

> +static struct uprobe_task *add_utask(struct task_struct *t,
> +					struct uprobe_process *uproc)
> +{
> +	struct uprobe_task *utask;
> +
> +	if (!t)
> +		return NULL;
> +	utask = kzalloc(sizeof *utask, GFP_USER);
> +	if (unlikely(utask == NULL))
> +		return ERR_PTR(-ENOMEM);
> +
> +	utask->uproc = uproc;
> +	utask->active_ppt = NULL;
> +	t->utask = utask;
> +	atomic_inc(&uproc->refcount);
> +
> +	return utask;
> +}

This is called by create_uprocess(). Who will free t->utask if t has
already passed tracehook_report_exit() ?

> +static struct task_struct *find_next_thread(struct uprobe_process *uproc,
> +						struct task_struct *start)
> +{
> +	struct task_struct *next_t = NULL;
> +
> +	rcu_read_lock();
> +	if (start) {
> +		struct task_struct *t = start;
> +
> +		do {
> +			if (unlikely(t->flags & PF_EXITING))
> +				goto dont_add;

not sure I understand this check. Somehow we should prevent the races
with tracehook_report_exit/tracehook_report_exec, but PF_EXITING can't
help ?

> +dont_add:
> +			t = next_thread(t);
> +		} while (t != start);

again, this doesn't look right. Btw, I'd suggest to use while_each_thread().

> +static struct uprobe_process *create_uprocess(struct pid *tg_leader)
> +{
> ...
> +	/*
> +	 * Create and populate one utask per thread in this process.  We
> +	 * can't call add_utask() while holding RCU lock, so we:
> +	 *	1. rcu_read_lock()
> +	 *	2. Find the next thread, add_me, in this process that's not
> +	 *	having a utask struct allocated.
> +	 *	3. rcu_read_unlock()
> +	 *	4. add_utask(add_me, uproc)
> +	 *	Repeat 1-4 'til we have utasks for all threads.
> +	 */
> +	cur_t = get_pid_task(tg_leader, PIDTYPE_PID);
> +	do {
> +		utask = add_utask(cur_t, uproc);
> +		if (IS_ERR(utask)) {
> +			err = PTR_ERR(utask);
> +			goto fail;
> +		}
> +		add_me = find_next_thread(uproc, cur_t);
> +		put_task_struct(cur_t);
> +		cur_t = add_me;
> +	} while (add_me != NULL);

can't we race with clone(CLONE_THREAD) and miss the new thread? Probably
I missed something, but afaics we need some barriers to ensure that either
tracehook_report_clone() sees current->utask != NULL or find_next_thread()
sees the new thread in ->thread_group.

> +static struct pid *get_tg_leader(pid_t p)
> +{
> +	struct pid *pid = NULL;
> +
> +	rcu_read_lock();
> +	if (current->nsproxy)
> +		pid = find_vpid(p);

Is it really possible to call register/unregister with nsproxy == NULL?

> +	if (pid) {
> +		struct task_struct *t = pid_task(pid, PIDTYPE_PID);
> +
> +		if (!t || unlikely(t->flags & PF_EXITING))

Why do we check PF_EXITING?

> +int register_uprobe(struct uprobe *u)
> +{
> +	struct uprobe_process *uproc;
> +	struct uprobe_probept *ppt;
> +	struct pid *p;
> +	int ret = 0;
> +
> +	if (!u || !u->handler)
> +		return -EINVAL;
> +
> +	p = get_tg_leader(u->pid);
> +	if (!p)
> +		return -ESRCH;
> +
> +
> +	/* Get the uprobe_process for this pid, or make a new one. */
> +	mutex_lock(&uprobe_mutex);
> +	uproc = find_uprocess(p);
> +
> +	if (!uproc) {
> +		uproc = create_uprocess(p);
> +		if (IS_ERR(uproc)) {
> +			ret = (int) PTR_ERR(uproc);
> +			mutex_unlock(&uprobe_mutex);
> +			goto fail_tsk;
> +		}
> +	}
> +	mutex_unlock(&uprobe_mutex);
> +	mutex_lock(&uproc->mutex);
> +
> +	if (uproc->n_ppts >= MAX_USER_BKPT_XOL_SLOTS)
> +		goto fail_uproc;
> +
> +	ret = xol_validate_vaddr(p, u->vaddr, uproc->xol_area);

OK, uproc and p can't go away. But why it is safe to use pid_task(p) ?

I am looking at 6th patch http://marc.info/?l=linux-kernel&m=127005086102256
and xol_validate_vaddr() calls pid_task() without rcu and doesn't check
the result is not NULL.

We already dropped uprobe_mutex, can't this task exit?

> +void uprobe_handle_clone(unsigned long clone_flags,
> +				struct task_struct *child)
> +{
> +	struct uprobe_process *uproc;
> +	struct uprobe_task *ptask, *ctask;
> +
> +	ptask = current->utask;
> +	if (!ptask)
> +		return;
> +
> +	uproc = ptask->uproc;
> +
> +	if (clone_flags & CLONE_THREAD) {
> +		mutex_lock(&uprobe_mutex);
> +		/* New thread in the same process. */
> +		ctask = child->utask;
> +		if (unlikely(ctask)) {
> +			/*
> +			 * create_uprocess() ran just as this clone
> +			 * happened, and has already accounted for the
> +			 * new child.
> +			 */
> +		} else
> +			ctask = add_utask(child, uproc);

This looks a bit strange. Why do we need "ctask" at all? It is not used,
you could just do

	if (likely(!child->utask))
		add_utask(child, uproc);

The same for "else" branch.

> +	} else {
> +		struct uprobe_probept *ppt;
> +		int ret;
> +
> +		/*
> +		 * New process spawned by parent.  Remove the probepoints
> +		 * in the child's text.
> +		 *
> +		 * We also hold the uproc->mutex for the parent - so no
> +		 * new uprobes will be registered 'til we return.
> +		 */
> +		mutex_lock(&uproc->mutex);
> +		ctask = child->utask;
> +		if (unlikely(ctask)) {
> +			/*
> +			 * create_uprocess() ran just as this fork
> +			 * happened, and has already created a new utask.
> +			 */
> +			mutex_unlock(&uproc->mutex);
> +			return;
> +		}
> +		list_for_each_entry(ppt, &uproc->uprobe_list, ut_node) {
> +			ret = user_bkpt_remove_bkpt(child, &ppt->user_bkpt);

OK, iiuc this should restore the original instruction, right?

But what about clone(CLONE_VM)? In this case this child shares ->mm with
parent.

Oleg.


  reply	other threads:[~2010-04-13 18:38 UTC|newest]

Thread overview: 46+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2010-03-31 15:51 [PATCH v2 0/11] Uprobes patches Srikar Dronamraju
2010-03-31 15:51 ` [PATCH v2 1/11] Move Macro W to insn.h Srikar Dronamraju
2010-03-31 15:51 ` [PATCH v2 2/11] Move replace_page() to mm/memory.c Srikar Dronamraju
2010-03-31 15:51 ` [PATCH v2 3/11] Enhance replace_page() to support pagecache Srikar Dronamraju
2010-03-31 15:51 ` [PATCH v2 4/11] User Space Breakpoint Assistance Layer Srikar Dronamraju
2010-03-31 15:52 ` [PATCH v2 5/11] X86 details for user space breakpoint assistance Srikar Dronamraju
2010-03-31 15:52 ` [PATCH v2 6/11] Slot allocation for Execution out of line Srikar Dronamraju
2010-03-31 15:52 ` [PATCH v2 7/11] Uprobes Implementation Srikar Dronamraju
2010-04-13 18:35   ` Oleg Nesterov [this message]
2010-04-15  9:35     ` Srikar Dronamraju
2010-04-19 19:31       ` Oleg Nesterov
2010-04-20 12:43         ` Srikar Dronamraju
2010-04-20 15:30           ` Oleg Nesterov
2010-04-21  6:59             ` Srikar Dronamraju
2010-04-21 16:05               ` Oleg Nesterov
2010-04-22 13:31                 ` Srikar Dronamraju
2010-04-22 15:40                   ` Oleg Nesterov
2010-04-23 14:58                     ` Srikar Dronamraju
2010-04-23 18:53                       ` Oleg Nesterov
2010-05-11 20:47                       ` Peter Zijlstra
2010-05-11 20:44                     ` Peter Zijlstra
2010-05-11 20:45                     ` Peter Zijlstra
2010-05-12 10:31                       ` Srikar Dronamraju
2010-05-13 19:40                       ` Oleg Nesterov
2010-05-13 19:59                         ` Linus Torvalds
2010-05-13 22:12                           ` Andi Kleen
2010-05-13 22:25                             ` Linus Torvalds
2010-05-14  0:56                           ` Roland McGrath
2010-05-14  5:42                           ` Srikar Dronamraju
2010-05-11 20:43                   ` Peter Zijlstra
2010-05-12 10:41                     ` Srikar Dronamraju
2010-05-12 11:12                       ` Peter Zijlstra
2010-05-12 14:24                         ` Srikar Dronamraju
2010-05-11 20:32             ` Peter Zijlstra
2010-05-11 20:57               ` Frank Ch. Eigler
2010-05-11 21:01                 ` Peter Zijlstra
2010-03-31 15:52 ` [PATCH v2 8/11] X86 details for uprobes Srikar Dronamraju
2010-03-31 15:52 ` [PATCH v2 9/11] Uprobes Documentation patch Srikar Dronamraju
2010-03-31 15:52 ` [PATCH v2 10/11] Uprobes samples Srikar Dronamraju
2010-03-31 15:53 ` [PATCH v2 11/11] Uprobes traceevents patch Srikar Dronamraju
2010-03-31 21:24   ` Steven Rostedt
2010-04-01  4:16     ` Masami Hiramatsu
2010-05-12 14:57       ` Frederic Weisbecker
2010-05-12 11:02   ` Frederic Weisbecker
2010-05-12 14:34     ` Srikar Dronamraju
2010-05-12 15:15   ` Frederic Weisbecker

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20100413183537.GA17538@redhat.com \
    --to=oleg@redhat.com \
    --cc=akpm@linux-foundation.org \
    --cc=ananth@in.ibm.com \
    --cc=fche@redhat.com \
    --cc=fweisbec@gmail.com \
    --cc=jkenisto@linux.vnet.ibm.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=mhiramat@redhat.com \
    --cc=mingo@elte.hu \
    --cc=peterz@infradead.org \
    --cc=rdunlap@xenotime.net \
    --cc=roland@redhat.com \
    --cc=srikar@linux.vnet.ibm.com \
    --cc=torvalds@linux-foundation.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).