Re: [PATCH v3 7/8] dma-buf: transfer memcg charges via DMA_BUF_IOCTL_XFER_CHARGE

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 3 potential issue(s) to consider:
- [High] dma-buf: unprivileged cgroup escape and OOM via dma_buf_xfer_charge()
- [High] dma-buf: race in mem_cgroup_transfer_dmabuf_charge() double-charges memcg
- [Medium] dma-buf: broken disabled-memcg handling in dma_buf_xfer_charge()
--

commit 1660729e62b15fd915c69f26225530bd5ae30ce0
Author: Albert Esteve <aesteve@redhat.com>

dma-buf: transfer memcg charges via DMA_BUF_IOCTL_XFER_CHARGE

This commit adds DMA_BUF_IOCTL_XFER_CHARGE to allow transferring a dma-buf's
memory cgroup charge to a cgroup identified by a cgroupfd or pidfd after
allocation. It also introduces a memcg_account flag in dma_buf_export_info for
exporters to opt in to memcg accounting, enabling it for the system heap.

> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
> index f31f49ebd0af3..07513ecedc105 100644
> --- a/drivers/dma-buf/dma-buf.c
> +++ b/drivers/dma-buf/dma-buf.c
> @@ -540,6 +540,45 @@ static long dma_buf_import_sync_file(struct dma_buf *dmabuf,
>  }
>  #endif
>  
> +/* Transfer a dma-buf's memcg charge; see struct dma_buf_transfer_charge */
> +static long dma_buf_xfer_charge(struct dma_buf *dmabuf,
> +				const void __user *user_data)
> +{
> +	struct dma_buf_transfer_charge arg;
> +	struct mem_cgroup *memcg;
> +	int ret = 0;
> +
> +	if (copy_from_user(&arg, user_data, sizeof(arg)))
> +		return -EFAULT;
> +
> +	if (!IS_ENABLED(CONFIG_MEMCG))
> +		return 0;
> +
> +	if (arg.flags != 0)
> +		return -EINVAL;
> +
> +	if (!dmabuf->memcg_account)
> +		return -EOPNOTSUPP;

[Severity: Medium]
When CONFIG_MEMCG is disabled, should dma_buf_xfer_charge() still validate
arg.flags and dmabuf->memcg_account, or return -EOPNOTSUPP?

Returning 0 before checking arg.flags != 0 and !dmabuf->memcg_account causes
calls with invalid flags, non-accounting exporters, or invalid file
descriptors to silently succeed under CONFIG_MEMCG=n, while also leaving the
!CONFIG_MEMCG stubs for mem_cgroup_get_from_fd() (which returns
ERR_PTR(-EOPNOTSUPP)) and mem_cgroup_transfer_dmabuf_charge() in
include/linux/memcontrol.h unreachable.

Also, when CONFIG_MEMCG=y and cgroup_disable=memory is passed on the kernel
command line (so mem_cgroup_disabled() is true at runtime),
mem_cgroup_get_from_fd() behaves inconsistently across its two branches:

mm/memcontrol.c:mem_cgroup_get_from_fd() {
	cgrp = cgroup_get_from_fd(fd);
	if (!IS_ERR(cgrp)) {
		rcu_read_lock();
		css = cgroup_css(cgrp, &memory_cgrp_subsys);
		if (!css || !css_tryget_online(css)) {
			rcu_read_unlock();
			cgroup_put(cgrp);
			return ERR_PTR(-ENOENT);
		}
...
	memcg = get_mem_cgroup_from_mm(mm);
	mmput(mm);
	return memcg;
}

For a cgroupfd, cgroup_css() is NULL and it returns ERR_PTR(-ENOENT), whereas
for a pidfd, get_mem_cgroup_from_mm() returns NULL rather than an ERR_PTR, so
dma_buf_xfer_charge() returns 0 for pidfds and -ENOENT for cgroupfds.

> +
> +	memcg = mem_cgroup_get_from_fd(arg.charge_fd);
> +	if (IS_ERR(memcg))
> +		return PTR_ERR(memcg);
> +
> +	if (dmabuf->memcg == memcg)
> +		goto skip_transfer;
> +
> +	if (!mem_cgroup_transfer_dmabuf_charge(dmabuf, memcg)) {
> +		ret = -ENOMEM;
> +		goto skip_transfer;
> +	}

[Severity: High]
Can an unprivileged holder of a dma-buf file descriptor use
dma_buf_xfer_charge() to escape its own memory cgroup limit or trigger OOM
kills in another cgroup?

In the call chain:

  dma_buf_ioctl()
    -> dma_buf_xfer_charge()
      -> mem_cgroup_get_from_fd()
      -> mem_cgroup_transfer_dmabuf_charge()
        -> mem_cgroup_reserve_dmabuf(new_memcg, GFP_KERNEL, nr_pages)
          -> try_charge()

neither dma_buf_xfer_charge() nor mem_cgroup_get_from_fd() checks DAC or
cgroup delegation/write permissions (such as inode_permission() with MAY_WRITE
or cgroup_procs_write_permission()), capabilities, or ptrace_may_access() on
either the source cgroup (dmabuf->memcg) being uncharged or the target
cgroup/task (new_memcg) being charged.

Because /sys/fs/cgroup directories are normally world-readable (mode 0755, so
any user can open them with O_RDONLY | O_DIRECTORY) and pidfd_open() can open
a pidfd for PID 1 or any visible process without privilege, an unprivileged
process can resolve a handle to the root cgroup or to another cgroup on the
system.

When SELinux is not active, security_dma_buf_charge_cgroup() and
security_dma_buf_charge_task() return 0. Even with SELinux enabled,
selinux_dma_buf_charge_cgroup() returns 0 when the cgroup kernfs node has no
explicit security.selinux xattr set:

security/selinux/hooks.c:selinux_dma_buf_charge_cgroup() {
	rc = kernfs_xattr_get(cgrp->kn, XATTR_NAME_SELINUX, NULL, 0);
	if (rc == -ENODATA)
		return 0;
}

and neither hook checks permission to uncharge the existing dmabuf->memcg. As
a result, a confined client receiving a pre-charged dma-buf fd can move the
charge to the root cgroup to bypass its own memory.max limit, or move the
charge into a victim cgroup where mem_cgroup_reserve_dmabuf() with GFP_KERNEL
triggers direct reclaim and the memcg OOM killer.

> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index 7b841b9057594..3adc84fe6668d 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -5786,6 +5786,47 @@ void mem_cgroup_cancel_dmabuf(struct mem_cgroup *memcg, unsigned int nr_pages)
>  		refill_stock(memcg, nr_pages);
>  }
>  
> +/**
> + * mem_cgroup_transfer_dmabuf_charge - move a dma-buf charge
> + * @dmabuf: buffer to transfer charge
> + * @new_memcg: target memcg; caller must hold a reference
> + *
> + * Transfers the charge for @dmabuf to @new_memcg and, on
> + * success, stores @new_memcg in @dmabuf->memcg.
> + * Returns %true on success, %false if the charge against @new_memcg
> + * failed or a concurrent transfer raced on @dmabuf->memcg.
> + * On failure the caller retains its reference to @new_memcg.
> + */
> +bool mem_cgroup_transfer_dmabuf_charge(struct dma_buf *dmabuf,
> +				       struct mem_cgroup *new_memcg)
> +{
> +	struct mem_cgroup *old_memcg = READ_ONCE(dmabuf->memcg);
> +	unsigned int nr_pages = PAGE_ALIGN(dmabuf->size) >> PAGE_SHIFT;
> +
> +	if (mem_cgroup_disabled())
> +		return true;
> +
> +	if (!mem_cgroup_reserve_dmabuf(new_memcg, GFP_KERNEL, nr_pages))
> +		return false;
> +
> +	mod_memcg_state(new_memcg, MEMCG_DMABUF, nr_pages);
> +
> +	if (cmpxchg(&dmabuf->memcg, old_memcg, new_memcg) != old_memcg) {
> +		mod_memcg_state(new_memcg, MEMCG_DMABUF, -nr_pages);
> +		mem_cgroup_cancel_dmabuf(new_memcg, nr_pages);
> +		return false;
> +	}

[Severity: High]
What happens if two threads concurrently call DMA_BUF_IOCTL_XFER_CHARGE on the
same dma-buf?

In dma_buf_xfer_charge(), dmabuf->memcg == memcg is checked without
synchronization before calling mem_cgroup_transfer_dmabuf_charge(), which then
re-reads old_memcg = READ_ONCE(dmabuf->memcg) without checking whether
old_memcg == new_memcg.

If thread 1 transfers the buffer to new_memcg after thread 2 passes the
dmabuf->memcg == memcg check in dma_buf_xfer_charge() (or if both threads
concurrently transfer the buffer to the same new_memcg), thread 2 sees
old_memcg == new_memcg and still calls mem_cgroup_reserve_dmabuf(new_memcg,
GFP_KERNEL, nr_pages) before uncharging old_memcg. That transiently charges
2 * nr_pages against new_memcg for the same buffer, which can trigger reclaim,
memcg OOM kills, or -ENOMEM in new_memcg even though the buffer is already
charged to it.

In addition, old_memcg is read before mem_cgroup_reserve_dmabuf(new_memcg,
GFP_KERNEL, nr_pages), which can sleep in direct reclaim. If another thread
updates dmabuf->memcg while mem_cgroup_reserve_dmabuf() is sleeping, the
cmpxchg(&dmabuf->memcg, old_memcg, new_memcg) fails without retrying against
the updated old_memcg and returns false, which dma_buf_xfer_charge() reports
to userspace as -ENOMEM even though charging new_memcg succeeded.

Finally, if mem_cgroup_transfer_dmabuf_charge() is called when
mem_cgroup_disabled() is true, returning true at the top without storing
new_memcg in dmabuf->memcg or dropping new_memcg's reference would leak the
 caller's reference to new_memcg, since callers only call mem_cgroup_put() when
this function returns false.

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261027-v2_20230123_tjmercier_google_com-v3-0-78d2c4cb8647@redhat.com?part=7