summaryrefslogtreecommitdiffstats
path: root/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
diff options
context:
space:
mode:
authorLinus Torvalds <torvalds@linux-foundation.org>2021-09-01 20:26:46 +0200
committerLinus Torvalds <torvalds@linux-foundation.org>2021-09-01 20:26:46 +0200
commit477f70cd2a67904e04c2c2b9bd0fa2e95222f2f6 (patch)
tree1897dd1de49e1ea24897163533e2d8ead5dad0ad /drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
parentMerge tag 'media/v5.15-1' of git://git.kernel.org/pub/scm/linux/kernel/git/mc... (diff)
parentMerge tag 'amd-drm-next-5.15-2021-08-27' of https://gitlab.freedesktop.org/ag... (diff)
downloadlinux-477f70cd2a67904e04c2c2b9bd0fa2e95222f2f6.tar.xz
linux-477f70cd2a67904e04c2c2b9bd0fa2e95222f2f6.zip
Merge tag 'drm-next-2021-08-31-1' of git://anongit.freedesktop.org/drm/drm
Pull drm updates from Dave Airlie: "Highlights: - i915 has seen a lot of refactoring and uAPI cleanups due to a change in the upstream direction going forward This has all been audited with known userspace, but there may be some pitfalls that were missed. - i915 now uses common TTM to enable discrete memory on DG1/2 GPUs - i915 enables Jasper and Elkhart Lake by default and has preliminary XeHP/DG2 support - amdgpu adds support for Cyan Skillfish - lots of implicit fencing rules documented and fixed up in drivers - msm now uses the core scheduler - the irq midlayer has been removed for non-legacy drivers - the sysfb code now works on more than x86. Otherwise the usual smattering of stuff everywhere, panels, bridges, refactorings. Detailed summary: core: - extract i915 eDP backlight into core - DP aux bus support - drm_device.irq_enabled removed - port drivers to native irq interfaces - export gem shadow plane handling for vgem - print proper driver name in framebuffer registration - driver fixes for implicit fencing rules - ARM fixed rate compression modifier added - updated fb damage handling - rmfb ioctl logging/docs - drop drm_gem_object_put_locked - define DRM_FORMAT_MAX_PLANES - add gem fb vmap/vunmap helpers - add lockdep_assert(once) helpers - mark drm irq midlayer as legacy - use offset adjusted bo mapping conversion vgaarb: - cleanups fbdev: - extend efifb handling to all arches - div by 0 fixes for multiple drivers udmabuf: - add hugepage mapping support dma-buf: - non-dynamic exporter fixups - document implicit fencing rules amdgpu: - Initial Cyan Skillfish support - switch virtual DCE over to vkms based atomic - VCN/JPEG power down fixes - NAVI PCIE link handling fixes - AMD HDMI freesync fixes - Yellow Carp + Beige Goby fixes - Clockgating/S0ix/SMU/EEPROM fixes - embed hw fence in job - rework dma-resv handling - ensure eviction to system ram amdkfd: - uapi: SVM address range query added - sysfs leak fix - GPUVM TLB optimizations - vmfault/migration counters i915: - Enable JSL and EHL by default - preliminary XeHP/DG2 support - remove all CNL support (never shipped) - move to TTM for discrete memory support - allow mixed object mmap handling - GEM uAPI spring cleaning - add I915_MMAP_OBJECT_FIXED - reinstate ADL-P mmap ioctls - drop a bunch of unused by userspace features - disable and remove GPU relocations - revert some i915 misfeatures - major refactoring of GuC for Gen11+ - execbuffer object locking separate step - reject caching/set-domain on discrete - Enable pipe DMC loading on XE-LPD and ADL-P - add PSF GV point support - Refactor and fix DDI buffer translations - Clean up FBC CFB allocation code - Finish INTEL_GEN() and friends macro conversions nouveau: - add eDP backlight support - implicit fence fix msm: - a680/7c3 support - drm/scheduler conversion panfrost: - rework GPU reset virtio: - fix fencing for planes ast: - add detect support bochs: - move to tiny GPU driver vc4: - use hotplug irqs - HDMI codec support vmwgfx: - use internal vmware device headers ingenic: - demidlayering irq rcar-du: - shutdown fixes - convert to bridge connector helpers zynqmp-dsub: - misc fixes mgag200: - convert PLL handling to atomic mediatek: - MT8133 AAL support - gem mmap object support - MT8167 support etnaviv: - NXP Layerscape LS1028A SoC support - GEM mmap cleanups tegra: - new user API exynos: - missing unlock fix - build warning fix - use refcount_t" * tag 'drm-next-2021-08-31-1' of git://anongit.freedesktop.org/drm/drm: (1318 commits) drm/amd/display: Move AllowDRAMSelfRefreshOrDRAMClockChangeInVblank to bounding box drm/amd/display: Remove duplicate dml init drm/amd/display: Update bounding box states (v2) drm/amd/display: Update number of DCN3 clock states drm/amdgpu: disable GFX CGCG in aldebaran drm/amdgpu: Clear RAS interrupt status on aldebaran drm/amdgpu: Add support for RAS XGMI err query drm/amdkfd: Account for SH/SE count when setting up cu masks. drm/amdgpu: rename amdgpu_bo_get_preferred_pin_domain drm/amdgpu: drop redundant cancel_delayed_work_sync call drm/amdgpu: add missing cleanups for more ASICs on UVD/VCE suspend drm/amdgpu: add missing cleanups for Polaris12 UVD/VCE on suspend drm/amdkfd: map SVM range with correct access permission drm/amdkfd: check access permisson to restore retry fault drm/amdgpu: Update RAS XGMI Error Query drm/amdgpu: Add driver infrastructure for MCA RAS drm/amd/display: Add Logging for HDMI color depth information drm/amd/amdgpu: consolidate PSP TA init shared buf functions drm/amd/amdgpu: add name field back to ras_common_if drm/amdgpu: Fix build with missing pm_suspend_target_state module export ...
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c')
-rw-r--r--drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c84
1 files changed, 83 insertions, 1 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
index f9c01bdc3d4c..3003ee1c9487 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
@@ -21,6 +21,7 @@
*/
#include "amdgpu_amdkfd.h"
+#include "amd_pcie.h"
#include "amd_shared.h"
#include "amdgpu.h"
@@ -553,6 +554,88 @@ uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct kgd_dev *dst, struct kgd_dev *s
return (uint8_t)ret;
}
+int amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(struct kgd_dev *dst, struct kgd_dev *src, bool is_min)
+{
+ struct amdgpu_device *adev = (struct amdgpu_device *)dst, *peer_adev;
+ int num_links;
+
+ if (adev->asic_type != CHIP_ALDEBARAN)
+ return 0;
+
+ if (src)
+ peer_adev = (struct amdgpu_device *)src;
+
+ /* num links returns 0 for indirect peers since indirect route is unknown. */
+ num_links = is_min ? 1 : amdgpu_xgmi_get_num_links(adev, peer_adev);
+ if (num_links < 0) {
+ DRM_ERROR("amdgpu: failed to get xgmi num links between node %d and %d. ret = %d\n",
+ adev->gmc.xgmi.physical_node_id,
+ peer_adev->gmc.xgmi.physical_node_id, num_links);
+ num_links = 0;
+ }
+
+ /* Aldebaran xGMI DPM is defeatured so assume x16 x 25Gbps for bandwidth. */
+ return (num_links * 16 * 25000)/BITS_PER_BYTE;
+}
+
+int amdgpu_amdkfd_get_pcie_bandwidth_mbytes(struct kgd_dev *dev, bool is_min)
+{
+ struct amdgpu_device *adev = (struct amdgpu_device *)dev;
+ int num_lanes_shift = (is_min ? ffs(adev->pm.pcie_mlw_mask) :
+ fls(adev->pm.pcie_mlw_mask)) - 1;
+ int gen_speed_shift = (is_min ? ffs(adev->pm.pcie_gen_mask &
+ CAIL_PCIE_LINK_SPEED_SUPPORT_MASK) :
+ fls(adev->pm.pcie_gen_mask &
+ CAIL_PCIE_LINK_SPEED_SUPPORT_MASK)) - 1;
+ uint32_t num_lanes_mask = 1 << num_lanes_shift;
+ uint32_t gen_speed_mask = 1 << gen_speed_shift;
+ int num_lanes_factor = 0, gen_speed_mbits_factor = 0;
+
+ switch (num_lanes_mask) {
+ case CAIL_PCIE_LINK_WIDTH_SUPPORT_X1:
+ num_lanes_factor = 1;
+ break;
+ case CAIL_PCIE_LINK_WIDTH_SUPPORT_X2:
+ num_lanes_factor = 2;
+ break;
+ case CAIL_PCIE_LINK_WIDTH_SUPPORT_X4:
+ num_lanes_factor = 4;
+ break;
+ case CAIL_PCIE_LINK_WIDTH_SUPPORT_X8:
+ num_lanes_factor = 8;
+ break;
+ case CAIL_PCIE_LINK_WIDTH_SUPPORT_X12:
+ num_lanes_factor = 12;
+ break;
+ case CAIL_PCIE_LINK_WIDTH_SUPPORT_X16:
+ num_lanes_factor = 16;
+ break;
+ case CAIL_PCIE_LINK_WIDTH_SUPPORT_X32:
+ num_lanes_factor = 32;
+ break;
+ }
+
+ switch (gen_speed_mask) {
+ case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1:
+ gen_speed_mbits_factor = 2500;
+ break;
+ case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2:
+ gen_speed_mbits_factor = 5000;
+ break;
+ case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3:
+ gen_speed_mbits_factor = 8000;
+ break;
+ case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4:
+ gen_speed_mbits_factor = 16000;
+ break;
+ case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN5:
+ gen_speed_mbits_factor = 32000;
+ break;
+ }
+
+ return (num_lanes_factor * gen_speed_mbits_factor)/BITS_PER_BYTE;
+}
+
uint64_t amdgpu_amdkfd_get_mmio_remap_phys_addr(struct kgd_dev *kgd)
{
struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
@@ -631,7 +714,6 @@ int amdgpu_amdkfd_submit_ib(struct kgd_dev *kgd, enum kgd_engine_type engine,
ret = dma_fence_wait(f, false);
err_ib_sched:
- dma_fence_put(f);
amdgpu_job_free(job);
err:
return ret;