lkml.org 
[lkml]   [2011]   [Oct]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v5 3.1.0-rc4-tip 5/26] Uprobes: copy of the original instruction.
On 09/20, Srikar Dronamraju wrote:
>
> +static int __copy_insn(struct address_space *mapping,
> + struct vm_area_struct *vma, char *insn,
> + unsigned long nbytes, unsigned long offset)
> +{
> + struct file *filp = vma->vm_file;
> + struct page *page;
> + void *vaddr;
> + unsigned long off1;
> + unsigned long idx;
> +
> + if (!filp)
> + return -EINVAL;
> +
> + idx = (unsigned long) (offset >> PAGE_CACHE_SHIFT);
> + off1 = offset &= ~PAGE_MASK;
> +
> + /*
> + * Ensure that the page that has the original instruction is
> + * populated and in page-cache.
> + */

Hmm. But how we can ensure?

> + page_cache_sync_readahead(mapping, &filp->f_ra, filp, idx, 1);

This schedules the i/o,

> + page = grab_cache_page(mapping, idx);

This finds/locks the page in the page-cache,

> + if (!page)
> + return -ENOMEM;
> +
> + vaddr = kmap_atomic(page);
> + memcpy(insn, vaddr + off1, nbytes);

What if this page is not PageUptodate() ?

Somehow this assumes that the i/o was already completed, I don't
understand this.

But I am starting to think I simply do not understand this change.
To the point, I do not underestand why do we need copy_insn() at all.
We are going to replace this page, can't we save/analyze ->insn later
when we copy the content of the old page? Most probably I missed
something simple...


> +static struct task_struct *get_mm_owner(struct mm_struct *mm)
> +{
> + struct task_struct *tsk;
> +
> + rcu_read_lock();
> + tsk = rcu_dereference(mm->owner);
> + if (tsk)
> + get_task_struct(tsk);
> + rcu_read_unlock();
> + return tsk;
> +}

Hmm. Do we really need task_struct?

> -static int install_breakpoint(struct mm_struct *mm, struct uprobe *uprobe)
> +static int install_breakpoint(struct mm_struct *mm, struct uprobe *uprobe,
> + struct vm_area_struct *vma, loff_t vaddr)
> {
> - /* Placeholder: Yet to be implemented */
> + struct task_struct *tsk;
> + unsigned long addr;
> + int ret = -EINVAL;
> +
> if (!uprobe->consumers)
> return 0;
>
> - atomic_inc(&mm->mm_uprobes_count);
> - return 0;
> + tsk = get_mm_owner(mm);
> + if (!tsk) /* task is probably exiting; bail-out */
> + return -ESRCH;
> +
> + if (vaddr > TASK_SIZE_OF(tsk))
> + goto put_return;

But this should not be possible, no? How it can map this vaddr above
TASK_SIZE ?

get_user_pages(tsk => NULL) is fine. Why else do we need mm->owner ?

Probably used by the next patches... Say, is_32bit_app(tsk). This
can use mm->context.ia32_compat (hopefully will be replaced with
MMF_COMPAT).

Oleg.



\
 
 \ /
  Last update: 2011-10-03 18:37    [W:0.501 / U:13.032 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site