drm/amdkfd: force raven as "dgpu" path (v2)
Subsystems (3)
PATCH v2 1/3 drm/amdkfd: force raven as "dgpu" path (v2)
We still have a few iommu issues which need to address, so force raven as "dgpu" path for the moment.
This is to add the fallback path to bypass IOMMU if IOMMU v2 is disabled or ACPI CRAT table not correct.
v2: use ignore_crat parameter to decide whether it will go with IOMMUv2.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 5 +++- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 27 ++++++++++++++++++- drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c | 4 +-- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 2 +- .../amd/amdkfd/kfd_device_queue_manager_v9.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c | 4 +-- drivers/gpu/drm/amd/amdkfd/kfd_iommu.c | 10 +++---- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 7 ++--- 9 files changed, 46 insertions(+), 17 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4 -1
@@ -684,11 +684,14 @@ MODULE_PARM_DESC(debug_largebar,* Ignore CRAT table during KFD initialization. By default, KFD uses the ACPI CRAT * table to get information about AMD APUs. This option can serve as a workaround on * systems with a broken CRAT table. +* +* Default is auto (according to asic type, iommu_v2, and crat table, to decide +* whehter use CRAT) */ int ignore_crat;module_param (ignore_crat, int,0444 );MODULE_PARM_DESC (ignore_crat, -"Ignore CRAT table during KFD initialization (0 = use CRAT (default), 1 = ignore CRAT)" ); +"Ignore CRAT table during KFD initialization (0 = auto (default), 1 = ignore CRAT)" );/** * DOC: halt_if_hws_hang (int)
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -1
@@ -1258,7 +1258,7 @@ bool kfd_dev_is_large_bar(struct kfd_dev *dev)return true; } -if (dev->device_info->needs_iommu_device) +if (dev->device_info->needs_iommu_device && !ignore_crat)return false;amdgpu_amdkfd_get_local_mem_info (dev->kgd, &mem_info);
drivers/gpu/drm/amd/amdkfd/kfd_crat.c +26 -1
@@ -22,6 +22,7 @@#include <linux/pci.h> #include <linux/acpi.h> +#include <asm/processor.h> #include "kfd_crat.h" #include "kfd_priv.h" #include "kfd_topology.h"
@@ -740,6 +741,29 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev,return 0 ; } + +#ifdef CONFIG_ACPI +static void kfd_setup_ignore_crat_option (void ) +{ + +if (ignore_crat) +return ; + +#ifndef KFD_SUPPORT_IOMMU_V2 + ignore_crat =1 ; +#else + ignore_crat =0 ; +#endif + +/* Raven series will go with the fallback path to use virtual CRAT */ +if (boot_cpu_data.x86_vendor ==X86_VENDOR_AMD && + boot_cpu_data.x86 ==0x17 ) { + ignore_crat =1 ; + } + +return ; +} +/* * kfd_create_crat_image_acpi - Allocates memory for CRAT image and * copies CRAT from ACPI (if available).
@@ -751,7 +775,6 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev,* * Return 0 if successful else return error code */ -#ifdef CONFIG_ACPI int kfd_create_crat_image_acpi (void **crat_image,size_t *size) {struct acpi_table_header *crat_table;
@@ -775,6 +798,8 @@ int kfd_create_crat_image_acpi(void **crat_image, size_t *size)return -EINVAL ; } +kfd_setup_ignore_crat_option (); +if (ignore_crat) {pr_info ("CRAT table disabled by module option\n" );return -ENODATA ;
drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c +2 -2
@@ -309,7 +309,7 @@ static int dbgdev_address_watch_nodiq(struct kfd_dbgdev *dbgdev,for (i =0 ; i < adw_info->num_watch_points; i++) {dbgdev_address_watch_set_registers (adw_info, &addrHi, &addrLo, &cntl, i, pdd->qpd.vmid, - dbgdev->dev->device_info->needs_iommu_device); + dbgdev->dev->device_info->needs_iommu_device && !ignore_crat);pr_debug ("\t\t%30s\n" ,"* * * * * * * * * * * * * * * * * *" );pr_debug ("\t\t%20s %08x\n" ,"register index :" , i);
@@ -399,7 +399,7 @@ static int dbgdev_address_watch_diq(struct kfd_dbgdev *dbgdev,for (i =0 ; i < adw_info->num_watch_points; i++) {dbgdev_address_watch_set_registers (adw_info, &addrHi, &addrLo, &cntl, i, vmid, - dbgdev->dev->device_info->needs_iommu_device); + dbgdev->dev->device_info->needs_iommu_device && !ignore_crat);pr_debug ("\t\t%30s\n" ,"* * * * * * * * * * * * * * * * * *" );pr_debug ("\t\t%20s %08x\n" ,"register index :" , i);
drivers/gpu/drm/amd/amdkfd/kfd_device.c +1 -1
@@ -112,6 +112,7 @@ static const struct kfd_device_info carrizo_device_info = { .num_xgmi_sdma_engines = 0, .num_sdma_queues_per_engine = 2, }; +#endif static const struct kfd_device_info raven_device_info = { .asic_family =CHIP_RAVEN ,
@@ -130,7 +131,6 @@ static const struct kfd_device_info raven_device_info = { .num_xgmi_sdma_engines = 0, .num_sdma_queues_per_engine = 2, }; -#endif static const struct kfd_device_info hawaii_device_info = { .asic_family =CHIP_HAWAII ,
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c +1 -1
@@ -62,7 +62,7 @@ static int update_qpd_v9(struct device_queue_manager *dqm,SH_MEM_ALIGNMENT_MODE_UNALIGNED <<SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT ;if (amdgpu_noretry && - !dqm->dev->device_info->needs_iommu_device) + !(dqm->dev->device_info->needs_iommu_device && !ignore_crat)) qpd->sh_mem_config |=1 <<SH_MEM_CONFIG__RETRY_DISABLE__SHIFT ;
drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c +2 -2
@@ -321,7 +321,7 @@ static void kfd_init_apertures_vi(struct kfd_process_device *pdd, uint8_t id) pdd->lds_base =MAKE_LDS_APP_BASE_VI (); pdd->lds_limit =MAKE_LDS_APP_LIMIT (pdd->lds_base); -if (!pdd->dev->device_info->needs_iommu_device) { +if (!(pdd->dev->device_info->needs_iommu_device && !ignore_crat)) {/* dGPUs: SVM aperture starting at 0 * with small reserved space for kernel. * Set them to CANONICAL addresses.
@@ -423,7 +423,7 @@ int kfd_init_apertures(struct kfd_process *process)return -EINVAL ; } -if (!dev->device_info->needs_iommu_device) { +if (!(dev->device_info->needs_iommu_device && !ignore_crat)) {/* dGPUs: the reserved space for kernel * before SVM */
drivers/gpu/drm/amd/amdkfd/kfd_iommu.c +5 -5
@@ -41,7 +41,7 @@ int kfd_iommu_check_device(struct kfd_dev *kfd)struct amd_iommu_device_info iommu_info;int err; -if (!kfd->device_info->needs_iommu_device) +if (!(kfd->device_info->needs_iommu_device && !ignore_crat))return -ENODEV ; iommu_info.flags =0 ;
@@ -63,7 +63,7 @@ int kfd_iommu_device_init(struct kfd_dev *kfd)unsigned int pasid_limit;int err; -if (!kfd->device_info->needs_iommu_device) +if (!(kfd->device_info->needs_iommu_device && !ignore_crat))return 0 ; iommu_info.flags =0 ;
@@ -109,7 +109,7 @@ int kfd_iommu_bind_process_to_device(struct kfd_process_device *pdd)struct kfd_process *p = pdd->process;int err; -if (!dev->device_info->needs_iommu_device || pdd->bound ==PDD_BOUND ) +if (!(dev->device_info->needs_iommu_device && !ignore_crat) || pdd->bound ==PDD_BOUND )return 0 ;if (unlikely (pdd->bound ==PDD_BOUND_SUSPENDED )) {
@@ -284,7 +284,7 @@ static void kfd_unbind_processes_from_device(struct kfd_dev *kfd)*/ void kfd_iommu_suspend (struct kfd_dev *kfd) { -if (!kfd->device_info->needs_iommu_device) +if (!(kfd->device_info->needs_iommu_device && !ignore_crat))return ;kfd_unbind_processes_from_device (kfd);
@@ -304,7 +304,7 @@ int kfd_iommu_resume(struct kfd_dev *kfd)unsigned int pasid_limit;int err; -if (!kfd->device_info->needs_iommu_device) +if (!(kfd->device_info->needs_iommu_device && !ignore_crat))return 0 ; pasid_limit =kfd_get_pasid_limit ();
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +4 -3
@@ -545,7 +545,7 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr,* If the ASIC is APU except Kaveri, set local memory size * to 0 to disable local memory support */ -if (!dev->gpu->device_info->needs_iommu_device +if (!(dev->gpu->device_info->needs_iommu_device && !ignore_crat) || dev->gpu->device_info->asic_family ==CHIP_KAVERI ) {amdgpu_amdkfd_get_local_mem_info (dev->gpu->kgd, &local_mem_info);
@@ -1090,6 +1090,7 @@ int kfd_topology_init(void)COMPUTE_UNIT_CPU ,NULL , proximity_domain); cpu_only_node =1 ; + ignore_crat =1 ;if (ret) {pr_err ("Error creating VCRAT table for CPU\n" );return ret;
@@ -1197,7 +1198,7 @@ static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)/* Discrete GPUs need their own topology device list * entries. Don't assign them to CPU/APU nodes. */ -if (!gpu->device_info->needs_iommu_device && +if (!(gpu->device_info->needs_iommu_device && !ignore_crat) && dev->node_props.cpu_cores_count)continue ;
@@ -1452,7 +1453,7 @@ int kfd_topology_add_device(struct kfd_dev *gpu)* Overwrite ATS capability according to needs_iommu_device to fix * potential missing corresponding bit in CRAT of BIOS. */ -if (dev->gpu->device_info->needs_iommu_device) +if (dev->gpu->device_info->needs_iommu_device && !ignore_crat) dev->node_props.capability |=HSA_CAP_ATS_PRESENT ;else dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT ;
Signature · 6 lines
2.25.1 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
PATCH v2 2/3 drm/amdkfd: abstract the iommu device checking with ignore_crat
It's better to use inline function to wrap the iommu checking.
--- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c | 4 ++-- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c | 4 ++-- drivers/gpu/drm/amd/amdkfd/kfd_iommu.c | 10 +++++----- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 8 ++++++++ drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 6 +++--- 7 files changed, 22 insertions(+), 14 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -1
@@ -1258,7 +1258,7 @@ bool kfd_dev_is_large_bar(struct kfd_dev *dev)return true; } -if (dev->device_info->needs_iommu_device && !ignore_crat) +if (kfd_go_iommu_v2 (dev->device_info))return false;amdgpu_amdkfd_get_local_mem_info (dev->kgd, &mem_info);
drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c +2 -2
@@ -309,7 +309,7 @@ static int dbgdev_address_watch_nodiq(struct kfd_dbgdev *dbgdev,for (i =0 ; i < adw_info->num_watch_points; i++) {dbgdev_address_watch_set_registers (adw_info, &addrHi, &addrLo, &cntl, i, pdd->qpd.vmid, - dbgdev->dev->device_info->needs_iommu_device && !ignore_crat); +kfd_go_iommu_v2 (dbgdev->dev->device_info));pr_debug ("\t\t%30s\n" ,"* * * * * * * * * * * * * * * * * *" );pr_debug ("\t\t%20s %08x\n" ,"register index :" , i);
@@ -399,7 +399,7 @@ static int dbgdev_address_watch_diq(struct kfd_dbgdev *dbgdev,for (i =0 ; i < adw_info->num_watch_points; i++) {dbgdev_address_watch_set_registers (adw_info, &addrHi, &addrLo, &cntl, i, vmid, - dbgdev->dev->device_info->needs_iommu_device && !ignore_crat); +kfd_go_iommu_v2 (dbgdev->dev->device_info));pr_debug ("\t\t%30s\n" ,"* * * * * * * * * * * * * * * * * *" );pr_debug ("\t\t%20s %08x\n" ,"register index :" , i);
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c +1 -1
@@ -62,7 +62,7 @@ static int update_qpd_v9(struct device_queue_manager *dqm,SH_MEM_ALIGNMENT_MODE_UNALIGNED <<SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT ;if (amdgpu_noretry && - !(dqm->dev->device_info->needs_iommu_device && !ignore_crat)) + !(kfd_go_iommu_v2 (dqm->dev->device_info))) qpd->sh_mem_config |=1 <<SH_MEM_CONFIG__RETRY_DISABLE__SHIFT ;
drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c +2 -2
@@ -321,7 +321,7 @@ static void kfd_init_apertures_vi(struct kfd_process_device *pdd, uint8_t id) pdd->lds_base =MAKE_LDS_APP_BASE_VI (); pdd->lds_limit =MAKE_LDS_APP_LIMIT (pdd->lds_base); -if (!(pdd->dev->device_info->needs_iommu_device && !ignore_crat)) { +if (!kfd_go_iommu_v2 (pdd->dev->device_info)) {/* dGPUs: SVM aperture starting at 0 * with small reserved space for kernel. * Set them to CANONICAL addresses.
@@ -423,7 +423,7 @@ int kfd_init_apertures(struct kfd_process *process)return -EINVAL ; } -if (!(dev->device_info->needs_iommu_device && !ignore_crat)) { +if (!kfd_go_iommu_v2 (dev->device_info)) {/* dGPUs: the reserved space for kernel * before SVM */
drivers/gpu/drm/amd/amdkfd/kfd_iommu.c +5 -5
@@ -41,7 +41,7 @@ int kfd_iommu_check_device(struct kfd_dev *kfd)struct amd_iommu_device_info iommu_info;int err; -if (!(kfd->device_info->needs_iommu_device && !ignore_crat)) +if (!kfd_go_iommu_v2 (kfd->device_info))return -ENODEV ; iommu_info.flags =0 ;
@@ -63,7 +63,7 @@ int kfd_iommu_device_init(struct kfd_dev *kfd)unsigned int pasid_limit;int err; -if (!(kfd->device_info->needs_iommu_device && !ignore_crat)) +if (!kfd_go_iommu_v2 (kfd->device_info))return 0 ; iommu_info.flags =0 ;
@@ -109,7 +109,7 @@ int kfd_iommu_bind_process_to_device(struct kfd_process_device *pdd)struct kfd_process *p = pdd->process;int err; -if (!(dev->device_info->needs_iommu_device && !ignore_crat) || pdd->bound ==PDD_BOUND ) +if (!kfd_go_iommu_v2 (dev->device_info) || pdd->bound ==PDD_BOUND )return 0 ;if (unlikely (pdd->bound ==PDD_BOUND_SUSPENDED )) {
@@ -284,7 +284,7 @@ static void kfd_unbind_processes_from_device(struct kfd_dev *kfd)*/ void kfd_iommu_suspend (struct kfd_dev *kfd) { -if (!(kfd->device_info->needs_iommu_device && !ignore_crat)) +if (!kfd_go_iommu_v2 (kfd->device_info))return ;kfd_unbind_processes_from_device (kfd);
@@ -304,7 +304,7 @@ int kfd_iommu_resume(struct kfd_dev *kfd)unsigned int pasid_limit;int err; -if (!(kfd->device_info->needs_iommu_device && !ignore_crat)) +if (!kfd_go_iommu_v2 (kfd->device_info))return 0 ; pasid_limit =kfd_get_pasid_limit ();
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +8 -0
@@ -1232,6 +1232,14 @@ static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd)#endif } +static inline bool kfd_go_iommu_v2 (const struct kfd_device_info *info) +{ +if (!info) +return false; + +return info->needs_iommu_device && !ignore_crat; +} +/* Debugfs */ #if defined(CONFIG_DEBUG_FS )
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +3 -3
@@ -545,7 +545,7 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr,* If the ASIC is APU except Kaveri, set local memory size * to 0 to disable local memory support */ -if (!(dev->gpu->device_info->needs_iommu_device && !ignore_crat) +if (!kfd_go_iommu_v2 (dev->gpu->device_info) || dev->gpu->device_info->asic_family ==CHIP_KAVERI ) {amdgpu_amdkfd_get_local_mem_info (dev->gpu->kgd, &local_mem_info);
@@ -1198,7 +1198,7 @@ static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)/* Discrete GPUs need their own topology device list * entries. Don't assign them to CPU/APU nodes. */ -if (!(gpu->device_info->needs_iommu_device && !ignore_crat) && +if (!kfd_go_iommu_v2 (gpu->device_info) && dev->node_props.cpu_cores_count)continue ;
@@ -1453,7 +1453,7 @@ int kfd_topology_add_device(struct kfd_dev *gpu)* Overwrite ATS capability according to needs_iommu_device to fix * potential missing corresponding bit in CRAT of BIOS. */ -if (dev->gpu->device_info->needs_iommu_device && !ignore_crat) +if (kfd_go_iommu_v2 (dev->gpu->device_info)) dev->node_props.capability |=HSA_CAP_ATS_PRESENT ;else dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT ;
Signature · 6 lines
2.25.1 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Felix Kuehling
maintainerRe: [PATCH v2 2/3] drm/amdkfd: abstract the iommu device checking with ignore_crat
I'd recommend making this the first change in the series. Make 'drm/amdkfd: force raven as "dgpu" path' the second patch. That way it only needs to change one place.
A few more comments inline.
Am 2020-08-18 um 9:09 a.m. schrieb Huang Rui:
Quoted 137 lines
It's better to use inline function to wrap the iommu checking.
Signed-off-by Huang Rui <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c | 4 ++-- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c | 4 ++-- drivers/gpu/drm/amd/amdkfd/kfd_iommu.c | 10 +++++----- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 8 ++++++++ drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 6 +++--- 7 files changed, 22 insertions(+), 14 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1
-1@@ -1258,7 +1258,7 @@ bool kfd_dev_is_large_bar(struct kfd_dev *dev)return true; } -if (dev->device_info->needs_iommu_device && !ignore_crat) +if (kfd_go_iommu_v2 (dev->device_info))return false;amdgpu_amdkfd_get_local_mem_info (dev->kgd, &mem_info);drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c +2
-2@@ -309,7 +309,7 @@ static int dbgdev_address_watch_nodiq(struct kfd_dbgdev *dbgdev,for (i =0 ; i < adw_info->num_watch_points; i++) {dbgdev_address_watch_set_registers (adw_info, &addrHi, &addrLo, &cntl, i, pdd->qpd.vmid, - dbgdev->dev->device_info->needs_iommu_device && !ignore_crat); +kfd_go_iommu_v2 (dbgdev->dev->device_info));pr_debug ("\t\t%30s\n" ,"* * * * * * * * * * * * * * * * * *" );pr_debug ("\t\t%20s %08x\n" ,"register index :" , i);@@ -399,7 +399,7 @@ static int dbgdev_address_watch_diq(struct kfd_dbgdev *dbgdev,for (i =0 ; i < adw_info->num_watch_points; i++) {dbgdev_address_watch_set_registers (adw_info, &addrHi, &addrLo, &cntl, i, vmid, - dbgdev->dev->device_info->needs_iommu_device && !ignore_crat); +kfd_go_iommu_v2 (dbgdev->dev->device_info));pr_debug ("\t\t%30s\n" ,"* * * * * * * * * * * * * * * * * *" );pr_debug ("\t\t%20s %08x\n" ,"register index :" , i);drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c +1
-1@@ -62,7 +62,7 @@ static int update_qpd_v9(struct device_queue_manager *dqm,SH_MEM_ALIGNMENT_MODE_UNALIGNED <<SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT ;if (amdgpu_noretry && - !(dqm->dev->device_info->needs_iommu_device && !ignore_crat)) + !(kfd_go_iommu_v2 (dqm->dev->device_info))) qpd->sh_mem_config |=1 <<SH_MEM_CONFIG__RETRY_DISABLE__SHIFT ;drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c +2
-2@@ -321,7 +321,7 @@ static void kfd_init_apertures_vi(struct kfd_process_device *pdd, uint8_t id) pdd->lds_base =MAKE_LDS_APP_BASE_VI (); pdd->lds_limit =MAKE_LDS_APP_LIMIT (pdd->lds_base); -if (!(pdd->dev->device_info->needs_iommu_device && !ignore_crat)) { +if (!kfd_go_iommu_v2 (pdd->dev->device_info)) {/* dGPUs: SVM aperture starting at 0 * with small reserved space for kernel. * Set them to CANONICAL addresses. @@ -423,7 +423,7 @@ int kfd_init_apertures(struct kfd_process *process)return -EINVAL ; } -if (!(dev->device_info->needs_iommu_device && !ignore_crat)) { +if (!kfd_go_iommu_v2 (dev->device_info)) {/* dGPUs: the reserved space for kernel * before SVM */ drivers/gpu/drm/amd/amdkfd/kfd_iommu.c +5
-5@@ -41,7 +41,7 @@ int kfd_iommu_check_device(struct kfd_dev *kfd)struct amd_iommu_device_info iommu_info;int err; -if (!(kfd->device_info->needs_iommu_device && !ignore_crat)) +if (!kfd_go_iommu_v2 (kfd->device_info))return -ENODEV ; iommu_info.flags =0 ;@@ -63,7 +63,7 @@ int kfd_iommu_device_init(struct kfd_dev *kfd)unsigned int pasid_limit;int err; -if (!(kfd->device_info->needs_iommu_device && !ignore_crat)) +if (!kfd_go_iommu_v2 (kfd->device_info))return 0 ; iommu_info.flags =0 ;@@ -109,7 +109,7 @@ int kfd_iommu_bind_process_to_device(struct kfd_process_device *pdd)struct kfd_process *p = pdd->process;int err; -if (!(dev->device_info->needs_iommu_device && !ignore_crat) || pdd->bound ==PDD_BOUND ) +if (!kfd_go_iommu_v2 (dev->device_info) || pdd->bound ==PDD_BOUND )return 0 ;if (unlikely (pdd->bound ==PDD_BOUND_SUSPENDED )) {@@ -284,7 +284,7 @@ static void kfd_unbind_processes_from_device(struct kfd_dev *kfd)*/ void kfd_iommu_suspend (struct kfd_dev *kfd) { -if (!(kfd->device_info->needs_iommu_device && !ignore_crat)) +if (!kfd_go_iommu_v2 (kfd->device_info))return ;kfd_unbind_processes_from_device (kfd);@@ -304,7 +304,7 @@ int kfd_iommu_resume(struct kfd_dev *kfd)unsigned int pasid_limit;int err; -if (!(kfd->device_info->needs_iommu_device && !ignore_crat)) +if (!kfd_go_iommu_v2 (kfd->device_info))return 0 ; pasid_limit =kfd_get_pasid_limit ();drivers/gpu/drm/amd/amdkfd/kfd_priv.h +0
-0
drivers/gpu/drm/amd/amdkfd/kfd_priv.h static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd) · line 1235
Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +1
-0@@ -1232,6 +1232,14 @@ static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd)#endif } +static inline bool kfd_go_iommu_v2 (const struct kfd_device_info *info)
I'd prefer a different function name: kfd_device_use_iommu_v2.
Then make the parameter const struct kfd_device *dev.
drivers/gpu/drm/amd/amdkfd/kfd_priv.h static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd) · line 1241
Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +6
-0+{ +if (!info) +return false; + +return info->needs_iommu_device && !ignore_crat; +}
If you need a NULL check, you could still write it in a more compact way (I like inline functions to be one-liners):
return !ignore_crat && dev && dev->device_info->needs_iommu_device;
If you make this the first patch in the series, remove the !ignore_crat part. That would be added in the second change.
Regards,
Felix
Quoted 27 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +1
-0+/* Debugfs */ #if defined(CONFIG_DEBUG_FS )drivers/gpu/drm/amd/amdkfd/kfd_topology.c +2
-2@@ -545,7 +545,7 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr,* If the ASIC is APU except Kaveri, set local memory size * to 0 to disable local memory support */ -if (!(dev->gpu->device_info->needs_iommu_device && !ignore_crat) +if (!kfd_go_iommu_v2 (dev->gpu->device_info) || dev->gpu->device_info->asic_family ==CHIP_KAVERI ) {amdgpu_amdkfd_get_local_mem_info (dev->gpu->kgd, &local_mem_info);@@ -1198,7 +1198,7 @@ static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)/* Discrete GPUs need their own topology device list * entries. Don't assign them to CPU/APU nodes. */ -if (!(gpu->device_info->needs_iommu_device && !ignore_crat) && +if (!kfd_go_iommu_v2 (gpu->device_info) && dev->node_props.cpu_cores_count)continue ;
drivers/gpu/drm/amd/amdkfd/kfd_topology.c int kfd_topology_add_device(struct kfd_dev *gpu) · line 1459
Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +1
-1@@ -1453,7 +1453,7 @@ int kfd_topology_add_device(struct kfd_dev *gpu)* Overwrite ATS capability according to needs_iommu_device to fix * potential missing corresponding bit in CRAT of BIOS. */ -if (dev->gpu->device_info->needs_iommu_device && !ignore_crat) +if (kfd_go_iommu_v2 (dev->gpu->device_info)) dev->node_props.capability |=HSA_CAP_ATS_PRESENT ;else dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT ;
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH v2 2/3] drm/amdkfd: abstract the iommu device checking with ignore_crat
Quoted 4 lines · Felix
I'd recommend making this the first change in the series. Make 'drm/amdkfd: force raven as "dgpu" path' the second patch. That way it only needs to change one place.
Yes, right.
Quoted 160 lines
A few more comments inline.
Am 2020-08-18 um 9:09 a.m. schrieb Huang Rui:
Quoted 137 lines
It's better to use inline function to wrap the iommu checking.
Signed-off-by Huang Rui <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c | 4 ++-- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c | 4 ++-- drivers/gpu/drm/amd/amdkfd/kfd_iommu.c | 10 +++++----- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 8 ++++++++ drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 6 +++--- 7 files changed, 22 insertions(+), 14 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1
-1@@ -1258,7 +1258,7 @@ bool kfd_dev_is_large_bar(struct kfd_dev *dev)return true; } -if (dev->device_info->needs_iommu_device && !ignore_crat) +if (kfd_go_iommu_v2 (dev->device_info))return false;amdgpu_amdkfd_get_local_mem_info (dev->kgd, &mem_info);drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c +2
-2@@ -309,7 +309,7 @@ static int dbgdev_address_watch_nodiq(struct kfd_dbgdev *dbgdev,for (i =0 ; i < adw_info->num_watch_points; i++) {dbgdev_address_watch_set_registers (adw_info, &addrHi, &addrLo, &cntl, i, pdd->qpd.vmid, - dbgdev->dev->device_info->needs_iommu_device && !ignore_crat); +kfd_go_iommu_v2 (dbgdev->dev->device_info));pr_debug ("\t\t%30s\n" ,"* * * * * * * * * * * * * * * * * *" );pr_debug ("\t\t%20s %08x\n" ,"register index :" , i);@@ -399,7 +399,7 @@ static int dbgdev_address_watch_diq(struct kfd_dbgdev *dbgdev,for (i =0 ; i < adw_info->num_watch_points; i++) {dbgdev_address_watch_set_registers (adw_info, &addrHi, &addrLo, &cntl, i, vmid, - dbgdev->dev->device_info->needs_iommu_device && !ignore_crat); +kfd_go_iommu_v2 (dbgdev->dev->device_info));pr_debug ("\t\t%30s\n" ,"* * * * * * * * * * * * * * * * * *" );pr_debug ("\t\t%20s %08x\n" ,"register index :" , i);drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c +1
-1@@ -62,7 +62,7 @@ static int update_qpd_v9(struct device_queue_manager *dqm,SH_MEM_ALIGNMENT_MODE_UNALIGNED <<SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT ;if (amdgpu_noretry && - !(dqm->dev->device_info->needs_iommu_device && !ignore_crat)) + !(kfd_go_iommu_v2 (dqm->dev->device_info))) qpd->sh_mem_config |=1 <<SH_MEM_CONFIG__RETRY_DISABLE__SHIFT ;drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c +2
-2@@ -321,7 +321,7 @@ static void kfd_init_apertures_vi(struct kfd_process_device *pdd, uint8_t id) pdd->lds_base =MAKE_LDS_APP_BASE_VI (); pdd->lds_limit =MAKE_LDS_APP_LIMIT (pdd->lds_base); -if (!(pdd->dev->device_info->needs_iommu_device && !ignore_crat)) { +if (!kfd_go_iommu_v2 (pdd->dev->device_info)) {/* dGPUs: SVM aperture starting at 0 * with small reserved space for kernel. * Set them to CANONICAL addresses. @@ -423,7 +423,7 @@ int kfd_init_apertures(struct kfd_process *process)return -EINVAL ; } -if (!(dev->device_info->needs_iommu_device && !ignore_crat)) { +if (!kfd_go_iommu_v2 (dev->device_info)) {/* dGPUs: the reserved space for kernel * before SVM */ drivers/gpu/drm/amd/amdkfd/kfd_iommu.c +5
-5@@ -41,7 +41,7 @@ int kfd_iommu_check_device(struct kfd_dev *kfd)struct amd_iommu_device_info iommu_info;int err; -if (!(kfd->device_info->needs_iommu_device && !ignore_crat)) +if (!kfd_go_iommu_v2 (kfd->device_info))return -ENODEV ; iommu_info.flags =0 ;@@ -63,7 +63,7 @@ int kfd_iommu_device_init(struct kfd_dev *kfd)unsigned int pasid_limit;int err; -if (!(kfd->device_info->needs_iommu_device && !ignore_crat)) +if (!kfd_go_iommu_v2 (kfd->device_info))return 0 ; iommu_info.flags =0 ;@@ -109,7 +109,7 @@ int kfd_iommu_bind_process_to_device(struct kfd_process_device *pdd)struct kfd_process *p = pdd->process;int err; -if (!(dev->device_info->needs_iommu_device && !ignore_crat) || pdd->bound ==PDD_BOUND ) +if (!kfd_go_iommu_v2 (dev->device_info) || pdd->bound ==PDD_BOUND )return 0 ;if (unlikely (pdd->bound ==PDD_BOUND_SUSPENDED )) {@@ -284,7 +284,7 @@ static void kfd_unbind_processes_from_device(struct kfd_dev *kfd)*/ void kfd_iommu_suspend (struct kfd_dev *kfd) { -if (!(kfd->device_info->needs_iommu_device && !ignore_crat)) +if (!kfd_go_iommu_v2 (kfd->device_info))return ;kfd_unbind_processes_from_device (kfd);@@ -304,7 +304,7 @@ int kfd_iommu_resume(struct kfd_dev *kfd)unsigned int pasid_limit;int err; -if (!(kfd->device_info->needs_iommu_device && !ignore_crat)) +if (!kfd_go_iommu_v2 (kfd->device_info))return 0 ; pasid_limit =kfd_get_pasid_limit ();drivers/gpu/drm/amd/amdkfd/kfd_priv.h +0
-0drivers/gpu/drm/amd/amdkfd/kfd_priv.h static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd) · line 1235
Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +1
-0@@ -1232,6 +1232,14 @@ static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd)#endif } +static inline bool kfd_go_iommu_v2 (const struct kfd_device_info *info)I'd prefer a different function name:
kfd_device_use_iommu_v2.Then make the parameter const struct
kfd_device*dev.drivers/gpu/drm/amd/amdkfd/kfd_priv.h static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd) · line 1241
Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +6
-0+{ +if (!info) +return false; + +return info->needs_iommu_device && !ignore_crat; +}If you need a NULL check, you could still write it in a more compact way (I like inline functions to be one-liners):
Quoted 4 lines
return !ignore_crat && dev &&
dev->device_info->needs_iommu_device;If you make this the first patch in the series, remove the !ignore_crat part. That would be added in the second change.
That's fine. It looks better, thanks. I will update it in V2.
Thanks,
Ray
Quoted 31 lines
Regards,
FelixQuoted 26 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +1
-0+/* Debugfs */ #if defined(CONFIG_DEBUG_FS )drivers/gpu/drm/amd/amdkfd/kfd_topology.c +2
-2@@ -545,7 +545,7 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr,* If the ASIC is APU except Kaveri, set local memory size * to 0 to disable local memory support */ -if (!(dev->gpu->device_info->needs_iommu_device && !ignore_crat) +if (!kfd_go_iommu_v2 (dev->gpu->device_info) || dev->gpu->device_info->asic_family ==CHIP_KAVERI ) {amdgpu_amdkfd_get_local_mem_info (dev->gpu->kgd, &local_mem_info);@@ -1198,7 +1198,7 @@ static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)/* Discrete GPUs need their own topology device list * entries. Don't assign them to CPU/APU nodes. */ -if (!(gpu->device_info->needs_iommu_device && !ignore_crat) && +if (!kfd_go_iommu_v2 (gpu->device_info) && dev->node_props.cpu_cores_count)continue ;
Quoted 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_topology.c int kfd_topology_add_device(struct kfd_dev *gpu) · line 1459
Quoted 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +1
-1@@ -1453,7 +1453,7 @@ int kfd_topology_add_device(struct kfd_dev *gpu)* Overwrite ATS capability according to needs_iommu_device to fix * potential missing corresponding bit in CRAT of BIOS. */ -if (dev->gpu->device_info->needs_iommu_device && !ignore_crat) +if (kfd_go_iommu_v2 (dev->gpu->device_info)) dev->node_props.capability |=HSA_CAP_ATS_PRESENT ;else dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT ;
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
PATCH v2 3/3 drm/amdkfd: remove iommu v2 for old apu series
In reply to Huang RuiWe already support the fallback path, so it doesn't need IOMMU v2 flag anymore.
--- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 6 ------ 1 file changed, 6 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device.c +0 -6
@@ -48,13 +48,11 @@ extern const struct kfd2kgd_calls arcturus_kfd2kgd;extern const struct kfd2kgd_calls gfx_v10_kfd2kgd;static const struct kfd2kgd_calls *kfd2kgd_funcs[] = { -#ifdef KFD_SUPPORT_IOMMU_V2 #ifdef CONFIG_DRM_AMDGPU_CIK [CHIP_KAVERI ] = &gfx_v7_kfd2kgd, #endif [CHIP_CARRIZO ] = &gfx_v8_kfd2kgd, [CHIP_RAVEN ] = &gfx_v9_kfd2kgd, -#endif #ifdef CONFIG_DRM_AMDGPU_CIK [CHIP_HAWAII ] = &gfx_v7_kfd2kgd, #endif
@@ -74,7 +72,6 @@ static const struct kfd2kgd_calls *kfd2kgd_funcs[] = { [CHIP_NAVI14] = &gfx_v10_kfd2kgd, }; -#ifdef KFD_SUPPORT_IOMMU_V2 static const struct kfd_device_info kaveri_device_info = { .asic_family =CHIP_KAVERI , .asic_name ="kaveri" ,
@@ -112,7 +109,6 @@ static const struct kfd_device_info carrizo_device_info = { .num_xgmi_sdma_engines = 0, .num_sdma_queues_per_engine = 2, }; -#endif static const struct kfd_device_info raven_device_info = { .asic_family =CHIP_RAVEN ,
@@ -460,11 +456,9 @@ static const struct kfd_device_info navi14_device_info = {/* For each entry, [0] is regular and [1] is virtualisation device. */ static const struct kfd_device_info *kfd_supported_devices[][2 ] = { -#ifdef KFD_SUPPORT_IOMMU_V2 [CHIP_KAVERI ] = {&kaveri_device_info,NULL }, [CHIP_CARRIZO ] = {&carrizo_device_info,NULL }, [CHIP_RAVEN ] = {&raven_device_info,NULL }, -#endif [CHIP_HAWAII ] = {&hawaii_device_info,NULL }, [CHIP_TONGA ] = {&tonga_device_info,NULL }, [CHIP_FIJI ] = {&fiji_device_info, &fiji_vf_device_info},
Signature · 6 lines
2.25.1 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH v2 3/3] drm/amdkfd: remove iommu v2 for old apu series
Interesting. Does this actually work on Carrizo or Kaveri? I'd like to see any Thunk changes needed to support this before giving my R-b. For now this patch is
Am 2020-08-18 um 9:09 a.m. schrieb Huang Rui:
Quoted 44 lines
We already support the fallback path, so it doesn't need IOMMU v2 flag anymore.
Signed-off-by Huang Rui <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 6 ------ 1 file changed, 6 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_device.c +0
-4@@ -48,13 +48,11 @@ extern const struct kfd2kgd_calls arcturus_kfd2kgd;extern const struct kfd2kgd_calls gfx_v10_kfd2kgd;static const struct kfd2kgd_calls *kfd2kgd_funcs[] = { -#ifdef KFD_SUPPORT_IOMMU_V2 #ifdef CONFIG_DRM_AMDGPU_CIK [CHIP_KAVERI ] = &gfx_v7_kfd2kgd, #endif [CHIP_CARRIZO ] = &gfx_v8_kfd2kgd, [CHIP_RAVEN ] = &gfx_v9_kfd2kgd, -#endif #ifdef CONFIG_DRM_AMDGPU_CIK [CHIP_HAWAII ] = &gfx_v7_kfd2kgd, #endif@@ -74,7 +72,6 @@ static const struct kfd2kgd_calls *kfd2kgd_funcs[] = { [CHIP_NAVI14] = &gfx_v10_kfd2kgd, }; -#ifdef KFD_SUPPORT_IOMMU_V2 static const struct kfd_device_info kaveri_device_info = { .asic_family =CHIP_KAVERI , .asic_name ="kaveri" ,@@ -112,7 +109,6 @@ static const struct kfd_device_info carrizo_device_info = { .num_xgmi_sdma_engines = 0, .num_sdma_queues_per_engine = 2, }; -#endif static const struct kfd_device_info raven_device_info = { .asic_family =CHIP_RAVEN ,@@ -460,11 +456,9 @@ static const struct kfd_device_info navi14_device_info = {
drivers/gpu/drm/amd/amdkfd/kfd_device.c static const struct kfd_device_info navi14_device_info = { · line 464
Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_device.c +0
-2/* For each entry, [0] is regular and [1] is virtualisation device. */ static const struct kfd_device_info *kfd_supported_devices[][2 ] = { -#ifdef KFD_SUPPORT_IOMMU_V2 [CHIP_KAVERI ] = {&kaveri_device_info,NULL }, [CHIP_CARRIZO ] = {&carrizo_device_info,NULL }, [CHIP_RAVEN ] = {&raven_device_info,NULL }, -#endif [CHIP_HAWAII ] = {&hawaii_device_info,NULL }, [CHIP_TONGA ] = {&tonga_device_info,NULL }, [CHIP_FIJI ] = {&fiji_device_info, &fiji_vf_device_info},
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH v2 3/3] drm/amdkfd: remove iommu v2 for old apu series
Quoted 1 line · Felix
Interesting. Does this actually work on Carrizo or Kaveri? I'd like to
I just found a Carrizo board, rocm info works well, and clinfo is able to print most information. But it seems stuck in the ROCr (submit SubmitLinearCopyCommand). Let me take a look.
Thanks,
Ray
Quoted 51 lines
see any Thunk changes needed to support this before giving my R-b. For now this patch is
Acked-by Felix Kuehling <[email protected]>Am 2020-08-18 um 9:09 a.m. schrieb Huang Rui:
Quoted 44 lines
We already support the fallback path, so it doesn't need IOMMU v2 flag anymore.
Signed-off-by Huang Rui <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 6 ------ 1 file changed, 6 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_device.c +0
-4@@ -48,13 +48,11 @@ extern const struct kfd2kgd_calls arcturus_kfd2kgd;extern const struct kfd2kgd_calls gfx_v10_kfd2kgd;static const struct kfd2kgd_calls *kfd2kgd_funcs[] = { -#ifdef KFD_SUPPORT_IOMMU_V2 #ifdef CONFIG_DRM_AMDGPU_CIK [CHIP_KAVERI ] = &gfx_v7_kfd2kgd, #endif [CHIP_CARRIZO ] = &gfx_v8_kfd2kgd, [CHIP_RAVEN ] = &gfx_v9_kfd2kgd, -#endif #ifdef CONFIG_DRM_AMDGPU_CIK [CHIP_HAWAII ] = &gfx_v7_kfd2kgd, #endif@@ -74,7 +72,6 @@ static const struct kfd2kgd_calls *kfd2kgd_funcs[] = { [CHIP_NAVI14] = &gfx_v10_kfd2kgd, }; -#ifdef KFD_SUPPORT_IOMMU_V2 static const struct kfd_device_info kaveri_device_info = { .asic_family =CHIP_KAVERI , .asic_name ="kaveri" ,@@ -112,7 +109,6 @@ static const struct kfd_device_info carrizo_device_info = { .num_xgmi_sdma_engines = 0, .num_sdma_queues_per_engine = 2, }; -#endif static const struct kfd_device_info raven_device_info = { .asic_family =CHIP_RAVEN ,@@ -460,11 +456,9 @@ static const struct kfd_device_info navi14_device_info = {
Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_device.c static const struct kfd_device_info navi14_device_info = { · line 464
Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_device.c +0
-2/* For each entry, [0] is regular and [1] is virtualisation device. */ static const struct kfd_device_info *kfd_supported_devices[][2 ] = { -#ifdef KFD_SUPPORT_IOMMU_V2 [CHIP_KAVERI ] = {&kaveri_device_info,NULL }, [CHIP_CARRIZO ] = {&carrizo_device_info,NULL }, [CHIP_RAVEN ] = {&raven_device_info,NULL }, -#endif [CHIP_HAWAII ] = {&hawaii_device_info,NULL }, [CHIP_TONGA ] = {&tonga_device_info,NULL }, [CHIP_FIJI ] = {&fiji_device_info, &fiji_vf_device_info},
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Felix Kuehling
maintainerAm 2020-08-18 um 9:09 a.m. schrieb Huang Rui:
Quoted 78 lines
We still have a few iommu issues which need to address, so force raven as "dgpu" path for the moment.
This is to add the fallback path to bypass IOMMU if IOMMU v2 is disabled or ACPI CRAT table not correct.
v2: use
ignore_cratparameter to decide whether it will go with IOMMUv2.Signed-off-by Huang Rui <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 5 +++- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 27 ++++++++++++++++++- drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c | 4 +-- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 2 +- .../amd/amdkfd/kfd_device_queue_manager_v9.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c | 4 +-- drivers/gpu/drm/amd/amdkfd/kfd_iommu.c | 10 +++---- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 7 ++--- 9 files changed, 46 insertions(+), 17 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4
-1@@ -684,11 +684,14 @@ MODULE_PARM_DESC(debug_largebar,* Ignore CRAT table during KFD initialization. By default, KFD uses the ACPI CRAT * table to get information about AMD APUs. This option can serve as a workaround on * systems with a broken CRAT table. +* +* Default is auto (according to asic type, iommu_v2, and crat table, to decide +* whehter use CRAT) */ int ignore_crat;module_param (ignore_crat, int,0444 );MODULE_PARM_DESC (ignore_crat, -"Ignore CRAT table during KFD initialization (0 = use CRAT (default), 1 = ignore CRAT)" ); +"Ignore CRAT table during KFD initialization (0 = auto (default), 1 = ignore CRAT)" );/** * DOC: halt_if_hws_hang (int) drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1
-1@@ -1258,7 +1258,7 @@ bool kfd_dev_is_large_bar(struct kfd_dev *dev)return true; } -if (dev->device_info->needs_iommu_device) +if (dev->device_info->needs_iommu_device && !ignore_crat)return false;amdgpu_amdkfd_get_local_mem_info (dev->kgd, &mem_info);drivers/gpu/drm/amd/amdkfd/kfd_crat.c +9
-0@@ -22,6 +22,7 @@#include <linux/pci.h> #include <linux/acpi.h> +#include <asm/processor.h> #include "kfd_crat.h" #include "kfd_priv.h" #include "kfd_topology.h" @@ -740,6 +741,29 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev,return 0 ; } + +#ifdef CONFIG_ACPI +static void kfd_setup_ignore_crat_option (void ) +{ + +if (ignore_crat) +return ; +
drivers/gpu/drm/amd/amdkfd/kfd_crat.c static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev, · line 761
Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_crat.c +10
-0+#ifndef KFD_SUPPORT_IOMMU_V2 + ignore_crat =1 ; +#else + ignore_crat =0 ; +#endif + +/* Raven series will go with the fallback path to use virtual CRAT */ +if (boot_cpu_data.x86_vendor ==X86_VENDOR_AMD && + boot_cpu_data.x86 ==0x17 ) { + ignore_crat =1 ;
Does your change require Thunk changes? If yes, then changing this by default breaks existing user mode and cannot be applied upstream.
If you expect that this is a temporary workaround, then don't apply it here, apply it in your grub config or modprobe config.
Regards,
Felix
Quoted 182 lines
drivers/gpu/drm/amd/amdkfd/kfd_crat.c +7
-1+ } + +return ; +} +/* * kfd_create_crat_image_acpi - Allocates memory for CRAT image and * copies CRAT from ACPI (if available). @@ -751,7 +775,6 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev,* * Return 0 if successful else return error code */ -#ifdef CONFIG_ACPI int kfd_create_crat_image_acpi (void **crat_image,size_t *size) {struct acpi_table_header *crat_table;@@ -775,6 +798,8 @@ int kfd_create_crat_image_acpi(void **crat_image, size_t *size)return -EINVAL ; } +kfd_setup_ignore_crat_option (); +if (ignore_crat) {pr_info ("CRAT table disabled by module option\n" );return -ENODATA ;drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c +2
-2@@ -309,7 +309,7 @@ static int dbgdev_address_watch_nodiq(struct kfd_dbgdev *dbgdev,for (i =0 ; i < adw_info->num_watch_points; i++) {dbgdev_address_watch_set_registers (adw_info, &addrHi, &addrLo, &cntl, i, pdd->qpd.vmid, - dbgdev->dev->device_info->needs_iommu_device); + dbgdev->dev->device_info->needs_iommu_device && !ignore_crat);pr_debug ("\t\t%30s\n" ,"* * * * * * * * * * * * * * * * * *" );pr_debug ("\t\t%20s %08x\n" ,"register index :" , i);@@ -399,7 +399,7 @@ static int dbgdev_address_watch_diq(struct kfd_dbgdev *dbgdev,for (i =0 ; i < adw_info->num_watch_points; i++) {dbgdev_address_watch_set_registers (adw_info, &addrHi, &addrLo, &cntl, i, vmid, - dbgdev->dev->device_info->needs_iommu_device); + dbgdev->dev->device_info->needs_iommu_device && !ignore_crat);pr_debug ("\t\t%30s\n" ,"* * * * * * * * * * * * * * * * * *" );pr_debug ("\t\t%20s %08x\n" ,"register index :" , i);drivers/gpu/drm/amd/amdkfd/kfd_device.c +1
-1@@ -112,6 +112,7 @@ static const struct kfd_device_info carrizo_device_info = { .num_xgmi_sdma_engines = 0, .num_sdma_queues_per_engine = 2, }; +#endif static const struct kfd_device_info raven_device_info = { .asic_family =CHIP_RAVEN ,@@ -130,7 +131,6 @@ static const struct kfd_device_info raven_device_info = { .num_xgmi_sdma_engines = 0, .num_sdma_queues_per_engine = 2, }; -#endif static const struct kfd_device_info hawaii_device_info = { .asic_family =CHIP_HAWAII ,drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c +1
-1@@ -62,7 +62,7 @@ static int update_qpd_v9(struct device_queue_manager *dqm,SH_MEM_ALIGNMENT_MODE_UNALIGNED <<SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT ;if (amdgpu_noretry && - !dqm->dev->device_info->needs_iommu_device) + !(dqm->dev->device_info->needs_iommu_device && !ignore_crat)) qpd->sh_mem_config |=1 <<SH_MEM_CONFIG__RETRY_DISABLE__SHIFT ;drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c +2
-2@@ -321,7 +321,7 @@ static void kfd_init_apertures_vi(struct kfd_process_device *pdd, uint8_t id) pdd->lds_base =MAKE_LDS_APP_BASE_VI (); pdd->lds_limit =MAKE_LDS_APP_LIMIT (pdd->lds_base); -if (!pdd->dev->device_info->needs_iommu_device) { +if (!(pdd->dev->device_info->needs_iommu_device && !ignore_crat)) {/* dGPUs: SVM aperture starting at 0 * with small reserved space for kernel. * Set them to CANONICAL addresses. @@ -423,7 +423,7 @@ int kfd_init_apertures(struct kfd_process *process)return -EINVAL ; } -if (!dev->device_info->needs_iommu_device) { +if (!(dev->device_info->needs_iommu_device && !ignore_crat)) {/* dGPUs: the reserved space for kernel * before SVM */ drivers/gpu/drm/amd/amdkfd/kfd_iommu.c +5
-5@@ -41,7 +41,7 @@ int kfd_iommu_check_device(struct kfd_dev *kfd)struct amd_iommu_device_info iommu_info;int err; -if (!kfd->device_info->needs_iommu_device) +if (!(kfd->device_info->needs_iommu_device && !ignore_crat))return -ENODEV ; iommu_info.flags =0 ;@@ -63,7 +63,7 @@ int kfd_iommu_device_init(struct kfd_dev *kfd)unsigned int pasid_limit;int err; -if (!kfd->device_info->needs_iommu_device) +if (!(kfd->device_info->needs_iommu_device && !ignore_crat))return 0 ; iommu_info.flags =0 ;@@ -109,7 +109,7 @@ int kfd_iommu_bind_process_to_device(struct kfd_process_device *pdd)struct kfd_process *p = pdd->process;int err; -if (!dev->device_info->needs_iommu_device || pdd->bound ==PDD_BOUND ) +if (!(dev->device_info->needs_iommu_device && !ignore_crat) || pdd->bound ==PDD_BOUND )return 0 ;if (unlikely (pdd->bound ==PDD_BOUND_SUSPENDED )) {@@ -284,7 +284,7 @@ static void kfd_unbind_processes_from_device(struct kfd_dev *kfd)*/ void kfd_iommu_suspend (struct kfd_dev *kfd) { -if (!kfd->device_info->needs_iommu_device) +if (!(kfd->device_info->needs_iommu_device && !ignore_crat))return ;kfd_unbind_processes_from_device (kfd);@@ -304,7 +304,7 @@ int kfd_iommu_resume(struct kfd_dev *kfd)unsigned int pasid_limit;int err; -if (!kfd->device_info->needs_iommu_device) +if (!(kfd->device_info->needs_iommu_device && !ignore_crat))return 0 ; pasid_limit =kfd_get_pasid_limit ();drivers/gpu/drm/amd/amdkfd/kfd_topology.c +3
-2@@ -545,7 +545,7 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr,* If the ASIC is APU except Kaveri, set local memory size * to 0 to disable local memory support */ -if (!dev->gpu->device_info->needs_iommu_device +if (!(dev->gpu->device_info->needs_iommu_device && !ignore_crat) || dev->gpu->device_info->asic_family ==CHIP_KAVERI ) {amdgpu_amdkfd_get_local_mem_info (dev->gpu->kgd, &local_mem_info);@@ -1090,6 +1090,7 @@ int kfd_topology_init(void)COMPUTE_UNIT_CPU ,NULL , proximity_domain); cpu_only_node =1 ; + ignore_crat =1 ;if (ret) {pr_err ("Error creating VCRAT table for CPU\n" );return ret;@@ -1197,7 +1198,7 @@ static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)/* Discrete GPUs need their own topology device list * entries. Don't assign them to CPU/APU nodes. */ -if (!gpu->device_info->needs_iommu_device && +if (!(gpu->device_info->needs_iommu_device && !ignore_crat) && dev->node_props.cpu_cores_count)continue ;
drivers/gpu/drm/amd/amdkfd/kfd_topology.c int kfd_topology_add_device(struct kfd_dev *gpu) · line 1459
Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +1
-1@@ -1452,7 +1453,7 @@ int kfd_topology_add_device(struct kfd_dev *gpu)* Overwrite ATS capability according to needs_iommu_device to fix * potential missing corresponding bit in CRAT of BIOS. */ -if (dev->gpu->device_info->needs_iommu_device) +if (dev->gpu->device_info->needs_iommu_device && !ignore_crat) dev->node_props.capability |=HSA_CAP_ATS_PRESENT ;else dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT ;
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Quoted 93 lines · Felix
Am 2020-08-18 um 9:09 a.m. schrieb Huang Rui:
Quoted 78 lines
We still have a few iommu issues which need to address, so force raven as "dgpu" path for the moment.
This is to add the fallback path to bypass IOMMU if IOMMU v2 is disabled or ACPI CRAT table not correct.
v2: use
ignore_cratparameter to decide whether it will go with IOMMUv2.Signed-off-by Huang Rui <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 5 +++- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 27 ++++++++++++++++++- drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c | 4 +-- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 2 +- .../amd/amdkfd/kfd_device_queue_manager_v9.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c | 4 +-- drivers/gpu/drm/amd/amdkfd/kfd_iommu.c | 10 +++---- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 7 ++--- 9 files changed, 46 insertions(+), 17 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4
-1@@ -684,11 +684,14 @@ MODULE_PARM_DESC(debug_largebar,* Ignore CRAT table during KFD initialization. By default, KFD uses the ACPI CRAT * table to get information about AMD APUs. This option can serve as a workaround on * systems with a broken CRAT table. +* +* Default is auto (according to asic type, iommu_v2, and crat table, to decide +* whehter use CRAT) */ int ignore_crat;module_param (ignore_crat, int,0444 );MODULE_PARM_DESC (ignore_crat, -"Ignore CRAT table during KFD initialization (0 = use CRAT (default), 1 = ignore CRAT)" ); +"Ignore CRAT table during KFD initialization (0 = auto (default), 1 = ignore CRAT)" );/** * DOC: halt_if_hws_hang (int) drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1
-1@@ -1258,7 +1258,7 @@ bool kfd_dev_is_large_bar(struct kfd_dev *dev)return true; } -if (dev->device_info->needs_iommu_device) +if (dev->device_info->needs_iommu_device && !ignore_crat)return false;amdgpu_amdkfd_get_local_mem_info (dev->kgd, &mem_info);drivers/gpu/drm/amd/amdkfd/kfd_crat.c +9
-0@@ -22,6 +22,7 @@#include <linux/pci.h> #include <linux/acpi.h> +#include <asm/processor.h> #include "kfd_crat.h" #include "kfd_priv.h" #include "kfd_topology.h" @@ -740,6 +741,29 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev,return 0 ; } + +#ifdef CONFIG_ACPI +static void kfd_setup_ignore_crat_option (void ) +{ + +if (ignore_crat) +return ; +drivers/gpu/drm/amd/amdkfd/kfd_crat.c static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev, · line 761
Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_crat.c +10
-0+#ifndef KFD_SUPPORT_IOMMU_V2 + ignore_crat =1 ; +#else + ignore_crat =0 ; +#endif + +/* Raven series will go with the fallback path to use virtual CRAT */ +if (boot_cpu_data.x86_vendor ==X86_VENDOR_AMD && + boot_cpu_data.x86 ==0x17 ) { + ignore_crat =1 ;Does your change require Thunk changes? If yes, then changing this by default breaks existing user mode and cannot be applied upstream.
Quoted 2 lines
If you expect that this is a temporary workaround, then don't apply it here, apply it in your grub config or modprobe config.
I see. Will keep the default using CRAT for Raven as well to make sure the existing thunk workable.
Thanks,
Ray
Quoted 186 lines
Regards,
FelixQuoted 181 lines
drivers/gpu/drm/amd/amdkfd/kfd_crat.c +7
-1+ } + +return ; +} +/* * kfd_create_crat_image_acpi - Allocates memory for CRAT image and * copies CRAT from ACPI (if available). @@ -751,7 +775,6 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev,* * Return 0 if successful else return error code */ -#ifdef CONFIG_ACPI int kfd_create_crat_image_acpi (void **crat_image,size_t *size) {struct acpi_table_header *crat_table;@@ -775,6 +798,8 @@ int kfd_create_crat_image_acpi(void **crat_image, size_t *size)return -EINVAL ; } +kfd_setup_ignore_crat_option (); +if (ignore_crat) {pr_info ("CRAT table disabled by module option\n" );return -ENODATA ;drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c +2
-2@@ -309,7 +309,7 @@ static int dbgdev_address_watch_nodiq(struct kfd_dbgdev *dbgdev,for (i =0 ; i < adw_info->num_watch_points; i++) {dbgdev_address_watch_set_registers (adw_info, &addrHi, &addrLo, &cntl, i, pdd->qpd.vmid, - dbgdev->dev->device_info->needs_iommu_device); + dbgdev->dev->device_info->needs_iommu_device && !ignore_crat);pr_debug ("\t\t%30s\n" ,"* * * * * * * * * * * * * * * * * *" );pr_debug ("\t\t%20s %08x\n" ,"register index :" , i);@@ -399,7 +399,7 @@ static int dbgdev_address_watch_diq(struct kfd_dbgdev *dbgdev,for (i =0 ; i < adw_info->num_watch_points; i++) {dbgdev_address_watch_set_registers (adw_info, &addrHi, &addrLo, &cntl, i, vmid, - dbgdev->dev->device_info->needs_iommu_device); + dbgdev->dev->device_info->needs_iommu_device && !ignore_crat);pr_debug ("\t\t%30s\n" ,"* * * * * * * * * * * * * * * * * *" );pr_debug ("\t\t%20s %08x\n" ,"register index :" , i);drivers/gpu/drm/amd/amdkfd/kfd_device.c +1
-1@@ -112,6 +112,7 @@ static const struct kfd_device_info carrizo_device_info = { .num_xgmi_sdma_engines = 0, .num_sdma_queues_per_engine = 2, }; +#endif static const struct kfd_device_info raven_device_info = { .asic_family =CHIP_RAVEN ,@@ -130,7 +131,6 @@ static const struct kfd_device_info raven_device_info = { .num_xgmi_sdma_engines = 0, .num_sdma_queues_per_engine = 2, }; -#endif static const struct kfd_device_info hawaii_device_info = { .asic_family =CHIP_HAWAII ,drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c +1
-1@@ -62,7 +62,7 @@ static int update_qpd_v9(struct device_queue_manager *dqm,SH_MEM_ALIGNMENT_MODE_UNALIGNED <<SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT ;if (amdgpu_noretry && - !dqm->dev->device_info->needs_iommu_device) + !(dqm->dev->device_info->needs_iommu_device && !ignore_crat)) qpd->sh_mem_config |=1 <<SH_MEM_CONFIG__RETRY_DISABLE__SHIFT ;drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c +2
-2@@ -321,7 +321,7 @@ static void kfd_init_apertures_vi(struct kfd_process_device *pdd, uint8_t id) pdd->lds_base =MAKE_LDS_APP_BASE_VI (); pdd->lds_limit =MAKE_LDS_APP_LIMIT (pdd->lds_base); -if (!pdd->dev->device_info->needs_iommu_device) { +if (!(pdd->dev->device_info->needs_iommu_device && !ignore_crat)) {/* dGPUs: SVM aperture starting at 0 * with small reserved space for kernel. * Set them to CANONICAL addresses. @@ -423,7 +423,7 @@ int kfd_init_apertures(struct kfd_process *process)return -EINVAL ; } -if (!dev->device_info->needs_iommu_device) { +if (!(dev->device_info->needs_iommu_device && !ignore_crat)) {/* dGPUs: the reserved space for kernel * before SVM */ drivers/gpu/drm/amd/amdkfd/kfd_iommu.c +5
-5@@ -41,7 +41,7 @@ int kfd_iommu_check_device(struct kfd_dev *kfd)struct amd_iommu_device_info iommu_info;int err; -if (!kfd->device_info->needs_iommu_device) +if (!(kfd->device_info->needs_iommu_device && !ignore_crat))return -ENODEV ; iommu_info.flags =0 ;@@ -63,7 +63,7 @@ int kfd_iommu_device_init(struct kfd_dev *kfd)unsigned int pasid_limit;int err; -if (!kfd->device_info->needs_iommu_device) +if (!(kfd->device_info->needs_iommu_device && !ignore_crat))return 0 ; iommu_info.flags =0 ;@@ -109,7 +109,7 @@ int kfd_iommu_bind_process_to_device(struct kfd_process_device *pdd)struct kfd_process *p = pdd->process;int err; -if (!dev->device_info->needs_iommu_device || pdd->bound ==PDD_BOUND ) +if (!(dev->device_info->needs_iommu_device && !ignore_crat) || pdd->bound ==PDD_BOUND )return 0 ;if (unlikely (pdd->bound ==PDD_BOUND_SUSPENDED )) {@@ -284,7 +284,7 @@ static void kfd_unbind_processes_from_device(struct kfd_dev *kfd)*/ void kfd_iommu_suspend (struct kfd_dev *kfd) { -if (!kfd->device_info->needs_iommu_device) +if (!(kfd->device_info->needs_iommu_device && !ignore_crat))return ;kfd_unbind_processes_from_device (kfd);@@ -304,7 +304,7 @@ int kfd_iommu_resume(struct kfd_dev *kfd)unsigned int pasid_limit;int err; -if (!kfd->device_info->needs_iommu_device) +if (!(kfd->device_info->needs_iommu_device && !ignore_crat))return 0 ; pasid_limit =kfd_get_pasid_limit ();drivers/gpu/drm/amd/amdkfd/kfd_topology.c +3
-2@@ -545,7 +545,7 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr,* If the ASIC is APU except Kaveri, set local memory size * to 0 to disable local memory support */ -if (!dev->gpu->device_info->needs_iommu_device +if (!(dev->gpu->device_info->needs_iommu_device && !ignore_crat) || dev->gpu->device_info->asic_family ==CHIP_KAVERI ) {amdgpu_amdkfd_get_local_mem_info (dev->gpu->kgd, &local_mem_info);@@ -1090,6 +1090,7 @@ int kfd_topology_init(void)COMPUTE_UNIT_CPU ,NULL , proximity_domain); cpu_only_node =1 ; + ignore_crat =1 ;if (ret) {pr_err ("Error creating VCRAT table for CPU\n" );return ret;@@ -1197,7 +1198,7 @@ static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)/* Discrete GPUs need their own topology device list * entries. Don't assign them to CPU/APU nodes. */ -if (!gpu->device_info->needs_iommu_device && +if (!(gpu->device_info->needs_iommu_device && !ignore_crat) && dev->node_props.cpu_cores_count)continue ;
Quoted 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_topology.c int kfd_topology_add_device(struct kfd_dev *gpu) · line 1459
Quoted 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +1
-1@@ -1452,7 +1453,7 @@ int kfd_topology_add_device(struct kfd_dev *gpu)* Overwrite ATS capability according to needs_iommu_device to fix * potential missing corresponding bit in CRAT of BIOS. */ -if (dev->gpu->device_info->needs_iommu_device) +if (dev->gpu->device_info->needs_iommu_device && !ignore_crat) dev->node_props.capability |=HSA_CAP_ATS_PRESENT ;else dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT ;
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx