Re: [PATCH v3 0/10] Uprobes v3

[Srikar Dronamraju <srikar@linux.vnet.ibm.com>]

* Peter Zijlstra <peterz@infradead.org> [2010-05-11 22:59:45]:

> On Thu, 2010-05-06 at 23:31 +0530, Srikar Dronamraju wrote:
> >   - Addressed comments from Oleg, including removal of interrupt context
> >     handlers, reverting background page replacement in favour of
> >     access_process_vm(). 
> 
> 
> > +static int write_opcode(struct task_struct *tsk, unsigned long vaddr,
> > +                                               user_bkpt_opcode_t opcode)
> > +{
> > +       int ret;
> > +
> > +       if (!tsk)
> > +               return -EINVAL;
> > +
> > +       ret = access_process_vm(tsk, vaddr, &opcode, user_bkpt_opcode_sz, 1);
> > +       return (ret == user_bkpt_opcode_sz ? 0 : -EFAULT);
> > +}
> 
> Why!
> 
> That's not not the atomic sequence outlined.
> 


Yes, we had moved away from access_process_vm to background page
replacement in Version 1 and Version 2. 

One of the reasons being Mathieu suggesting to Jim in LFCS that 
for almost all architectures insertion of a breakpoint instruction on a
user page is an atomic operation, as far as the CPU is concerned.

Can you and other VM experts tell me if access_process_vm isnt going to
be atomic with respect to inserting/deleting a breakpoint instruction? 

Oleg had few questions which I didnt have answers. (Most of
which you have already answered yesterday). One thing that's still
missing is 

[ snipping from Oleg's mail: ]
-----
But suppose that the application does mprotect(PROT_WRITE)
after register_uprobe() installs the bp, now unregister_uprobe/etc
can't restore the original insn?  
---

Also I tried a write_opcode that uses background page replacement which 
addressed some of Oleg's comments.  The pseudo-code is here:
write_opcode()
{
	down_read(mmap_sem);

	get_user_pages(tsk, mm, vaddr, .. &old_page, &vma);

	anon_vma_prepare(vma);	

	new_page=alloc_page_vma(.., vma, vaddr);

	copy_user_page(new_page, old_page);

	kmap_atomic(new_page,...);

	memcpy(vaddr,..);

	kunmap_atomic(..);

	lock_page(new_page);

	old_pte = get_pte(mm,vaddr);

	replace_page(vma, new_page, old_page, old_pte);

	unlock_page(new_page);

	put_page(new_page);

	put_page(old_page);

	up_read(mmap);
}	


Will this work?

The Other VM quieries that I had were:

Is there any thing else needed for the parent process to pass on the anon_vma to
the child process. (I inserted a breakpoint in the parent and tried
removing the breakpoint in the child. 
However page_address_in_vma() (called by replace_page() returned
EFAULT because "vma->anon_vma != page_anon_vma(page)"

Do we need to take care of mem_cgroups?
Do we need to update mm counters?


--
Thanks and Regards
Srikar