2023-01-30 13:08:56

by Arnd Bergmann

[permalink] [raw]
Subject: [PATCH] dmaengine: dw-edma: reduce stack usage after debugfs rework

From: Arnd Bergmann <[email protected]>

After the dw_edma_debugfs_entry arrays are no longer compile-time
constant, they take up space on the stack, which exceeds the
warning limit after inlining:

drivers/dma/dw-edma/dw-edma-v0-debugfs.c:280:6: error: stack frame size (1784) exceeds limit (1400) in 'dw_edma_v0_debugfs_on' [-Werror,-Wframe-larger-than]
void dw_edma_v0_debugfs_on(struct dw_edma *dw)

Work around this by marking the functions with the largest arrays
as 'noinline_for_stack' to make them not use up space on the same
stack together.

Fixes: 5c0373eafd83 ("dmaengine: dw-edma: Move eDMA data pointer to debugfs node descriptor")
Signed-off-by: Arnd Bergmann <[email protected]>
---
drivers/dma/dw-edma/dw-edma-v0-debugfs.c | 13 ++++++++-----
1 file changed, 8 insertions(+), 5 deletions(-)

diff --git a/drivers/dma/dw-edma/dw-edma-v0-debugfs.c b/drivers/dma/dw-edma/dw-edma-v0-debugfs.c
index 7be23c26ac88..9dfcbbdfb27b 100644
--- a/drivers/dma/dw-edma/dw-edma-v0-debugfs.c
+++ b/drivers/dma/dw-edma/dw-edma-v0-debugfs.c
@@ -116,7 +116,8 @@ static void dw_edma_debugfs_create_x32(struct dw_edma *dw,
}
}

-static void dw_edma_debugfs_regs_ch(struct dw_edma *dw, enum dw_edma_dir dir,
+static noinline_for_stack void
+dw_edma_debugfs_regs_ch(struct dw_edma *dw, enum dw_edma_dir dir,
u16 ch, struct dentry *dent)
{
struct dw_edma_debugfs_entry debugfs_regs[] = {
@@ -136,9 +137,10 @@ static void dw_edma_debugfs_regs_ch(struct dw_edma *dw, enum dw_edma_dir dir,
dw_edma_debugfs_create_x32(dw, debugfs_regs, nr_entries, dent);
}

-static void dw_edma_debugfs_regs_wr(struct dw_edma *dw, struct dentry *dent)
+static noinline_for_stack void
+dw_edma_debugfs_regs_wr(struct dw_edma *dw, struct dentry *dent)
{
- static const struct dw_edma_debugfs_entry debugfs_regs[] = {
+ const struct dw_edma_debugfs_entry debugfs_regs[] = {
/* eDMA global registers */
WR_REGISTER(dw, engine_en),
WR_REGISTER(dw, doorbell),
@@ -159,7 +161,7 @@ static void dw_edma_debugfs_regs_wr(struct dw_edma *dw, struct dentry *dent)
WR_REGISTER(dw, ch67_imwr_data),
WR_REGISTER(dw, linked_list_err_en),
};
- static const struct dw_edma_debugfs_entry debugfs_unroll_regs[] = {
+ const struct dw_edma_debugfs_entry debugfs_unroll_regs[] = {
/* eDMA channel context grouping */
WR_REGISTER_UNROLL(dw, engine_chgroup),
WR_REGISTER_UNROLL(dw, engine_hshake_cnt.lsb),
@@ -197,7 +199,8 @@ static void dw_edma_debugfs_regs_wr(struct dw_edma *dw, struct dentry *dent)
}
}

-static void dw_edma_debugfs_regs_rd(struct dw_edma *dw, struct dentry *dent)
+static noinline void
+dw_edma_debugfs_regs_rd(struct dw_edma *dw, struct dentry *dent)
{
const struct dw_edma_debugfs_entry debugfs_regs[] = {
/* eDMA global registers */
--
2.39.0



2023-01-30 17:13:00

by Serge Semin

[permalink] [raw]
Subject: Re: [PATCH] dmaengine: dw-edma: reduce stack usage after debugfs rework

Hi Arnd

On Mon, Jan 30, 2023 at 02:08:10PM +0100, Arnd Bergmann wrote:
> From: Arnd Bergmann <[email protected]>
>
> After the dw_edma_debugfs_entry arrays are no longer compile-time
> constant, they take up space on the stack, which exceeds the
> warning limit after inlining:
>
> drivers/dma/dw-edma/dw-edma-v0-debugfs.c:280:6: error: stack frame size (1784) exceeds limit (1400) in 'dw_edma_v0_debugfs_on' [-Werror,-Wframe-larger-than]
> void dw_edma_v0_debugfs_on(struct dw_edma *dw)
>

> Work around this by marking the functions with the largest arrays
> as 'noinline_for_stack' to make them not use up space on the same
> stack together.

You mean the same stack !frame!, right?

>
> Fixes: 5c0373eafd83 ("dmaengine: dw-edma: Move eDMA data pointer to debugfs node descriptor")
> Signed-off-by: Arnd Bergmann <[email protected]>
> ---
> drivers/dma/dw-edma/dw-edma-v0-debugfs.c | 13 ++++++++-----
> 1 file changed, 8 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/dma/dw-edma/dw-edma-v0-debugfs.c b/drivers/dma/dw-edma/dw-edma-v0-debugfs.c
> index 7be23c26ac88..9dfcbbdfb27b 100644
> --- a/drivers/dma/dw-edma/dw-edma-v0-debugfs.c
> +++ b/drivers/dma/dw-edma/dw-edma-v0-debugfs.c
> @@ -116,7 +116,8 @@ static void dw_edma_debugfs_create_x32(struct dw_edma *dw,
> }
> }
>

> -static void dw_edma_debugfs_regs_ch(struct dw_edma *dw, enum dw_edma_dir dir,
> +static noinline_for_stack void
> +dw_edma_debugfs_regs_ch(struct dw_edma *dw, enum dw_edma_dir dir,
> u16 ch, struct dentry *dent)

This doesn't seem like required. Does it? This method is called from
two functions: dw_edma_debugfs_regs_wr() and dw_edma_debugfs_regs_rd()
and if I am not mistaken will unlikely be inlined. Even if compiler will
inline it the stack memory consumption won't change much since the
functions aren't called from each other.

> {
> struct dw_edma_debugfs_entry debugfs_regs[] = {
> @@ -136,9 +137,10 @@ static void dw_edma_debugfs_regs_ch(struct dw_edma *dw, enum dw_edma_dir dir,
> dw_edma_debugfs_create_x32(dw, debugfs_regs, nr_entries, dent);
> }
>
> -static void dw_edma_debugfs_regs_wr(struct dw_edma *dw, struct dentry *dent)
> +static noinline_for_stack void
> +dw_edma_debugfs_regs_wr(struct dw_edma *dw, struct dentry *dent)
> {

> - static const struct dw_edma_debugfs_entry debugfs_regs[] = {
> + const struct dw_edma_debugfs_entry debugfs_regs[] = {
> /* eDMA global registers */
> WR_REGISTER(dw, engine_en),
> WR_REGISTER(dw, doorbell),
> @@ -159,7 +161,7 @@ static void dw_edma_debugfs_regs_wr(struct dw_edma *dw, struct dentry *dent)
> WR_REGISTER(dw, ch67_imwr_data),
> WR_REGISTER(dw, linked_list_err_en),
> };
> - static const struct dw_edma_debugfs_entry debugfs_unroll_regs[] = {
> + const struct dw_edma_debugfs_entry debugfs_unroll_regs[] = {

I am confused how come these locals've turned to be static const?
Neither my patches nor the Bjorn repo provide such changes:
https://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci.git/tree/drivers/dma/dw-edma/dw-edma-v0-debugfs.c?h=next#n139
they were supposed to be just const.

What repo is this patch based on?

> /* eDMA channel context grouping */
> WR_REGISTER_UNROLL(dw, engine_chgroup),
> WR_REGISTER_UNROLL(dw, engine_hshake_cnt.lsb),
> @@ -197,7 +199,8 @@ static void dw_edma_debugfs_regs_wr(struct dw_edma *dw, struct dentry *dent)
> }
> }
>

> -static void dw_edma_debugfs_regs_rd(struct dw_edma *dw, struct dentry *dent)
> +static noinline void

noinline_for_stack, right?

-Serge(y)

> +dw_edma_debugfs_regs_rd(struct dw_edma *dw, struct dentry *dent)
> {
> const struct dw_edma_debugfs_entry debugfs_regs[] = {
> /* eDMA global registers */
> --
> 2.39.0
>

2023-01-30 18:51:05

by Arnd Bergmann

[permalink] [raw]
Subject: Re: [PATCH] dmaengine: dw-edma: reduce stack usage after debugfs rework

On Mon, Jan 30, 2023, at 18:12, Serge Semin wrote:
> Hi Arnd
>
> On Mon, Jan 30, 2023 at 02:08:10PM +0100, Arnd Bergmann wrote:
>> From: Arnd Bergmann <[email protected]>
>>
>> After the dw_edma_debugfs_entry arrays are no longer compile-time
>> constant, they take up space on the stack, which exceeds the
>> warning limit after inlining:
>>
>> drivers/dma/dw-edma/dw-edma-v0-debugfs.c:280:6: error: stack frame size (1784) exceeds limit (1400) in 'dw_edma_v0_debugfs_on' [-Werror,-Wframe-larger-than]
>> void dw_edma_v0_debugfs_on(struct dw_edma *dw)
>>
>
>> Work around this by marking the functions with the largest arrays
>> as 'noinline_for_stack' to make them not use up space on the same
>> stack together.
>
> You mean the same stack !frame!, right?

Right.

>> -static void dw_edma_debugfs_regs_ch(struct dw_edma *dw, enum dw_edma_dir dir,
>> +static noinline_for_stack void
>> +dw_edma_debugfs_regs_ch(struct dw_edma *dw, enum dw_edma_dir dir,
>> u16 ch, struct dentry *dent)
>
> This doesn't seem like required. Does it? This method is called from
> two functions: dw_edma_debugfs_regs_wr() and dw_edma_debugfs_regs_rd()
> and if I am not mistaken will unlikely be inlined. Even if compiler will
> inline it the stack memory consumption won't change much since the
> functions aren't called from each other.

I did this mainly for consistency, marking all the functions that
have an array of dw_edma_debugfs_entry. I have retested without this
change now and adapted the changelog accordingly.

>> @@ -159,7 +161,7 @@ static void dw_edma_debugfs_regs_wr(struct dw_edma *dw, struct dentry *dent)
>> WR_REGISTER(dw, ch67_imwr_data),
>> WR_REGISTER(dw, linked_list_err_en),
>> };
>> - static const struct dw_edma_debugfs_entry debugfs_unroll_regs[] = {
>> + const struct dw_edma_debugfs_entry debugfs_unroll_regs[] = {
>
> I am confused how come these locals've turned to be static const?
> Neither my patches nor the Bjorn repo provide such changes:
> https://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci.git/tree/drivers/dma/dw-edma/dw-edma-v0-debugfs.c?h=next#n139
> they were supposed to be just const.
>
> What repo is this patch based on?

My mistake, that was a revert from an earlier broken attempt to
fix the issue. It's fixed now after reverting the other patch
and rebasing.

>> /* eDMA channel context grouping */
>> WR_REGISTER_UNROLL(dw, engine_chgroup),
>> WR_REGISTER_UNROLL(dw, engine_hshake_cnt.lsb),
>> @@ -197,7 +199,8 @@ static void dw_edma_debugfs_regs_wr(struct dw_edma *dw, struct dentry *dent)
>> }
>> }
>>
>
>> -static void dw_edma_debugfs_regs_rd(struct dw_edma *dw, struct dentry *dent)
>> +static noinline void
>
> noinline_for_stack, right?

Fixed.

Arnd