Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <20220629082541.118-1-xieyongji@bytedance.com> <20220629082541.118-6-xieyongji@bytedance.com>
 <20220629043539-mutt-send-email-mst@kernel.org> <CACycT3sAcH-b40hORjSOQb67jZ0Fd-fxdzmZNwt=4iZdX6gLeA@mail.gmail.com>
 <20220629055241-mutt-send-email-mst@kernel.org>
In-Reply-To: <20220629055241-mutt-send-email-mst@kernel.org>
From:   Yongji Xie <xieyongji@bytedance.com>
Date:   Wed, 29 Jun 2022 18:19:31 +0800
Message-ID: <CACycT3vaNLYRid5SsT11LuVCaGXbBfV=q7c7SUp1+r9BcRpwkw@mail.gmail.com>
Subject: Re: [PATCH 5/6] vduse: Support registering userspace memory for IOTLB
To:     "Michael S. Tsirkin" <mst@redhat.com>
Cc:     Jason Wang <jasowang@redhat.com>,
        Liu Xiaodong <xiaodong.liu@intel.com>,
        Maxime Coquelin <maxime.coquelin@redhat.com>,
        Stefan Hajnoczi <stefanha@redhat.com>,
        virtualization <virtualization@lists.linux-foundation.org>,
        linux-kernel <linux-kernel@vger.kernel.org>,
        songmuchun@bytedance.com
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Wed, Jun 29, 2022 at 5:54 PM Michael S. Tsirkin <mst@redhat.com> wrote:
>
> On Wed, Jun 29, 2022 at 05:26:04PM +0800, Yongji Xie wrote:
> > On Wed, Jun 29, 2022 at 4:43 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> > >
> > > On Wed, Jun 29, 2022 at 04:25:40PM +0800, Xie Yongji wrote:
> > > > Introduce two ioctls: VDUSE_IOTLB_REG_UMEM and
> > > > VDUSE_IOTLB_DEREG_UMEM to support registering
> > > > and de-registering userspace memory for IOTLB
> > > > in virtio-vdpa case.
> > > >
> > > > Now it only supports registering userspace memory
> > > > for IOTLB as bounce buffer.
> > > >
> > > > Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
> > > > ---
> > > >  drivers/vdpa/vdpa_user/vduse_dev.c | 138 +++++++++++++++++++++++++++++
> > > >  include/uapi/linux/vduse.h         |  28 ++++++
> > > >  2 files changed, 166 insertions(+)
> > > >
> > > > diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > index c47a5d9765cf..7b2ea7612da9 100644
> > > > --- a/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > +++ b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > @@ -21,6 +21,7 @@
> > > >  #include <linux/uio.h>
> > > >  #include <linux/vdpa.h>
> > > >  #include <linux/nospec.h>
> > > > +#include <linux/sched/mm.h>
> > > >  #include <uapi/linux/vduse.h>
> > > >  #include <uapi/linux/vdpa.h>
> > > >  #include <uapi/linux/virtio_config.h>
> > > > @@ -64,6 +65,13 @@ struct vduse_vdpa {
> > > >       struct vduse_dev *dev;
> > > >  };
> > > >
> > > > +struct vduse_iotlb_mem {
> > > > +     unsigned long iova;
> > > > +     unsigned long npages;
> > > > +     struct page **pages;
> > > > +     struct mm_struct *mm;
> > > > +};
> > > > +
> > > >  struct vduse_dev {
> > > >       struct vduse_vdpa *vdev;
> > > >       struct device *dev;
> > > > @@ -95,6 +103,8 @@ struct vduse_dev {
> > > >       u8 status;
> > > >       u32 vq_num;
> > > >       u32 vq_align;
> > > > +     struct vduse_iotlb_mem *iotlb_mem;
> > > > +     struct mutex mem_lock;
> > > >  };
> > > >
> > > >  struct vduse_dev_msg {
> > > > @@ -917,6 +927,100 @@ static int vduse_dev_queue_irq_work(struct vduse_dev *dev,
> > > >       return ret;
> > > >  }
> > > >
> > > > +static int vduse_dev_dereg_iotlb_mem(struct vduse_dev *dev,
> > > > +                                  u64 iova, u64 size)
> > > > +{
> > > > +     int ret;
> > > > +
> > > > +     mutex_lock(&dev->mem_lock);
> > > > +     ret = -ENOENT;
> > > > +     if (!dev->iotlb_mem)
> > > > +             goto unlock;
> > > > +
> > > > +     ret = -EINVAL;
> > > > +     if (dev->iotlb_mem->iova != iova || size != dev->domain->bounce_size)
> > > > +             goto unlock;
> > > > +
> > > > +     vduse_domain_remove_user_bounce_pages(dev->domain);
> > > > +     unpin_user_pages(dev->iotlb_mem->pages, dev->iotlb_mem->npages);
> > >
> > > I notice you don't mark the pages dirty. This is going to be a problem.
> > >
> >
> > Thanks for pointing out this, I will use unpin_user_pages_dirty_lock() instead.
> >
> > > > +     atomic64_sub(dev->iotlb_mem->npages, &dev->iotlb_mem->mm->pinned_vm);
> > > > +     mmdrop(dev->iotlb_mem->mm);
> > > > +     vfree(dev->iotlb_mem->pages);
> > > > +     kfree(dev->iotlb_mem);
> > > > +     dev->iotlb_mem = NULL;
> > > > +     ret = 0;
> > > > +unlock:
> > > > +     mutex_unlock(&dev->mem_lock);
> > > > +     return ret;
> > > > +}
> > > > +
> > > > +static int vduse_dev_reg_iotlb_mem(struct vduse_dev *dev,
> > > > +                                u64 iova, u64 uaddr, u64 size)
> > > > +{
> > > > +     struct page **page_list = NULL;
> > > > +     struct vduse_iotlb_mem *mem = NULL;
> > > > +     long pinned = 0;
> > > > +     unsigned long npages, lock_limit;
> > > > +     int ret;
> > > > +
> > > > +     if (size != dev->domain->bounce_size ||
> > > > +         iova != 0 || uaddr & ~PAGE_MASK)
> > > > +             return -EINVAL;
> > > > +
> > > > +     mutex_lock(&dev->mem_lock);
> > > > +     ret = -EEXIST;
> > > > +     if (dev->iotlb_mem)
> > > > +             goto unlock;
> > > > +
> > > > +     ret = -ENOMEM;
> > > > +     npages = size >> PAGE_SHIFT;
> > > > +     page_list = vmalloc(array_size(npages,
> > > > +                         sizeof(struct page *)));
> > >
> > > Is this basically trying to do a vmalloc with userspace-controlled size?
> > > That's an easy DOS vector.
> > >
> >
> > We already checked the size before. The size must equal to (64MB >>
> > PAGE_SHIFT) now.
>
> That's not a small amount. Can this be accounted e.g. through cgroups at least?
>

Make sense, will use __vmalloc(__GFP_ACCOUNT) instead.

> > > > +     mem = kzalloc(sizeof(*mem), GFP_KERNEL);
> > > > +     if (!page_list || !mem)
> > > > +             goto unlock;
> > > > +
> > > > +     mmap_read_lock(current->mm);
> > > > +
> > > > +     lock_limit = PFN_DOWN(rlimit(RLIMIT_MEMLOCK));
> > > > +     if (npages + atomic64_read(&current->mm->pinned_vm) > lock_limit)
> > > > +             goto out;
> > > > +
> > > > +     pinned = pin_user_pages(uaddr, npages, FOLL_LONGTERM | FOLL_WRITE,
> > > > +                             page_list, NULL);
> > > > +     if (pinned != npages) {
> > > > +             ret = pinned < 0 ? pinned : -ENOMEM;
> > > > +             goto out;
> > > > +     }
> > >
> > >
> > > This is a popular approach but it's problematic if multiple
> > > devices try to pin the same page.
> >
> > Do you mean the data would be corrupted if multiple devices use the
> > same page as bounce buffer? This is indeed a problem.
>
> No i mean you decrement the lock twice. Question is can two bounce
> buffers share a page?
>

I think we can't. I will find a way to prevent it.

Thanks,
Yongji