Quoting Wolfram Sang (2018-09-18 15:16:46)
>
> > This fixes a problem where the kernel oopses cleaning pages for a buffer
> > that's mapped into the vmalloc space. The pages are returned from
> > request_firmware() and passed down directly to the i2c master to write
> > to the i2c touchscreen device. Mapping vmalloc buffers with
> > dma_map_single() won't work reliably, causing an oops like below:
>
> Exactly the reason why I implemented I2C_M_DMA_SAFE. Did you also notice
> the helper i2c_get_dma_safe_msg_buf() which you maybe could use for
> len > 32?
>
Yes I noticed that after sending the patch, thanks for pointing it out.
But now when I try to use it I'm not exicted when the buffer is bounced
but we fail to map the buffer with the DMA APIs. For an I2C_M_RD
message, presumably we would call i2c_release_dma_safe_msg_buf() in this
error case, but that will cause the original buffer to be copied over
which seems wasteful to do, but I guess it's OK. I suppose we could have
another function like:
void i2c_release_dma_safe_msg_buf_on_err(struct i2c_msg *msg, u8 *buf)
{
if (!buf || buf == msg->buf)
return;
kfree(buf);
}
so that we don't copy over the buffer on failure and still properly free
the buffer that we setup. Or we can pass an argument to
i2c_release_dma_safe_msg_buf() to indicate if we should do the memcpy or
not? Removing the I2C_M_RD flag from the message on failure doesn't
sound like a good idea.
Either way, I can resend the patch with the releasing and duplicate
memcpy and we can discuss this minor optimization.
