Re: [RFC PATCH] cw1200: use kmalloc() allocation instead of stack

On Wed, Jun 30, 2021 at 11:56 AM Ulf Hansson [off-list ref] wrote:

On Tue, 22 Jun 2021 at 22:33, Arnd Bergmann [off-list ref] wrote:

quoted

On Tue, Jun 22, 2021 at 10:24 PM Jernej Skrabec
[off-list ref] wrote:

quoted

It turns out that if CONFIG_VMAP_STACK is enabled and src or dst is
memory allocated on stack, SDIO operations fail due to invalid memory
address conversion:

Thank you for sending this!

It's worth pointing out that even without CONFIG_VMAP_STACK, using
dma_map_sg() on a stack variable is broken, though it will appear to
work most of the time but rarely cause a stack data corruption when
the cache management goes wrong.

This clearly needs to be fixed somewhere, if not with your patch, then
a similar one.

quoted

diff --git a/drivers/net/wireless/st/cw1200/hwio.c b/drivers/net/wireless/st/cw1200/hwio.c
index 3ba462de8e91..5521cb7f2233 100644
--- a/drivers/net/wireless/st/cw1200/hwio.c
+++ b/drivers/net/wireless/st/cw1200/hwio.c

@@ -66,33 +66,65 @@ static int __cw1200_reg_write(struct cw1200_common *priv, u16 addr,
 static inline int __cw1200_reg_read_32(struct cw1200_common *priv,
                                        u16 addr, u32 *val)
 {
-       __le32 tmp;
-       int i = __cw1200_reg_read(priv, addr, &tmp, sizeof(tmp), 0);
-       *val = le32_to_cpu(tmp);
+       __le32 *tmp;
+       int i;
+
+       tmp = kmalloc(sizeof(*tmp), GFP_KERNEL);
+       if (!tmp)
+               return -ENOMEM;
+
+       i = __cw1200_reg_read(priv, addr, tmp, sizeof(*tmp), 0);
+       *val = le32_to_cpu(*tmp);
+       kfree(tmp);
        return i;
 }

There is a possible problem here when the function gets called from
atomic context, so it might need to use GFP_ATOMIC instead of
GFP_KERNEL. If it's never called from atomic context, then this patch
looks correct to me.

I would be surprised if this is called from atomic context (when IRQs
are turned off), because in most cases, to complete the read/write
request the mmc controller driver relies on IRQs being delivered.

I thought I had seen a spinlock in the forked driver, but I don't see
it now, so I probably misremembered that bit.

quoted

The alternative would be to add a bounce buffer check based on
is_vmalloc_or_module_addr() in sdio_io_rw_ext_helper(), which would
add a small bit of complexity there but solve the problem for
all drivers at once. In this case, it would probably have to use
GFP_ATOMIC regardless of whether __cw1200_reg_read_32()
is allowed to sleep, since other callers might not.

I like the idea, but...

I don't think we should see this as an alternative, but rather as a
complement which would have performance issues. A warning should be
printed, if the buffer isn't properly allocated.

Fair enough. I found the function call I was looking for: object_is_on_stack(),
the patch below should print a warning once when a driver passes
a bad buffer, but I did not test that.

There are some possible variations on that: an on-stack buffer by
itself can work as long as the DMA is cache-coherent and stacks
are not vmapped. For the is_vmalloc_or_module_addr() case,
we may decide to just return an error, rather than running into
a kernel oops.

Additionally, I don't think GFT_ATOMIC should be needed.

Ok, I now see the mmc_wait_for_req() in mmc_io_rw_extended()
that probably means it can not be called in atomic context at all,
and that GFP_KERNEL is safe, and that any driver calling it with
a spinlock held is already broken.

       Arnd

8<---

diff --git a/drivers/mmc/core/sdio_ops.c b/drivers/mmc/core/sdio_ops.c
index 4c229dd2b6e5..845f9ca3b200 100644
--- a/drivers/mmc/core/sdio_ops.c
+++ b/drivers/mmc/core/sdio_ops.c

@@ -124,6 +124,7 @@ int mmc_io_rw_extended(struct mmc_card *card, int

write, unsigned fn,
        int err;

        WARN_ON(blksz == 0);
+       WARN_ON_ONCE(is_vmalloc_or_module_addr(buf) || object_is_on_stack(buf));

        /* sanity check */
        if (addr & ~0x1FFFF)