linux/drivers/gpu/drm/radeon/rs600.c

694 lines
19 KiB
C
Raw Normal View History

drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
/*
* Copyright 2008 Advanced Micro Devices, Inc.
* Copyright 2008 Red Hat Inc.
* Copyright 2009 Jerome Glisse.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
* OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
* ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
* OTHER DEALINGS IN THE SOFTWARE.
*
* Authors: Dave Airlie
* Alex Deucher
* Jerome Glisse
*/
/* RS600 / Radeon X1250/X1270 integrated GPU
*
* This file gather function specific to RS600 which is the IGP of
* the X1250/X1270 family supporting intel CPU (while RS690/RS740
* is the X1250/X1270 supporting AMD CPU). The display engine are
* the avivo one, bios is an atombios, 3D block are the one of the
* R4XX family. The GART is different from the RS400 one and is very
* close to the one of the R600 family (R600 likely being an evolution
* of the RS600 GART block).
*/
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
#include "drmP.h"
#include "radeon.h"
#include "atom.h"
#include "rs600d.h"
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
#include "rs600_reg_safe.h"
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
void rs600_gpu_init(struct radeon_device *rdev);
int rs600_mc_wait_for_idle(struct radeon_device *rdev);
int rs600_mc_init(struct radeon_device *rdev)
{
/* read back the MC value from the hw */
int r;
u32 tmp;
/* Setup GPU memory space */
tmp = RREG32_MC(R_000004_MC_FB_LOCATION);
rdev->mc.vram_location = G_000004_MC_FB_START(tmp) << 16;
rdev->mc.gtt_location = 0xffffffffUL;
r = radeon_mc_setup(rdev);
if (r)
return r;
return 0;
}
/* hpd for digital panel detect/disconnect */
bool rs600_hpd_sense(struct radeon_device *rdev, enum radeon_hpd_id hpd)
{
u32 tmp;
bool connected = false;
switch (hpd) {
case RADEON_HPD_1:
tmp = RREG32(R_007D04_DC_HOT_PLUG_DETECT1_INT_STATUS);
if (G_007D04_DC_HOT_PLUG_DETECT1_SENSE(tmp))
connected = true;
break;
case RADEON_HPD_2:
tmp = RREG32(R_007D14_DC_HOT_PLUG_DETECT2_INT_STATUS);
if (G_007D14_DC_HOT_PLUG_DETECT2_SENSE(tmp))
connected = true;
break;
default:
break;
}
return connected;
}
void rs600_hpd_set_polarity(struct radeon_device *rdev,
enum radeon_hpd_id hpd)
{
u32 tmp;
bool connected = rs600_hpd_sense(rdev, hpd);
switch (hpd) {
case RADEON_HPD_1:
tmp = RREG32(R_007D08_DC_HOT_PLUG_DETECT1_INT_CONTROL);
if (connected)
tmp &= ~S_007D08_DC_HOT_PLUG_DETECT1_INT_POLARITY(1);
else
tmp |= S_007D08_DC_HOT_PLUG_DETECT1_INT_POLARITY(1);
WREG32(R_007D08_DC_HOT_PLUG_DETECT1_INT_CONTROL, tmp);
break;
case RADEON_HPD_2:
tmp = RREG32(R_007D18_DC_HOT_PLUG_DETECT2_INT_CONTROL);
if (connected)
tmp &= ~S_007D18_DC_HOT_PLUG_DETECT2_INT_POLARITY(1);
else
tmp |= S_007D18_DC_HOT_PLUG_DETECT2_INT_POLARITY(1);
WREG32(R_007D18_DC_HOT_PLUG_DETECT2_INT_CONTROL, tmp);
break;
default:
break;
}
}
void rs600_hpd_init(struct radeon_device *rdev)
{
struct drm_device *dev = rdev->ddev;
struct drm_connector *connector;
list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
struct radeon_connector *radeon_connector = to_radeon_connector(connector);
switch (radeon_connector->hpd.hpd) {
case RADEON_HPD_1:
WREG32(R_007D00_DC_HOT_PLUG_DETECT1_CONTROL,
S_007D00_DC_HOT_PLUG_DETECT1_EN(1));
rdev->irq.hpd[0] = true;
break;
case RADEON_HPD_2:
WREG32(R_007D10_DC_HOT_PLUG_DETECT2_CONTROL,
S_007D10_DC_HOT_PLUG_DETECT2_EN(1));
rdev->irq.hpd[1] = true;
break;
default:
break;
}
}
rs600_irq_set(rdev);
}
void rs600_hpd_fini(struct radeon_device *rdev)
{
struct drm_device *dev = rdev->ddev;
struct drm_connector *connector;
list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
struct radeon_connector *radeon_connector = to_radeon_connector(connector);
switch (radeon_connector->hpd.hpd) {
case RADEON_HPD_1:
WREG32(R_007D00_DC_HOT_PLUG_DETECT1_CONTROL,
S_007D00_DC_HOT_PLUG_DETECT1_EN(0));
rdev->irq.hpd[0] = false;
break;
case RADEON_HPD_2:
WREG32(R_007D10_DC_HOT_PLUG_DETECT2_CONTROL,
S_007D10_DC_HOT_PLUG_DETECT2_EN(0));
rdev->irq.hpd[1] = false;
break;
default:
break;
}
}
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
/*
* GART.
*/
void rs600_gart_tlb_flush(struct radeon_device *rdev)
{
uint32_t tmp;
tmp = RREG32_MC(R_000100_MC_PT0_CNTL);
tmp &= C_000100_INVALIDATE_ALL_L1_TLBS & C_000100_INVALIDATE_L2_CACHE;
WREG32_MC(R_000100_MC_PT0_CNTL, tmp);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
tmp = RREG32_MC(R_000100_MC_PT0_CNTL);
tmp |= S_000100_INVALIDATE_ALL_L1_TLBS(1) & S_000100_INVALIDATE_L2_CACHE(1);
WREG32_MC(R_000100_MC_PT0_CNTL, tmp);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
tmp = RREG32_MC(R_000100_MC_PT0_CNTL);
tmp &= C_000100_INVALIDATE_ALL_L1_TLBS & C_000100_INVALIDATE_L2_CACHE;
WREG32_MC(R_000100_MC_PT0_CNTL, tmp);
tmp = RREG32_MC(R_000100_MC_PT0_CNTL);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
}
int rs600_gart_init(struct radeon_device *rdev)
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
{
int r;
if (rdev->gart.table.vram.robj) {
WARN(1, "RS600 GART already initialized.\n");
return 0;
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
/* Initialize common gart structure */
r = radeon_gart_init(rdev);
if (r) {
return r;
}
rdev->gart.table_size = rdev->gart.num_gpu_pages * 8;
return radeon_gart_table_vram_alloc(rdev);
}
int rs600_gart_enable(struct radeon_device *rdev)
{
u32 tmp;
int r, i;
if (rdev->gart.table.vram.robj == NULL) {
dev_err(rdev->dev, "No VRAM object for PCIE GART.\n");
return -EINVAL;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
}
r = radeon_gart_table_vram_pin(rdev);
if (r)
return r;
/* Enable bus master */
tmp = RREG32(R_00004C_BUS_CNTL) & C_00004C_BUS_MASTER_DIS;
WREG32(R_00004C_BUS_CNTL, tmp);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
/* FIXME: setup default page */
WREG32_MC(R_000100_MC_PT0_CNTL,
(S_000100_EFFECTIVE_L2_CACHE_SIZE(6) |
S_000100_EFFECTIVE_L2_QUEUE_SIZE(6)));
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
for (i = 0; i < 19; i++) {
WREG32_MC(R_00016C_MC_PT0_CLIENT0_CNTL + i,
S_00016C_ENABLE_TRANSLATION_MODE_OVERRIDE(1) |
S_00016C_SYSTEM_ACCESS_MODE_MASK(
V_00016C_SYSTEM_ACCESS_MODE_NOT_IN_SYS) |
S_00016C_SYSTEM_APERTURE_UNMAPPED_ACCESS(
V_00016C_SYSTEM_APERTURE_UNMAPPED_PASSTHROUGH) |
S_00016C_EFFECTIVE_L1_CACHE_SIZE(3) |
S_00016C_ENABLE_FRAGMENT_PROCESSING(1) |
S_00016C_EFFECTIVE_L1_QUEUE_SIZE(3));
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
}
/* enable first context */
WREG32_MC(R_000102_MC_PT0_CONTEXT0_CNTL,
S_000102_ENABLE_PAGE_TABLE(1) |
S_000102_PAGE_TABLE_DEPTH(V_000102_PAGE_TABLE_FLAT));
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
/* disable all other contexts */
for (i = 1; i < 8; i++)
WREG32_MC(R_000102_MC_PT0_CONTEXT0_CNTL + i, 0);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
/* setup the page table */
WREG32_MC(R_00012C_MC_PT0_CONTEXT0_FLAT_BASE_ADDR,
rdev->gart.table_addr);
WREG32_MC(R_00013C_MC_PT0_CONTEXT0_FLAT_START_ADDR, rdev->mc.gtt_start);
WREG32_MC(R_00014C_MC_PT0_CONTEXT0_FLAT_END_ADDR, rdev->mc.gtt_end);
WREG32_MC(R_00011C_MC_PT0_CONTEXT0_DEFAULT_READ_ADDR, 0);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
/* System context maps to VRAM space */
WREG32_MC(R_000112_MC_PT0_SYSTEM_APERTURE_LOW_ADDR, rdev->mc.vram_start);
WREG32_MC(R_000114_MC_PT0_SYSTEM_APERTURE_HIGH_ADDR, rdev->mc.vram_end);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
/* enable page tables */
tmp = RREG32_MC(R_000100_MC_PT0_CNTL);
WREG32_MC(R_000100_MC_PT0_CNTL, (tmp | S_000100_ENABLE_PT(1)));
tmp = RREG32_MC(R_000009_MC_CNTL1);
WREG32_MC(R_000009_MC_CNTL1, (tmp | S_000009_ENABLE_PAGE_TABLES(1)));
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
rs600_gart_tlb_flush(rdev);
rdev->gart.ready = true;
return 0;
}
void rs600_gart_disable(struct radeon_device *rdev)
{
u32 tmp;
int r;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
/* FIXME: disable out of gart access */
WREG32_MC(R_000100_MC_PT0_CNTL, 0);
tmp = RREG32_MC(R_000009_MC_CNTL1);
WREG32_MC(R_000009_MC_CNTL1, tmp & C_000009_ENABLE_PAGE_TABLES);
if (rdev->gart.table.vram.robj) {
r = radeon_bo_reserve(rdev->gart.table.vram.robj, false);
if (r == 0) {
radeon_bo_kunmap(rdev->gart.table.vram.robj);
radeon_bo_unpin(rdev->gart.table.vram.robj);
radeon_bo_unreserve(rdev->gart.table.vram.robj);
}
}
}
void rs600_gart_fini(struct radeon_device *rdev)
{
rs600_gart_disable(rdev);
radeon_gart_table_vram_free(rdev);
radeon_gart_fini(rdev);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
}
#define R600_PTE_VALID (1 << 0)
#define R600_PTE_SYSTEM (1 << 1)
#define R600_PTE_SNOOPED (1 << 2)
#define R600_PTE_READABLE (1 << 5)
#define R600_PTE_WRITEABLE (1 << 6)
int rs600_gart_set_page(struct radeon_device *rdev, int i, uint64_t addr)
{
void __iomem *ptr = (void *)rdev->gart.table.vram.ptr;
if (i < 0 || i > rdev->gart.num_gpu_pages) {
return -EINVAL;
}
addr = addr & 0xFFFFFFFFFFFFF000ULL;
addr |= R600_PTE_VALID | R600_PTE_SYSTEM | R600_PTE_SNOOPED;
addr |= R600_PTE_READABLE | R600_PTE_WRITEABLE;
writeq(addr, ((void __iomem *)ptr) + (i * 8));
return 0;
}
int rs600_irq_set(struct radeon_device *rdev)
{
uint32_t tmp = 0;
uint32_t mode_int = 0;
u32 hpd1 = RREG32(R_007D08_DC_HOT_PLUG_DETECT1_INT_CONTROL) &
~S_007D08_DC_HOT_PLUG_DETECT1_INT_EN(1);
u32 hpd2 = RREG32(R_007D18_DC_HOT_PLUG_DETECT2_INT_CONTROL) &
~S_007D18_DC_HOT_PLUG_DETECT2_INT_EN(1);
if (rdev->irq.sw_int) {
tmp |= S_000040_SW_INT_EN(1);
}
if (rdev->irq.crtc_vblank_int[0]) {
mode_int |= S_006540_D1MODE_VBLANK_INT_MASK(1);
}
if (rdev->irq.crtc_vblank_int[1]) {
mode_int |= S_006540_D2MODE_VBLANK_INT_MASK(1);
}
if (rdev->irq.hpd[0]) {
hpd1 |= S_007D08_DC_HOT_PLUG_DETECT1_INT_EN(1);
}
if (rdev->irq.hpd[1]) {
hpd2 |= S_007D18_DC_HOT_PLUG_DETECT2_INT_EN(1);
}
WREG32(R_000040_GEN_INT_CNTL, tmp);
WREG32(R_006540_DxMODE_INT_MASK, mode_int);
WREG32(R_007D08_DC_HOT_PLUG_DETECT1_INT_CONTROL, hpd1);
WREG32(R_007D18_DC_HOT_PLUG_DETECT2_INT_CONTROL, hpd2);
return 0;
}
static inline uint32_t rs600_irq_ack(struct radeon_device *rdev, u32 *r500_disp_int)
{
uint32_t irqs = RREG32(R_000044_GEN_INT_STATUS);
uint32_t irq_mask = ~C_000044_SW_INT;
u32 tmp;
if (G_000044_DISPLAY_INT_STAT(irqs)) {
*r500_disp_int = RREG32(R_007EDC_DISP_INTERRUPT_STATUS);
if (G_007EDC_LB_D1_VBLANK_INTERRUPT(*r500_disp_int)) {
WREG32(R_006534_D1MODE_VBLANK_STATUS,
S_006534_D1MODE_VBLANK_ACK(1));
}
if (G_007EDC_LB_D2_VBLANK_INTERRUPT(*r500_disp_int)) {
WREG32(R_006D34_D2MODE_VBLANK_STATUS,
S_006D34_D2MODE_VBLANK_ACK(1));
}
if (G_007EDC_DC_HOT_PLUG_DETECT1_INTERRUPT(*r500_disp_int)) {
tmp = RREG32(R_007D08_DC_HOT_PLUG_DETECT1_INT_CONTROL);
tmp |= S_007D08_DC_HOT_PLUG_DETECT1_INT_ACK(1);
WREG32(R_007D08_DC_HOT_PLUG_DETECT1_INT_CONTROL, tmp);
}
if (G_007EDC_DC_HOT_PLUG_DETECT2_INTERRUPT(*r500_disp_int)) {
tmp = RREG32(R_007D18_DC_HOT_PLUG_DETECT2_INT_CONTROL);
tmp |= S_007D18_DC_HOT_PLUG_DETECT2_INT_ACK(1);
WREG32(R_007D18_DC_HOT_PLUG_DETECT2_INT_CONTROL, tmp);
}
} else {
*r500_disp_int = 0;
}
if (irqs) {
WREG32(R_000044_GEN_INT_STATUS, irqs);
}
return irqs & irq_mask;
}
void rs600_irq_disable(struct radeon_device *rdev)
{
u32 tmp;
WREG32(R_000040_GEN_INT_CNTL, 0);
WREG32(R_006540_DxMODE_INT_MASK, 0);
/* Wait and acknowledge irq */
mdelay(1);
rs600_irq_ack(rdev, &tmp);
}
int rs600_irq_process(struct radeon_device *rdev)
{
uint32_t status, msi_rearm;
uint32_t r500_disp_int;
bool queue_hotplug = false;
status = rs600_irq_ack(rdev, &r500_disp_int);
if (!status && !r500_disp_int) {
return IRQ_NONE;
}
while (status || r500_disp_int) {
/* SW interrupt */
if (G_000044_SW_INT(status))
radeon_fence_process(rdev);
/* Vertical blank interrupts */
if (G_007EDC_LB_D1_VBLANK_INTERRUPT(r500_disp_int))
drm_handle_vblank(rdev->ddev, 0);
if (G_007EDC_LB_D2_VBLANK_INTERRUPT(r500_disp_int))
drm_handle_vblank(rdev->ddev, 1);
if (G_007EDC_DC_HOT_PLUG_DETECT1_INTERRUPT(r500_disp_int)) {
queue_hotplug = true;
DRM_DEBUG("HPD1\n");
}
if (G_007EDC_DC_HOT_PLUG_DETECT2_INTERRUPT(r500_disp_int)) {
queue_hotplug = true;
DRM_DEBUG("HPD2\n");
}
status = rs600_irq_ack(rdev, &r500_disp_int);
}
if (queue_hotplug)
queue_work(rdev->wq, &rdev->hotplug_work);
if (rdev->msi_enabled) {
switch (rdev->family) {
case CHIP_RS600:
case CHIP_RS690:
case CHIP_RS740:
msi_rearm = RREG32(RADEON_BUS_CNTL) & ~RS600_MSI_REARM;
WREG32(RADEON_BUS_CNTL, msi_rearm);
WREG32(RADEON_BUS_CNTL, msi_rearm | RS600_MSI_REARM);
break;
default:
msi_rearm = RREG32(RADEON_MSI_REARM_EN) & ~RV370_MSI_REARM_EN;
WREG32(RADEON_MSI_REARM_EN, msi_rearm);
WREG32(RADEON_MSI_REARM_EN, msi_rearm | RV370_MSI_REARM_EN);
break;
}
}
return IRQ_HANDLED;
}
u32 rs600_get_vblank_counter(struct radeon_device *rdev, int crtc)
{
if (crtc == 0)
return RREG32(R_0060A4_D1CRTC_STATUS_FRAME_COUNT);
else
return RREG32(R_0068A4_D2CRTC_STATUS_FRAME_COUNT);
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
int rs600_mc_wait_for_idle(struct radeon_device *rdev)
{
unsigned i;
for (i = 0; i < rdev->usec_timeout; i++) {
if (G_000000_MC_IDLE(RREG32_MC(R_000000_MC_STATUS)))
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
return 0;
udelay(1);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
}
return -1;
}
void rs600_gpu_init(struct radeon_device *rdev)
{
r100_hdp_reset(rdev);
r420_pipes_init(rdev);
/* Wait for mc idle */
if (rs600_mc_wait_for_idle(rdev))
dev_warn(rdev->dev, "Wait MC idle timeout before updating MC.\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
}
void rs600_vram_info(struct radeon_device *rdev)
{
rdev->mc.vram_is_ddr = true;
rdev->mc.vram_width = 128;
rdev->mc.real_vram_size = RREG32(RADEON_CONFIG_MEMSIZE);
rdev->mc.mc_vram_size = rdev->mc.real_vram_size;
rdev->mc.aper_base = drm_get_resource_start(rdev->ddev, 0);
rdev->mc.aper_size = drm_get_resource_len(rdev->ddev, 0);
if (rdev->mc.mc_vram_size > rdev->mc.aper_size)
rdev->mc.mc_vram_size = rdev->mc.aper_size;
if (rdev->mc.real_vram_size > rdev->mc.aper_size)
rdev->mc.real_vram_size = rdev->mc.aper_size;
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
}
void rs600_bandwidth_update(struct radeon_device *rdev)
{
/* FIXME: implement, should this be like rs690 ? */
}
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
uint32_t rs600_mc_rreg(struct radeon_device *rdev, uint32_t reg)
{
WREG32(R_000070_MC_IND_INDEX, S_000070_MC_IND_ADDR(reg) |
S_000070_MC_IND_CITF_ARB0(1));
return RREG32(R_000074_MC_IND_DATA);
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
}
void rs600_mc_wreg(struct radeon_device *rdev, uint32_t reg, uint32_t v)
{
WREG32(R_000070_MC_IND_INDEX, S_000070_MC_IND_ADDR(reg) |
S_000070_MC_IND_CITF_ARB0(1) | S_000070_MC_IND_WR_EN(1));
WREG32(R_000074_MC_IND_DATA, v);
}
void rs600_debugfs(struct radeon_device *rdev)
{
if (r100_debugfs_rbbm_init(rdev))
DRM_ERROR("Failed to register debugfs file for RBBM !\n");
drm/radeon: introduce kernel modesetting for radeon hardware Add kernel modesetting support to radeon driver, use the ttm memory manager to manage memory and DRM/GEM to provide userspace API. In order to avoid backward compatibility issue and to allow clean design and code the radeon kernel modesetting use different code path than old radeon/drm driver. When kernel modesetting is enabled the IOCTL of radeon/drm driver are considered as invalid and an error message is printed in the log and they return failure. KMS enabled userspace will use new API to talk with the radeon/drm driver. The new API provide functions to create/destroy/share/mmap buffer object which are then managed by the kernel memory manager (here TTM). In order to submit command to the GPU the userspace provide a buffer holding the command stream, along this buffer userspace have to provide a list of buffer object used by the command stream. The kernel radeon driver will then place buffer in GPU accessible memory and will update command stream to reflect the position of the different buffers. The kernel will also perform security check on command stream provided by the user, we want to catch and forbid any illegal use of the GPU such as DMA into random system memory or into memory not owned by the process supplying the command stream. This part of the code is still incomplete and this why we propose that patch as a staging driver addition, future security might forbid current experimental userspace to run. This code support the following hardware : R1XX,R2XX,R3XX,R4XX,R5XX (radeon up to X1950). Works is underway to provide support for R6XX, R7XX and newer hardware (radeon from HD2XXX to HD4XXX). Authors: Jerome Glisse <jglisse@redhat.com> Dave Airlie <airlied@redhat.com> Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Jerome Glisse <jglisse@redhat.com> Signed-off-by: Dave Airlie <airlied@redhat.com> Signed-off-by: Alex Deucher <alexdeucher@gmail.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2009-06-05 12:42:42 +00:00
}
void rs600_set_safe_registers(struct radeon_device *rdev)
{
rdev->config.r300.reg_safe_bm = rs600_reg_safe_bm;
rdev->config.r300.reg_safe_bm_size = ARRAY_SIZE(rs600_reg_safe_bm);
}
static void rs600_mc_program(struct radeon_device *rdev)
{
struct rv515_mc_save save;
/* Stops all mc clients */
rv515_mc_stop(rdev, &save);
/* Wait for mc idle */
if (rs600_mc_wait_for_idle(rdev))
dev_warn(rdev->dev, "Wait MC idle timeout before updating MC.\n");
/* FIXME: What does AGP means for such chipset ? */
WREG32_MC(R_000005_MC_AGP_LOCATION, 0x0FFFFFFF);
WREG32_MC(R_000006_AGP_BASE, 0);
WREG32_MC(R_000007_AGP_BASE_2, 0);
/* Program MC */
WREG32_MC(R_000004_MC_FB_LOCATION,
S_000004_MC_FB_START(rdev->mc.vram_start >> 16) |
S_000004_MC_FB_TOP(rdev->mc.vram_end >> 16));
WREG32(R_000134_HDP_FB_LOCATION,
S_000134_HDP_FB_START(rdev->mc.vram_start >> 16));
rv515_mc_resume(rdev, &save);
}
static int rs600_startup(struct radeon_device *rdev)
{
int r;
rs600_mc_program(rdev);
/* Resume clock */
rv515_clock_startup(rdev);
/* Initialize GPU configuration (# pipes, ...) */
rs600_gpu_init(rdev);
/* Initialize GART (initialize after TTM so we can allocate
* memory through TTM but finalize after TTM) */
r = rs600_gart_enable(rdev);
if (r)
return r;
/* Enable IRQ */
rs600_irq_set(rdev);
/* 1M ring buffer */
r = r100_cp_init(rdev, 1024 * 1024);
if (r) {
dev_err(rdev->dev, "failled initializing CP (%d).\n", r);
return r;
}
r = r100_wb_init(rdev);
if (r)
dev_err(rdev->dev, "failled initializing WB (%d).\n", r);
r = r100_ib_init(rdev);
if (r) {
dev_err(rdev->dev, "failled initializing IB (%d).\n", r);
return r;
}
return 0;
}
int rs600_resume(struct radeon_device *rdev)
{
/* Make sur GART are not working */
rs600_gart_disable(rdev);
/* Resume clock before doing reset */
rv515_clock_startup(rdev);
/* Reset gpu before posting otherwise ATOM will enter infinite loop */
if (radeon_gpu_reset(rdev)) {
dev_warn(rdev->dev, "GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
RREG32(R_000E40_RBBM_STATUS),
RREG32(R_0007C0_CP_STAT));
}
/* post */
atom_asic_init(rdev->mode_info.atom_context);
/* Resume clock after posting */
rv515_clock_startup(rdev);
/* Initialize surface registers */
radeon_surface_init(rdev);
return rs600_startup(rdev);
}
int rs600_suspend(struct radeon_device *rdev)
{
r100_cp_disable(rdev);
r100_wb_disable(rdev);
rs600_irq_disable(rdev);
rs600_gart_disable(rdev);
return 0;
}
void rs600_fini(struct radeon_device *rdev)
{
rs600_suspend(rdev);
r100_cp_fini(rdev);
r100_wb_fini(rdev);
r100_ib_fini(rdev);
radeon_gem_fini(rdev);
rs600_gart_fini(rdev);
radeon_irq_kms_fini(rdev);
radeon_fence_driver_fini(rdev);
radeon_bo_fini(rdev);
radeon_atombios_fini(rdev);
kfree(rdev->bios);
rdev->bios = NULL;
}
int rs600_init(struct radeon_device *rdev)
{
int r;
/* Disable VGA */
rv515_vga_render_disable(rdev);
/* Initialize scratch registers */
radeon_scratch_init(rdev);
/* Initialize surface registers */
radeon_surface_init(rdev);
/* BIOS */
if (!radeon_get_bios(rdev)) {
if (ASIC_IS_AVIVO(rdev))
return -EINVAL;
}
if (rdev->is_atom_bios) {
r = radeon_atombios_init(rdev);
if (r)
return r;
} else {
dev_err(rdev->dev, "Expecting atombios for RS600 GPU\n");
return -EINVAL;
}
/* Reset gpu before posting otherwise ATOM will enter infinite loop */
if (radeon_gpu_reset(rdev)) {
dev_warn(rdev->dev,
"GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
RREG32(R_000E40_RBBM_STATUS),
RREG32(R_0007C0_CP_STAT));
}
/* check if cards are posted or not */
if (radeon_boot_test_post_card(rdev) == false)
return -EINVAL;
/* Initialize clocks */
radeon_get_clock_info(rdev->ddev);
/* Initialize power management */
radeon_pm_init(rdev);
/* Get vram informations */
rs600_vram_info(rdev);
/* Initialize memory controller (also test AGP) */
r = rs600_mc_init(rdev);
if (r)
return r;
rs600_debugfs(rdev);
/* Fence driver */
r = radeon_fence_driver_init(rdev);
if (r)
return r;
r = radeon_irq_kms_init(rdev);
if (r)
return r;
/* Memory manager */
r = radeon_bo_init(rdev);
if (r)
return r;
r = rs600_gart_init(rdev);
if (r)
return r;
rs600_set_safe_registers(rdev);
rdev->accel_working = true;
r = rs600_startup(rdev);
if (r) {
/* Somethings want wront with the accel init stop accel */
dev_err(rdev->dev, "Disabling GPU acceleration\n");
rs600_suspend(rdev);
r100_cp_fini(rdev);
r100_wb_fini(rdev);
r100_ib_fini(rdev);
rs600_gart_fini(rdev);
radeon_irq_kms_fini(rdev);
rdev->accel_working = false;
}
return 0;
}