From mboxrd@z Thu Jan  1 00:00:00 1970
From: Jason Gunthorpe <jgg@ziepe.ca>
Subject: Re: [PATCH v2] IB/umem: ib_ucontext already have tgid, remove pid
 from ib_umem structure
Date: Tue, 12 Jun 2018 22:36:20 -0600
Message-ID: <20180613043620.GA22578@ziepe.ca>
References: <1525769416-14596-1-git-send-email-lidongchen@tencent.com>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Return-path: <linux-kernel-owner@vger.kernel.org>
Content-Disposition: inline
In-Reply-To: <1525769416-14596-1-git-send-email-lidongchen@tencent.com>
Sender: linux-kernel-owner@vger.kernel.org
To: Lidong Chen <jemmy858585@gmail.com>
Cc: dledford@redhat.com, akpm@linux-foundation.org, qing.huang@oracle.com, leon@kernel.org, artemyko@mellanox.com, dan.j.williams@intel.com, linux-rdma@vger.kernel.org, linux-kernel@vger.kernel.org, adido@mellanox.com, galsha@mellanox.com, aviadye@mellanox.com, Lidong Chen <lidongchen@tencent.com>
List-Id: linux-rdma@vger.kernel.org

On Tue, May 08, 2018 at 04:50:16PM +0800, Lidong Chen wrote:
> The userspace may invoke ibv_reg_mr and ibv_dereg_mr by different threads.
> If when ibv_dereg_mr invoke and the thread which invoked ibv_reg_mr has
> exited, get_pid_task will return NULL, ib_umem_release does not decrease
> mm->pinned_vm. This patch fixes it by use tgid in ib_ucontext struct.
> 
> Signed-off-by: Lidong Chen <lidongchen@tencent.com>
> ---
>  [v2]
>  - use ib_ucontext tgid instread of tgid in ib_umem structure

I'm looking at this again, and it doesn't seem quite right..

> diff --git a/drivers/infiniband/core/umem.c b/drivers/infiniband/core/umem.c
> index 9a4e899..2b6c9b5 100644
> --- a/drivers/infiniband/core/umem.c
> +++ b/drivers/infiniband/core/umem.c
> @@ -119,7 +119,6 @@ struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
>  	umem->length     = size;
>  	umem->address    = addr;
>  	umem->page_shift = PAGE_SHIFT;
> -	umem->pid	 = get_task_pid(current, PIDTYPE_PID);
>  	/*
>  	 * We ask for writable memory if any of the following
>  	 * access flags are set.  "Local write" and "remote write"
> @@ -132,7 +131,6 @@ struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
>  		 IB_ACCESS_REMOTE_ATOMIC | IB_ACCESS_MW_BIND));
>  
>  	if (access & IB_ACCESS_ON_DEMAND) {
> -		put_pid(umem->pid);
>  		ret = ib_umem_odp_get(context, umem, access);
>  		if (ret) {
>  			kfree(umem);
> @@ -148,7 +146,6 @@ struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr,
>  
>  	page_list = (struct page **) __get_free_page(GFP_KERNEL);
>  	if (!page_list) {
> -		put_pid(umem->pid);
>  		kfree(umem);
>  		return ERR_PTR(-ENOMEM);
>  	}

in ib_umem_get we are doing this:

	down_write(&current->mm->mmap_sem);
	locked     = npages + current->mm->pinned_vm;

And then in release we now do:

	task = get_pid_task(umem->context->tgid, PIDTYPE_PID);
	if (!task)
		goto out;
	mm = get_task_mm(task);
	mm->pinned_vm -= diff;

But there is no guarantee that context->tgid and 'current' are the
same thing during ib_umem_get..

So in the dysfunctional case where someone forks and keeps the context
FD open on both sides of the fork they can cause the pinned_vm
counter to become wrong in the processes. Sounds bad..

Thus, I think we need to go back to storing the tgid in the ib_umem
and just fix it to store the group leader not the thread PID?

And then even more we need the ib_get_mr_mm() helper to make sense of
this, because all the drivers are doing the wrong thing by using the
context->tgid too.

Is that all right?

Jason