From: Oleg Nesterov <oleg@redhat.com>
To: Khalid Aziz <khalid.aziz@oracle.com>
Cc: tglx@linutronix.de, mingo@redhat.com, hpa@zytor.com,
peterz@infradead.org, akpm@linux-foundation.org,
andi.kleen@intel.com, rob@landley.net, viro@zeniv.linux.org.uk,
venki@google.com, linux-kernel@vger.kernel.org
Subject: Re: [RFC] [PATCH] Pre-emption control for userspace
Date: Tue, 4 Mar 2014 14:56:24 +0100 [thread overview]
Message-ID: <20140304135624.GA6846@redhat.com> (raw)
In-Reply-To: <1393870033-31076-1-git-send-email-khalid.aziz@oracle.com>
On 03/03, Khalid Aziz wrote:
>
> This queueing
> and subsequent CPU cycle wastage can be avoided if the locking thread
> could request to be granted an additional timeslice if its current
> timeslice runs out before it gives up the lock.
Well. I am in no position to discuss the changes in sched/fair.c. I have
to admit that I am skeptical about the whole idea/implementation, but I
leave this to sched/ maintainers.
However, at least the proc/mmap changes do not look right. I didn't read
the whole patch, just picked a "random" ->mmap function, see below.
> kernel/sched/preempt_delay.c | 39 ++++++
Why? This can go into proc/ as well.
> +static void
> +close_preempt_delay_vmops(struct vm_area_struct *vma)
> +{
> + struct preemp_delay_mmap_state *state;
> +
> + state = (struct preemp_delay_mmap_state *) vma->vm_private_data;
> + BUG_ON(!state || !state->task);
> +
> + state->page->mapping = NULL;
> + /* point delay request flag pointer back to old flag in task_struct */
> + state->task->sched_preempt_delay.delay_req =
> + &state->task->sched_preempt_delay.delay_flag;
> + state->task->sched_preempt_delay.mmap_state = NULL;
> + vfree(state->kaddr);
> + kfree(state);
> + vma->vm_private_data = NULL;
> +}
Suppose that state->task != current. Then this can race with do_exit()
which cleanups ->mmap_state too. OTOH do_exit() unmaps this region, it
is not clear why it can't rely in vm_ops->close().
Hmm. In fact I think do_exit() should crash after munmap? ->mmap_state
should be NULL ?? Perhaps I misread this patch completely...
> +static int
> +tid_preempt_delay_mmap(struct file *file, struct vm_area_struct *vma)
> +{
> + int retval = 0;
> + void *kaddr = NULL;
> + struct preemp_delay_mmap_state *state = NULL;
> + struct inode *inode = file_inode(file);
> + struct task_struct *task;
> + struct page *page;
> +
> + /*
> + * Validate args:
> + * - Only offset 0 support for now
> + * - size should be PAGE_SIZE
> + */
> + if (vma->vm_pgoff != 0 || (vma->vm_end - vma->vm_start) != PAGE_SIZE) {
> + retval = -EINVAL;
> + goto error;
> + }
> +
> + /*
> + * Only one mmap allowed at a time
> + */
> + if (current->sched_preempt_delay.mmap_state != NULL) {
> + retval = -EEXIST;
> + goto error;
This assumes that we are going to setup current->sched_preempt_delay.mmap_state,
but what if the task opens /proc/random_tid/sched_preempt_delay ?
> + state = kzalloc(sizeof(struct preemp_delay_mmap_state), GFP_KERNEL);
> + kaddr = vmalloc_user(PAGE_SIZE);
Why vmalloc() ? We only need a single page?
> + task = get_proc_task(inode);
And it seems that nobody does put_task_struct(state->task);
> + state->page = page;
> + state->kaddr = kaddr;
> + state->uaddr = (void *)vma->vm_start;
This is used by do_exit(). But ->vm_start can be changed by mremap() ?
Hmm. And mremap() can do vm_ops->close() too. But the new vma will
have the same vm_ops/vm_private_data, so exit_mmap() will try to do
this again... Perhaps I missed something, but I bet this all can't be
right.
> + state->task = task;
> +
> + /* Clear the current delay request flag */
> + task->sched_preempt_delay.delay_flag = 0;
> +
> + /* Point delay request flag pointer to the newly allocated memory */
> + task->sched_preempt_delay.delay_req = (unsigned char *)kaddr;
> +
> + task->sched_preempt_delay.mmap_state = state;
> + vma->vm_private_data = state;
> + vma->vm_ops = &preempt_delay_vmops;
> + vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_SHARED | VM_WRITE;
This probably also needs VM_IO, to protect from madvise(MADV_DOFORK).
VM_SHARED/VM_WRITE doesn't look right.
Oleg.
next prev parent reply other threads:[~2014-03-04 13:57 UTC|newest]
Thread overview: 67+ messages / expand[flat|nested] mbox.gz Atom feed top
2014-03-03 18:07 [RFC] [PATCH] Pre-emption control for userspace Khalid Aziz
2014-03-03 21:51 ` Davidlohr Bueso
2014-03-03 23:29 ` Khalid Aziz
2014-03-04 13:56 ` Oleg Nesterov [this message]
2014-03-04 17:44 ` Khalid Aziz
2014-03-04 18:38 ` Al Viro
2014-03-04 19:01 ` Khalid Aziz
2014-03-04 19:03 ` Oleg Nesterov
2014-03-04 20:14 ` Khalid Aziz
2014-03-05 14:38 ` Oleg Nesterov
2014-03-05 16:12 ` Oleg Nesterov
2014-03-05 17:10 ` Khalid Aziz
2014-03-04 21:12 ` H. Peter Anvin
2014-03-04 21:39 ` Khalid Aziz
2014-03-04 22:23 ` One Thousand Gnomes
2014-03-04 22:44 ` Khalid Aziz
2014-03-05 0:39 ` Thomas Gleixner
2014-03-05 0:51 ` Andi Kleen
2014-03-05 11:10 ` Peter Zijlstra
2014-03-05 17:29 ` Khalid Aziz
2014-03-05 19:58 ` Khalid Aziz
2014-03-06 9:57 ` Peter Zijlstra
2014-03-06 16:08 ` Khalid Aziz
2014-03-06 11:14 ` Thomas Gleixner
2014-03-06 16:32 ` Khalid Aziz
2014-03-05 14:54 ` Oleg Nesterov
2014-03-05 15:56 ` Andi Kleen
2014-03-05 16:36 ` Oleg Nesterov
2014-03-05 17:22 ` Khalid Aziz
2014-03-05 23:13 ` David Lang
2014-03-05 23:48 ` Khalid Aziz
2014-03-05 23:56 ` H. Peter Anvin
2014-03-06 0:02 ` Khalid Aziz
2014-03-06 0:13 ` H. Peter Anvin
2014-03-05 23:59 ` David Lang
2014-03-06 0:17 ` Khalid Aziz
2014-03-06 0:36 ` David Lang
2014-03-06 1:22 ` Khalid Aziz
2014-03-06 14:23 ` David Lang
2014-03-06 12:13 ` Kevin Easton
2014-03-06 13:59 ` Peter Zijlstra
2014-03-06 22:41 ` Andi Kleen
2014-03-06 14:25 ` David Lang
2014-03-06 16:12 ` Khalid Aziz
2014-03-06 13:24 ` Rasmus Villemoes
2014-03-06 13:34 ` Peter Zijlstra
2014-03-06 13:45 ` Rasmus Villemoes
2014-03-06 14:02 ` Peter Zijlstra
2014-03-06 14:33 ` Thomas Gleixner
2014-03-06 14:34 ` H. Peter Anvin
2014-03-06 14:04 ` Thomas Gleixner
2014-03-25 17:17 ` [PATCH v2] " Khalid Aziz
2014-03-25 17:44 ` Andrew Morton
2014-03-25 17:56 ` Khalid Aziz
2014-03-25 18:14 ` Andrew Morton
2014-03-25 17:46 ` Oleg Nesterov
2014-03-25 17:59 ` Khalid Aziz
2014-03-25 18:20 ` Andi Kleen
2014-03-25 18:47 ` Khalid Aziz
2014-03-25 19:47 ` Andi Kleen
2014-03-25 18:59 ` Eric W. Biederman
2014-03-25 19:15 ` Khalid Aziz
2014-03-25 20:31 ` Eric W. Biederman
2014-03-25 21:37 ` Khalid Aziz
2014-03-26 6:03 ` Mike Galbraith
2014-03-25 23:01 ` [RFC] [PATCH] " Davidlohr Bueso
2014-03-25 23:29 ` Khalid Aziz
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=20140304135624.GA6846@redhat.com \
--to=oleg@redhat.com \
--cc=akpm@linux-foundation.org \
--cc=andi.kleen@intel.com \
--cc=hpa@zytor.com \
--cc=khalid.aziz@oracle.com \
--cc=linux-kernel@vger.kernel.org \
--cc=mingo@redhat.com \
--cc=peterz@infradead.org \
--cc=rob@landley.net \
--cc=tglx@linutronix.de \
--cc=venki@google.com \
--cc=viro@zeniv.linux.org.uk \
/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).