[PATCH 09/83] hsa/radeon: Add code base of hsa driver for AMD's GPUs
Jerome Glisse
j.glisse at gmail.com
Fri Jul 11 10:04:12 PDT 2014
On Fri, Jul 11, 2014 at 12:50:09AM +0300, Oded Gabbay wrote:
> This patch adds the code base of the hsa driver for
> AMD's GPUs.
>
> This driver is called kfd.
>
> This initial version supports the first HSA chip, Kaveri.
>
> This driver is located in a new directory structure under drivers/gpu.
>
> Signed-off-by: Oded Gabbay <oded.gabbay at amd.com>
There is too coding style issues. While we have been lax on the enforcing the
scripts/checkpatch.pl rules i think there is a limit to that. I am not strict
on the 80chars per line but others things needs fixing so we stay inline.
Also i am a bit worried about the license, given top comment in each of the
files i am not sure this is GPL2 compatible. I would need to ask lawyer to
review that.
Others comment inline.
> ---
> drivers/Kconfig | 2 +
> drivers/gpu/Makefile | 1 +
> drivers/gpu/hsa/Kconfig | 20 +
> drivers/gpu/hsa/Makefile | 1 +
> drivers/gpu/hsa/radeon/Makefile | 8 +
> drivers/gpu/hsa/radeon/kfd_chardev.c | 133 ++++
> drivers/gpu/hsa/radeon/kfd_crat.h | 292 ++++++++
> drivers/gpu/hsa/radeon/kfd_device.c | 162 +++++
> drivers/gpu/hsa/radeon/kfd_module.c | 117 ++++
> drivers/gpu/hsa/radeon/kfd_pasid.c | 92 +++
> drivers/gpu/hsa/radeon/kfd_priv.h | 232 ++++++
> drivers/gpu/hsa/radeon/kfd_process.c | 400 +++++++++++
> drivers/gpu/hsa/radeon/kfd_scheduler.h | 62 ++
> drivers/gpu/hsa/radeon/kfd_topology.c | 1201 ++++++++++++++++++++++++++++++++
> drivers/gpu/hsa/radeon/kfd_topology.h | 168 +++++
> 15 files changed, 2891 insertions(+)
> create mode 100644 drivers/gpu/hsa/Kconfig
> create mode 100644 drivers/gpu/hsa/Makefile
> create mode 100644 drivers/gpu/hsa/radeon/Makefile
> create mode 100644 drivers/gpu/hsa/radeon/kfd_chardev.c
> create mode 100644 drivers/gpu/hsa/radeon/kfd_crat.h
> create mode 100644 drivers/gpu/hsa/radeon/kfd_device.c
> create mode 100644 drivers/gpu/hsa/radeon/kfd_module.c
> create mode 100644 drivers/gpu/hsa/radeon/kfd_pasid.c
> create mode 100644 drivers/gpu/hsa/radeon/kfd_priv.h
> create mode 100644 drivers/gpu/hsa/radeon/kfd_process.c
> create mode 100644 drivers/gpu/hsa/radeon/kfd_scheduler.h
> create mode 100644 drivers/gpu/hsa/radeon/kfd_topology.c
> create mode 100644 drivers/gpu/hsa/radeon/kfd_topology.h
>
> diff --git a/drivers/Kconfig b/drivers/Kconfig
> index 9b2dcc2..c1ac8f8 100644
> --- a/drivers/Kconfig
> +++ b/drivers/Kconfig
> @@ -178,4 +178,6 @@ source "drivers/mcb/Kconfig"
>
> source "drivers/thunderbolt/Kconfig"
>
> +source "drivers/gpu/hsa/Kconfig"
> +
> endmenu
> diff --git a/drivers/gpu/Makefile b/drivers/gpu/Makefile
> index 70da9eb..749a7ea 100644
> --- a/drivers/gpu/Makefile
> +++ b/drivers/gpu/Makefile
> @@ -1,3 +1,4 @@
> obj-y += drm/ vga/
> obj-$(CONFIG_TEGRA_HOST1X) += host1x/
> obj-$(CONFIG_IMX_IPUV3_CORE) += ipu-v3/
> +obj-$(CONFIG_HSA) += hsa/
> \ No newline at end of file
> diff --git a/drivers/gpu/hsa/Kconfig b/drivers/gpu/hsa/Kconfig
> new file mode 100644
> index 0000000..ee7bb28
> --- /dev/null
> +++ b/drivers/gpu/hsa/Kconfig
> @@ -0,0 +1,20 @@
> +#
> +# Heterogenous system architecture configuration
> +#
> +
> +menuconfig HSA
> + bool "Heterogenous System Architecture"
> + default y
> + help
> + Say Y here if you want Heterogenous System Architecture support.
Maybe a bit more chatty here, there is already enough kernel option that
are cryptic even to kernel developer. Not everyone is well aware of all
the fence 3 letter accronym GPU uses :)
> +
> +if HSA
> +
> +config HSA_RADEON
> + tristate "HSA kernel driver for AMD Radeon devices"
> + depends on HSA && AMD_IOMMU_V2 && X86_64
> + default m
> + help
> + Enable this if you want to support HSA on AMD Radeon devices.
> +
> +endif # HSA
> diff --git a/drivers/gpu/hsa/Makefile b/drivers/gpu/hsa/Makefile
> new file mode 100644
> index 0000000..0951584
> --- /dev/null
> +++ b/drivers/gpu/hsa/Makefile
> @@ -0,0 +1 @@
> +obj-$(CONFIG_HSA_RADEON) += radeon/
> diff --git a/drivers/gpu/hsa/radeon/Makefile b/drivers/gpu/hsa/radeon/Makefile
> new file mode 100644
> index 0000000..ba16a09
> --- /dev/null
> +++ b/drivers/gpu/hsa/radeon/Makefile
> @@ -0,0 +1,8 @@
> +#
> +# Makefile for Heterogenous System Architecture support for AMD Radeon devices
> +#
> +
> +radeon_kfd-y := kfd_module.o kfd_device.o kfd_chardev.o \
> + kfd_pasid.o kfd_topology.o kfd_process.o
> +
> +obj-$(CONFIG_HSA_RADEON) += radeon_kfd.o
> diff --git a/drivers/gpu/hsa/radeon/kfd_chardev.c b/drivers/gpu/hsa/radeon/kfd_chardev.c
> new file mode 100644
> index 0000000..7a56a8f
> --- /dev/null
> +++ b/drivers/gpu/hsa/radeon/kfd_chardev.c
> @@ -0,0 +1,133 @@
> +/*
> + * Copyright 2014 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#include <linux/device.h>
> +#include <linux/export.h>
> +#include <linux/err.h>
> +#include <linux/fs.h>
> +#include <linux/sched.h>
> +#include <linux/slab.h>
> +#include <linux/uaccess.h>
> +#include "kfd_priv.h"
> +#include "kfd_scheduler.h"
> +
> +static long kfd_ioctl(struct file *, unsigned int, unsigned long);
Nitpick, avoid unsigned int just use unsigned.
> +static int kfd_open(struct inode *, struct file *);
> +
> +static const char kfd_dev_name[] = "kfd";
> +
> +static const struct file_operations kfd_fops = {
> + .owner = THIS_MODULE,
> + .unlocked_ioctl = kfd_ioctl,
> + .open = kfd_open,
> +};
> +
> +static int kfd_char_dev_major = -1;
> +static struct class *kfd_class;
> +struct device *kfd_device;
> +
> +int
> +radeon_kfd_chardev_init(void)
> +{
> + int err = 0;
> +
> + kfd_char_dev_major = register_chrdev(0, kfd_dev_name, &kfd_fops);
> + err = kfd_char_dev_major;
> + if (err < 0)
> + goto err_register_chrdev;
> +
> + kfd_class = class_create(THIS_MODULE, kfd_dev_name);
> + err = PTR_ERR(kfd_class);
> + if (IS_ERR(kfd_class))
> + goto err_class_create;
> +
> + kfd_device = device_create(kfd_class, NULL, MKDEV(kfd_char_dev_major, 0), NULL, kfd_dev_name);
> + err = PTR_ERR(kfd_device);
> + if (IS_ERR(kfd_device))
> + goto err_device_create;
> +
> + return 0;
> +
> +err_device_create:
> + class_destroy(kfd_class);
> +err_class_create:
> + unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
> +err_register_chrdev:
> + return err;
> +}
> +
> +void
> +radeon_kfd_chardev_exit(void)
> +{
> + device_destroy(kfd_class, MKDEV(kfd_char_dev_major, 0));
> + class_destroy(kfd_class);
> + unregister_chrdev(kfd_char_dev_major, kfd_dev_name);
> +}
> +
> +struct device*
> +radeon_kfd_chardev(void)
> +{
> + return kfd_device;
> +}
> +
> +
> +static int
> +kfd_open(struct inode *inode, struct file *filep)
> +{
> + struct kfd_process *process;
> +
> + if (iminor(inode) != 0)
> + return -ENODEV;
> +
> + process = radeon_kfd_create_process(current);
> + if (IS_ERR(process))
> + return PTR_ERR(process);
> +
> + pr_debug("\nkfd: process %d opened dev/kfd", process->pasid);
> +
> + return 0;
> +}
> +
> +
> +static long
> +kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
> +{
> + long err = -EINVAL;
> +
> + dev_info(kfd_device,
> + "ioctl cmd 0x%x (#%d), arg 0x%lx\n",
> + cmd, _IOC_NR(cmd), arg);
> +
> + switch (cmd) {
> + default:
> + dev_err(kfd_device,
> + "unknown ioctl cmd 0x%x, arg 0x%lx)\n",
> + cmd, arg);
> + err = -EINVAL;
> + break;
> + }
> +
> + if (err < 0)
> + dev_err(kfd_device, "ioctl error %ld\n", err);
> +
> + return err;
> +}
> diff --git a/drivers/gpu/hsa/radeon/kfd_crat.h b/drivers/gpu/hsa/radeon/kfd_crat.h
> new file mode 100644
> index 0000000..587455d
> --- /dev/null
> +++ b/drivers/gpu/hsa/radeon/kfd_crat.h
> @@ -0,0 +1,292 @@
> +/*
> + * Copyright 2014 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#ifndef KFD_CRAT_H_INCLUDED
> +#define KFD_CRAT_H_INCLUDED
> +
> +#include <linux/types.h>
> +
> +#pragma pack(1)
> +
> +/*
> + * 4CC signature values for the CRAT and CDIT ACPI tables
> + */
> +
> +#define CRAT_SIGNATURE "CRAT"
> +#define CDIT_SIGNATURE "CDIT"
> +
> +/*
> + * Component Resource Association Table (CRAT)
> + */
> +
> +#define CRAT_OEMID_LENGTH 6
> +#define CRAT_OEMTABLEID_LENGTH 8
> +#define CRAT_RESERVED_LENGTH 6
> +
> +struct crat_header {
> + uint32_t signature;
> + uint32_t length;
> + uint8_t revision;
> + uint8_t checksum;
> + uint8_t oem_id[CRAT_OEMID_LENGTH];
> + uint8_t oem_table_id[CRAT_OEMTABLEID_LENGTH];
> + uint32_t oem_revision;
> + uint32_t creator_id;
> + uint32_t creator_revision;
> + uint32_t total_entries;
> + uint16_t num_domains;
> + uint8_t reserved[CRAT_RESERVED_LENGTH];
> +};
> +
> +/*
> + * The header structure is immediately followed by total_entries of the
> + * data definitions
> + */
> +
> +/*
> + * The currently defined subtype entries in the CRAT
> + */
> +#define CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY 0
> +#define CRAT_SUBTYPE_MEMORY_AFFINITY 1
> +#define CRAT_SUBTYPE_CACHE_AFFINITY 2
> +#define CRAT_SUBTYPE_TLB_AFFINITY 3
> +#define CRAT_SUBTYPE_CCOMPUTE_AFFINITY 4
> +#define CRAT_SUBTYPE_IOLINK_AFFINITY 5
> +#define CRAT_SUBTYPE_MAX 6
> +
> +#define CRAT_SIBLINGMAP_SIZE 32
> +
> +/*
> + * ComputeUnit Affinity structure and definitions
> + */
> +#define CRAT_CU_FLAGS_ENABLED 0x00000001
> +#define CRAT_CU_FLAGS_HOT_PLUGGABLE 0x00000002
> +#define CRAT_CU_FLAGS_CPU_PRESENT 0x00000004
> +#define CRAT_CU_FLAGS_GPU_PRESENT 0x00000008
> +#define CRAT_CU_FLAGS_IOMMU_PRESENT 0x00000010
> +#define CRAT_CU_FLAGS_RESERVED 0xffffffe0
> +
> +#define CRAT_COMPUTEUNIT_RESERVED_LENGTH 4
> +
> +struct crat_subtype_computeunit {
> + uint8_t type;
> + uint8_t length;
> + uint16_t reserved;
> + uint32_t flags;
> + uint32_t proximity_domain;
> + uint32_t processor_id_low;
> + uint16_t num_cpu_cores;
> + uint16_t num_simd_cores;
> + uint16_t max_waves_simd;
> + uint16_t io_count;
> + uint16_t hsa_capability;
> + uint16_t lds_size_in_kb;
> + uint8_t wave_front_size;
> + uint8_t num_banks;
> + uint16_t micro_engine_id;
> + uint8_t num_arrays;
> + uint8_t num_cu_per_array;
> + uint8_t num_simd_per_cu;
> + uint8_t max_slots_scatch_cu;
> + uint8_t reserved2[CRAT_COMPUTEUNIT_RESERVED_LENGTH];
> +};
> +
> +/*
> + * HSA Memory Affinity structure and definitions
> + */
> +#define CRAT_MEM_FLAGS_ENABLED 0x00000001
> +#define CRAT_MEM_FLAGS_HOT_PLUGGABLE 0x00000002
> +#define CRAT_MEM_FLAGS_NON_VOLATILE 0x00000004
> +#define CRAT_MEM_FLAGS_RESERVED 0xfffffff8
> +
> +#define CRAT_MEMORY_RESERVED_LENGTH 8
> +
> +struct crat_subtype_memory {
> + uint8_t type;
> + uint8_t length;
> + uint16_t reserved;
> + uint32_t flags;
> + uint32_t promixity_domain;
> + uint32_t base_addr_low;
> + uint32_t base_addr_high;
> + uint32_t length_low;
> + uint32_t length_high;
> + uint32_t width;
> + uint8_t reserved2[CRAT_MEMORY_RESERVED_LENGTH];
> +};
> +
> +/*
> + * HSA Cache Affinity structure and definitions
> + */
> +#define CRAT_CACHE_FLAGS_ENABLED 0x00000001
> +#define CRAT_CACHE_FLAGS_DATA_CACHE 0x00000002
> +#define CRAT_CACHE_FLAGS_INST_CACHE 0x00000004
> +#define CRAT_CACHE_FLAGS_CPU_CACHE 0x00000008
> +#define CRAT_CACHE_FLAGS_SIMD_CACHE 0x00000010
> +#define CRAT_CACHE_FLAGS_RESERVED 0xffffffe0
> +
> +#define CRAT_CACHE_RESERVED_LENGTH 8
> +
> +struct crat_subtype_cache {
> + uint8_t type;
> + uint8_t length;
> + uint16_t reserved;
> + uint32_t flags;
> + uint32_t processor_id_low;
> + uint8_t sibling_map[CRAT_SIBLINGMAP_SIZE];
> + uint32_t cache_size;
> + uint8_t cache_level;
> + uint8_t lines_per_tag;
> + uint16_t cache_line_size;
> + uint8_t associativity;
> + uint8_t cache_properties;
> + uint16_t cache_latency;
> + uint8_t reserved2[CRAT_CACHE_RESERVED_LENGTH];
> +};
> +
> +/*
> + * HSA TLB Affinity structure and definitions
> + */
> +#define CRAT_TLB_FLAGS_ENABLED 0x00000001
> +#define CRAT_TLB_FLAGS_DATA_TLB 0x00000002
> +#define CRAT_TLB_FLAGS_INST_TLB 0x00000004
> +#define CRAT_TLB_FLAGS_CPU_TLB 0x00000008
> +#define CRAT_TLB_FLAGS_SIMD_TLB 0x00000010
> +#define CRAT_TLB_FLAGS_RESERVED 0xffffffe0
> +
> +#define CRAT_TLB_RESERVED_LENGTH 4
> +
> +struct crat_subtype_tlb {
> + uint8_t type;
> + uint8_t length;
> + uint16_t reserved;
> + uint32_t flags;
> + uint32_t processor_id_low;
> + uint8_t sibling_map[CRAT_SIBLINGMAP_SIZE];
> + uint32_t tlb_level;
> + uint8_t data_tlb_associativity_2mb;
> + uint8_t data_tlb_size_2mb;
> + uint8_t instruction_tlb_associativity_2mb;
> + uint8_t instruction_tlb_size_2mb;
> + uint8_t data_tlb_associativity_4k;
> + uint8_t data_tlb_size_4k;
> + uint8_t instruction_tlb_associativity_4k;
> + uint8_t instruction_tlb_size_4k;
> + uint8_t data_tlb_associativity_1gb;
> + uint8_t data_tlb_size_1gb;
> + uint8_t instruction_tlb_associativity_1gb;
> + uint8_t instruction_tlb_size_1gb;
> + uint8_t reserved2[CRAT_TLB_RESERVED_LENGTH];
> +};
> +
> +/*
> + * HSA CCompute/APU Affinity structure and definitions
> + */
> +#define CRAT_CCOMPUTE_FLAGS_ENABLED 0x00000001
> +#define CRAT_CCOMPUTE_FLAGS_RESERVED 0xfffffffe
> +
> +#define CRAT_CCOMPUTE_RESERVED_LENGTH 16
> +
> +struct crat_subtype_ccompute {
> + uint8_t type;
> + uint8_t length;
> + uint16_t reserved;
> + uint32_t flags;
> + uint32_t processor_id_low;
> + uint8_t sibling_map[CRAT_SIBLINGMAP_SIZE];
> + uint32_t apu_size;
> + uint8_t reserved2[CRAT_CCOMPUTE_RESERVED_LENGTH];
> +};
> +
> +/*
> + * HSA IO Link Affinity structure and definitions
> + */
> +#define CRAT_IOLINK_FLAGS_ENABLED 0x00000001
> +#define CRAT_IOLINK_FLAGS_COHERENCY 0x00000002
> +#define CRAT_IOLINK_FLAGS_RESERVED 0xfffffffc
> +
> +/*
> + * IO interface types
> + */
> +#define CRAT_IOLINK_TYPE_UNDEFINED 0
> +#define CRAT_IOLINK_TYPE_HYPERTRANSPORT 1
> +#define CRAT_IOLINK_TYPE_PCIEXPRESS 2
> +#define CRAT_IOLINK_TYPE_OTHER 3
> +#define CRAT_IOLINK_TYPE_MAX 255
> +
> +#define CRAT_IOLINK_RESERVED_LENGTH 24
> +
> +struct crat_subtype_iolink {
> + uint8_t type;
> + uint8_t length;
> + uint16_t reserved;
> + uint32_t flags;
> + uint32_t proximity_domain_from;
> + uint32_t proximity_domain_to;
> + uint8_t io_interface_type;
> + uint8_t version_major;
> + uint16_t version_minor;
> + uint32_t minimum_latency;
> + uint32_t maximum_latency;
> + uint32_t minimum_bandwidth_mbs;
> + uint32_t maximum_bandwidth_mbs;
> + uint32_t recommended_transfer_size;
> + uint8_t reserved2[CRAT_IOLINK_RESERVED_LENGTH];
> +};
> +
> +/*
> + * HSA generic sub-type header
> + */
> +
> +#define CRAT_SUBTYPE_FLAGS_ENABLED 0x00000001
> +
> +struct crat_subtype_generic {
> + uint8_t type;
> + uint8_t length;
> + uint16_t reserved;
> + uint32_t flags;
> +};
> +
> +/*
> + * Component Locality Distance Information Table (CDIT)
> + */
> +#define CDIT_OEMID_LENGTH 6
> +#define CDIT_OEMTABLEID_LENGTH 8
> +
> +struct cdit_header {
> + uint32_t signature;
> + uint32_t length;
> + uint8_t revision;
> + uint8_t checksum;
> + uint8_t oem_id[CDIT_OEMID_LENGTH];
> + uint8_t oem_table_id[CDIT_OEMTABLEID_LENGTH];
> + uint32_t oem_revision;
> + uint32_t creator_id;
> + uint32_t creator_revision;
> + uint32_t total_entries;
> + uint16_t num_domains;
> + uint8_t entry[1];
> +};
> +
> +#pragma pack()
> +
> +#endif /* KFD_CRAT_H_INCLUDED */
> diff --git a/drivers/gpu/hsa/radeon/kfd_device.c b/drivers/gpu/hsa/radeon/kfd_device.c
> new file mode 100644
> index 0000000..d122920
> --- /dev/null
> +++ b/drivers/gpu/hsa/radeon/kfd_device.c
> @@ -0,0 +1,162 @@
> +/*
> + * Copyright 2014 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#include <linux/amd-iommu.h>
> +#include <linux/bsearch.h>
> +#include <linux/pci.h>
> +#include <linux/slab.h>
> +#include "kfd_priv.h"
> +#include "kfd_scheduler.h"
> +
> +static const struct kfd_device_info bonaire_device_info = {
> + .max_pasid_bits = 16,
> +};
> +
> +struct kfd_deviceid {
> + unsigned short did;
> + const struct kfd_device_info *device_info;
> +};
> +
> +/* Please keep this sorted by increasing device id. */
> +static const struct kfd_deviceid supported_devices[] = {
> + { 0x1305, &bonaire_device_info }, /* Kaveri */
> + { 0x1307, &bonaire_device_info }, /* Kaveri */
> + { 0x130F, &bonaire_device_info }, /* Kaveri */
> + { 0x665C, &bonaire_device_info }, /* Bonaire */
> +};
> +
> +static const struct kfd_device_info *
> +lookup_device_info(unsigned short did)
> +{
> + size_t i;
> +
> + for (i = 0; i < ARRAY_SIZE(supported_devices); i++) {
> + if (supported_devices[i].did == did) {
> + BUG_ON(supported_devices[i].device_info == NULL);
> + return supported_devices[i].device_info;
> + }
> + }
> +
> + return NULL;
> +}
> +
> +struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev)
> +{
> + struct kfd_dev *kfd;
> +
> + const struct kfd_device_info *device_info = lookup_device_info(pdev->device);
> +
> + if (!device_info)
> + return NULL;
> +
> + kfd = kzalloc(sizeof(*kfd), GFP_KERNEL);
> + kfd->kgd = kgd;
> + kfd->device_info = device_info;
> + kfd->pdev = pdev;
> +
> + return kfd;
> +}
> +
> +static bool
> +device_iommu_pasid_init(struct kfd_dev *kfd)
> +{
> + const u32 required_iommu_flags = AMD_IOMMU_DEVICE_FLAG_ATS_SUP | AMD_IOMMU_DEVICE_FLAG_PRI_SUP
> + | AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
> +
> + struct amd_iommu_device_info iommu_info;
> + pasid_t pasid_limit;
> + int err;
> +
> + err = amd_iommu_device_info(kfd->pdev, &iommu_info);
> + if (err < 0)
> + return false;
> +
> + if ((iommu_info.flags & required_iommu_flags) != required_iommu_flags)
> + return false;
> +
> + pasid_limit = min_t(pasid_t, (pasid_t)1 << kfd->device_info->max_pasid_bits, iommu_info.max_pasids);
> + pasid_limit = min_t(pasid_t, pasid_limit, kfd->doorbell_process_limit);
> +
> + err = amd_iommu_init_device(kfd->pdev, pasid_limit);
> + if (err < 0)
> + return false;
> +
> + if (!radeon_kfd_set_pasid_limit(pasid_limit)) {
> + amd_iommu_free_device(kfd->pdev);
> + return false;
> + }
> +
> + return true;
> +}
> +
> +static void iommu_pasid_shutdown_callback(struct pci_dev *pdev, int pasid)
> +{
> + struct kfd_dev *dev = radeon_kfd_device_by_pci_dev(pdev);
> +
> + if (dev)
> + radeon_kfd_unbind_process_from_device(dev, pasid);
> +}
> +
> +bool kgd2kfd_device_init(struct kfd_dev *kfd,
> + const struct kgd2kfd_shared_resources *gpu_resources)
> +{
> + kfd->shared_resources = *gpu_resources;
> +
> + kfd->regs = gpu_resources->mmio_registers;
> +
> + if (!device_iommu_pasid_init(kfd))
> + return false;
> +
> + if (kfd_topology_add_device(kfd) != 0) {
> + amd_iommu_free_device(kfd->pdev);
> + return false;
> + }
> +
> + amd_iommu_set_invalidate_ctx_cb(kfd->pdev, iommu_pasid_shutdown_callback);
> +
> + if (kfd->device_info->scheduler_class->create(kfd, &kfd->scheduler)) {
> + amd_iommu_free_device(kfd->pdev);
> + return false;
> + }
> +
> + kfd->device_info->scheduler_class->start(kfd->scheduler);
> +
> + kfd->init_complete = true;
> +
> + return true;
> +}
> +
> +void kgd2kfd_device_exit(struct kfd_dev *kfd)
> +{
> + int err = kfd_topology_remove_device(kfd);
> +
> + BUG_ON(err != 0);
> +
> + if (kfd->init_complete) {
> + kfd->device_info->scheduler_class->stop(kfd->scheduler);
> + kfd->device_info->scheduler_class->destroy(kfd->scheduler);
> +
> + amd_iommu_free_device(kfd->pdev);
> + }
> +
> + kfree(kfd);
> +}
> diff --git a/drivers/gpu/hsa/radeon/kfd_module.c b/drivers/gpu/hsa/radeon/kfd_module.c
> new file mode 100644
> index 0000000..6978bc0
> --- /dev/null
> +++ b/drivers/gpu/hsa/radeon/kfd_module.c
> @@ -0,0 +1,117 @@
> +/*
> + * Copyright 2014 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#include <linux/module.h>
> +#include <linux/sched.h>
> +#include <linux/notifier.h>
> +
> +#include "kfd_priv.h"
> +
> +#define DRIVER_AUTHOR "Andrew Lewycky, Oded Gabbay, Evgeny Pinchuk, others."
> +
> +#define DRIVER_NAME "kfd"
> +#define DRIVER_DESC "AMD HSA Kernel Fusion Driver"
> +#define DRIVER_DATE "20140127"
> +
> +const struct kfd2kgd_calls *kfd2kgd;
> +static const struct kgd2kfd_calls kgd2kfd = {
> + .exit = kgd2kfd_exit,
> + .probe = kgd2kfd_probe,
> + .device_init = kgd2kfd_device_init,
> + .device_exit = kgd2kfd_device_exit,
> +};
> +
> +bool kgd2kfd_init(unsigned interface_version,
> + const struct kfd2kgd_calls *f2g,
> + const struct kgd2kfd_calls **g2f)
> +{
> + /* Only one interface version is supported, no kfd/kgd version skew allowed. */
> + if (interface_version != KFD_INTERFACE_VERSION)
> + return false;
> +
> + kfd2kgd = f2g;
> + *g2f = &kgd2kfd;
> +
> + return true;
> +}
> +EXPORT_SYMBOL(kgd2kfd_init);
> +
> +void kgd2kfd_exit(void)
> +{
> +}
> +
> +extern int kfd_process_exit(struct notifier_block *nb,
> + unsigned long action, void *data);
> +
> +static struct notifier_block kfd_mmput_nb = {
> + .notifier_call = kfd_process_exit,
> + .priority = 3,
> +};
> +
> +static int __init kfd_module_init(void)
> +{
> + int err;
> +
> + err = radeon_kfd_pasid_init();
> + if (err < 0)
> + goto err_pasid;
> +
> + err = radeon_kfd_chardev_init();
> + if (err < 0)
> + goto err_ioctl;
> +
> + err = mmput_register_notifier(&kfd_mmput_nb);
> + if (err)
> + goto err_mmu_notifier;
> +
> + err = kfd_topology_init();
> + if (err < 0)
> + goto err_topology;
> +
> + pr_info("[hsa] Initialized kfd module");
> +
> + return 0;
> +err_topology:
> + mmput_unregister_notifier(&kfd_mmput_nb);
> +err_mmu_notifier:
> + radeon_kfd_chardev_exit();
> +err_ioctl:
> + radeon_kfd_pasid_exit();
> +err_pasid:
> + return err;
> +}
> +
> +static void __exit kfd_module_exit(void)
> +{
> + kfd_topology_shutdown();
> + mmput_unregister_notifier(&kfd_mmput_nb);
> + radeon_kfd_chardev_exit();
> + radeon_kfd_pasid_exit();
> + pr_info("[hsa] Removed kfd module");
> +}
> +
> +module_init(kfd_module_init);
> +module_exit(kfd_module_exit);
> +
> +MODULE_AUTHOR(DRIVER_AUTHOR);
> +MODULE_DESCRIPTION(DRIVER_DESC);
> +MODULE_LICENSE("GPL");
If it is GPL then comment at the top of all files must reflect that
and not use some special worded license.
> diff --git a/drivers/gpu/hsa/radeon/kfd_pasid.c b/drivers/gpu/hsa/radeon/kfd_pasid.c
> new file mode 100644
> index 0000000..d78bd00
> --- /dev/null
> +++ b/drivers/gpu/hsa/radeon/kfd_pasid.c
> @@ -0,0 +1,92 @@
> +/*
> + * Copyright 2014 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#include <linux/slab.h>
> +#include <linux/types.h>
> +#include "kfd_priv.h"
> +
> +#define INITIAL_PASID_LIMIT (1<<20)
> +
> +static unsigned long *pasid_bitmap;
> +static pasid_t pasid_limit;
> +static DEFINE_MUTEX(pasid_mutex);
> +
> +int radeon_kfd_pasid_init(void)
> +{
> + pasid_limit = INITIAL_PASID_LIMIT;
> +
> + pasid_bitmap = kzalloc(DIV_ROUND_UP(INITIAL_PASID_LIMIT, BITS_PER_BYTE), GFP_KERNEL);
> + if (!pasid_bitmap)
> + return -ENOMEM;
> +
> + set_bit(0, pasid_bitmap); /* PASID 0 is reserved. */
> +
> + return 0;
> +}
> +
> +void radeon_kfd_pasid_exit(void)
> +{
> + kfree(pasid_bitmap);
> +}
> +
> +bool radeon_kfd_set_pasid_limit(pasid_t new_limit)
> +{
> + if (new_limit < pasid_limit) {
> + bool ok;
> +
> + mutex_lock(&pasid_mutex);
> +
> + /* ensure that no pasids >= new_limit are in-use */
> + ok = (find_next_bit(pasid_bitmap, pasid_limit, new_limit) == pasid_limit);
> + if (ok)
> + pasid_limit = new_limit;
> +
> + mutex_unlock(&pasid_mutex);
> +
> + return ok;
> + }
> +
> + return true;
> +}
> +
> +pasid_t radeon_kfd_pasid_alloc(void)
> +{
> + pasid_t found;
> +
> + mutex_lock(&pasid_mutex);
> +
> + found = find_first_zero_bit(pasid_bitmap, pasid_limit);
> + if (found == pasid_limit)
> + found = 0;
> + else
> + set_bit(found, pasid_bitmap);
> +
> + mutex_unlock(&pasid_mutex);
> +
> + return found;
> +}
> +
> +void radeon_kfd_pasid_free(pasid_t pasid)
> +{
> + BUG_ON(pasid == 0 || pasid >= pasid_limit);
> + clear_bit(pasid, pasid_bitmap);
> +}
> diff --git a/drivers/gpu/hsa/radeon/kfd_priv.h b/drivers/gpu/hsa/radeon/kfd_priv.h
> new file mode 100644
> index 0000000..1d1dbcf
> --- /dev/null
> +++ b/drivers/gpu/hsa/radeon/kfd_priv.h
> @@ -0,0 +1,232 @@
> +/*
> + * Copyright 2014 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#ifndef KFD_PRIV_H_INCLUDED
> +#define KFD_PRIV_H_INCLUDED
> +
> +#include <linux/hashtable.h>
> +#include <linux/mmu_notifier.h>
> +#include <linux/mutex.h>
> +#include <linux/radeon_kfd.h>
> +#include <linux/types.h>
> +
> +struct kfd_scheduler_class;
> +
> +#define MAX_KFD_DEVICES 16 /* Global limit - only MAX_KFD_DEVICES will be supported by KFD. */
> +
> +/*
> + * Per-process limit. Each process can only
> + * create MAX_PROCESS_QUEUES across all devices
> + */
> +#define MAX_PROCESS_QUEUES 1024
> +
> +#define MAX_DOORBELL_INDEX MAX_PROCESS_QUEUES
> +#define KFD_SYSFS_FILE_MODE 0444
> +
> +/* We multiplex different sorts of mmap-able memory onto /dev/kfd.
> +** We figure out what type of memory the caller wanted by comparing the mmap page offset to known ranges. */
> +#define KFD_MMAP_DOORBELL_START (((1ULL << 32)*1) >> PAGE_SHIFT)
> +#define KFD_MMAP_DOORBELL_END (((1ULL << 32)*2) >> PAGE_SHIFT)
> +
> +/* GPU ID hash width in bits */
> +#define KFD_GPU_ID_HASH_WIDTH 16
> +
> +/* Macro for allocating structures */
> +#define kfd_alloc_struct(ptr_to_struct) ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL))
> +
> +/* Large enough to hold the maximum usable pasid + 1.
> +** It must also be able to store the number of doorbells reported by a KFD device. */
> +typedef unsigned int pasid_t;
Same on unsigned int.
> +
> +/* Type that represents a HW doorbell slot. */
> +typedef u32 doorbell_t;
> +
> +struct kfd_device_info {
> + const struct kfd_scheduler_class *scheduler_class;
> + unsigned int max_pasid_bits;
> +};
> +
> +struct kfd_dev {
> + struct kgd_dev *kgd;
> +
> + const struct kfd_device_info *device_info;
> + struct pci_dev *pdev;
> +
> + void __iomem *regs;
> +
> + bool init_complete;
> +
> + unsigned int id; /* topology stub index */
> +
> + phys_addr_t doorbell_base; /* Start of actual doorbells used by
> + * KFD. It is aligned for mapping
> + * into user mode
> + */
> + size_t doorbell_id_offset; /* Doorbell offset (from KFD doorbell
> + * to HW doorbell, GFX reserved some
> + * at the start)
> + */
> + size_t doorbell_process_limit; /* Number of processes we have doorbell space for. */
> +
> + struct kgd2kfd_shared_resources shared_resources;
> +
> + struct kfd_scheduler *scheduler;
> +};
> +
> +/* KGD2KFD callbacks */
> +void kgd2kfd_exit(void);
> +struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev);
> +bool kgd2kfd_device_init(struct kfd_dev *kfd,
> + const struct kgd2kfd_shared_resources *gpu_resources);
> +void kgd2kfd_device_exit(struct kfd_dev *kfd);
> +
> +extern const struct kfd2kgd_calls *kfd2kgd;
> +
> +
> +/* KFD2KGD callback wrappers */
> +void radeon_kfd_lock_srbm_index(struct kfd_dev *kfd);
> +void radeon_kfd_unlock_srbm_index(struct kfd_dev *kfd);
> +
> +enum kfd_mempool {
> + KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
> + KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
> + KFD_MEMPOOL_FRAMEBUFFER = 3,
> +};
> +
> +struct kfd_mem_obj_s; /* Dummy struct just to make kfd_mem_obj* a unique pointer type. */
> +typedef struct kfd_mem_obj_s *kfd_mem_obj;
> +
> +int radeon_kfd_vidmem_alloc(struct kfd_dev *kfd, size_t size, size_t alignment,
> + enum kfd_mempool pool, kfd_mem_obj *mem_obj);
> +void radeon_kfd_vidmem_free(struct kfd_dev *kfd, kfd_mem_obj mem_obj);
> +int radeon_kfd_vidmem_gpumap(struct kfd_dev *kfd, kfd_mem_obj mem_obj, uint64_t *vmid0_address);
> +void radeon_kfd_vidmem_ungpumap(struct kfd_dev *kfd, kfd_mem_obj mem_obj);
> +int radeon_kfd_vidmem_kmap(struct kfd_dev *kfd, kfd_mem_obj mem_obj, void **ptr);
> +void radeon_kfd_vidmem_unkmap(struct kfd_dev *kfd, kfd_mem_obj mem_obj);
> +
> +/* Character device interface */
> +int radeon_kfd_chardev_init(void);
> +void radeon_kfd_chardev_exit(void);
> +struct device *radeon_kfd_chardev(void);
> +
> +/* Scheduler */
> +struct kfd_scheduler;
> +struct kfd_scheduler_process;
> +struct kfd_scheduler_queue {
> + uint64_t dummy;
> +};
> +
> +struct kfd_queue {
> + struct kfd_dev *dev;
> +
> + /* scheduler_queue must be last. It is variable sized (dev->device_info->scheduler_class->queue_size) */
> + struct kfd_scheduler_queue scheduler_queue;
> +};
> +
> +/* Data that is per-process-per device. */
> +struct kfd_process_device {
> + /* List of all per-device data for a process. Starts from kfd_process.per_device_data. */
> + struct list_head per_device_list;
> +
> + /* The device that owns this data. */
> + struct kfd_dev *dev;
> +
> + /* The user-mode address of the doorbell mapping for this device. */
> + doorbell_t __user *doorbell_mapping;
> +
> + /* The number of queues created by this process for this device. */
> + uint32_t queue_count;
> +
> + /* Scheduler process data for this device. */
> + struct kfd_scheduler_process *scheduler_process;
> +
> + /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
> + bool bound;
> +};
> +
> +/* Process data */
> +struct kfd_process {
> + struct list_head processes_list;
> +
> + struct mm_struct *mm;
> +
> + struct mutex mutex;
> +
> + /* In any process, the thread that started main() is the lead thread and outlives the rest.
> + * It is here because amd_iommu_bind_pasid wants a task_struct. */
> + struct task_struct *lead_thread;
> +
> + pasid_t pasid;
> +
> + /* List of kfd_process_device structures, one for each device the process is using. */
> + struct list_head per_device_data;
> +
> + /* The process's queues. */
> + size_t queue_array_size;
> + struct kfd_queue **queues; /* Size is queue_array_size, up to MAX_PROCESS_QUEUES. */
> + unsigned long allocated_queue_bitmap[DIV_ROUND_UP(MAX_PROCESS_QUEUES, BITS_PER_LONG)];
> +};
> +
> +struct kfd_process *radeon_kfd_create_process(const struct task_struct *);
> +struct kfd_process *radeon_kfd_get_process(const struct task_struct *);
> +
> +struct kfd_process_device *radeon_kfd_bind_process_to_device(struct kfd_dev *dev, struct kfd_process *p);
> +void radeon_kfd_unbind_process_from_device(struct kfd_dev *dev, pasid_t pasid);
> +struct kfd_process_device *radeon_kfd_get_process_device_data(struct kfd_dev *dev, struct kfd_process *p);
> +
> +bool radeon_kfd_allocate_queue_id(struct kfd_process *p, unsigned int *queue_id);
> +void radeon_kfd_install_queue(struct kfd_process *p, unsigned int queue_id, struct kfd_queue *queue);
> +void radeon_kfd_remove_queue(struct kfd_process *p, unsigned int queue_id);
> +struct kfd_queue *radeon_kfd_get_queue(struct kfd_process *p, unsigned int queue_id);
> +
> +
> +/* PASIDs */
> +int radeon_kfd_pasid_init(void);
> +void radeon_kfd_pasid_exit(void);
> +bool radeon_kfd_set_pasid_limit(pasid_t new_limit);
> +pasid_t radeon_kfd_pasid_alloc(void);
> +void radeon_kfd_pasid_free(pasid_t pasid);
> +
> +/* Doorbells */
> +void radeon_kfd_doorbell_init(struct kfd_dev *kfd);
> +int radeon_kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma);
> +doorbell_t __user *radeon_kfd_get_doorbell(struct file *devkfd, struct kfd_process *process, struct kfd_dev *dev,
> + unsigned int doorbell_index);
> +unsigned int radeon_kfd_queue_id_to_doorbell(struct kfd_dev *kfd, struct kfd_process *process, unsigned int queue_id);
> +
> +extern struct device *kfd_device;
> +
> +/* Topology */
> +int kfd_topology_init(void);
> +void kfd_topology_shutdown(void);
> +int kfd_topology_add_device(struct kfd_dev *gpu);
> +int kfd_topology_remove_device(struct kfd_dev *gpu);
> +struct kfd_dev *radeon_kfd_device_by_id(uint32_t gpu_id);
> +struct kfd_dev *radeon_kfd_device_by_pci_dev(const struct pci_dev *pdev);
> +
> +/* MMIO registers */
> +#define WRITE_REG(dev, reg, value) radeon_kfd_write_reg((dev), (reg), (value))
> +#define READ_REG(dev, reg) radeon_kfd_read_reg((dev), (reg))
> +void radeon_kfd_write_reg(struct kfd_dev *dev, uint32_t reg, uint32_t value);
> +uint32_t radeon_kfd_read_reg(struct kfd_dev *dev, uint32_t reg);
> +
> +#endif
> diff --git a/drivers/gpu/hsa/radeon/kfd_process.c b/drivers/gpu/hsa/radeon/kfd_process.c
> new file mode 100644
> index 0000000..145ee38
> --- /dev/null
> +++ b/drivers/gpu/hsa/radeon/kfd_process.c
> @@ -0,0 +1,400 @@
> +/*
> + * Copyright 2014 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#include <linux/mutex.h>
> +#include <linux/log2.h>
> +#include <linux/sched.h>
> +#include <linux/slab.h>
> +#include <linux/amd-iommu.h>
> +#include <linux/notifier.h>
> +struct mm_struct;
> +
> +#include "kfd_priv.h"
> +#include "kfd_scheduler.h"
> +
> +/* Initial size for the array of queues.
> + * The allocated size is doubled each time it is exceeded up to MAX_PROCESS_QUEUES. */
> +#define INITIAL_QUEUE_ARRAY_SIZE 16
> +
> +/* List of struct kfd_process */
> +static struct list_head kfd_processes_list = LIST_HEAD_INIT(kfd_processes_list);
> +
> +static DEFINE_MUTEX(kfd_processes_mutex);
> +
> +static struct kfd_process *create_process(const struct task_struct *thread);
> +
> +struct kfd_process*
> +radeon_kfd_create_process(const struct task_struct *thread)
> +{
> + struct kfd_process *process;
> +
> + if (thread->mm == NULL)
> + return ERR_PTR(-EINVAL);
> +
> + /* Only the pthreads threading model is supported. */
> + if (thread->group_leader->mm != thread->mm)
> + return ERR_PTR(-EINVAL);
> +
> + /*
> + * take kfd processes mutex before starting of process creation
> + * so there won't be a case where two threads of the same process
> + * create two kfd_process structures
> + */
> + mutex_lock(&kfd_processes_mutex);
Given that this is to protect mm->kfd_process i would rather that you
use some mm lock so that if another non kfd code ever need to check
this variable in a sensible way then it could protect itself with a
mm lock.
But again i believe that mm_struct should not have a new kfd field but
rather some generic iommu pasid field that can then forward through
generic iommu code things to kfd.
> +
> + /* A prior open of /dev/kfd could have already created the process. */
> + process = thread->mm->kfd_process;
> + if (process)
> + pr_debug("kfd: process already found\n");
> +
> + if (!process)
> + process = create_process(thread);
> +
> + mutex_unlock(&kfd_processes_mutex);
> +
> + return process;
> +}
> +
> +struct kfd_process*
> +radeon_kfd_get_process(const struct task_struct *thread)
> +{
> + struct kfd_process *process;
> +
> + if (thread->mm == NULL)
> + return ERR_PTR(-EINVAL);
> +
> + /* Only the pthreads threading model is supported. */
> + if (thread->group_leader->mm != thread->mm)
> + return ERR_PTR(-EINVAL);
> +
> + process = thread->mm->kfd_process;
> +
> + return process;
> +}
> +
> +/* Assumes that the kfd_process mutex is held.
> + * (Or that it doesn't need to be held because the process is exiting.)
> + *
> + * dev_filter can be set to only destroy queues for one device.
> + * Otherwise all queues for the process are destroyed.
> + */
> +static void
> +destroy_queues(struct kfd_process *p, struct kfd_dev *dev_filter)
> +{
> + unsigned long queue_id;
> +
> + for_each_set_bit(queue_id, p->allocated_queue_bitmap, MAX_PROCESS_QUEUES) {
> +
> + struct kfd_queue *queue = radeon_kfd_get_queue(p, queue_id);
> + struct kfd_dev *dev;
> +
> + BUG_ON(queue == NULL);
> +
> + dev = queue->dev;
> +
> + if (!dev_filter || dev == dev_filter) {
> + struct kfd_process_device *pdd = radeon_kfd_get_process_device_data(dev, p);
> +
> + BUG_ON(pdd == NULL); /* A queue exists so pdd must. */
> +
> + radeon_kfd_remove_queue(p, queue_id);
> + dev->device_info->scheduler_class->destroy_queue(dev->scheduler, &queue->scheduler_queue);
> +
> + kfree(queue);
> +
> + BUG_ON(pdd->queue_count == 0);
> + BUG_ON(pdd->scheduler_process == NULL);
> +
> + if (--pdd->queue_count == 0) {
> + dev->device_info->scheduler_class->deregister_process(dev->scheduler,
> + pdd->scheduler_process);
> + pdd->scheduler_process = NULL;
> + }
> + }
> + }
> +}
> +
> +static void free_process(struct kfd_process *p)
> +{
> + struct kfd_process_device *pdd, *temp;
> +
> + BUG_ON(p == NULL);
> +
> + destroy_queues(p, NULL);
> +
> + /* doorbell mappings: automatic */
> +
> + list_for_each_entry_safe(pdd, temp, &p->per_device_data, per_device_list) {
> + amd_iommu_unbind_pasid(pdd->dev->pdev, p->pasid);
> + list_del(&pdd->per_device_list);
> + kfree(pdd);
> + }
> +
> + radeon_kfd_pasid_free(p->pasid);
> +
> + mutex_destroy(&p->mutex);
> +
> + kfree(p->queues);
> +
> + list_del(&p->processes_list);
> +
> + kfree(p);
> +}
> +
> +int kfd_process_exit(struct notifier_block *nb,
> + unsigned long action, void *data)
> +{
> + struct mm_struct *mm = data;
> + struct kfd_process *p;
> +
> + mutex_lock(&kfd_processes_mutex);
> +
> + p = mm->kfd_process;
> + if (p) {
> + free_process(p);
> + mm->kfd_process = NULL;
> + }
> +
> + mutex_unlock(&kfd_processes_mutex);
> +
> + return 0;
> +}
> +
> +static struct kfd_process *create_process(const struct task_struct *thread)
> +{
> + struct kfd_process *process;
> + int err = -ENOMEM;
> +
> + process = kzalloc(sizeof(*process), GFP_KERNEL);
> +
> + if (!process)
> + goto err_alloc;
> +
> + process->queues = kmalloc_array(INITIAL_QUEUE_ARRAY_SIZE, sizeof(process->queues[0]), GFP_KERNEL);
> + if (!process->queues)
> + goto err_alloc;
> +
> + process->pasid = radeon_kfd_pasid_alloc();
> + if (process->pasid == 0)
> + goto err_alloc;
> +
> + mutex_init(&process->mutex);
> +
> + process->mm = thread->mm;
> + thread->mm->kfd_process = process;
> + list_add_tail(&process->processes_list, &kfd_processes_list);
> +
> + process->lead_thread = thread->group_leader;
> +
> + process->queue_array_size = INITIAL_QUEUE_ARRAY_SIZE;
> +
> + INIT_LIST_HEAD(&process->per_device_data);
> +
> + return process;
> +
> +err_alloc:
> + kfree(process->queues);
> + kfree(process);
> + return ERR_PTR(err);
> +}
> +
> +struct kfd_process_device *
> +radeon_kfd_get_process_device_data(struct kfd_dev *dev, struct kfd_process *p)
> +{
> + struct kfd_process_device *pdd;
> +
> + list_for_each_entry(pdd, &p->per_device_data, per_device_list)
> + if (pdd->dev == dev)
> + return pdd;
> +
> + pdd = kzalloc(sizeof(*pdd), GFP_KERNEL);
> + if (pdd != NULL) {
> + pdd->dev = dev;
> + list_add(&pdd->per_device_list, &p->per_device_data);
> + }
> +
> + return pdd;
> +}
> +
> +/* Direct the IOMMU to bind the process (specifically the pasid->mm) to the device.
> + * Unbinding occurs when the process dies or the device is removed.
> + *
> + * Assumes that the process lock is held.
> + */
> +struct kfd_process_device *radeon_kfd_bind_process_to_device(struct kfd_dev *dev, struct kfd_process *p)
> +{
> + struct kfd_process_device *pdd = radeon_kfd_get_process_device_data(dev, p);
> + int err;
> +
> + if (pdd == NULL)
> + return ERR_PTR(-ENOMEM);
> +
> + if (pdd->bound)
> + return pdd;
> +
> + err = amd_iommu_bind_pasid(dev->pdev, p->pasid, p->lead_thread);
Are we to assume that for eternity this will not work on iommu that do support
PASID/ATS but are not from AMD ? If it was an APU specific function i would
understand but it seems that the IOMMU API needs to grow. I am pretty sure
Intel will have an ATS/PASID IOMMU.
> + if (err < 0)
> + return ERR_PTR(err);
> +
> + pdd->bound = true;
> +
> + return pdd;
> +}
> +
> +void radeon_kfd_unbind_process_from_device(struct kfd_dev *dev, pasid_t pasid)
> +{
> + struct kfd_process *p;
> + struct kfd_process_device *pdd;
> +
> + BUG_ON(dev == NULL);
> +
> + mutex_lock(&kfd_processes_mutex);
> +
> + list_for_each_entry(p, &kfd_processes_list, processes_list)
> + if (p->pasid == pasid)
> + break;
> +
> + mutex_unlock(&kfd_processes_mutex);
> +
> + BUG_ON(p->pasid != pasid);
> +
> + pdd = radeon_kfd_get_process_device_data(dev, p);
> +
> + BUG_ON(pdd == NULL);
> +
> + mutex_lock(&p->mutex);
> +
> + destroy_queues(p, dev);
> +
> + /* All queues just got destroyed so this should be gone. */
> + BUG_ON(pdd->scheduler_process != NULL);
> +
> + /*
> + * Just mark pdd as unbound, because we still need it to call
> + * amd_iommu_unbind_pasid() in when the process exits.
> + * We don't call amd_iommu_unbind_pasid() here
> + * because the IOMMU called us.
> + */
> + pdd->bound = false;
> +
> + mutex_unlock(&p->mutex);
> +}
> +
> +/* Ensure that the process's queue array is large enough to hold the queue at queue_id.
> + * Assumes that the process lock is held. */
> +static bool ensure_queue_array_size(struct kfd_process *p, unsigned int queue_id)
> +{
> + size_t desired_size;
> + struct kfd_queue **new_queues;
> +
> + compiletime_assert(INITIAL_QUEUE_ARRAY_SIZE > 0, "INITIAL_QUEUE_ARRAY_SIZE must not be 0");
> + compiletime_assert(INITIAL_QUEUE_ARRAY_SIZE <= MAX_PROCESS_QUEUES,
> + "INITIAL_QUEUE_ARRAY_SIZE must be less than MAX_PROCESS_QUEUES");
> + /* Ensure that doubling the current size won't ever overflow. */
> + compiletime_assert(MAX_PROCESS_QUEUES < SIZE_MAX / 2, "MAX_PROCESS_QUEUES must be less than SIZE_MAX/2");
> +
> + /*
> + * These & queue_id < MAX_PROCESS_QUEUES guarantee that
> + * the desired_size calculation will end up <= MAX_PROCESS_QUEUES
> + */
> + compiletime_assert(is_power_of_2(INITIAL_QUEUE_ARRAY_SIZE), "INITIAL_QUEUE_ARRAY_SIZE must be power of 2.");
> + compiletime_assert(MAX_PROCESS_QUEUES % INITIAL_QUEUE_ARRAY_SIZE == 0,
> + "MAX_PROCESS_QUEUES must be multiple of INITIAL_QUEUE_ARRAY_SIZE.");
> + compiletime_assert(is_power_of_2(MAX_PROCESS_QUEUES / INITIAL_QUEUE_ARRAY_SIZE),
> + "MAX_PROCESS_QUEUES must be a power-of-2 multiple of INITIAL_QUEUE_ARRAY_SIZE.");
> +
> + if (queue_id < p->queue_array_size)
> + return true;
> +
> + if (queue_id >= MAX_PROCESS_QUEUES)
> + return false;
> +
> + desired_size = p->queue_array_size;
> + while (desired_size <= queue_id)
> + desired_size *= 2;
> +
> + BUG_ON(desired_size < queue_id || desired_size > MAX_PROCESS_QUEUES);
> + BUG_ON(desired_size % INITIAL_QUEUE_ARRAY_SIZE != 0 || !is_power_of_2(desired_size / INITIAL_QUEUE_ARRAY_SIZE));
> +
> + new_queues = kmalloc_array(desired_size, sizeof(p->queues[0]), GFP_KERNEL);
> + if (!new_queues)
> + return false;
> +
> + memcpy(new_queues, p->queues, p->queue_array_size * sizeof(p->queues[0]));
> +
> + kfree(p->queues);
> + p->queues = new_queues;
> + p->queue_array_size = desired_size;
> +
> + return true;
> +}
> +
> +/* Assumes that the process lock is held. */
> +bool radeon_kfd_allocate_queue_id(struct kfd_process *p, unsigned int *queue_id)
> +{
> + unsigned int qid = find_first_zero_bit(p->allocated_queue_bitmap, MAX_PROCESS_QUEUES);
> +
> + if (qid >= MAX_PROCESS_QUEUES)
> + return false;
> +
> + if (!ensure_queue_array_size(p, qid))
> + return false;
> +
> + __set_bit(qid, p->allocated_queue_bitmap);
> +
> + p->queues[qid] = NULL;
> + *queue_id = qid;
> +
> + return true;
> +}
> +
> +/* Install a queue into a previously-allocated queue id.
> + * Assumes that the process lock is held. */
> +void radeon_kfd_install_queue(struct kfd_process *p, unsigned int queue_id, struct kfd_queue *queue)
> +{
> + BUG_ON(queue_id >= p->queue_array_size); /* Have to call allocate_queue_id before install_queue. */
> + BUG_ON(queue == NULL);
> +
> + p->queues[queue_id] = queue;
> +}
> +
> +/* Remove a queue from the open queue list and deallocate the queue id.
> + * This can be called whether or not a queue was installed.
> + * Assumes that the process lock is held. */
> +void radeon_kfd_remove_queue(struct kfd_process *p, unsigned int queue_id)
> +{
> + BUG_ON(!test_bit(queue_id, p->allocated_queue_bitmap));
> + BUG_ON(queue_id >= p->queue_array_size);
> +
> + __clear_bit(queue_id, p->allocated_queue_bitmap);
> +}
> +
> +/* Assumes that the process lock is held. */
> +struct kfd_queue *radeon_kfd_get_queue(struct kfd_process *p, unsigned int queue_id)
> +{
> + /* test_bit because the contents of unallocated queue slots are undefined.
> + * Otherwise ensure_queue_array_size would have to clear new entries and
> + * remove_queue would have to NULL removed queues. */
> + return (queue_id < p->queue_array_size &&
> + test_bit(queue_id, p->allocated_queue_bitmap)) ?
> + p->queues[queue_id] : NULL;
> +}
> diff --git a/drivers/gpu/hsa/radeon/kfd_scheduler.h b/drivers/gpu/hsa/radeon/kfd_scheduler.h
> new file mode 100644
> index 0000000..48a032f
> --- /dev/null
> +++ b/drivers/gpu/hsa/radeon/kfd_scheduler.h
> @@ -0,0 +1,62 @@
> +/*
> + * Copyright 2014 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#ifndef KFD_SCHEDULER_H_INCLUDED
> +#define KFD_SCHEDULER_H_INCLUDED
> +
> +#include <linux/types.h>
> +struct kfd_process;
> +
> +/* Opaque types for scheduler private data. */
> +struct kfd_scheduler;
> +struct kfd_scheduler_process;
> +struct kfd_scheduler_queue;
> +
> +struct kfd_scheduler_class {
> + const char *name;
> +
> + int (*create)(struct kfd_dev *, struct kfd_scheduler **);
> + void (*destroy)(struct kfd_scheduler *);
> +
> + void (*start)(struct kfd_scheduler *);
> + void (*stop)(struct kfd_scheduler *);
> +
> + int (*register_process)(struct kfd_scheduler *, struct kfd_process *, struct kfd_scheduler_process **);
> + void (*deregister_process)(struct kfd_scheduler *, struct kfd_scheduler_process *);
> +
> + size_t queue_size;
> +
> + int (*create_queue)(struct kfd_scheduler *scheduler,
> + struct kfd_scheduler_process *process,
> + struct kfd_scheduler_queue *queue,
> + void __user *ring_address,
> + uint64_t ring_size,
> + void __user *rptr_address,
> + void __user *wptr_address,
> + unsigned int doorbell);
> +
> + void (*destroy_queue)(struct kfd_scheduler *, struct kfd_scheduler_queue *);
> +};
> +
> +extern const struct kfd_scheduler_class radeon_kfd_cik_static_scheduler_class;
> +
> +#endif
> diff --git a/drivers/gpu/hsa/radeon/kfd_topology.c b/drivers/gpu/hsa/radeon/kfd_topology.c
> new file mode 100644
> index 0000000..6acac25
> --- /dev/null
> +++ b/drivers/gpu/hsa/radeon/kfd_topology.c
> @@ -0,0 +1,1201 @@
> +/*
> + * Copyright 2014 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#include <linux/types.h>
> +#include <linux/kernel.h>
> +#include <linux/pci.h>
> +#include <linux/errno.h>
> +#include <linux/acpi.h>
> +#include <linux/hash.h>
> +
> +#include "kfd_priv.h"
> +#include "kfd_crat.h"
> +#include "kfd_topology.h"
> +
> +static struct list_head topology_device_list;
> +static int topology_crat_parsed;
> +static struct kfd_system_properties sys_props;
> +
> +static DECLARE_RWSEM(topology_lock);
> +
> +
> +static uint8_t checksum_image(const void *buf, size_t len)
> +{
> + uint8_t *p = (uint8_t *)buf;
> + uint8_t sum = 0;
> +
> + if (!buf)
> + return 0;
> +
> + while (len-- > 0)
> + sum += *p++;
> +
> + return sum;
> + }
> +
> +struct kfd_dev *radeon_kfd_device_by_id(uint32_t gpu_id)
> +{
> + struct kfd_topology_device *top_dev;
> + struct kfd_dev *device = NULL;
> +
> + down_read(&topology_lock);
> +
> + list_for_each_entry(top_dev, &topology_device_list, list)
> + if (top_dev->gpu_id == gpu_id) {
> + device = top_dev->gpu;
> + break;
> + }
> +
> + up_read(&topology_lock);
> +
> + return device;
> +}
> +
> +struct kfd_dev *radeon_kfd_device_by_pci_dev(const struct pci_dev *pdev)
> +{
> + struct kfd_topology_device *top_dev;
> + struct kfd_dev *device = NULL;
> +
> + down_read(&topology_lock);
> +
> + list_for_each_entry(top_dev, &topology_device_list, list)
> + if (top_dev->gpu->pdev == pdev) {
> + device = top_dev->gpu;
> + break;
> + }
> +
> + up_read(&topology_lock);
> +
> + return device;
> +}
> +
> +static int kfd_topology_get_crat_acpi(void *crat_image, size_t *size)
> +{
> + struct acpi_table_header *crat_table;
> + acpi_status status;
> +
> + if (!size)
> + return -EINVAL;
> +
> +/*
> + * Fetch the CRAT table from ACPI
> + */
> + status = acpi_get_table(CRAT_SIGNATURE, 0, &crat_table);
> + if (status == AE_NOT_FOUND) {
> + pr_warn("CRAT table not found\n");
> + return -ENODATA;
> + } else if (ACPI_FAILURE(status)) {
> + const char *err = acpi_format_exception(status);
> +
> + pr_err("CRAT table error: %s\n", err);
> + return -EINVAL;
> + }
> +
> + /*
> + * The checksum of the table should be verified
> + */
> + if (checksum_image(crat_table, crat_table->length) ==
> + crat_table->checksum) {
> + pr_err("Bad checksum for the CRAT table\n");
> + return -EINVAL;
> +}
> +
> +
> + if (*size >= crat_table->length && crat_image != 0)
> + memcpy(crat_image, crat_table, crat_table->length);
> +
> + *size = crat_table->length;
> +
> + return 0;
> +}
> +
> +static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev,
> + struct crat_subtype_computeunit *cu)
> +{
> + BUG_ON(!dev);
> + BUG_ON(!cu);
> +
> + dev->node_props.cpu_cores_count = cu->num_cpu_cores;
> + dev->node_props.cpu_core_id_base = cu->processor_id_low;
> + if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT)
> + dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
> +
> + pr_info("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores,
> + cu->processor_id_low);
> +}
> +
> +static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev,
> + struct crat_subtype_computeunit *cu)
> +{
> + BUG_ON(!dev);
> + BUG_ON(!cu);
> +
> + dev->node_props.simd_id_base = cu->processor_id_low;
> + dev->node_props.simd_count = cu->num_simd_cores;
> + dev->node_props.lds_size_in_kb = cu->lds_size_in_kb;
> + dev->node_props.max_waves_per_simd = cu->max_waves_simd;
> + dev->node_props.wave_front_size = cu->wave_front_size;
> + dev->node_props.mem_banks_count = cu->num_banks;
> + dev->node_props.array_count = cu->num_arrays;
> + dev->node_props.cu_per_simd_array = cu->num_cu_per_array;
> + dev->node_props.simd_per_cu = cu->num_simd_per_cu;
> + dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu;
> + if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE)
> + dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE;
> + pr_info("CU GPU: simds=%d id_base=%d\n", cu->num_simd_cores,
> + cu->processor_id_low);
> +}
> +
> +/* kfd_parse_subtype_cu is called when the topology mutex is already acquired */
> +static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu)
> +{
> + struct kfd_topology_device *dev;
> + int i = 0;
> +
> + BUG_ON(!cu);
> +
> + pr_info("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n",
> + cu->proximity_domain, cu->hsa_capability);
> + list_for_each_entry(dev, &topology_device_list, list) {
> + if (cu->proximity_domain == i) {
> + if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT)
> + kfd_populated_cu_info_cpu(dev, cu);
> +
> + if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT)
> + kfd_populated_cu_info_gpu(dev, cu);
> + break;
> + }
> + i++;
> + }
> +
> + return 0;
> +}
> +
> +/* kfd_parse_subtype_mem is called when the topology mutex is already acquired */
> +static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem)
> +{
> + struct kfd_mem_properties *props;
> + struct kfd_topology_device *dev;
> + int i = 0;
> +
> + BUG_ON(!mem);
> +
> + pr_info("Found memory entry in CRAT table with proximity_domain=%d\n",
> + mem->promixity_domain);
> + list_for_each_entry(dev, &topology_device_list, list) {
> + if (mem->promixity_domain == i) {
> + props = kfd_alloc_struct(props);
> + if (props == 0)
> + return -ENOMEM;
> +
> + if (dev->node_props.cpu_cores_count == 0)
> + props->heap_type = HSA_MEM_HEAP_TYPE_FB_PRIVATE;
> + else
> + props->heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
> +
> + if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE)
> + props->flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
> + if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE)
> + props->flags |= HSA_MEM_FLAGS_NON_VOLATILE;
> +
> + props->size_in_bytes = ((uint64_t)mem->length_high << 32) +
> + mem->length_low;
> + props->width = mem->width;
> +
> + dev->mem_bank_count++;
> + list_add_tail(&props->list, &dev->mem_props);
> +
> + break;
> + }
> + i++;
> + }
> +
> + return 0;
> +}
> +
> +/* kfd_parse_subtype_cache is called when the topology mutex is already acquired */
> +static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache)
> +{
> + struct kfd_cache_properties *props;
> + struct kfd_topology_device *dev;
> + uint32_t id;
> +
> + BUG_ON(!cache);
> +
> + id = cache->processor_id_low;
> +
> + pr_info("Found cache entry in CRAT table with processor_id=%d\n", id);
> + list_for_each_entry(dev, &topology_device_list, list)
> + if (id == dev->node_props.cpu_core_id_base ||
> + id == dev->node_props.simd_id_base) {
> + props = kfd_alloc_struct(props);
> + if (props == 0)
> + return -ENOMEM;
> +
> + props->processor_id_low = id;
> + props->cache_level = cache->cache_level;
> + props->cache_size = cache->cache_size;
> + props->cacheline_size = cache->cache_line_size;
> + props->cachelines_per_tag = cache->lines_per_tag;
> + props->cache_assoc = cache->associativity;
> + props->cache_latency = cache->cache_latency;
> +
> + if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE)
> + props->cache_type |= HSA_CACHE_TYPE_DATA;
> + if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE)
> + props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
> + if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE)
> + props->cache_type |= HSA_CACHE_TYPE_CPU;
> + if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
> + props->cache_type |= HSA_CACHE_TYPE_HSACU;
> +
> + dev->cache_count++;
> + dev->node_props.caches_count++;
> + list_add_tail(&props->list, &dev->cache_props);
> +
> + break;
> + }
> +
> + return 0;
> +}
> +
> +/* kfd_parse_subtype_iolink is called when the topology mutex is already acquired */
> +static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink)
> +{
> + struct kfd_iolink_properties *props;
> + struct kfd_topology_device *dev;
> + uint32_t i = 0;
> + uint32_t id_from;
> + uint32_t id_to;
> +
> + BUG_ON(!iolink);
> +
> + id_from = iolink->proximity_domain_from;
> + id_to = iolink->proximity_domain_to;
> +
> + pr_info("Found IO link entry in CRAT table with id_from=%d\n", id_from);
> + list_for_each_entry(dev, &topology_device_list, list) {
> + if (id_from == i) {
> + props = kfd_alloc_struct(props);
> + if (props == 0)
> + return -ENOMEM;
> +
> + props->node_from = id_from;
> + props->node_to = id_to;
> + props->ver_maj = iolink->version_major;
> + props->ver_min = iolink->version_minor;
> +
> + /*
> + * weight factor (derived from CDIR), currently always 1
> + */
> + props->weight = 1;
> +
> + props->min_latency = iolink->minimum_latency;
> + props->max_latency = iolink->maximum_latency;
> + props->min_bandwidth = iolink->minimum_bandwidth_mbs;
> + props->max_bandwidth = iolink->maximum_bandwidth_mbs;
> + props->rec_transfer_size =
> + iolink->recommended_transfer_size;
> +
> + dev->io_link_count++;
> + dev->node_props.io_links_count++;
> + list_add_tail(&props->list, &dev->io_link_props);
> +
> + break;
> + }
> + i++;
> + }
> +
> + return 0;
> +}
> +
> +static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr)
> +{
> + struct crat_subtype_computeunit *cu;
> + struct crat_subtype_memory *mem;
> + struct crat_subtype_cache *cache;
> + struct crat_subtype_iolink *iolink;
> + int ret = 0;
> +
> + BUG_ON(!sub_type_hdr);
> +
> + switch (sub_type_hdr->type) {
> + case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY:
> + cu = (struct crat_subtype_computeunit *)sub_type_hdr;
> + ret = kfd_parse_subtype_cu(cu);
> + break;
> + case CRAT_SUBTYPE_MEMORY_AFFINITY:
> + mem = (struct crat_subtype_memory *)sub_type_hdr;
> + ret = kfd_parse_subtype_mem(mem);
> + break;
> + case CRAT_SUBTYPE_CACHE_AFFINITY:
> + cache = (struct crat_subtype_cache *)sub_type_hdr;
> + ret = kfd_parse_subtype_cache(cache);
> + break;
> + case CRAT_SUBTYPE_TLB_AFFINITY:
> + /*
> + * For now, nothing to do here
> + */
> + pr_info("Found TLB entry in CRAT table (not processing)\n");
> + break;
> + case CRAT_SUBTYPE_CCOMPUTE_AFFINITY:
> + /*
> + * For now, nothing to do here
> + */
> + pr_info("Found CCOMPUTE entry in CRAT table (not processing)\n");
> + break;
> + case CRAT_SUBTYPE_IOLINK_AFFINITY:
> + iolink = (struct crat_subtype_iolink *)sub_type_hdr;
> + ret = kfd_parse_subtype_iolink(iolink);
> + break;
> + default:
> + pr_warn("Unknown subtype (%d) in CRAT\n",
> + sub_type_hdr->type);
> + }
> +
> + return ret;
> +}
> +
> +static void kfd_release_topology_device(struct kfd_topology_device *dev)
> +{
> + struct kfd_mem_properties *mem;
> + struct kfd_cache_properties *cache;
> + struct kfd_iolink_properties *iolink;
> +
> + BUG_ON(!dev);
> +
> + list_del(&dev->list);
> +
> + while (dev->mem_props.next != &dev->mem_props) {
> + mem = container_of(dev->mem_props.next,
> + struct kfd_mem_properties, list);
> + list_del(&mem->list);
> + kfree(mem);
> + }
> +
> + while (dev->cache_props.next != &dev->cache_props) {
> + cache = container_of(dev->cache_props.next,
> + struct kfd_cache_properties, list);
> + list_del(&cache->list);
> + kfree(cache);
> + }
> +
> + while (dev->io_link_props.next != &dev->io_link_props) {
> + iolink = container_of(dev->io_link_props.next,
> + struct kfd_iolink_properties, list);
> + list_del(&iolink->list);
> + kfree(iolink);
> + }
> +
> + kfree(dev);
> +
> + sys_props.num_devices--;
> +}
> +
> +static void kfd_release_live_view(void)
> +{
> + struct kfd_topology_device *dev;
> +
> + while (topology_device_list.next != &topology_device_list) {
> + dev = container_of(topology_device_list.next,
> + struct kfd_topology_device, list);
> + kfd_release_topology_device(dev);
> +}
> +
> + memset(&sys_props, 0, sizeof(sys_props));
> +}
> +
> +static struct kfd_topology_device *kfd_create_topology_device(void)
> +{
> + struct kfd_topology_device *dev;
> +
> + dev = kfd_alloc_struct(dev);
> + if (dev == 0) {
> + pr_err("No memory to allocate a topology device");
> + return 0;
> + }
> +
> + INIT_LIST_HEAD(&dev->mem_props);
> + INIT_LIST_HEAD(&dev->cache_props);
> + INIT_LIST_HEAD(&dev->io_link_props);
> +
> + list_add_tail(&dev->list, &topology_device_list);
> + sys_props.num_devices++;
> +
> + return dev;
> + }
> +
> +static int kfd_parse_crat_table(void *crat_image)
> +{
> + struct kfd_topology_device *top_dev;
> + struct crat_subtype_generic *sub_type_hdr;
> + uint16_t node_id;
> + int ret;
> + struct crat_header *crat_table = (struct crat_header *)crat_image;
> + uint16_t num_nodes;
> + uint32_t image_len;
> +
> + if (!crat_image)
> + return -EINVAL;
> +
> + num_nodes = crat_table->num_domains;
> + image_len = crat_table->length;
> +
> + pr_info("Parsing CRAT table with %d nodes\n", num_nodes);
> +
> + for (node_id = 0; node_id < num_nodes; node_id++) {
> + top_dev = kfd_create_topology_device();
> + if (!top_dev) {
> + kfd_release_live_view();
> + return -ENOMEM;
> + }
> +}
> +
> + sys_props.platform_id = *((uint64_t *)crat_table->oem_id);
> + sys_props.platform_oem = *((uint64_t *)crat_table->oem_table_id);
> + sys_props.platform_rev = crat_table->revision;
> +
> + sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
> + while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) <
> + ((char *)crat_image) + image_len) {
> + if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) {
> + ret = kfd_parse_subtype(sub_type_hdr);
> + if (ret != 0) {
> + kfd_release_live_view();
> + return ret;
> + }
> + }
> +
> + sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
> + sub_type_hdr->length);
> + }
> +
> + sys_props.generation_count++;
> + topology_crat_parsed = 1;
> +
> + return 0;
> +}
> +
> +
> +#define sysfs_show_gen_prop(buffer, fmt, ...) \
> + snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__)
> +#define sysfs_show_32bit_prop(buffer, name, value) \
> + sysfs_show_gen_prop(buffer, "%s %u\n", name, value)
> +#define sysfs_show_64bit_prop(buffer, name, value) \
> + sysfs_show_gen_prop(buffer, "%s %llu\n", name, value)
> +#define sysfs_show_32bit_val(buffer, value) \
> + sysfs_show_gen_prop(buffer, "%u\n", value)
> +#define sysfs_show_str_val(buffer, value) \
> + sysfs_show_gen_prop(buffer, "%s\n", value)
> +
> +static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
> + char *buffer)
> +{
> + ssize_t ret;
> +
> + /* Making sure that the buffer is an empty string */
> + buffer[0] = 0;
> +
> + if (attr == &sys_props.attr_genid) {
> + ret = sysfs_show_32bit_val(buffer, sys_props.generation_count);
> + } else if (attr == &sys_props.attr_props) {
> + sysfs_show_64bit_prop(buffer, "platform_oem",
> + sys_props.platform_oem);
> + sysfs_show_64bit_prop(buffer, "platform_id",
> + sys_props.platform_id);
> + ret = sysfs_show_64bit_prop(buffer, "platform_rev",
> + sys_props.platform_rev);
> + } else {
> + ret = -EINVAL;
> + }
> +
> + return ret;
> +}
> +
> +static const struct sysfs_ops sysprops_ops = {
> + .show = sysprops_show,
> +};
> +
> +static struct kobj_type sysprops_type = {
> + .sysfs_ops = &sysprops_ops,
> +};
> +
> +static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
> + char *buffer)
> +{
> + ssize_t ret;
> + struct kfd_iolink_properties *iolink;
> +
> + /* Making sure that the buffer is an empty string */
> + buffer[0] = 0;
> +
> + iolink = container_of(attr, struct kfd_iolink_properties, attr);
> + sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type);
> + sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj);
> + sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min);
> + sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from);
> + sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to);
> + sysfs_show_32bit_prop(buffer, "weight", iolink->weight);
> + sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency);
> + sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency);
> + sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth);
> + sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth);
> + sysfs_show_32bit_prop(buffer, "recommended_transfer_size",
> + iolink->rec_transfer_size);
> + ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags);
> +
> + return ret;
> +}
> +
> +static const struct sysfs_ops iolink_ops = {
> + .show = iolink_show,
> +};
> +
> +static struct kobj_type iolink_type = {
> + .sysfs_ops = &iolink_ops,
> +};
> +
> +static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
> + char *buffer)
> +{
> + ssize_t ret;
> + struct kfd_mem_properties *mem;
> +
> + /* Making sure that the buffer is an empty string */
> + buffer[0] = 0;
> +
> + mem = container_of(attr, struct kfd_mem_properties, attr);
> + sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type);
> + sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes);
> + sysfs_show_32bit_prop(buffer, "flags", mem->flags);
> + sysfs_show_32bit_prop(buffer, "width", mem->width);
> + ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max);
> +
> + return ret;
> +}
> +
> +static const struct sysfs_ops mem_ops = {
> + .show = mem_show,
> +};
> +
> +static struct kobj_type mem_type = {
> + .sysfs_ops = &mem_ops,
> +};
> +
> +static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
> + char *buffer)
> +{
> + ssize_t ret;
> + uint32_t i;
> + struct kfd_cache_properties *cache;
> +
> + /* Making sure that the buffer is an empty string */
> + buffer[0] = 0;
> +
> + cache = container_of(attr, struct kfd_cache_properties, attr);
> + sysfs_show_32bit_prop(buffer, "processor_id_low",
> + cache->processor_id_low);
> + sysfs_show_32bit_prop(buffer, "level", cache->cache_level);
> + sysfs_show_32bit_prop(buffer, "size", cache->cache_size);
> + sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size);
> + sysfs_show_32bit_prop(buffer, "cache_lines_per_tag",
> + cache->cachelines_per_tag);
> + sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc);
> + sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency);
> + sysfs_show_32bit_prop(buffer, "type", cache->cache_type);
> + snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer);
> + for (i = 0; i < KFD_TOPOLOGY_CPU_SIBLINGS; i++)
> + ret = snprintf(buffer, PAGE_SIZE, "%s%d%s",
> + buffer, cache->sibling_map[i],
> + (i == KFD_TOPOLOGY_CPU_SIBLINGS-1) ?
> + "\n" : ",");
> +
> + return ret;
> +}
> +
> +static const struct sysfs_ops cache_ops = {
> + .show = kfd_cache_show,
> +};
> +
> +static struct kobj_type cache_type = {
> + .sysfs_ops = &cache_ops,
> +};
> +
> +static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
> + char *buffer)
> +{
> + ssize_t ret;
> + struct kfd_topology_device *dev;
> + char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE];
> + uint32_t i;
> +
> + /* Making sure that the buffer is an empty string */
> + buffer[0] = 0;
> +
> + if (strcmp(attr->name, "gpu_id") == 0) {
> + dev = container_of(attr, struct kfd_topology_device,
> + attr_gpuid);
> + ret = sysfs_show_32bit_val(buffer, dev->gpu_id);
> + } else if (strcmp(attr->name, "name") == 0) {
> + dev = container_of(attr, struct kfd_topology_device,
> + attr_name);
> + for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) {
> + public_name[i] =
> + (char)dev->node_props.marketing_name[i];
> + if (dev->node_props.marketing_name[i] == 0)
> + break;
> + }
> + public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0;
> + ret = sysfs_show_str_val(buffer, public_name);
> + } else {
> + dev = container_of(attr, struct kfd_topology_device,
> + attr_props);
> + sysfs_show_32bit_prop(buffer, "cpu_cores_count",
> + dev->node_props.cpu_cores_count);
> + sysfs_show_32bit_prop(buffer, "simd_count",
> + dev->node_props.simd_count);
> + sysfs_show_32bit_prop(buffer, "mem_banks_count",
> + dev->node_props.mem_banks_count);
> + sysfs_show_32bit_prop(buffer, "caches_count",
> + dev->node_props.caches_count);
> + sysfs_show_32bit_prop(buffer, "io_links_count",
> + dev->node_props.io_links_count);
> + sysfs_show_32bit_prop(buffer, "cpu_core_id_base",
> + dev->node_props.cpu_core_id_base);
> + sysfs_show_32bit_prop(buffer, "simd_id_base",
> + dev->node_props.simd_id_base);
> + sysfs_show_32bit_prop(buffer, "capability",
> + dev->node_props.capability);
> + sysfs_show_32bit_prop(buffer, "max_waves_per_simd",
> + dev->node_props.max_waves_per_simd);
> + sysfs_show_32bit_prop(buffer, "lds_size_in_kb",
> + dev->node_props.lds_size_in_kb);
> + sysfs_show_32bit_prop(buffer, "gds_size_in_kb",
> + dev->node_props.gds_size_in_kb);
> + sysfs_show_32bit_prop(buffer, "wave_front_size",
> + dev->node_props.wave_front_size);
> + sysfs_show_32bit_prop(buffer, "array_count",
> + dev->node_props.array_count);
> + sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine",
> + dev->node_props.simd_arrays_per_engine);
> + sysfs_show_32bit_prop(buffer, "cu_per_simd_array",
> + dev->node_props.cu_per_simd_array);
> + sysfs_show_32bit_prop(buffer, "simd_per_cu",
> + dev->node_props.simd_per_cu);
> + sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu",
> + dev->node_props.max_slots_scratch_cu);
> + sysfs_show_32bit_prop(buffer, "engine_id",
> + dev->node_props.engine_id);
> + sysfs_show_32bit_prop(buffer, "vendor_id",
> + dev->node_props.vendor_id);
> + sysfs_show_32bit_prop(buffer, "device_id",
> + dev->node_props.device_id);
> + sysfs_show_32bit_prop(buffer, "location_id",
> + dev->node_props.location_id);
> + sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute",
> + dev->node_props.max_engine_clk_fcompute);
> + ret = sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute",
> + dev->node_props.max_engine_clk_ccompute);
> + }
> +
> + return ret;
> +}
> +
> +static const struct sysfs_ops node_ops = {
> + .show = node_show,
> +};
> +
> +static struct kobj_type node_type = {
> + .sysfs_ops = &node_ops,
> +};
> +
> +static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
> +{
> + sysfs_remove_file(kobj, attr);
> + kobject_del(kobj);
> + kobject_put(kobj);
> +}
> +
> +static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
> +{
> + struct kfd_iolink_properties *iolink;
> + struct kfd_cache_properties *cache;
> + struct kfd_mem_properties *mem;
> +
> + BUG_ON(!dev);
> +
> + if (dev->kobj_iolink) {
> + list_for_each_entry(iolink, &dev->io_link_props, list)
> + if (iolink->kobj) {
> + kfd_remove_sysfs_file(iolink->kobj, &iolink->attr);
> + iolink->kobj = 0;
> + }
> + kobject_del(dev->kobj_iolink);
> + kobject_put(dev->kobj_iolink);
> + dev->kobj_iolink = 0;
> + }
> +
> + if (dev->kobj_cache) {
> + list_for_each_entry(cache, &dev->cache_props, list)
> + if (cache->kobj) {
> + kfd_remove_sysfs_file(cache->kobj, &cache->attr);
> + cache->kobj = 0;
> + }
> + kobject_del(dev->kobj_cache);
> + kobject_put(dev->kobj_cache);
> + dev->kobj_cache = 0;
> + }
> +
> + if (dev->kobj_mem) {
> + list_for_each_entry(mem, &dev->mem_props, list)
> + if (mem->kobj) {
> + kfd_remove_sysfs_file(mem->kobj, &mem->attr);
> + mem->kobj = 0;
> + }
> + kobject_del(dev->kobj_mem);
> + kobject_put(dev->kobj_mem);
> + dev->kobj_mem = 0;
> + }
> +
> + if (dev->kobj_node) {
> + sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
> + sysfs_remove_file(dev->kobj_node, &dev->attr_name);
> + sysfs_remove_file(dev->kobj_node, &dev->attr_props);
> + kobject_del(dev->kobj_node);
> + kobject_put(dev->kobj_node);
> + dev->kobj_node = 0;
> + }
> +}
> +
> +static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
> + uint32_t id)
> +{
> + struct kfd_iolink_properties *iolink;
> + struct kfd_cache_properties *cache;
> + struct kfd_mem_properties *mem;
> + int ret;
> + uint32_t i;
> +
> + BUG_ON(!dev);
> +
> + /*
> + * Creating the sysfs folders
> + */
> + BUG_ON(dev->kobj_node);
> + dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
> + if (!dev->kobj_node)
> + return -ENOMEM;
> +
> + ret = kobject_init_and_add(dev->kobj_node, &node_type,
> + sys_props.kobj_nodes, "%d", id);
> + if (ret < 0)
> + return ret;
> +
> + dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
> + if (!dev->kobj_mem)
> + return -ENOMEM;
> +
> + dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
> + if (!dev->kobj_cache)
> + return -ENOMEM;
> +
> + dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
> + if (!dev->kobj_iolink)
> + return -ENOMEM;
> +
> + /*
> + * Creating sysfs files for node properties
> + */
> + dev->attr_gpuid.name = "gpu_id";
> + dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
> + sysfs_attr_init(&dev->attr_gpuid);
> + dev->attr_name.name = "name";
> + dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
> + sysfs_attr_init(&dev->attr_name);
> + dev->attr_props.name = "properties";
> + dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
> + sysfs_attr_init(&dev->attr_props);
> + ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
> + if (ret < 0)
> + return ret;
> + ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
> + if (ret < 0)
> + return ret;
> + ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
> + if (ret < 0)
> + return ret;
> +
> + i = 0;
> + list_for_each_entry(mem, &dev->mem_props, list) {
> + mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
> + if (!mem->kobj)
> + return -ENOMEM;
> + ret = kobject_init_and_add(mem->kobj, &mem_type,
> + dev->kobj_mem, "%d", i);
> + if (ret < 0)
> + return ret;
> +
> + mem->attr.name = "properties";
> + mem->attr.mode = KFD_SYSFS_FILE_MODE;
> + sysfs_attr_init(&mem->attr);
> + ret = sysfs_create_file(mem->kobj, &mem->attr);
> + if (ret < 0)
> + return ret;
> + i++;
> + }
> +
> + i = 0;
> + list_for_each_entry(cache, &dev->cache_props, list) {
> + cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
> + if (!cache->kobj)
> + return -ENOMEM;
> + ret = kobject_init_and_add(cache->kobj, &cache_type,
> + dev->kobj_cache, "%d", i);
> + if (ret < 0)
> + return ret;
> +
> + cache->attr.name = "properties";
> + cache->attr.mode = KFD_SYSFS_FILE_MODE;
> + sysfs_attr_init(&cache->attr);
> + ret = sysfs_create_file(cache->kobj, &cache->attr);
> + if (ret < 0)
> + return ret;
> + i++;
> + }
> +
> + i = 0;
> + list_for_each_entry(iolink, &dev->io_link_props, list) {
> + iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
> + if (!iolink->kobj)
> + return -ENOMEM;
> + ret = kobject_init_and_add(iolink->kobj, &iolink_type,
> + dev->kobj_iolink, "%d", i);
> + if (ret < 0)
> + return ret;
> +
> + iolink->attr.name = "properties";
> + iolink->attr.mode = KFD_SYSFS_FILE_MODE;
> + sysfs_attr_init(&iolink->attr);
> + ret = sysfs_create_file(iolink->kobj, &iolink->attr);
> + if (ret < 0)
> + return ret;
> + i++;
> +}
> +
> + return 0;
> +}
> +
> +static int kfd_build_sysfs_node_tree(void)
> +{
> + struct kfd_topology_device *dev;
> + int ret;
> + uint32_t i = 0;
> +
> + list_for_each_entry(dev, &topology_device_list, list) {
> + ret = kfd_build_sysfs_node_entry(dev, 0);
> + if (ret < 0)
> + return ret;
> + i++;
> + }
> +
> + return 0;
> +}
> +
> +static void kfd_remove_sysfs_node_tree(void)
> +{
> + struct kfd_topology_device *dev;
> +
> + list_for_each_entry(dev, &topology_device_list, list)
> + kfd_remove_sysfs_node_entry(dev);
> +}
> +
> +static int kfd_topology_update_sysfs(void)
> +{
> + int ret;
> +
> + pr_info("Creating topology SYSFS entries\n");
> + if (sys_props.kobj_topology == 0) {
> + sys_props.kobj_topology = kfd_alloc_struct(sys_props.kobj_topology);
> + if (!sys_props.kobj_topology)
> + return -ENOMEM;
> +
> + ret = kobject_init_and_add(sys_props.kobj_topology,
> + &sysprops_type, &kfd_device->kobj,
> + "topology");
> + if (ret < 0)
> + return ret;
> +
> + sys_props.kobj_nodes = kobject_create_and_add("nodes",
> + sys_props.kobj_topology);
> + if (!sys_props.kobj_nodes)
> + return -ENOMEM;
> +
> + sys_props.attr_genid.name = "generation_id";
> + sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
> + sysfs_attr_init(&sys_props.attr_genid);
> + ret = sysfs_create_file(sys_props.kobj_topology,
> + &sys_props.attr_genid);
> + if (ret < 0)
> + return ret;
> +
> + sys_props.attr_props.name = "system_properties";
> + sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
> + sysfs_attr_init(&sys_props.attr_props);
> + ret = sysfs_create_file(sys_props.kobj_topology,
> + &sys_props.attr_props);
> + if (ret < 0)
> + return ret;
> + }
> +
> + kfd_remove_sysfs_node_tree();
> +
> + return kfd_build_sysfs_node_tree();
> +}
> +
> +static void kfd_topology_release_sysfs(void)
> +{
> + kfd_remove_sysfs_node_tree();
> + if (sys_props.kobj_topology) {
> + sysfs_remove_file(sys_props.kobj_topology,
> + &sys_props.attr_genid);
> + sysfs_remove_file(sys_props.kobj_topology,
> + &sys_props.attr_props);
> + if (sys_props.kobj_nodes) {
> + kobject_del(sys_props.kobj_nodes);
> + kobject_put(sys_props.kobj_nodes);
> + sys_props.kobj_nodes = 0;
> + }
> + kobject_del(sys_props.kobj_topology);
> + kobject_put(sys_props.kobj_topology);
> + sys_props.kobj_topology = 0;
> + }
> +}
> +
> +int kfd_topology_init(void)
> +{
> + void *crat_image = 0;
> + size_t image_size = 0;
> + int ret;
> +
> + /*
> + * Initialize the head for the topology device list
> + */
> + INIT_LIST_HEAD(&topology_device_list);
> + init_rwsem(&topology_lock);
> + topology_crat_parsed = 0;
> +
> + memset(&sys_props, 0, sizeof(sys_props));
> +
> + /*
> + * Get the CRAT image from the ACPI
> + */
> + ret = kfd_topology_get_crat_acpi(crat_image, &image_size);
> + if (ret == 0 && image_size > 0) {
> + pr_info("Found CRAT image with size=%zd\n", image_size);
> + crat_image = kmalloc(image_size, GFP_KERNEL);
> + if (!crat_image) {
> + ret = -ENOMEM;
> + pr_err("No memory for allocating CRAT image\n");
> + goto err;
> + }
> + ret = kfd_topology_get_crat_acpi(crat_image, &image_size);
> +
> + if (ret == 0) {
> + down_write(&topology_lock);
> + ret = kfd_parse_crat_table(crat_image);
> + if (ret == 0)
> + ret = kfd_topology_update_sysfs();
> + up_write(&topology_lock);
> + } else {
> + pr_err("Couldn't get CRAT table size from ACPI\n");
> + }
> + kfree(crat_image);
> + } else if (ret == -ENODATA) {
> + ret = 0;
> + } else {
> + pr_err("Couldn't get CRAT table size from ACPI\n");
> + }
> +
> +err:
> + pr_info("Finished initializing topology ret=%d\n", ret);
> + return ret;
> +}
> +
> +void kfd_topology_shutdown(void)
> +{
> + kfd_topology_release_sysfs();
> + kfd_release_live_view();
> +}
> +
> +static void kfd_debug_print_topology(void)
> +{
> + struct kfd_topology_device *dev;
> + uint32_t i = 0;
> +
> + pr_info("DEBUG PRINT OF TOPOLOGY:");
> + list_for_each_entry(dev, &topology_device_list, list) {
> + pr_info("Node: %d\n", i);
> + pr_info("\tGPU assigned: %s\n", (dev->gpu ? "yes" : "no"));
> + pr_info("\tCPU count: %d\n", dev->node_props.cpu_cores_count);
> + pr_info("\tSIMD count: %d", dev->node_props.simd_count);
> + i++;
> + }
> +}
> +
> +static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
> +{
> + uint32_t hashout;
> + uint32_t buf[7];
> + int i;
> +
> + if (!gpu)
> + return 0;
> +
> + buf[0] = gpu->pdev->devfn;
> + buf[1] = gpu->pdev->subsystem_vendor;
> + buf[2] = gpu->pdev->subsystem_device;
> + buf[3] = gpu->pdev->device;
> + buf[4] = gpu->pdev->bus->number;
> + buf[5] = (uint32_t)(kfd2kgd->get_vmem_size(gpu->kgd) & 0xffffffff);
> + buf[6] = (uint32_t)(kfd2kgd->get_vmem_size(gpu->kgd) >> 32);
> +
> + for (i = 0, hashout = 0; i < 7; i++)
> + hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
> +
> + return hashout;
> +}
> +
> +static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
> +{
> + struct kfd_topology_device *dev;
> + struct kfd_topology_device *out_dev = 0;
> +
> + BUG_ON(!gpu);
> +
> + list_for_each_entry(dev, &topology_device_list, list)
> + if (dev->gpu == 0 && dev->node_props.simd_count > 0) {
> + dev->gpu = gpu;
> + out_dev = dev;
> + break;
> + }
> +
> + return out_dev;
> +}
> +
> +static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
> +{
> + /*
> + * TODO: Generate an event for thunk about the arrival/removal
> + * of the GPU
> + */
> +}
> +
> +int kfd_topology_add_device(struct kfd_dev *gpu)
> +{
> + uint32_t gpu_id;
> + struct kfd_topology_device *dev;
> + int res;
> +
> + BUG_ON(!gpu);
> +
> + gpu_id = kfd_generate_gpu_id(gpu);
> +
> + pr_info("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
> +
> + down_write(&topology_lock);
> + /*
> + * Try to assign the GPU to existing topology device (generated from
> + * CRAT table
> + */
> + dev = kfd_assign_gpu(gpu);
> + if (!dev) {
> + pr_info("GPU was not found in the current topology. Extending.\n");
> + kfd_debug_print_topology();
> + dev = kfd_create_topology_device();
> + if (!dev) {
> + res = -ENOMEM;
> + goto err;
> + }
> + dev->gpu = gpu;
> +
> + /*
> + * TODO: Make a call to retrieve topology information from the
> + * GPU vBIOS
> + */
> +
> + /*
> + * Update the SYSFS tree, since we added another topology device
> + */
> + if (kfd_topology_update_sysfs() < 0)
> + kfd_topology_release_sysfs();
> +
> + }
> +
> + dev->gpu_id = gpu_id;
> + gpu->id = gpu_id;
> + dev->node_props.vendor_id = gpu->pdev->vendor;
> + dev->node_props.device_id = gpu->pdev->device;
> + dev->node_props.location_id = (gpu->pdev->bus->number << 24) +
> + (gpu->pdev->devfn & 0xffffff);
> + /*
> + * TODO: Retrieve max engine clock values from KGD
> + */
> +
> + res = 0;
> +
> +err:
> + up_write(&topology_lock);
> +
> + if (res == 0)
> + kfd_notify_gpu_change(gpu_id, 1);
> +
> + return res;
> +}
> +
> +int kfd_topology_remove_device(struct kfd_dev *gpu)
> +{
> + struct kfd_topology_device *dev;
> + uint32_t gpu_id;
> + int res = -ENODEV;
> +
> + BUG_ON(!gpu);
> +
> + down_write(&topology_lock);
> +
> + list_for_each_entry(dev, &topology_device_list, list)
> + if (dev->gpu == gpu) {
> + gpu_id = dev->gpu_id;
> + kfd_remove_sysfs_node_entry(dev);
> + kfd_release_topology_device(dev);
> + res = 0;
> + if (kfd_topology_update_sysfs() < 0)
> + kfd_topology_release_sysfs();
> + break;
> + }
> +
> + up_write(&topology_lock);
> +
> + if (res == 0)
> + kfd_notify_gpu_change(gpu_id, 0);
> +
> + return res;
> +}
I am not convince that sysfs is the right place to expose this.
I need to think on that a bit.
> diff --git a/drivers/gpu/hsa/radeon/kfd_topology.h b/drivers/gpu/hsa/radeon/kfd_topology.h
> new file mode 100644
> index 0000000..989624b
> --- /dev/null
> +++ b/drivers/gpu/hsa/radeon/kfd_topology.h
> @@ -0,0 +1,168 @@
> +/*
> + * Copyright 2014 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#ifndef __KFD_TOPOLOGY_H__
> +#define __KFD_TOPOLOGY_H__
> +
> +#include <linux/types.h>
> +#include <linux/list.h>
> +#include "kfd_priv.h"
> +
> +#define KFD_TOPOLOGY_PUBLIC_NAME_SIZE 128
> +
> +#define HSA_CAP_HOT_PLUGGABLE 0x00000001
> +#define HSA_CAP_ATS_PRESENT 0x00000002
> +#define HSA_CAP_SHARED_WITH_GRAPHICS 0x00000004
> +#define HSA_CAP_QUEUE_SIZE_POW2 0x00000008
> +#define HSA_CAP_QUEUE_SIZE_32BIT 0x00000010
> +#define HSA_CAP_QUEUE_IDLE_EVENT 0x00000020
> +#define HSA_CAP_VA_LIMIT 0x00000040
> +#define HSA_CAP_WATCH_POINTS_SUPPORTED 0x00000080
> +#define HSA_CAP_WATCH_POINTS_TOTALBITS_MASK 0x00000f00
> +#define HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT 8
> +#define HSA_CAP_RESERVED 0xfffff000
> +
> +struct kfd_node_properties {
> + uint32_t cpu_cores_count;
> + uint32_t simd_count;
> + uint32_t mem_banks_count;
> + uint32_t caches_count;
> + uint32_t io_links_count;
> + uint32_t cpu_core_id_base;
> + uint32_t simd_id_base;
> + uint32_t capability;
> + uint32_t max_waves_per_simd;
> + uint32_t lds_size_in_kb;
> + uint32_t gds_size_in_kb;
> + uint32_t wave_front_size;
> + uint32_t array_count;
> + uint32_t simd_arrays_per_engine;
> + uint32_t cu_per_simd_array;
> + uint32_t simd_per_cu;
> + uint32_t max_slots_scratch_cu;
> + uint32_t engine_id;
> + uint32_t vendor_id;
> + uint32_t device_id;
> + uint32_t location_id;
> + uint32_t max_engine_clk_fcompute;
> + uint32_t max_engine_clk_ccompute;
> + uint16_t marketing_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE];
> +};
> +
> +#define HSA_MEM_HEAP_TYPE_SYSTEM 0
> +#define HSA_MEM_HEAP_TYPE_FB_PUBLIC 1
> +#define HSA_MEM_HEAP_TYPE_FB_PRIVATE 2
> +#define HSA_MEM_HEAP_TYPE_GPU_GDS 3
> +#define HSA_MEM_HEAP_TYPE_GPU_LDS 4
> +#define HSA_MEM_HEAP_TYPE_GPU_SCRATCH 5
> +
> +#define HSA_MEM_FLAGS_HOT_PLUGGABLE 0x00000001
> +#define HSA_MEM_FLAGS_NON_VOLATILE 0x00000002
> +#define HSA_MEM_FLAGS_RESERVED 0xfffffffc
> +
> +struct kfd_mem_properties {
> + struct list_head list;
> + uint32_t heap_type;
> + uint64_t size_in_bytes;
> + uint32_t flags;
> + uint32_t width;
> + uint32_t mem_clk_max;
> + struct kobject *kobj;
> + struct attribute attr;
> +};
> +
> +#define KFD_TOPOLOGY_CPU_SIBLINGS 256
> +
> +#define HSA_CACHE_TYPE_DATA 0x00000001
> +#define HSA_CACHE_TYPE_INSTRUCTION 0x00000002
> +#define HSA_CACHE_TYPE_CPU 0x00000004
> +#define HSA_CACHE_TYPE_HSACU 0x00000008
> +#define HSA_CACHE_TYPE_RESERVED 0xfffffff0
> +
> +struct kfd_cache_properties {
> + struct list_head list;
> + uint32_t processor_id_low;
> + uint32_t cache_level;
> + uint32_t cache_size;
> + uint32_t cacheline_size;
> + uint32_t cachelines_per_tag;
> + uint32_t cache_assoc;
> + uint32_t cache_latency;
> + uint32_t cache_type;
> + uint8_t sibling_map[KFD_TOPOLOGY_CPU_SIBLINGS];
> + struct kobject *kobj;
> + struct attribute attr;
> +};
> +
> +struct kfd_iolink_properties {
> + struct list_head list;
> + uint32_t iolink_type;
> + uint32_t ver_maj;
> + uint32_t ver_min;
> + uint32_t node_from;
> + uint32_t node_to;
> + uint32_t weight;
> + uint32_t min_latency;
> + uint32_t max_latency;
> + uint32_t min_bandwidth;
> + uint32_t max_bandwidth;
> + uint32_t rec_transfer_size;
> + uint32_t flags;
> + struct kobject *kobj;
> + struct attribute attr;
> +};
> +
> +struct kfd_topology_device {
> + struct list_head list;
> + uint32_t gpu_id;
> + struct kfd_node_properties node_props;
> + uint32_t mem_bank_count;
> + struct list_head mem_props;
> + uint32_t cache_count;
> + struct list_head cache_props;
> + uint32_t io_link_count;
> + struct list_head io_link_props;
> + struct kfd_dev *gpu;
> + struct kobject *kobj_node;
> + struct kobject *kobj_mem;
> + struct kobject *kobj_cache;
> + struct kobject *kobj_iolink;
> + struct attribute attr_gpuid;
> + struct attribute attr_name;
> + struct attribute attr_props;
> +};
> +
> +struct kfd_system_properties {
> + uint32_t num_devices; /* Number of H-NUMA nodes */
> + uint32_t generation_count;
> + uint64_t platform_oem;
> + uint64_t platform_id;
> + uint64_t platform_rev;
> + struct kobject *kobj_topology;
> + struct kobject *kobj_nodes;
> + struct attribute attr_genid;
> + struct attribute attr_props;
> +};
> +
> +
> +
> +#endif /* __KFD_TOPOLOGY_H__ */
> --
> 1.9.1
>
More information about the dri-devel
mailing list