drm/amdkfd: abstract the iommu device checking with ignore_crat (v2)
Subsystems (3)
PATCH v4 1/2 drm/amdkfd: abstract the iommu device checking with ignore_crat (v2)
It's better to use inline function to wrap the iommu checking.
v2: rename the function and use another input.
--- 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 | 5 +++++ drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 6 +++--- 7 files changed, 19 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) +if (kfd_device_use_iommu_v2 (dev))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); +kfd_device_use_iommu_v2 (dbgdev->dev));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); +kfd_device_use_iommu_v2 (dbgdev->dev));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) + !(kfd_device_use_iommu_v2 (dqm->dev))) 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 (!kfd_device_use_iommu_v2 (pdd->dev)) {/* 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 (!kfd_device_use_iommu_v2 (dev)) {/* 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_use_iommu_v2 (kfd))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_use_iommu_v2 (kfd))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 (!kfd_device_use_iommu_v2 (dev) || 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_use_iommu_v2 (kfd))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_use_iommu_v2 (kfd))return 0 ; pasid_limit =kfd_get_pasid_limit ();
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +5 -0
@@ -1232,6 +1232,11 @@ static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd)#endif } +static inline bool kfd_device_use_iommu_v2 (const struct kfd_dev *dev) +{ +return dev && dev->device_info->needs_iommu_device; +} +/* 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 +if (!kfd_device_use_iommu_v2 (dev->gpu) || dev->gpu->device_info->asic_family ==CHIP_KAVERI ) {amdgpu_amdkfd_get_local_mem_info (dev->gpu->kgd, &local_mem_info);
@@ -1197,7 +1197,7 @@ 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 (!kfd_device_use_iommu_v2 (gpu) && dev->node_props.cpu_cores_count)continue ;
@@ -1452,7 +1452,7 @@ 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 (kfd_device_use_iommu_v2 (dev->gpu)) 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 v4 2/2 drm/amdkfd: force raven as "dgpu" path (v4)
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.
v3: Align with existed thunk, don't change the way of raven, only renoir
will use "dgpu" path by default.
v4: don't update global ignore_crat in the driver, and revise fallback
function if CRAT is broken.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 5 ++++- drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 23 +++++++++++++++++++++-- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 5 ++++- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 10 ++++++++-- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 20 ++++++++++++++++++++ 5 files changed, 57 insertions(+), 6 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_crat.c +21 -2
@@ -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 +741,25 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev,return 0 ; } + +#ifdef CONFIG_ACPI + +bool kfd_ignore_crat (void ) +{ +bool ret; + +if (ignore_crat) +return true; + +#ifndef KFD_SUPPORT_IOMMU_V2 + ret = true; +#else + ret = false; +#endif + +return ret; +} +/* * kfd_create_crat_image_acpi - Allocates memory for CRAT image and * copies CRAT from ACPI (if available).
@@ -751,7 +771,6 @@ 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;
@@ -775,7 +794,7 @@ int kfd_create_crat_image_acpi(void **crat_image, size_t *size)return -EINVAL ; } -if (ignore_crat) { +if (kfd_ignore_crat ()) {pr_info ("CRAT table disabled by module option\n" );return -ENODATA ; }
drivers/gpu/drm/amd/amdkfd/kfd_device.c +4 -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 ,
@@ -688,6 +688,9 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, goto gws_error; } +/* If CRAT is broken, won't set iommu enabled */ +kfd_double_confirm_iommu_support (kfd); +if (kfd_iommu_device_init (kfd)) {dev_err (kfd_device,"Error initializing iommuv2\n" ); goto device_iommu_error;
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +8 -2
@@ -308,12 +308,14 @@ struct kfd_dev {/* xGMI */ uint64_t hive_id; -/* UUID */ uint64_t unique_id;bool pci_atomic_requested; +/* Use IOMMU v2 flag */ +bool use_iommu_v2; +/* SRAM ECC flag */ atomic_t sram_ecc_flag;
@@ -1009,6 +1011,7 @@ struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);struct kfd_dev *kfd_device_by_kgd (const struct kgd_dev *kgd);int kfd_topology_enum_kfd_devices (uint8_t idx,struct kfd_dev **kdev);int kfd_numa_node_to_apic_id (int numa_node_id); +void kfd_double_confirm_iommu_support (struct kfd_dev *gpu);/* Interrupts */ int kfd_interrupt_init (struct kfd_dev *dev);
@@ -1232,9 +1235,12 @@ static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd)#endif } +bool kfd_ignore_crat (void ); +static inline bool kfd_device_use_iommu_v2 (const struct kfd_dev *dev) { -return dev && dev->device_info->needs_iommu_device; +return !kfd_ignore_crat () && dev && dev->use_iommu_v2 && + dev->device_info->needs_iommu_device; }/* Debugfs */
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +20 -0
@@ -1579,6 +1579,26 @@ int kfd_numa_node_to_apic_id(int numa_node_id)return kfd_cpumask_to_apic_id (cpumask_of_node (numa_node_id)); } +void kfd_double_confirm_iommu_support (struct kfd_dev *gpu) +{ +struct kfd_topology_device *dev; + +unsigned temp =0 ; + +down_read (&topology_lock); + +/* The cpu_cores_count and simd_count aren't zero at the same time in +* APU node. +*/ +list_for_each_entry (dev, &topology_device_list, list) + temp |= dev->node_props.cpu_cores_count * + dev->node_props.simd_count; + +up_read (&topology_lock); + + gpu->use_iommu_v2 = temp ? true : false; +} +#if defined(CONFIG_DEBUG_FS )int kfd_debugfs_hqds_by_device (struct seq_file *m,void *data)
Signature · 6 lines
2.25.1 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Felix Kuehling
maintainerRe: [PATCH v4 2/2] drm/amdkfd: force raven as "dgpu" path (v4)
Am 2020-08-20 um 4:40 a.m. schrieb Huang Rui:
Quoted 155 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.
v3: Align with existed thunk, don't change the way of raven, only renoir
will use "dgpu" path by default.
v4: don't update globalignore_cratin the driver, and revise fallback
function if CRAT is broken.Signed-off-by Huang Rui <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 5 ++++- drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 23 +++++++++++++++++++++-- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 5 ++++- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 10 ++++++++-- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 20 ++++++++++++++++++++ 5 files changed, 57 insertions(+), 6 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_crat.c +21
-2@@ -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 +741,25 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev,return 0 ; } + +#ifdef CONFIG_ACPI + +bool kfd_ignore_crat (void ) +{ +bool ret; + +if (ignore_crat) +return true; + +#ifndef KFD_SUPPORT_IOMMU_V2 + ret = true; +#else + ret = false; +#endif + +return ret; +} +/* * kfd_create_crat_image_acpi - Allocates memory for CRAT image and * copies CRAT from ACPI (if available). @@ -751,7 +771,6 @@ 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;@@ -775,7 +794,7 @@ int kfd_create_crat_image_acpi(void **crat_image, size_t *size)return -EINVAL ; } -if (ignore_crat) { +if (kfd_ignore_crat ()) {pr_info ("CRAT table disabled by module option\n" );return -ENODATA ; }drivers/gpu/drm/amd/amdkfd/kfd_device.c +4
-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 ,@@ -688,6 +688,9 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, goto gws_error; } +/* If CRAT is broken, won't set iommu enabled */ +kfd_double_confirm_iommu_support (kfd); +if (kfd_iommu_device_init (kfd)) {dev_err (kfd_device,"Error initializing iommuv2\n" ); goto device_iommu_error;drivers/gpu/drm/amd/amdkfd/kfd_priv.h +4
-1@@ -308,12 +308,14 @@ struct kfd_dev {/* xGMI */ uint64_t hive_id; -/* UUID */ uint64_t unique_id;bool pci_atomic_requested; +/* Use IOMMU v2 flag */ +bool use_iommu_v2; +/* SRAM ECC flag */ atomic_t sram_ecc_flag;@@ -1009,6 +1011,7 @@ struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);struct kfd_dev *kfd_device_by_kgd (const struct kgd_dev *kgd);int kfd_topology_enum_kfd_devices (uint8_t idx,struct kfd_dev **kdev);int kfd_numa_node_to_apic_id (int numa_node_id); +void kfd_double_confirm_iommu_support (struct kfd_dev *gpu);/* Interrupts */ int kfd_interrupt_init (struct kfd_dev *dev);
drivers/gpu/drm/amd/amdkfd/kfd_priv.h static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd) · line 1243
Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +4
-1@@ -1232,9 +1235,12 @@ static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd)#endif } +bool kfd_ignore_crat (void ); +static inline bool kfd_device_use_iommu_v2 (const struct kfd_dev *dev) { -return dev && dev->device_info->needs_iommu_device; +return !kfd_ignore_crat () && dev && dev->use_iommu_v2 && + dev->device_info->needs_iommu_device;
I think this could now be simplified:
return dev && dev->use_iommu_v2;
So maybe you don't need this function any more.
Quoted 15 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +0
-0}/* Debugfs */ drivers/gpu/drm/amd/amdkfd/kfd_topology.c +4
-0@@ -1579,6 +1579,26 @@ int kfd_numa_node_to_apic_id(int numa_node_id)return kfd_cpumask_to_apic_id (cpumask_of_node (numa_node_id)); } +void kfd_double_confirm_iommu_support (struct kfd_dev *gpu) +{ +struct kfd_topology_device *dev; +
drivers/gpu/drm/amd/amdkfd/kfd_topology.c int kfd_numa_node_to_apic_id(int numa_node_id) · line 1595
Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +10
-0+unsigned temp =0 ; + +down_read (&topology_lock); + +/* The cpu_cores_count and simd_count aren't zero at the same time in +* APU node. +*/ +list_for_each_entry (dev, &topology_device_list, list) + temp |= dev->node_props.cpu_cores_count * + dev->node_props.simd_count;
You shouldn't look at all GPUs, only at the GPU currently being initialized. Otherwise all your dGPUs in an APU system will also have use_iommu_v2 == true, which would be confusing.
I'd do this in kfd_assign_gpu, because at that point you have access to the kfd_topology_device and the kfd_dev at the same time without having to add another loop.
...
list_for_each_entry(dev, &topology_device_list, list) {
/* Discrete GPUs need their own topology device list
* entries. Don't assign them to CPU/APU nodes.
*/
if (!gpu->device_info->needs_iommu_device &&
dev->node_props.cpu_cores_count)
continue;
if (!dev->gpu && (dev->node_props.simd_count > 0)) {
+ if (dev->node_props.cpu_cores_count)
+ dev->use_iommu_v2 = true;
...Regards,
Felix
drivers/gpu/drm/amd/amdkfd/kfd_topology.c int kfd_numa_node_to_apic_id(int numa_node_id) · line 1604
Quoted 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +6
-0+ +up_read (&topology_lock); + + gpu->use_iommu_v2 = temp ? true : false; +} +#if defined(CONFIG_DEBUG_FS )int kfd_debugfs_hqds_by_device (struct seq_file *m,void *data)
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH v4 2/2] drm/amdkfd: force raven as "dgpu" path (v4)
Quoted 170 lines · Felix
Am 2020-08-20 um 4:40 a.m. schrieb Huang Rui:
Quoted 155 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.
v3: Align with existed thunk, don't change the way of raven, only renoir
will use "dgpu" path by default.
v4: don't update globalignore_cratin the driver, and revise fallback
function if CRAT is broken.Signed-off-by Huang Rui <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 5 ++++- drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 23 +++++++++++++++++++++-- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 5 ++++- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 10 ++++++++-- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 20 ++++++++++++++++++++ 5 files changed, 57 insertions(+), 6 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_crat.c +21
-2@@ -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 +741,25 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev,return 0 ; } + +#ifdef CONFIG_ACPI + +bool kfd_ignore_crat (void ) +{ +bool ret; + +if (ignore_crat) +return true; + +#ifndef KFD_SUPPORT_IOMMU_V2 + ret = true; +#else + ret = false; +#endif + +return ret; +} +/* * kfd_create_crat_image_acpi - Allocates memory for CRAT image and * copies CRAT from ACPI (if available). @@ -751,7 +771,6 @@ 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;@@ -775,7 +794,7 @@ int kfd_create_crat_image_acpi(void **crat_image, size_t *size)return -EINVAL ; } -if (ignore_crat) { +if (kfd_ignore_crat ()) {pr_info ("CRAT table disabled by module option\n" );return -ENODATA ; }drivers/gpu/drm/amd/amdkfd/kfd_device.c +4
-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 ,@@ -688,6 +688,9 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, goto gws_error; } +/* If CRAT is broken, won't set iommu enabled */ +kfd_double_confirm_iommu_support (kfd); +if (kfd_iommu_device_init (kfd)) {dev_err (kfd_device,"Error initializing iommuv2\n" ); goto device_iommu_error;drivers/gpu/drm/amd/amdkfd/kfd_priv.h +4
-1@@ -308,12 +308,14 @@ struct kfd_dev {/* xGMI */ uint64_t hive_id; -/* UUID */ uint64_t unique_id;bool pci_atomic_requested; +/* Use IOMMU v2 flag */ +bool use_iommu_v2; +/* SRAM ECC flag */ atomic_t sram_ecc_flag;@@ -1009,6 +1011,7 @@ struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);struct kfd_dev *kfd_device_by_kgd (const struct kgd_dev *kgd);int kfd_topology_enum_kfd_devices (uint8_t idx,struct kfd_dev **kdev);int kfd_numa_node_to_apic_id (int numa_node_id); +void kfd_double_confirm_iommu_support (struct kfd_dev *gpu);/* Interrupts */ int kfd_interrupt_init (struct kfd_dev *dev);drivers/gpu/drm/amd/amdkfd/kfd_priv.h static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd) · line 1243
Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +4
-1@@ -1232,9 +1235,12 @@ static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd)#endif } +bool kfd_ignore_crat (void ); +static inline bool kfd_device_use_iommu_v2 (const struct kfd_dev *dev) { -return dev && dev->device_info->needs_iommu_device; +return !kfd_ignore_crat () && dev && dev->use_iommu_v2 && + dev->device_info->needs_iommu_device;I think this could now be simplified:
Quoted 3 lines
return dev && dev->use_iommu_v2;So maybe you don't need this function any more.
In Renoir, if ACPI CRAT from SBIOS is good, we may still use dev->device_info->needs_iommu_device to confirm whether we should go dGPU.
Quoted 32 lines
Quoted 17 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +0
-0}/* Debugfs */ drivers/gpu/drm/amd/amdkfd/kfd_topology.c +4
-0@@ -1579,6 +1579,26 @@ int kfd_numa_node_to_apic_id(int numa_node_id)return kfd_cpumask_to_apic_id (cpumask_of_node (numa_node_id)); } +void kfd_double_confirm_iommu_support (struct kfd_dev *gpu) +{ +struct kfd_topology_device *dev; +drivers/gpu/drm/amd/amdkfd/kfd_topology.c int kfd_numa_node_to_apic_id(int numa_node_id) · line 1595
Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +10
-0+unsigned temp =0 ; + +down_read (&topology_lock); + +/* The cpu_cores_count and simd_count aren't zero at the same time in +* APU node. +*/ +list_for_each_entry (dev, &topology_device_list, list) + temp |= dev->node_props.cpu_cores_count * + dev->node_props.simd_count;You shouldn't look at all GPUs, only at the GPU currently being initialized. Otherwise all your dGPUs in an APU system will also have
use_iommu_v2== true, which would be confusing.
Quoted 4 lines
I'd do this in
kfd_assign_gpu, because at that point you have access to thekfd_topology_deviceand thekfd_devat the same time without having to add another loop.
Actually, I follow your comment to do it like this, however, we have to set the use_iommu_v2 before kfd_iommu_device_init(). kfd_assign_gpu in kfd_topology_add_device() is a little late.
Thanks,
Ray
Quoted 26 lines
... list_for_each_entry(dev, &topology_device_list, list) { /* Discrete GPUs need their own topology device list * entries. Don't assign them to CPU/APU nodes. */ if (!gpu->device_info->needs_iommu_device && dev->node_props.cpu_cores_count) continue; if (!dev->gpu && (dev->node_props.simd_count > 0)) { + if (dev->node_props.cpu_cores_count) + dev->use_iommu_v2 = true; ...Regards,
FelixQuoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +6
-0+ +up_read (&topology_lock); + + gpu->use_iommu_v2 = temp ? true : false; +} +#if defined(CONFIG_DEBUG_FS )
Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_topology.c int kfd_numa_node_to_apic_id(int numa_node_id) · line 1604
Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +0
-0int kfd_debugfs_hqds_by_device (struct seq_file *m,void *data)
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Felix Kuehling
maintainerRe: [PATCH v4 2/2] drm/amdkfd: force raven as "dgpu" path (v4)
Am 2020-08-20 um 11:53 p.m. schrieb Huang Rui:
Quoted 169 lines · Felix
Quoted 168 lines
Am 2020-08-20 um 4:40 a.m. schrieb Huang Rui:
Quoted 155 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.
v3: Align with existed thunk, don't change the way of raven, only renoir
will use "dgpu" path by default.
v4: don't update globalignore_cratin the driver, and revise fallback
function if CRAT is broken.Signed-off-by Huang Rui <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 5 ++++- drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 23 +++++++++++++++++++++-- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 5 ++++- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 10 ++++++++-- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 20 ++++++++++++++++++++ 5 files changed, 57 insertions(+), 6 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_crat.c +21
-2@@ -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 +741,25 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev,return 0 ; } + +#ifdef CONFIG_ACPI + +bool kfd_ignore_crat (void ) +{ +bool ret; + +if (ignore_crat) +return true; + +#ifndef KFD_SUPPORT_IOMMU_V2 + ret = true; +#else + ret = false; +#endif + +return ret; +} +/* * kfd_create_crat_image_acpi - Allocates memory for CRAT image and * copies CRAT from ACPI (if available). @@ -751,7 +771,6 @@ 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;@@ -775,7 +794,7 @@ int kfd_create_crat_image_acpi(void **crat_image, size_t *size)return -EINVAL ; } -if (ignore_crat) { +if (kfd_ignore_crat ()) {pr_info ("CRAT table disabled by module option\n" );return -ENODATA ; }drivers/gpu/drm/amd/amdkfd/kfd_device.c +4
-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 ,@@ -688,6 +688,9 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, goto gws_error; } +/* If CRAT is broken, won't set iommu enabled */ +kfd_double_confirm_iommu_support (kfd); +if (kfd_iommu_device_init (kfd)) {dev_err (kfd_device,"Error initializing iommuv2\n" ); goto device_iommu_error;drivers/gpu/drm/amd/amdkfd/kfd_priv.h +4
-1@@ -308,12 +308,14 @@ struct kfd_dev {/* xGMI */ uint64_t hive_id; -/* UUID */ uint64_t unique_id;bool pci_atomic_requested; +/* Use IOMMU v2 flag */ +bool use_iommu_v2; +/* SRAM ECC flag */ atomic_t sram_ecc_flag;@@ -1009,6 +1011,7 @@ struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);struct kfd_dev *kfd_device_by_kgd (const struct kgd_dev *kgd);int kfd_topology_enum_kfd_devices (uint8_t idx,struct kfd_dev **kdev);int kfd_numa_node_to_apic_id (int numa_node_id); +void kfd_double_confirm_iommu_support (struct kfd_dev *gpu);/* Interrupts */ int kfd_interrupt_init (struct kfd_dev *dev);drivers/gpu/drm/amd/amdkfd/kfd_priv.h static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd) · line 1243
Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +4
-1@@ -1232,9 +1235,12 @@ static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd)#endif } +bool kfd_ignore_crat (void ); +static inline bool kfd_device_use_iommu_v2 (const struct kfd_dev *dev) { -return dev && dev->device_info->needs_iommu_device; +return !kfd_ignore_crat () && dev && dev->use_iommu_v2 && + dev->device_info->needs_iommu_device;I think this could now be simplified:
Quoted 5 lines
Quoted 3 lines
return dev && dev->use_iommu_v2;So maybe you don't need this function any more.
In Renoir, if ACPI CRAT from SBIOS is good, we may still use
dev->device_info->needs_iommu_deviceto confirm whether we should go dGPU.
That should be reflected correctly in the dev->use_iommu_v2 flag. dev->use_iommu_v2 should not be set on any GPU that doesn't have dev->device_info->needs_iommu_device. See below.
Quoted 31 lines
Quoted 31 lines
Quoted 17 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +0
-0}/* Debugfs */ drivers/gpu/drm/amd/amdkfd/kfd_topology.c +4
-0@@ -1579,6 +1579,26 @@ int kfd_numa_node_to_apic_id(int numa_node_id)return kfd_cpumask_to_apic_id (cpumask_of_node (numa_node_id)); } +void kfd_double_confirm_iommu_support (struct kfd_dev *gpu) +{ +struct kfd_topology_device *dev; +drivers/gpu/drm/amd/amdkfd/kfd_topology.c int kfd_numa_node_to_apic_id(int numa_node_id) · line 1595
Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +10
-0+unsigned temp =0 ; + +down_read (&topology_lock); + +/* The cpu_cores_count and simd_count aren't zero at the same time in +* APU node. +*/ +list_for_each_entry (dev, &topology_device_list, list) + temp |= dev->node_props.cpu_cores_count * + dev->node_props.simd_count;You shouldn't look at all GPUs, only at the GPU currently being initialized. Otherwise all your dGPUs in an APU system will also have
use_iommu_v2== true, which would be confusing.
Quoted 7 lines
Quoted 4 lines
I'd do this in
kfd_assign_gpu, because at that point you have access to thekfd_topology_deviceand thekfd_devat the same time without having to add another loop.Actually, I follow your comment to do it like this, however, we have to set the
use_iommu_v2beforekfd_iommu_device_init().kfd_assign_gpuinkfd_topology_add_device()is a little late.
I see, I missed that. But we still should make sure we set gpu->use_iommu_v2 only on devices that have gpu->device_info->needs_iommu_device, and only if this GPU can be assigned to an APU node in the topology. So then kfd_double_confirm_iommu_support would look like this:
+ if (!gpu->device_info->needs_iommu_device)
+ return;
+ down_read(&topology_lock);
+ /* Only use IOMMUv2 if there is an APU topology node with no GPU assigned yet.
+ * This GPU will be assigned to it.
+ */
+ list_for_each_entry(dev, &topology_device_list, list)
+ if (dev->node_props.cpu_cores_count &&
+ dev->node_props.simd_count &&
+ !dev->gpu) {
+ gpu->use_iommu_v2 = true;
+ break;
+ }
+ up_read(&topology_lock);
Regards,
Felix
Quoted 30 lines
Thanks,
RayQuoted 26 lines
... list_for_each_entry(dev, &topology_device_list, list) { /* Discrete GPUs need their own topology device list * entries. Don't assign them to CPU/APU nodes. */ if (!gpu->device_info->needs_iommu_device && dev->node_props.cpu_cores_count) continue; if (!dev->gpu && (dev->node_props.simd_count > 0)) { + if (dev->node_props.cpu_cores_count) + dev->use_iommu_v2 = true; ...Regards,
FelixQuoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +6
-0+ +up_read (&topology_lock); + + gpu->use_iommu_v2 = temp ? true : false; +} +#if defined(CONFIG_DEBUG_FS )
Quoted 1 line
Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_topology.c int kfd_numa_node_to_apic_id(int numa_node_id) · line 1604
Quoted 1 line
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +0
-0int kfd_debugfs_hqds_by_device (struct seq_file *m,void *data)
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx