| Message ID | 20200108135116.3687988-1-arnd@arndb.de |
|---|---|
| State | Accepted |
| Commit | dba9bf0a98c5fcfb3f616bf2446d5aeef67b632c |
| Headers | show |
| Series | [v2] drm: panel: fix excessive stack usage in td028ttec1_prepare | expand |
Hi Arnd. On Wed, Jan 08, 2020 at 02:51:05PM +0100, Arnd Bergmann wrote: > With gcc -O3 in combination with the structleak plug, the compiler can > inline very aggressively, leading to rather large stack usage: > > drivers/gpu/drm/panel/panel-tpo-td028ttec1.c: In function 'td028ttec1_prepare': > drivers/gpu/drm/panel/panel-tpo-td028ttec1.c:233:1: error: the frame size of 2768 bytes is larger than 2048 bytes [-Werror=frame-larger-than=] > } > > Marking jbt_reg_write_*() as noinline avoids the case where > multiple instances of this function get inlined into the same > stack frame and each one adds a copy of 'tx_buf'. > > The compiler is clearly making some bad decisions here, but I > did not open a new bug report as this only happens in combination > with the structleak plugin. > > Link: https://lore.kernel.org/lkml/CAK8P3a3jAnFZA3GFRtdYdg1-i-oih3pOQzkkrK-X3BGsFrMiZQ@mail.gmail.com/ > Fixes: mmtom ("init/Kconfig: enable -O3 for all arches") > Signed-off-by: Arnd Bergmann <arnd@arndb.de> > --- > v2: > - mark all three functions as noinlien > - add code comment > - add link to more detailed analysis Thanks for the updated patch. Applied to drm-misc-next. Sam > --- > drivers/gpu/drm/panel/panel-tpo-td028ttec1.c | 13 ++++++++++--- > 1 file changed, 10 insertions(+), 3 deletions(-) > > diff --git a/drivers/gpu/drm/panel/panel-tpo-td028ttec1.c b/drivers/gpu/drm/panel/panel-tpo-td028ttec1.c > index cf29405a2dbe..5034db8b55de 100644 > --- a/drivers/gpu/drm/panel/panel-tpo-td028ttec1.c > +++ b/drivers/gpu/drm/panel/panel-tpo-td028ttec1.c > @@ -86,7 +86,12 @@ struct td028ttec1_panel { > > #define to_td028ttec1_device(p) container_of(p, struct td028ttec1_panel, panel) > > -static int jbt_ret_write_0(struct td028ttec1_panel *lcd, u8 reg, int *err) > +/* > + * noinline_for_stack so we don't get multiple copies of tx_buf > + * on the stack in case of gcc-plugin-structleak > + */ > +static int noinline_for_stack > +jbt_ret_write_0(struct td028ttec1_panel *lcd, u8 reg, int *err) > { > struct spi_device *spi = lcd->spi; > u16 tx_buf = JBT_COMMAND | reg; > @@ -105,7 +110,8 @@ static int jbt_ret_write_0(struct td028ttec1_panel *lcd, u8 reg, int *err) > return ret; > } > > -static int jbt_reg_write_1(struct td028ttec1_panel *lcd, > +static int noinline_for_stack > +jbt_reg_write_1(struct td028ttec1_panel *lcd, > u8 reg, u8 data, int *err) > { > struct spi_device *spi = lcd->spi; > @@ -128,7 +134,8 @@ static int jbt_reg_write_1(struct td028ttec1_panel *lcd, > return ret; > } > > -static int jbt_reg_write_2(struct td028ttec1_panel *lcd, > +static int noinline_for_stack > +jbt_reg_write_2(struct td028ttec1_panel *lcd, > u8 reg, u16 data, int *err) > { > struct spi_device *spi = lcd->spi; > -- > 2.20.0 _______________________________________________ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
diff --git a/drivers/gpu/drm/panel/panel-tpo-td028ttec1.c b/drivers/gpu/drm/panel/panel-tpo-td028ttec1.c index cf29405a2dbe..5034db8b55de 100644 --- a/drivers/gpu/drm/panel/panel-tpo-td028ttec1.c +++ b/drivers/gpu/drm/panel/panel-tpo-td028ttec1.c @@ -86,7 +86,12 @@ struct td028ttec1_panel { #define to_td028ttec1_device(p) container_of(p, struct td028ttec1_panel, panel) -static int jbt_ret_write_0(struct td028ttec1_panel *lcd, u8 reg, int *err) +/* + * noinline_for_stack so we don't get multiple copies of tx_buf + * on the stack in case of gcc-plugin-structleak + */ +static int noinline_for_stack +jbt_ret_write_0(struct td028ttec1_panel *lcd, u8 reg, int *err) { struct spi_device *spi = lcd->spi; u16 tx_buf = JBT_COMMAND | reg; @@ -105,7 +110,8 @@ static int jbt_ret_write_0(struct td028ttec1_panel *lcd, u8 reg, int *err) return ret; } -static int jbt_reg_write_1(struct td028ttec1_panel *lcd, +static int noinline_for_stack +jbt_reg_write_1(struct td028ttec1_panel *lcd, u8 reg, u8 data, int *err) { struct spi_device *spi = lcd->spi; @@ -128,7 +134,8 @@ static int jbt_reg_write_1(struct td028ttec1_panel *lcd, return ret; } -static int jbt_reg_write_2(struct td028ttec1_panel *lcd, +static int noinline_for_stack +jbt_reg_write_2(struct td028ttec1_panel *lcd, u8 reg, u16 data, int *err) { struct spi_device *spi = lcd->spi;
With gcc -O3 in combination with the structleak plug, the compiler can inline very aggressively, leading to rather large stack usage: drivers/gpu/drm/panel/panel-tpo-td028ttec1.c: In function 'td028ttec1_prepare': drivers/gpu/drm/panel/panel-tpo-td028ttec1.c:233:1: error: the frame size of 2768 bytes is larger than 2048 bytes [-Werror=frame-larger-than=] } Marking jbt_reg_write_*() as noinline avoids the case where multiple instances of this function get inlined into the same stack frame and each one adds a copy of 'tx_buf'. The compiler is clearly making some bad decisions here, but I did not open a new bug report as this only happens in combination with the structleak plugin. Link: https://lore.kernel.org/lkml/CAK8P3a3jAnFZA3GFRtdYdg1-i-oih3pOQzkkrK-X3BGsFrMiZQ@mail.gmail.com/ Fixes: mmtom ("init/Kconfig: enable -O3 for all arches") Signed-off-by: Arnd Bergmann <arnd@arndb.de> --- v2: - mark all three functions as noinlien - add code comment - add link to more detailed analysis --- drivers/gpu/drm/panel/panel-tpo-td028ttec1.c | 13 ++++++++++--- 1 file changed, 10 insertions(+), 3 deletions(-) -- 2.20.0 _______________________________________________ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel