[Devel] Re: [PATCH] c/r: do not hold mmap_sem while checkpointing vma's

[Oren Laadan]

Matt Helsley wrote:
> On Sun, Oct 25, 2009 at 06:23:29PM -0400, Oren Laadan wrote:
>> This patch modifies the memory checkpoint code to _not_ hold the
>> mmap_sem while dumping out the vma's.
>>
>> The problem with holding the mmap_sem is that it first takes the
>> mmap_sem and then takes the file's inode semaphore. This violates the
>> normal locking order, e,g, when taking a page fault during a copyout,
>> which is inode sem and then the mmap_sem.
>>
>> Normally this reverse locking order won't cause a lockup because a the
>> output file for the checkpoint image isn't used by the checkpointee.
>> However, there a couple of cases where it may be a problem, e.g. when
>> some async-IO happens to complete and triggers a page fault at the
>> wrong time.
>>
>> This fixes complaints from the lockdep about this reverse ordering.
>>
>> Signed-off-by: Oren Laadan <orenl at cs.columbia.edu>
>> ---
>>  checkpoint/memory.c |  133 ++++++++++++++++++++++++++++++++++++---------------
>>  1 files changed, 94 insertions(+), 39 deletions(-)

[...]

>> @@ -1288,9 +1343,9 @@ static struct mm_struct *do_restore_mm(struct ckpt_ctx *ctx)
>>  		}
>>  		set_mm_exe_file(mm, file);
>>  	}
>> +	up_write(&mm->mmap_sem);
>>
>>  	ret = _ckpt_read_buffer(ctx, mm->saved_auxv, sizeof(mm->saved_auxv));
>> -	up_write(&mm->mmap_sem);
>>  	if (ret < 0)
>>  		goto out;
>>
>> -- 
> 
> At least in the restart path it's interesting to see how Alexey did it
> without mmap_sem, at least for part of it:
> 
> http://patchwork.kernel.org/patch/25337/
> 
> (search for kstate_restore_mm_struct())

He's allocating a new mm. And he must do so because all tasks in the
tree are created sharing their parent's mm (unless it's a thread).

> Is that a feasible and more-suitable approach for the initial portions
> of mm restore?

Feasible ?  yes.
More-suitable ?  why ?

In our case, processes (unless threads) already have their "new" mm,
so you are suggesting to drop it and allocate a new one.

I'm unsure what is the issue with the current approach.

Oren.

_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers