Re: [PATCH 01/12] nvmem: imx-iim: use stack for nvmem_config instead of malloc'ing it

From: Greg KH
Date: Fri Oct 20 2017 - 09:54:03 EST


On Fri, Oct 20, 2017 at 10:47:16PM +0900, Masahiro Yamada wrote:
> Hi Greg,
>
> 2017-10-20 22:32 GMT+09:00 Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx>:
> > On Mon, Oct 09, 2017 at 03:26:30PM +0200, srinivas.kandagatla@xxxxxxxxxx wrote:
> >> From: Masahiro Yamada <yamada.masahiro@xxxxxxxxxxxxx>
> >>
> >> nvmem_register() copies all the members of nvmem_config to
> >> nvmem_device. So, nvmem_config is one-time use data during
> >> probing. There is no point to keep it until the driver detach.
> >> Using stack should be no problem because nvmem_config is pretty
> >> small.
> >>
> >> Signed-off-by: Masahiro Yamada <yamada.masahiro@xxxxxxxxxxxxx>
> >> Signed-off-by: Srinivas Kandagatla <srinivas.kandagatla@xxxxxxxxxx>
> >> ---
> >> drivers/nvmem/imx-iim.c | 27 ++++++++++++---------------
> >> 1 file changed, 12 insertions(+), 15 deletions(-)
> >>
> >> diff --git a/drivers/nvmem/imx-iim.c b/drivers/nvmem/imx-iim.c
> >> index 52ff65e0673f..a5992602709a 100644
> >> --- a/drivers/nvmem/imx-iim.c
> >> +++ b/drivers/nvmem/imx-iim.c
> >> @@ -34,7 +34,6 @@ struct imx_iim_drvdata {
> >> struct iim_priv {
> >> void __iomem *base;
> >> struct clk *clk;
> >> - struct nvmem_config nvmem;
> >> };
> >>
> >> static int imx_iim_read(void *context, unsigned int offset,
> >> @@ -108,7 +107,7 @@ static int imx_iim_probe(struct platform_device *pdev)
> >> struct resource *res;
> >> struct iim_priv *iim;
> >> struct nvmem_device *nvmem;
> >> - struct nvmem_config *cfg;
> >> + struct nvmem_config cfg = {};
> >
> > You do realize you are now not zeroing out this structure, and have to
> > explicitly initialize all of the fields, right?
>
> Why?
>
> I am surely zeroing out the structure.
>
> Did you miss "= {};" in my code?

Are you sure that does zero it out? I know we have had issues with this
in the past...

> > What is the real problem with doing a dynamic allocation for this?
> > Putting structures on the stack is a "bad idea" for all of the obvious
> > reasons (small stack in the kernel, initialized data, lower layers
> > expect it to be dma-able, etc.)
>
>
> Why is this a problem?
>
> Did you really understand this patch?
>
> - This structure is very small.
> struct uart_8250_port is five times bigger
> and it is allocated in the stack and it is fine.
>
> - All data are initialized.
>
> - Why DMA?
> Please do not exaggerate things by introducing unrelated topic.

I just want you to realize the change, the initialized is the big thing.

And keeping structures off of the stack is a good thing, if this is not
a performance issue, I suggest keeping it as-is, right?

thanks,

greg k-h