Re: [Nouveau] Fix unaligned accesses for SPARC
On Tue, Mar 12, 2013 at 10:18:41PM -0500, Patrick Baggett wrote:
The nouveau driver makes a number of unaligned accesses via the
ROM16(), ROM32() and ROM64() macros which fault on SPARC (but would be
transparently handled by x86 hardware). Making use of get_unaligned()
macro fixes the problem for me.
diff --git a/drivers/gpu/drm/nouveau/nouveau_bios.h
b/drivers/gpu/drm/nouveau/nouveau_bios.h
index 7ccd28f..92031f6 100644
--- a/drivers/gpu/drm/nouveau/nouveau_bios.h
+++ b/drivers/gpu/drm/nouveau/nouveau_bios.h
@@ -26,6 +26,8 @@
#include nvreg.h
+#include asm/unaligned.h
+
#define DCB_MAX_NUM_ENTRIES 16
#define DCB_MAX_NUM_I2C_ENTRIES 16
#define DCB_MAX_NUM_GPIO_ENTRIES 32
@@ -33,10 +35,10 @@
#define DCB_LOC_ON_CHIP 0
-#define ROM16(x) le16_to_cpu(*(u16 *)(x))
-#define ROM32(x) le32_to_cpu(*(u32 *)(x))
+#define ROM16(x) le16_to_cpu(get_unaligned((u16 *)(x)))
+#define ROM32(x) le32_to_cpu(get_unaligned((u32 *)(x)))
#define ROM48(x) ({ u8 *p = (x); (u64)ROM16(p[4]) 32 |
ROM32(p[0]); })
-#define ROM64(x) le64_to_cpu(*(u64 *)(x))
+#define ROM64(x) le64_to_cpu(get_unaligned((u64 *)(x)))
#define ROMPTR(d,x) ({\
struct nouveau_drm *drm = nouveau_drm((d)); \
ROM16(x) ? drm-vbios.data[ROM16(x)] : NULL; \
Please take a look at Documentation/SubmittingPatches and resubmit after
adapting.
IMHO, I think you should submit this patch only once you figure out how to
correctly boot and use at least basic accelerated operations, because I'm not
100% sure this is correct - 1. you have other severe issues with nouveau which
may or may not be connected to this code, 2. sparc defines get_unaligned to
__get_unaligned_be, so maybe you need to use _le variant? or maybe use
get_unaligned_le* without le*_to_cpu?.
Marcin
___
Nouveau mailing list
Nouveau@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/nouveau
Re: [Nouveau] Fix unaligned accesses for SPARC
Marcin,
I had the same messages on NV17, which worked extremely well (OpenGL games
worked fine). The kernel will properly emulate the unaligned access, it
just causes a trap which is slow. If the unaligned access caused an error,
the NV17 card would have been unable to function given the number of
erroneous accesses.
Using __get_unaligned_le would probably remove one byteswap though.
Patrick
On Wed, Mar 13, 2013 at 4:29 PM, Marcin Slusarz marcin.slus...@gmail.comwrote:
On Tue, Mar 12, 2013 at 10:18:41PM -0500, Patrick Baggett wrote:
The nouveau driver makes a number of unaligned accesses via the
ROM16(), ROM32() and ROM64() macros which fault on SPARC (but would be
transparently handled by x86 hardware). Making use of get_unaligned()
macro fixes the problem for me.
diff --git a/drivers/gpu/drm/nouveau/nouveau_bios.h
b/drivers/gpu/drm/nouveau/nouveau_bios.h
index 7ccd28f..92031f6 100644
--- a/drivers/gpu/drm/nouveau/nouveau_bios.h
+++ b/drivers/gpu/drm/nouveau/nouveau_bios.h
@@ -26,6 +26,8 @@
#include nvreg.h
+#include asm/unaligned.h
+
#define DCB_MAX_NUM_ENTRIES 16
#define DCB_MAX_NUM_I2C_ENTRIES 16
#define DCB_MAX_NUM_GPIO_ENTRIES 32
@@ -33,10 +35,10 @@
#define DCB_LOC_ON_CHIP 0
-#define ROM16(x) le16_to_cpu(*(u16 *)(x))
-#define ROM32(x) le32_to_cpu(*(u32 *)(x))
+#define ROM16(x) le16_to_cpu(get_unaligned((u16 *)(x)))
+#define ROM32(x) le32_to_cpu(get_unaligned((u32 *)(x)))
#define ROM48(x) ({ u8 *p = (x); (u64)ROM16(p[4]) 32 |
ROM32(p[0]); })
-#define ROM64(x) le64_to_cpu(*(u64 *)(x))
+#define ROM64(x) le64_to_cpu(get_unaligned((u64 *)(x)))
#define ROMPTR(d,x) ({\
struct nouveau_drm *drm = nouveau_drm((d)); \
ROM16(x) ? drm-vbios.data[ROM16(x)] : NULL; \
Please take a look at Documentation/SubmittingPatches and resubmit after
adapting.
IMHO, I think you should submit this patch only once you figure out how to
correctly boot and use at least basic accelerated operations, because I'm
not
100% sure this is correct - 1. you have other severe issues with nouveau
which
may or may not be connected to this code, 2. sparc defines get_unaligned to
__get_unaligned_be, so maybe you need to use _le variant? or maybe use
get_unaligned_le* without le*_to_cpu?.
Marcin
___
Nouveau mailing list
Nouveau@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/nouveau
[Nouveau] Fix unaligned accesses for SPARC
The nouveau driver makes a number of unaligned accesses via the ROM16(),
ROM32() and ROM64() macros which fault on SPARC (but would be transparently
handled by x86 hardware). Making use of get_unaligned() macro fixes the
problem for me.
diff --git a/drivers/gpu/drm/nouveau/nouveau_bios.h
b/drivers/gpu/drm/nouveau/nouveau_bios.h
index 7ccd28f..92031f6 100644
--- a/drivers/gpu/drm/nouveau/nouveau_bios.h
+++ b/drivers/gpu/drm/nouveau/nouveau_bios.h
@@ -26,6 +26,8 @@
#include nvreg.h
+#include asm/unaligned.h
+
#define DCB_MAX_NUM_ENTRIES 16
#define DCB_MAX_NUM_I2C_ENTRIES 16
#define DCB_MAX_NUM_GPIO_ENTRIES 32
@@ -33,10 +35,10 @@
#define DCB_LOC_ON_CHIP 0
-#define ROM16(x) le16_to_cpu(*(u16 *)(x))
-#define ROM32(x) le32_to_cpu(*(u32 *)(x))
+#define ROM16(x) le16_to_cpu(get_unaligned((u16 *)(x)))
+#define ROM32(x) le32_to_cpu(get_unaligned((u32 *)(x)))
#define ROM48(x) ({ u8 *p = (x); (u64)ROM16(p[4]) 32 | ROM32(p[0]); })
-#define ROM64(x) le64_to_cpu(*(u64 *)(x))
+#define ROM64(x) le64_to_cpu(get_unaligned((u64 *)(x)))
#define ROMPTR(d,x) ({\
struct nouveau_drm *drm = nouveau_drm((d)); \
ROM16(x) ? drm-vbios.data[ROM16(x)] : NULL; \
___
Nouveau mailing list
Nouveau@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/nouveau
