diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdkfd/kfd_topology.c')
| -rw-r--r-- | drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 2413 |
1 files changed, 1779 insertions, 634 deletions
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index 1e5064749959..811636af14ea 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -1,5 +1,6 @@ +// SPDX-License-Identifier: GPL-2.0 OR MIT /* - * Copyright 2014 Advanced Micro Devices, Inc. + * Copyright 2014-2022 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"), @@ -28,347 +29,93 @@ #include <linux/hash.h> #include <linux/cpufreq.h> #include <linux/log2.h> +#include <linux/dmi.h> +#include <linux/atomic.h> +#include <linux/crc16.h> #include "kfd_priv.h" #include "kfd_crat.h" #include "kfd_topology.h" - +#include "kfd_device_queue_manager.h" +#include "kfd_svm.h" +#include "kfd_debug.h" +#include "amdgpu_amdkfd.h" +#include "amdgpu_ras.h" +#include "amdgpu.h" + +/* topology_device_list - Master list of all topology devices */ 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 uint32_t topology_crat_proximity_domain; -struct kfd_dev *kfd_device_by_id(uint32_t gpu_id) +struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock( + uint32_t proximity_domain) { struct kfd_topology_device *top_dev; - struct kfd_dev *device = NULL; - - down_read(&topology_lock); + struct kfd_topology_device *device = NULL; list_for_each_entry(top_dev, &topology_device_list, list) - if (top_dev->gpu_id == gpu_id) { - device = top_dev->gpu; + if (top_dev->proximity_domain == proximity_domain) { + device = top_dev; break; } - up_read(&topology_lock); - return device; } -struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev) +struct kfd_topology_device *kfd_topology_device_by_proximity_domain( + uint32_t proximity_domain) { - struct kfd_topology_device *top_dev; - struct kfd_dev *device = NULL; + struct kfd_topology_device *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; - } - + device = kfd_topology_device_by_proximity_domain_no_lock( + proximity_domain); 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; - } - - if (*size >= crat_table->length && crat_image != NULL) - 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_topology_device *kfd_topology_device_by_id(uint32_t gpu_id) { - struct kfd_mem_properties *props; - struct kfd_topology_device *dev; - int i = 0; + struct kfd_topology_device *top_dev = NULL; + struct kfd_topology_device *ret = NULL; - 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 == NULL) - 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 == NULL) - 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); + down_read(&topology_lock); + list_for_each_entry(top_dev, &topology_device_list, list) + if (top_dev->gpu_id == gpu_id) { + ret = top_dev; 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 == NULL) - 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++; - } + up_read(&topology_lock); - return 0; + return ret; } -static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr) +struct kfd_node *kfd_device_by_id(uint32_t gpu_id) { - 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); + struct kfd_topology_device *top_dev; - 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); - } + top_dev = kfd_topology_device_by_id(gpu_id); + if (!top_dev) + return NULL; - return ret; + return top_dev->gpu; } +/* Called with write topology_lock acquired */ 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); + struct kfd_iolink_properties *p2plink; + struct kfd_perf_properties *perf; list_del(&dev->list); @@ -393,30 +140,47 @@ static void kfd_release_topology_device(struct kfd_topology_device *dev) kfree(iolink); } - kfree(dev); + while (dev->p2p_link_props.next != &dev->p2p_link_props) { + p2plink = container_of(dev->p2p_link_props.next, + struct kfd_iolink_properties, list); + list_del(&p2plink->list); + kfree(p2plink); + } + + while (dev->perf_props.next != &dev->perf_props) { + perf = container_of(dev->perf_props.next, + struct kfd_perf_properties, list); + list_del(&perf->list); + kfree(perf); + } - sys_props.num_devices--; + kfree(dev); } -static void kfd_release_live_view(void) +void kfd_release_topology_device_list(struct list_head *device_list) { 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); + while (!list_empty(device_list)) { + dev = list_first_entry(device_list, + struct kfd_topology_device, list); kfd_release_topology_device(dev); + } } +static void kfd_release_live_view(void) +{ + kfd_release_topology_device_list(&topology_device_list); memset(&sys_props, 0, sizeof(sys_props)); } -static struct kfd_topology_device *kfd_create_topology_device(void) +struct kfd_topology_device *kfd_create_topology_device( + struct list_head *device_list) { struct kfd_topology_device *dev; dev = kfd_alloc_struct(dev); - if (dev == NULL) { + if (!dev) { pr_err("No memory to allocate a topology device"); return NULL; } @@ -424,216 +188,224 @@ static struct kfd_topology_device *kfd_create_topology_device(void) INIT_LIST_HEAD(&dev->mem_props); INIT_LIST_HEAD(&dev->cache_props); INIT_LIST_HEAD(&dev->io_link_props); + INIT_LIST_HEAD(&dev->p2p_link_props); + INIT_LIST_HEAD(&dev->perf_props); - list_add_tail(&dev->list, &topology_device_list); - sys_props.num_devices++; + list_add_tail(&dev->list, device_list); 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)) & CRAT_OEMID_64BIT_MASK; - 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) +#define sysfs_show_gen_prop(buffer, offs, fmt, ...) \ + (offs += snprintf(buffer+offs, PAGE_SIZE-offs, \ + fmt, __VA_ARGS__)) +#define sysfs_show_32bit_prop(buffer, offs, name, value) \ + sysfs_show_gen_prop(buffer, offs, "%s %u\n", name, value) +#define sysfs_show_64bit_prop(buffer, offs, name, value) \ + sysfs_show_gen_prop(buffer, offs, "%s %llu\n", name, value) +#define sysfs_show_32bit_val(buffer, offs, value) \ + sysfs_show_gen_prop(buffer, offs, "%u\n", value) +#define sysfs_show_str_val(buffer, offs, value) \ + sysfs_show_gen_prop(buffer, offs, "%s\n", value) static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr, char *buffer) { - ssize_t ret; + int offs = 0; /* 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); + sysfs_show_32bit_val(buffer, offs, + 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); + sysfs_show_64bit_prop(buffer, offs, "platform_oem", + sys_props.platform_oem); + sysfs_show_64bit_prop(buffer, offs, "platform_id", + sys_props.platform_id); + sysfs_show_64bit_prop(buffer, offs, "platform_rev", + sys_props.platform_rev); } else { - ret = -EINVAL; + offs = -EINVAL; } - return ret; + return offs; +} + +static void kfd_topology_kobj_release(struct kobject *kobj) +{ + kfree(kobj); } static const struct sysfs_ops sysprops_ops = { .show = sysprops_show, }; -static struct kobj_type sysprops_type = { +static const struct kobj_type sysprops_type = { + .release = kfd_topology_kobj_release, .sysfs_ops = &sysprops_ops, }; static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr, char *buffer) { - ssize_t ret; + int offs = 0; 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; + if (iolink->gpu && kfd_devcgroup_check_permission(iolink->gpu)) + return -EPERM; + sysfs_show_32bit_prop(buffer, offs, "type", iolink->iolink_type); + sysfs_show_32bit_prop(buffer, offs, "version_major", iolink->ver_maj); + sysfs_show_32bit_prop(buffer, offs, "version_minor", iolink->ver_min); + sysfs_show_32bit_prop(buffer, offs, "node_from", iolink->node_from); + sysfs_show_32bit_prop(buffer, offs, "node_to", iolink->node_to); + sysfs_show_32bit_prop(buffer, offs, "weight", iolink->weight); + sysfs_show_32bit_prop(buffer, offs, "min_latency", iolink->min_latency); + sysfs_show_32bit_prop(buffer, offs, "max_latency", iolink->max_latency); + sysfs_show_32bit_prop(buffer, offs, "min_bandwidth", + iolink->min_bandwidth); + sysfs_show_32bit_prop(buffer, offs, "max_bandwidth", + iolink->max_bandwidth); + sysfs_show_32bit_prop(buffer, offs, "recommended_transfer_size", + iolink->rec_transfer_size); + sysfs_show_32bit_prop(buffer, offs, "recommended_sdma_engine_id_mask", + iolink->rec_sdma_eng_id_mask); + sysfs_show_32bit_prop(buffer, offs, "flags", iolink->flags); + + return offs; } static const struct sysfs_ops iolink_ops = { .show = iolink_show, }; -static struct kobj_type iolink_type = { +static const struct kobj_type iolink_type = { + .release = kfd_topology_kobj_release, .sysfs_ops = &iolink_ops, }; static ssize_t mem_show(struct kobject *kobj, struct attribute *attr, char *buffer) { - ssize_t ret; + int offs = 0; 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; + if (mem->gpu && kfd_devcgroup_check_permission(mem->gpu)) + return -EPERM; + sysfs_show_32bit_prop(buffer, offs, "heap_type", mem->heap_type); + sysfs_show_64bit_prop(buffer, offs, "size_in_bytes", + mem->size_in_bytes); + sysfs_show_32bit_prop(buffer, offs, "flags", mem->flags); + sysfs_show_32bit_prop(buffer, offs, "width", mem->width); + sysfs_show_32bit_prop(buffer, offs, "mem_clk_max", + mem->mem_clk_max); + + return offs; } static const struct sysfs_ops mem_ops = { .show = mem_show, }; -static struct kobj_type mem_type = { +static const struct kobj_type mem_type = { + .release = kfd_topology_kobj_release, .sysfs_ops = &mem_ops, }; static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr, char *buffer) { - ssize_t ret; - uint32_t i; + int offs = 0; + uint32_t i, j; 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", + if (cache->gpu && kfd_devcgroup_check_permission(cache->gpu)) + return -EPERM; + sysfs_show_32bit_prop(buffer, offs, "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; + sysfs_show_32bit_prop(buffer, offs, "level", cache->cache_level); + sysfs_show_32bit_prop(buffer, offs, "size", cache->cache_size); + sysfs_show_32bit_prop(buffer, offs, "cache_line_size", + cache->cacheline_size); + sysfs_show_32bit_prop(buffer, offs, "cache_lines_per_tag", + cache->cachelines_per_tag); + sysfs_show_32bit_prop(buffer, offs, "association", cache->cache_assoc); + sysfs_show_32bit_prop(buffer, offs, "latency", cache->cache_latency); + sysfs_show_32bit_prop(buffer, offs, "type", cache->cache_type); + + offs += snprintf(buffer+offs, PAGE_SIZE-offs, "sibling_map "); + for (i = 0; i < cache->sibling_map_size; i++) + for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++) + /* Check each bit */ + offs += snprintf(buffer+offs, PAGE_SIZE-offs, "%d,", + (cache->sibling_map[i] >> j) & 1); + + /* Replace the last "," with end of line */ + buffer[offs-1] = '\n'; + return offs; } static const struct sysfs_ops cache_ops = { .show = kfd_cache_show, }; -static struct kobj_type cache_type = { +static const struct kobj_type cache_type = { + .release = kfd_topology_kobj_release, .sysfs_ops = &cache_ops, }; +/****** Sysfs of Performance Counters ******/ + +struct kfd_perf_attr { + struct kobj_attribute attr; + uint32_t data; +}; + +static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs, + char *buf) +{ + int offs = 0; + struct kfd_perf_attr *attr; + + buf[0] = 0; + attr = container_of(attrs, struct kfd_perf_attr, attr); + if (!attr->data) /* invalid data for PMC */ + return 0; + else + return sysfs_show_32bit_val(buf, offs, attr->data); +} + +#define KFD_PERF_DESC(_name, _data) \ +{ \ + .attr = __ATTR(_name, 0444, perf_show, NULL), \ + .data = _data, \ +} + +static struct kfd_perf_attr perf_attr_iommu[] = { + KFD_PERF_DESC(max_concurrent, 0), + KFD_PERF_DESC(num_counters, 0), + KFD_PERF_DESC(counter_ids, 0), +}; +/****************************************/ + static ssize_t node_show(struct kobject *kobj, struct attribute *attr, char *buffer) { + int offs = 0; struct kfd_topology_device *dev; - char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE]; - uint32_t i; uint32_t log_max_watch_addr; /* Making sure that the buffer is an empty string */ @@ -642,76 +414,87 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr, if (strcmp(attr->name, "gpu_id") == 0) { dev = container_of(attr, struct kfd_topology_device, attr_gpuid); - return sysfs_show_32bit_val(buffer, dev->gpu_id); + if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu)) + return -EPERM; + return sysfs_show_32bit_val(buffer, offs, dev->gpu_id); } 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; - return sysfs_show_str_val(buffer, public_name); + + if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu)) + return -EPERM; + return sysfs_show_str_val(buffer, offs, dev->node_props.name); } 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); - - if (dev->mem_bank_count < dev->node_props.mem_banks_count) { - pr_info_once("kfd: mem_banks_count truncated from %d to %d\n", - dev->node_props.mem_banks_count, - dev->mem_bank_count); - sysfs_show_32bit_prop(buffer, "mem_banks_count", - dev->mem_bank_count); - } else { - 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, "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, "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); + if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu)) + return -EPERM; + sysfs_show_32bit_prop(buffer, offs, "cpu_cores_count", + dev->node_props.cpu_cores_count); + sysfs_show_32bit_prop(buffer, offs, "simd_count", + dev->gpu ? dev->node_props.simd_count : 0); + sysfs_show_32bit_prop(buffer, offs, "mem_banks_count", + dev->node_props.mem_banks_count); + sysfs_show_32bit_prop(buffer, offs, "caches_count", + dev->node_props.caches_count); + sysfs_show_32bit_prop(buffer, offs, "io_links_count", + dev->node_props.io_links_count); + sysfs_show_32bit_prop(buffer, offs, "p2p_links_count", + dev->node_props.p2p_links_count); + sysfs_show_32bit_prop(buffer, offs, "cpu_core_id_base", + dev->node_props.cpu_core_id_base); + sysfs_show_32bit_prop(buffer, offs, "simd_id_base", + dev->node_props.simd_id_base); + sysfs_show_32bit_prop(buffer, offs, "max_waves_per_simd", + dev->node_props.max_waves_per_simd); + sysfs_show_32bit_prop(buffer, offs, "lds_size_in_kb", + dev->node_props.lds_size_in_kb); + sysfs_show_32bit_prop(buffer, offs, "gds_size_in_kb", + dev->node_props.gds_size_in_kb); + sysfs_show_32bit_prop(buffer, offs, "num_gws", + dev->node_props.num_gws); + sysfs_show_32bit_prop(buffer, offs, "wave_front_size", + dev->node_props.wave_front_size); + sysfs_show_32bit_prop(buffer, offs, "array_count", + dev->gpu ? (dev->node_props.array_count * + NUM_XCC(dev->gpu->xcc_mask)) : 0); + sysfs_show_32bit_prop(buffer, offs, "simd_arrays_per_engine", + dev->node_props.simd_arrays_per_engine); + sysfs_show_32bit_prop(buffer, offs, "cu_per_simd_array", + dev->node_props.cu_per_simd_array); + sysfs_show_32bit_prop(buffer, offs, "simd_per_cu", + dev->node_props.simd_per_cu); + sysfs_show_32bit_prop(buffer, offs, "max_slots_scratch_cu", + dev->node_props.max_slots_scratch_cu); + sysfs_show_32bit_prop(buffer, offs, "gfx_target_version", + dev->node_props.gfx_target_version); + sysfs_show_32bit_prop(buffer, offs, "vendor_id", + dev->node_props.vendor_id); + sysfs_show_32bit_prop(buffer, offs, "device_id", + dev->node_props.device_id); + sysfs_show_32bit_prop(buffer, offs, "location_id", + dev->node_props.location_id); + sysfs_show_32bit_prop(buffer, offs, "domain", + dev->node_props.domain); + sysfs_show_32bit_prop(buffer, offs, "drm_render_minor", + dev->node_props.drm_render_minor); + sysfs_show_64bit_prop(buffer, offs, "hive_id", + dev->node_props.hive_id); + sysfs_show_32bit_prop(buffer, offs, "num_sdma_engines", + dev->node_props.num_sdma_engines); + sysfs_show_32bit_prop(buffer, offs, "num_sdma_xgmi_engines", + dev->node_props.num_sdma_xgmi_engines); + sysfs_show_32bit_prop(buffer, offs, "num_sdma_queues_per_engine", + dev->node_props.num_sdma_queues_per_engine); + sysfs_show_32bit_prop(buffer, offs, "num_cp_queues", + dev->node_props.num_cp_queues); if (dev->gpu) { log_max_watch_addr = - __ilog2_u32(dev->gpu->device_info->num_of_watch_points); + __ilog2_u32(dev->gpu->kfd->device_info.num_of_watch_points); if (log_max_watch_addr) { dev->node_props.capability |= @@ -723,30 +506,49 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr, HSA_CAP_WATCH_POINTS_TOTALBITS_MASK); } - sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute", - dev->gpu->kfd2kgd->get_max_engine_clock_in_mhz( - dev->gpu->kgd)); - - sysfs_show_64bit_prop(buffer, "local_mem_size", - (unsigned long long int) 0); - - sysfs_show_32bit_prop(buffer, "fw_version", - dev->gpu->kfd2kgd->get_fw_version( - dev->gpu->kgd, - KGD_ENGINE_MEC1)); - sysfs_show_32bit_prop(buffer, "capability", - dev->node_props.capability); + if (dev->gpu->adev->asic_type == CHIP_TONGA) + dev->node_props.capability |= + HSA_CAP_AQL_QUEUE_DOUBLE_MAP; + + if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(10, 0, 0) && + (dev->gpu->adev->sdma.supported_reset & AMDGPU_RESET_TYPE_PER_QUEUE)) + dev->node_props.capability2 |= HSA_CAP2_PER_SDMA_QUEUE_RESET_SUPPORTED; + + sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_fcompute", + dev->node_props.max_engine_clk_fcompute); + + sysfs_show_64bit_prop(buffer, offs, "local_mem_size", 0ULL); + + sysfs_show_32bit_prop(buffer, offs, "fw_version", + dev->gpu->kfd->mec_fw_version); + sysfs_show_32bit_prop(buffer, offs, "capability", + dev->node_props.capability); + sysfs_show_32bit_prop(buffer, offs, "capability2", + dev->node_props.capability2); + sysfs_show_64bit_prop(buffer, offs, "debug_prop", + dev->node_props.debug_prop); + sysfs_show_32bit_prop(buffer, offs, "sdma_fw_version", + dev->gpu->kfd->sdma_fw_version); + sysfs_show_64bit_prop(buffer, offs, "unique_id", + dev->gpu->xcp && + (dev->gpu->xcp->xcp_mgr->mode != + AMDGPU_SPX_PARTITION_MODE) ? + dev->gpu->xcp->unique_id : + dev->gpu->adev->unique_id); + sysfs_show_32bit_prop(buffer, offs, "num_xcc", + NUM_XCC(dev->gpu->xcc_mask)); } - return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute", - cpufreq_quick_get_max(0)/1000); + return sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_ccompute", + cpufreq_quick_get_max(0)/1000); } static const struct sysfs_ops node_ops = { .show = node_show, }; -static struct kobj_type node_type = { +static const struct kobj_type node_type = { + .release = kfd_topology_kobj_release, .sysfs_ops = &node_ops, }; @@ -759,11 +561,11 @@ static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr) static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev) { + struct kfd_iolink_properties *p2plink; struct kfd_iolink_properties *iolink; struct kfd_cache_properties *cache; struct kfd_mem_properties *mem; - - BUG_ON(!dev); + struct kfd_perf_properties *perf; if (dev->kobj_iolink) { list_for_each_entry(iolink, &dev->io_link_props, list) @@ -777,6 +579,18 @@ static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev) dev->kobj_iolink = NULL; } + if (dev->kobj_p2plink) { + list_for_each_entry(p2plink, &dev->p2p_link_props, list) + if (p2plink->kobj) { + kfd_remove_sysfs_file(p2plink->kobj, + &p2plink->attr); + p2plink->kobj = NULL; + } + kobject_del(dev->kobj_p2plink); + kobject_put(dev->kobj_p2plink); + dev->kobj_p2plink = NULL; + } + if (dev->kobj_cache) { list_for_each_entry(cache, &dev->cache_props, list) if (cache->kobj) { @@ -800,6 +614,16 @@ static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev) dev->kobj_mem = NULL; } + if (dev->kobj_perf) { + list_for_each_entry(perf, &dev->perf_props, list) { + kfree(perf->attr_group); + perf->attr_group = NULL; + } + kobject_del(dev->kobj_perf); + kobject_put(dev->kobj_perf); + dev->kobj_perf = NULL; + } + if (dev->kobj_node) { sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid); sysfs_remove_file(dev->kobj_node, &dev->attr_name); @@ -813,26 +637,31 @@ static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev) static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev, uint32_t id) { + struct kfd_iolink_properties *p2plink; struct kfd_iolink_properties *iolink; struct kfd_cache_properties *cache; struct kfd_mem_properties *mem; + struct kfd_perf_properties *perf; int ret; - uint32_t i; + uint32_t i, num_attrs; + struct attribute **attrs; - BUG_ON(!dev); + if (WARN_ON(dev->kobj_node)) + return -EEXIST; /* * 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) + if (ret < 0) { + kobject_put(dev->kobj_node); return ret; + } dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node); if (!dev->kobj_mem) @@ -846,6 +675,14 @@ static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev, if (!dev->kobj_iolink) return -ENOMEM; + dev->kobj_p2plink = kobject_create_and_add("p2p_links", dev->kobj_node); + if (!dev->kobj_p2plink) + return -ENOMEM; + + dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node); + if (!dev->kobj_perf) + return -ENOMEM; + /* * Creating sysfs files for node properties */ @@ -875,8 +712,10 @@ static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev, return -ENOMEM; ret = kobject_init_and_add(mem->kobj, &mem_type, dev->kobj_mem, "%d", i); - if (ret < 0) + if (ret < 0) { + kobject_put(mem->kobj); return ret; + } mem->attr.name = "properties"; mem->attr.mode = KFD_SYSFS_FILE_MODE; @@ -894,8 +733,10 @@ static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev, return -ENOMEM; ret = kobject_init_and_add(cache->kobj, &cache_type, dev->kobj_cache, "%d", i); - if (ret < 0) + if (ret < 0) { + kobject_put(cache->kobj); return ret; + } cache->attr.name = "properties"; cache->attr.mode = KFD_SYSFS_FILE_MODE; @@ -913,8 +754,10 @@ static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev, return -ENOMEM; ret = kobject_init_and_add(iolink->kobj, &iolink_type, dev->kobj_iolink, "%d", i); - if (ret < 0) + if (ret < 0) { + kobject_put(iolink->kobj); return ret; + } iolink->attr.name = "properties"; iolink->attr.mode = KFD_SYSFS_FILE_MODE; @@ -923,11 +766,59 @@ static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev, if (ret < 0) return ret; i++; -} + } + + i = 0; + list_for_each_entry(p2plink, &dev->p2p_link_props, list) { + p2plink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); + if (!p2plink->kobj) + return -ENOMEM; + ret = kobject_init_and_add(p2plink->kobj, &iolink_type, + dev->kobj_p2plink, "%d", i); + if (ret < 0) { + kobject_put(p2plink->kobj); + return ret; + } + + p2plink->attr.name = "properties"; + p2plink->attr.mode = KFD_SYSFS_FILE_MODE; + sysfs_attr_init(&p2plink->attr); + ret = sysfs_create_file(p2plink->kobj, &p2plink->attr); + if (ret < 0) + return ret; + i++; + } + + /* All hardware blocks have the same number of attributes. */ + num_attrs = ARRAY_SIZE(perf_attr_iommu); + list_for_each_entry(perf, &dev->perf_props, list) { + perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr) + * num_attrs + sizeof(struct attribute_group), + GFP_KERNEL); + if (!perf->attr_group) + return -ENOMEM; + + attrs = (struct attribute **)(perf->attr_group + 1); + if (!strcmp(perf->block_name, "iommu")) { + /* Information of IOMMU's num_counters and counter_ids is shown + * under /sys/bus/event_source/devices/amd_iommu. We don't + * duplicate here. + */ + perf_attr_iommu[0].data = perf->max_concurrent; + for (i = 0; i < num_attrs; i++) + attrs[i] = &perf_attr_iommu[i].attr.attr; + } + perf->attr_group->name = perf->block_name; + perf->attr_group->attrs = attrs; + ret = sysfs_create_group(dev->kobj_perf, perf->attr_group); + if (ret < 0) + return ret; + } return 0; } +/* Called with write topology lock acquired */ static int kfd_build_sysfs_node_tree(void) { struct kfd_topology_device *dev; @@ -944,6 +835,7 @@ static int kfd_build_sysfs_node_tree(void) return 0; } +/* Called with write topology lock acquired */ static void kfd_remove_sysfs_node_tree(void) { struct kfd_topology_device *dev; @@ -956,8 +848,7 @@ static int kfd_topology_update_sysfs(void) { int ret; - pr_info("Creating topology SYSFS entries\n"); - if (sys_props.kobj_topology == NULL) { + if (!sys_props.kobj_topology) { sys_props.kobj_topology = kfd_alloc_struct(sys_props.kobj_topology); if (!sys_props.kobj_topology) @@ -966,8 +857,10 @@ static int kfd_topology_update_sysfs(void) ret = kobject_init_and_add(sys_props.kobj_topology, &sysprops_type, &kfd_device->kobj, "topology"); - if (ret < 0) + if (ret < 0) { + kobject_put(sys_props.kobj_topology); return ret; + } sys_props.kobj_nodes = kobject_create_and_add("nodes", sys_props.kobj_topology); @@ -1015,117 +908,271 @@ static void kfd_topology_release_sysfs(void) } } +/* Called with write topology_lock acquired */ +static void kfd_topology_update_device_list(struct list_head *temp_list, + struct list_head *master_list) +{ + while (!list_empty(temp_list)) { + list_move_tail(temp_list->next, master_list); + sys_props.num_devices++; + } +} + +static void kfd_debug_print_topology(void) +{ + struct kfd_topology_device *dev; + + down_read(&topology_lock); + + dev = list_last_entry(&topology_device_list, + struct kfd_topology_device, list); + if (dev) { + if (dev->node_props.cpu_cores_count && + dev->node_props.simd_count) { + pr_info("Topology: Add APU node [0x%0x:0x%0x]\n", + dev->node_props.device_id, + dev->node_props.vendor_id); + } else if (dev->node_props.cpu_cores_count) + pr_info("Topology: Add CPU node\n"); + else if (dev->node_props.simd_count) + pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n", + dev->node_props.device_id, + dev->node_props.vendor_id); + } + up_read(&topology_lock); +} + +/* Helper function for intializing platform_xx members of + * kfd_system_properties. Uses OEM info from the last CPU/APU node. + */ +static void kfd_update_system_properties(void) +{ + struct kfd_topology_device *dev; + + down_read(&topology_lock); + dev = list_last_entry(&topology_device_list, + struct kfd_topology_device, list); + if (dev) { + sys_props.platform_id = dev->oem_id64; + sys_props.platform_oem = *((uint64_t *)dev->oem_table_id); + sys_props.platform_rev = dev->oem_revision; + } + up_read(&topology_lock); +} + +static void find_system_memory(const struct dmi_header *dm, void *private) +{ + struct dmi_mem_device *memdev = container_of(dm, struct dmi_mem_device, header); + struct kfd_mem_properties *mem; + struct kfd_topology_device *kdev = + (struct kfd_topology_device *)private; + + if (memdev->header.type != DMI_ENTRY_MEM_DEVICE) + return; + if (memdev->header.length < sizeof(struct dmi_mem_device)) + return; + + list_for_each_entry(mem, &kdev->mem_props, list) { + if (memdev->total_width != 0xFFFF && memdev->total_width != 0) + mem->width = memdev->total_width; + if (memdev->speed != 0) + mem->mem_clk_max = memdev->speed; + } +} + +/* kfd_add_non_crat_information - Add information that is not currently + * defined in CRAT but is necessary for KFD topology + * @dev - topology device to which addition info is added + */ +static void kfd_add_non_crat_information(struct kfd_topology_device *kdev) +{ + /* Check if CPU only node. */ + if (!kdev->gpu) { + /* Add system memory information */ + dmi_walk(find_system_memory, kdev); + } + /* TODO: For GPU node, rearrange code from kfd_topology_add_device */ +} + int kfd_topology_init(void) { void *crat_image = NULL; size_t image_size = 0; int ret; - - /* - * Initialize the head for the topology device list + struct list_head temp_topology_device_list; + int cpu_only_node = 0; + struct kfd_topology_device *kdev; + int proximity_domain; + + /* topology_device_list - Master list of all topology devices + * temp_topology_device_list - temporary list created while parsing CRAT + * or VCRAT. Once parsing is complete the contents of list is moved to + * topology_device_list */ + + /* Initialize the head for the both the lists */ INIT_LIST_HEAD(&topology_device_list); + INIT_LIST_HEAD(&temp_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 + /* Proximity domains in ACPI CRAT tables start counting at + * 0. The same should be true for virtual CRAT tables created + * at this stage. GPUs added later in kfd_topology_add_device + * use a counter. */ - 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"); + proximity_domain = 0; + + ret = kfd_create_crat_image_virtual(&crat_image, &image_size, + COMPUTE_UNIT_CPU, NULL, + proximity_domain); + cpu_only_node = 1; + if (ret) { + pr_err("Error creating VCRAT table for CPU\n"); + return ret; + } + + ret = kfd_parse_crat_table(crat_image, + &temp_topology_device_list, + proximity_domain); + if (ret) { + pr_err("Error parsing VCRAT table for CPU\n"); + goto err; + } + + kdev = list_first_entry(&temp_topology_device_list, + struct kfd_topology_device, list); + + down_write(&topology_lock); + kfd_topology_update_device_list(&temp_topology_device_list, + &topology_device_list); + topology_crat_proximity_domain = sys_props.num_devices-1; + ret = kfd_topology_update_sysfs(); + up_write(&topology_lock); + + if (!ret) { + sys_props.generation_count++; + kfd_update_system_properties(); + kfd_debug_print_topology(); + } else + pr_err("Failed to update topology in sysfs ret=%d\n", ret); + + /* For nodes with GPU, this information gets added + * when GPU is detected (kfd_topology_add_device). + */ + if (cpu_only_node) { + /* Add additional information to CPU only node created above */ + down_write(&topology_lock); + kdev = list_first_entry(&topology_device_list, + struct kfd_topology_device, list); + up_write(&topology_lock); + kfd_add_non_crat_information(kdev); } err: - pr_info("Finished initializing topology ret=%d\n", ret); + kfd_destroy_crat_image(crat_image); return ret; } void kfd_topology_shutdown(void) { + down_write(&topology_lock); kfd_topology_release_sysfs(); kfd_release_live_view(); + up_write(&topology_lock); } -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) +static uint32_t kfd_generate_gpu_id(struct kfd_node *gpu) { - uint32_t hashout; - uint32_t buf[7]; + uint32_t gpu_id; + uint32_t buf[8]; uint64_t local_mem_size; - int i; + struct kfd_topology_device *dev; + bool is_unique; + uint8_t *crc_buf; if (!gpu) return 0; - local_mem_size = gpu->kfd2kgd->get_vmem_size(gpu->kgd); - - 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; + crc_buf = (uint8_t *)&buf; + local_mem_size = gpu->local_mem_info.local_mem_size_private + + gpu->local_mem_info.local_mem_size_public; + buf[0] = gpu->adev->pdev->devfn; + buf[1] = gpu->adev->pdev->subsystem_vendor | + (gpu->adev->pdev->subsystem_device << 16); + buf[2] = pci_domain_nr(gpu->adev->pdev->bus); + buf[3] = gpu->adev->pdev->device; + buf[4] = gpu->adev->pdev->bus->number; buf[5] = lower_32_bits(local_mem_size); buf[6] = upper_32_bits(local_mem_size); + buf[7] = (ffs(gpu->xcc_mask) - 1) | (NUM_XCC(gpu->xcc_mask) << 16); - for (i = 0, hashout = 0; i < 7; i++) - hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH); + gpu_id = crc16(0, crc_buf, sizeof(buf)) & + ((1 << KFD_GPU_ID_HASH_WIDTH) - 1); - return hashout; -} + /* There is a very small possibility when generating a + * 16 (KFD_GPU_ID_HASH_WIDTH) bit value from 8 word buffer + * that the value could be 0 or non-unique. So, check if + * it is unique and non-zero. If not unique increment till + * unique one is found. In case of overflow, restart from 1 + */ + + down_read(&topology_lock); + do { + is_unique = true; + if (!gpu_id) + gpu_id = 1; + list_for_each_entry(dev, &topology_device_list, list) { + if (dev->gpu && dev->gpu_id == gpu_id) { + is_unique = false; + break; + } + } + if (unlikely(!is_unique)) + gpu_id = (gpu_id + 1) & + ((1 << KFD_GPU_ID_HASH_WIDTH) - 1); + } while (!is_unique); + up_read(&topology_lock); -static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu) + return gpu_id; +} +/* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If + * the GPU device is not already present in the topology device + * list then return NULL. This means a new topology device has to + * be created for this GPU. + */ +static struct kfd_topology_device *kfd_assign_gpu(struct kfd_node *gpu) { struct kfd_topology_device *dev; struct kfd_topology_device *out_dev = NULL; + struct kfd_mem_properties *mem; + struct kfd_cache_properties *cache; + struct kfd_iolink_properties *iolink; + struct kfd_iolink_properties *p2plink; - BUG_ON(!gpu); + list_for_each_entry(dev, &topology_device_list, list) { + /* Discrete GPUs need their own topology device list + * entries. Don't assign them to CPU/APU nodes. + */ + if (dev->node_props.cpu_cores_count) + continue; - list_for_each_entry(dev, &topology_device_list, list) - if (dev->gpu == NULL && dev->node_props.simd_count > 0) { + if (!dev->gpu && (dev->node_props.simd_count > 0)) { dev->gpu = gpu; out_dev = dev; + + list_for_each_entry(mem, &dev->mem_props, list) + mem->gpu = dev->gpu; + list_for_each_entry(cache, &dev->cache_props, list) + cache->gpu = dev->gpu; + list_for_each_entry(iolink, &dev->io_link_props, list) + iolink->gpu = dev->gpu; + list_for_each_entry(p2plink, &dev->p2p_link_props, list) + p2plink->gpu = dev->gpu; break; } - + } return out_dev; } @@ -1137,118 +1184,1133 @@ static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival) */ } -int kfd_topology_add_device(struct kfd_dev *gpu) +/* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info, + * patch this after CRAT parsing. + */ +static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev) +{ + struct kfd_mem_properties *mem; + struct kfd_local_mem_info local_mem_info; + + if (!dev) + return; + + /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with + * single bank of VRAM local memory. + * for dGPUs - VCRAT reports only one bank of Local Memory + * for APUs - If CRAT from ACPI reports more than one bank, then + * all the banks will report the same mem_clk_max information + */ + amdgpu_amdkfd_get_local_mem_info(dev->gpu->adev, &local_mem_info, + dev->gpu->xcp); + + list_for_each_entry(mem, &dev->mem_props, list) + mem->mem_clk_max = local_mem_info.mem_clk_max; +} + +static void kfd_set_iolink_no_atomics(struct kfd_topology_device *dev, + struct kfd_topology_device *target_gpu_dev, + struct kfd_iolink_properties *link) +{ + /* xgmi always supports atomics between links. */ + if (link->iolink_type == CRAT_IOLINK_TYPE_XGMI) + return; + + /* check pcie support to set cpu(dev) flags for target_gpu_dev link. */ + if (target_gpu_dev) { + uint32_t cap; + + pcie_capability_read_dword(target_gpu_dev->gpu->adev->pdev, + PCI_EXP_DEVCAP2, &cap); + + if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 | + PCI_EXP_DEVCAP2_ATOMIC_COMP64))) + link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT | + CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT; + /* set gpu (dev) flags. */ + } else { + if (!dev->gpu->kfd->pci_atomic_requested || + dev->gpu->adev->asic_type == CHIP_HAWAII) + link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT | + CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT; + } +} + +static void kfd_set_iolink_non_coherent(struct kfd_topology_device *to_dev, + struct kfd_iolink_properties *outbound_link, + struct kfd_iolink_properties *inbound_link) +{ + /* CPU -> GPU with PCIe */ + if (!to_dev->gpu && + inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS) + inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT; + + if (to_dev->gpu) { + /* GPU <-> GPU with PCIe and + * Vega20 with XGMI + */ + if (inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS || + (inbound_link->iolink_type == CRAT_IOLINK_TYPE_XGMI && + KFD_GC_VERSION(to_dev->gpu) == IP_VERSION(9, 4, 0))) { + outbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT; + inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT; + } + } +} + +#define REC_SDMA_NUM_GPU 8 +static const int rec_sdma_eng_map[REC_SDMA_NUM_GPU][REC_SDMA_NUM_GPU] = { + { -1, 14, 12, 2, 4, 8, 10, 6 }, + { 14, -1, 2, 10, 8, 4, 6, 12 }, + { 10, 2, -1, 12, 14, 6, 4, 8 }, + { 2, 12, 10, -1, 6, 14, 8, 4 }, + { 4, 8, 14, 6, -1, 10, 12, 2 }, + { 8, 4, 6, 14, 12, -1, 2, 10 }, + { 10, 6, 4, 8, 12, 2, -1, 14 }, + { 6, 12, 8, 4, 2, 10, 14, -1 }}; + +static void kfd_set_recommended_sdma_engines(struct kfd_topology_device *to_dev, + struct kfd_iolink_properties *outbound_link, + struct kfd_iolink_properties *inbound_link) +{ + struct kfd_node *gpu = outbound_link->gpu; + struct amdgpu_device *adev = gpu->adev; + unsigned int num_xgmi_nodes = adev->gmc.xgmi.num_physical_nodes; + unsigned int num_xgmi_sdma_engines = kfd_get_num_xgmi_sdma_engines(gpu); + unsigned int num_sdma_engines = kfd_get_num_sdma_engines(gpu); + uint32_t sdma_eng_id_mask = (1 << num_sdma_engines) - 1; + uint32_t xgmi_sdma_eng_id_mask = + ((1 << num_xgmi_sdma_engines) - 1) << num_sdma_engines; + + bool support_rec_eng = !amdgpu_sriov_vf(adev) && to_dev->gpu && + adev->aid_mask && num_xgmi_nodes && gpu->kfd->num_nodes == 1 && + num_xgmi_sdma_engines >= 6 && (!(adev->flags & AMD_IS_APU) && + num_xgmi_nodes == 8); + + if (support_rec_eng) { + int src_socket_id = adev->gmc.xgmi.physical_node_id; + int dst_socket_id = to_dev->gpu->adev->gmc.xgmi.physical_node_id; + unsigned int reshift = num_xgmi_sdma_engines == 6 ? 1 : 0; + + outbound_link->rec_sdma_eng_id_mask = + 1 << (rec_sdma_eng_map[src_socket_id][dst_socket_id] >> reshift); + inbound_link->rec_sdma_eng_id_mask = + 1 << (rec_sdma_eng_map[dst_socket_id][src_socket_id] >> reshift); + + /* If recommended engine is out of range, need to reset the mask */ + if (outbound_link->rec_sdma_eng_id_mask & sdma_eng_id_mask) + outbound_link->rec_sdma_eng_id_mask = xgmi_sdma_eng_id_mask; + if (inbound_link->rec_sdma_eng_id_mask & sdma_eng_id_mask) + inbound_link->rec_sdma_eng_id_mask = xgmi_sdma_eng_id_mask; + + } else { + uint32_t engine_mask = (outbound_link->iolink_type == CRAT_IOLINK_TYPE_XGMI && + num_xgmi_sdma_engines && to_dev->gpu) ? xgmi_sdma_eng_id_mask : + sdma_eng_id_mask; + + outbound_link->rec_sdma_eng_id_mask = engine_mask; + inbound_link->rec_sdma_eng_id_mask = engine_mask; + } +} + +static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev) +{ + struct kfd_iolink_properties *link, *inbound_link; + struct kfd_topology_device *peer_dev; + + if (!dev || !dev->gpu) + return; + + /* GPU only creates direct links so apply flags setting to all */ + list_for_each_entry(link, &dev->io_link_props, list) { + link->flags = CRAT_IOLINK_FLAGS_ENABLED; + kfd_set_iolink_no_atomics(dev, NULL, link); + peer_dev = kfd_topology_device_by_proximity_domain( + link->node_to); + + if (!peer_dev) + continue; + + /* Include the CPU peer in GPU hive if connected over xGMI. */ + if (!peer_dev->gpu && + link->iolink_type == CRAT_IOLINK_TYPE_XGMI) { + /* + * If the GPU is not part of a GPU hive, use its pci + * device location as the hive ID to bind with the CPU. + */ + if (!dev->node_props.hive_id) + dev->node_props.hive_id = pci_dev_id(dev->gpu->adev->pdev); + peer_dev->node_props.hive_id = dev->node_props.hive_id; + } + + list_for_each_entry(inbound_link, &peer_dev->io_link_props, + list) { + if (inbound_link->node_to != link->node_from) + continue; + + inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED; + kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link); + kfd_set_iolink_non_coherent(peer_dev, link, inbound_link); + kfd_set_recommended_sdma_engines(peer_dev, link, inbound_link); + } + } + + /* Create indirect links so apply flags setting to all */ + list_for_each_entry(link, &dev->p2p_link_props, list) { + link->flags = CRAT_IOLINK_FLAGS_ENABLED; + kfd_set_iolink_no_atomics(dev, NULL, link); + peer_dev = kfd_topology_device_by_proximity_domain( + link->node_to); + + if (!peer_dev) + continue; + + list_for_each_entry(inbound_link, &peer_dev->p2p_link_props, + list) { + if (inbound_link->node_to != link->node_from) + continue; + + inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED; + kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link); + kfd_set_iolink_non_coherent(peer_dev, link, inbound_link); + } + } +} + +static int kfd_build_p2p_node_entry(struct kfd_topology_device *dev, + struct kfd_iolink_properties *p2plink) +{ + int ret; + + p2plink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); + if (!p2plink->kobj) + return -ENOMEM; + + ret = kobject_init_and_add(p2plink->kobj, &iolink_type, + dev->kobj_p2plink, "%d", dev->node_props.p2p_links_count - 1); + if (ret < 0) { + kobject_put(p2plink->kobj); + return ret; + } + + p2plink->attr.name = "properties"; + p2plink->attr.mode = KFD_SYSFS_FILE_MODE; + sysfs_attr_init(&p2plink->attr); + ret = sysfs_create_file(p2plink->kobj, &p2plink->attr); + if (ret < 0) + return ret; + + return 0; +} + +static int kfd_create_indirect_link_prop(struct kfd_topology_device *kdev, int gpu_node) +{ + struct kfd_iolink_properties *gpu_link, *tmp_link, *cpu_link; + struct kfd_iolink_properties *props = NULL, *props2 = NULL; + struct kfd_topology_device *cpu_dev; + int ret = 0; + int i, num_cpu; + + num_cpu = 0; + list_for_each_entry(cpu_dev, &topology_device_list, list) { + if (cpu_dev->gpu) + break; + num_cpu++; + } + + if (list_empty(&kdev->io_link_props)) + return -ENODATA; + + gpu_link = list_first_entry(&kdev->io_link_props, + struct kfd_iolink_properties, list); + + for (i = 0; i < num_cpu; i++) { + /* CPU <--> GPU */ + if (gpu_link->node_to == i) + continue; + + /* find CPU <--> CPU links */ + cpu_link = NULL; + cpu_dev = kfd_topology_device_by_proximity_domain(i); + if (cpu_dev) { + list_for_each_entry(tmp_link, + &cpu_dev->io_link_props, list) { + if (tmp_link->node_to == gpu_link->node_to) { + cpu_link = tmp_link; + break; + } + } + } + + if (!cpu_link) + return -ENOMEM; + + /* CPU <--> CPU <--> GPU, GPU node*/ + props = kfd_alloc_struct(props); + if (!props) + return -ENOMEM; + + memcpy(props, gpu_link, sizeof(struct kfd_iolink_properties)); + props->weight = gpu_link->weight + cpu_link->weight; + props->min_latency = gpu_link->min_latency + cpu_link->min_latency; + props->max_latency = gpu_link->max_latency + cpu_link->max_latency; + props->min_bandwidth = min(gpu_link->min_bandwidth, cpu_link->min_bandwidth); + props->max_bandwidth = min(gpu_link->max_bandwidth, cpu_link->max_bandwidth); + + props->node_from = gpu_node; + props->node_to = i; + kdev->node_props.p2p_links_count++; + list_add_tail(&props->list, &kdev->p2p_link_props); + ret = kfd_build_p2p_node_entry(kdev, props); + if (ret < 0) + return ret; + + /* for small Bar, no CPU --> GPU in-direct links */ + if (kfd_dev_is_large_bar(kdev->gpu)) { + /* CPU <--> CPU <--> GPU, CPU node*/ + props2 = kfd_alloc_struct(props2); + if (!props2) + return -ENOMEM; + + memcpy(props2, props, sizeof(struct kfd_iolink_properties)); + props2->node_from = i; + props2->node_to = gpu_node; + props2->kobj = NULL; + cpu_dev->node_props.p2p_links_count++; + list_add_tail(&props2->list, &cpu_dev->p2p_link_props); + ret = kfd_build_p2p_node_entry(cpu_dev, props2); + if (ret < 0) + return ret; + } + } + return ret; +} + +#if defined(CONFIG_HSA_AMD_P2P) +static int kfd_add_peer_prop(struct kfd_topology_device *kdev, + struct kfd_topology_device *peer, int from, int to) +{ + struct kfd_iolink_properties *props = NULL; + struct kfd_iolink_properties *iolink1, *iolink2, *iolink3; + struct kfd_topology_device *cpu_dev; + int ret = 0; + + if (!amdgpu_device_is_peer_accessible( + kdev->gpu->adev, + peer->gpu->adev)) + return ret; + + if (list_empty(&kdev->io_link_props)) + return -ENODATA; + + iolink1 = list_first_entry(&kdev->io_link_props, + struct kfd_iolink_properties, list); + + if (list_empty(&peer->io_link_props)) + return -ENODATA; + + iolink2 = list_first_entry(&peer->io_link_props, + struct kfd_iolink_properties, list); + + props = kfd_alloc_struct(props); + if (!props) + return -ENOMEM; + + memcpy(props, iolink1, sizeof(struct kfd_iolink_properties)); + + props->weight = iolink1->weight + iolink2->weight; + props->min_latency = iolink1->min_latency + iolink2->min_latency; + props->max_latency = iolink1->max_latency + iolink2->max_latency; + props->min_bandwidth = min(iolink1->min_bandwidth, iolink2->min_bandwidth); + props->max_bandwidth = min(iolink2->max_bandwidth, iolink2->max_bandwidth); + + if (iolink1->node_to != iolink2->node_to) { + /* CPU->CPU link*/ + cpu_dev = kfd_topology_device_by_proximity_domain(iolink1->node_to); + if (cpu_dev) { + list_for_each_entry(iolink3, &cpu_dev->io_link_props, list) { + if (iolink3->node_to != iolink2->node_to) + continue; + + props->weight += iolink3->weight; + props->min_latency += iolink3->min_latency; + props->max_latency += iolink3->max_latency; + props->min_bandwidth = min(props->min_bandwidth, + iolink3->min_bandwidth); + props->max_bandwidth = min(props->max_bandwidth, + iolink3->max_bandwidth); + break; + } + } else { + WARN(1, "CPU node not found"); + } + } + + props->node_from = from; + props->node_to = to; + peer->node_props.p2p_links_count++; + list_add_tail(&props->list, &peer->p2p_link_props); + ret = kfd_build_p2p_node_entry(peer, props); + + return ret; +} +#endif + +static int kfd_dev_create_p2p_links(void) { - uint32_t gpu_id; struct kfd_topology_device *dev; - int res; + struct kfd_topology_device *new_dev; +#if defined(CONFIG_HSA_AMD_P2P) + uint32_t i; +#endif + uint32_t k; + int ret = 0; - BUG_ON(!gpu); + k = 0; + list_for_each_entry(dev, &topology_device_list, list) + k++; + if (k < 2) + return 0; - gpu_id = kfd_generate_gpu_id(gpu); + new_dev = list_last_entry(&topology_device_list, struct kfd_topology_device, list); + if (WARN_ON(!new_dev->gpu)) + return 0; - pr_debug("kfd: Adding new GPU (ID: 0x%x) to topology\n", gpu_id); + k--; - down_write(&topology_lock); - /* - * Try to assign the GPU to existing topology device (generated from - * CRAT table + /* create in-direct links */ + ret = kfd_create_indirect_link_prop(new_dev, k); + if (ret < 0) + goto out; + + /* create p2p links */ +#if defined(CONFIG_HSA_AMD_P2P) + i = 0; + list_for_each_entry(dev, &topology_device_list, list) { + if (dev == new_dev) + break; + if (!dev->gpu || !dev->gpu->adev || + (dev->gpu->kfd->hive_id && + dev->gpu->kfd->hive_id == new_dev->gpu->kfd->hive_id && + amdgpu_xgmi_get_is_sharing_enabled(dev->gpu->adev, new_dev->gpu->adev))) + goto next; + + /* check if node(s) is/are peer accessible in one direction or bi-direction */ + ret = kfd_add_peer_prop(new_dev, dev, i, k); + if (ret < 0) + goto out; + + ret = kfd_add_peer_prop(dev, new_dev, k, i); + if (ret < 0) + goto out; +next: + i++; + } +#endif + +out: + return ret; +} + +/* Helper function. See kfd_fill_gpu_cache_info for parameter description */ +static int fill_in_l1_pcache(struct kfd_cache_properties **props_ext, + struct kfd_gpu_cache_info *pcache_info, + int cu_bitmask, + int cache_type, unsigned int cu_processor_id, + int cu_block) +{ + unsigned int cu_sibling_map_mask; + int first_active_cu; + struct kfd_cache_properties *pcache = NULL; + + cu_sibling_map_mask = cu_bitmask; + cu_sibling_map_mask >>= cu_block; + cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1); + first_active_cu = ffs(cu_sibling_map_mask); + + /* CU could be inactive. In case of shared cache find the first active + * CU. and incase of non-shared cache check if the CU is inactive. If + * inactive active skip it */ - 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; + if (first_active_cu) { + pcache = kfd_alloc_struct(pcache); + if (!pcache) + return -ENOMEM; - /* - * TODO: Make a call to retrieve topology information from the - * GPU vBIOS + memset(pcache, 0, sizeof(struct kfd_cache_properties)); + pcache->processor_id_low = cu_processor_id + (first_active_cu - 1); + pcache->cache_level = pcache_info[cache_type].cache_level; + pcache->cache_size = pcache_info[cache_type].cache_size; + pcache->cacheline_size = pcache_info[cache_type].cache_line_size; + + if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_DATA_CACHE) + pcache->cache_type |= HSA_CACHE_TYPE_DATA; + if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_INST_CACHE) + pcache->cache_type |= HSA_CACHE_TYPE_INSTRUCTION; + if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_CPU_CACHE) + pcache->cache_type |= HSA_CACHE_TYPE_CPU; + if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_SIMD_CACHE) + pcache->cache_type |= HSA_CACHE_TYPE_HSACU; + + /* Sibling map is w.r.t processor_id_low, so shift out + * inactive CU */ + cu_sibling_map_mask = + cu_sibling_map_mask >> (first_active_cu - 1); + + pcache->sibling_map[0] = (uint8_t)(cu_sibling_map_mask & 0xFF); + pcache->sibling_map[1] = + (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF); + pcache->sibling_map[2] = + (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF); + pcache->sibling_map[3] = + (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF); + + pcache->sibling_map_size = 4; + *props_ext = pcache; + + return 0; + } + return 1; +} + +/* Helper function. See kfd_fill_gpu_cache_info for parameter description */ +static int fill_in_l2_l3_pcache(struct kfd_cache_properties **props_ext, + struct kfd_gpu_cache_info *pcache_info, + struct amdgpu_cu_info *cu_info, + struct amdgpu_gfx_config *gfx_info, + int cache_type, unsigned int cu_processor_id, + struct kfd_node *knode) +{ + unsigned int cu_sibling_map_mask = 0; + int first_active_cu; + int i, j, k, xcc, start, end; + int num_xcc = NUM_XCC(knode->xcc_mask); + struct kfd_cache_properties *pcache = NULL; + enum amdgpu_memory_partition mode; + struct amdgpu_device *adev = knode->adev; + bool found = false; + + start = ffs(knode->xcc_mask) - 1; + end = start + num_xcc; + + /* To find the bitmap in the first active cu in the first + * xcc, it is based on the assumption that evrey xcc must + * have at least one active cu. + */ + for (i = 0; i < gfx_info->max_shader_engines && !found; i++) { + for (j = 0; j < gfx_info->max_sh_per_se && !found; j++) { + if (cu_info->bitmap[start][i % 4][j % 4]) { + cu_sibling_map_mask = + cu_info->bitmap[start][i % 4][j % 4]; + found = true; + } + } + } + + cu_sibling_map_mask &= + ((1 << pcache_info[cache_type].num_cu_shared) - 1); + first_active_cu = ffs(cu_sibling_map_mask); - /* - * Update the SYSFS tree, since we added another topology device + /* CU could be inactive. In case of shared cache find the first active + * CU. and incase of non-shared cache check if the CU is inactive. If + * inactive active skip it + */ + if (first_active_cu) { + pcache = kfd_alloc_struct(pcache); + if (!pcache) + return -ENOMEM; + + memset(pcache, 0, sizeof(struct kfd_cache_properties)); + pcache->processor_id_low = cu_processor_id + + (first_active_cu - 1); + pcache->cache_level = pcache_info[cache_type].cache_level; + pcache->cacheline_size = pcache_info[cache_type].cache_line_size; + + if (KFD_GC_VERSION(knode) == IP_VERSION(9, 4, 3) || + KFD_GC_VERSION(knode) == IP_VERSION(9, 4, 4) || + KFD_GC_VERSION(knode) == IP_VERSION(9, 5, 0)) + mode = adev->gmc.gmc_funcs->query_mem_partition_mode(adev); + else + mode = UNKNOWN_MEMORY_PARTITION_MODE; + + pcache->cache_size = pcache_info[cache_type].cache_size; + /* Partition mode only affects L3 cache size */ + if (mode && pcache->cache_level == 3) + pcache->cache_size /= mode; + + if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_DATA_CACHE) + pcache->cache_type |= HSA_CACHE_TYPE_DATA; + if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_INST_CACHE) + pcache->cache_type |= HSA_CACHE_TYPE_INSTRUCTION; + if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_CPU_CACHE) + pcache->cache_type |= HSA_CACHE_TYPE_CPU; + if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_SIMD_CACHE) + pcache->cache_type |= HSA_CACHE_TYPE_HSACU; + + /* Sibling map is w.r.t processor_id_low, so shift out + * inactive CU */ - if (kfd_topology_update_sysfs() < 0) - kfd_topology_release_sysfs(); + cu_sibling_map_mask = cu_sibling_map_mask >> (first_active_cu - 1); + k = 0; + + for (xcc = start; xcc < end; xcc++) { + for (i = 0; i < gfx_info->max_shader_engines; i++) { + for (j = 0; j < gfx_info->max_sh_per_se; j++) { + pcache->sibling_map[k] = (uint8_t)(cu_sibling_map_mask & 0xFF); + pcache->sibling_map[k+1] = (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF); + pcache->sibling_map[k+2] = (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF); + pcache->sibling_map[k+3] = (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF); + k += 4; + + cu_sibling_map_mask = cu_info->bitmap[xcc][i % 4][j + i / 4]; + cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1); + } + } + } + pcache->sibling_map_size = k; + *props_ext = pcache; + return 0; + } + return 1; +} +#define KFD_MAX_CACHE_TYPES 6 + +/* kfd_fill_cache_non_crat_info - Fill GPU cache info using kfd_gpu_cache_info + * tables + */ +static void kfd_fill_cache_non_crat_info(struct kfd_topology_device *dev, struct kfd_node *kdev) +{ + struct kfd_gpu_cache_info *pcache_info = NULL; + int i, j, k, xcc, start, end; + int ct = 0; + unsigned int cu_processor_id; + int ret; + unsigned int num_cu_shared; + struct amdgpu_cu_info *cu_info = &kdev->adev->gfx.cu_info; + struct amdgpu_gfx_config *gfx_info = &kdev->adev->gfx.config; + int gpu_processor_id; + struct kfd_cache_properties *props_ext = NULL; + int num_of_entries = 0; + int num_of_cache_types = 0; + struct kfd_gpu_cache_info cache_info[KFD_MAX_CACHE_TYPES]; + + + gpu_processor_id = dev->node_props.simd_id_base; + + memset(cache_info, 0, sizeof(cache_info)); + pcache_info = cache_info; + num_of_cache_types = kfd_get_gpu_cache_info(kdev, &pcache_info); + if (!num_of_cache_types) { + pr_warn("no cache info found\n"); + return; } + /* For each type of cache listed in the kfd_gpu_cache_info table, + * go through all available Compute Units. + * The [i,j,k] loop will + * if kfd_gpu_cache_info.num_cu_shared = 1 + * will parse through all available CU + * If (kfd_gpu_cache_info.num_cu_shared != 1) + * then it will consider only one CU from + * the shared unit + */ + start = ffs(kdev->xcc_mask) - 1; + end = start + NUM_XCC(kdev->xcc_mask); + + for (ct = 0; ct < num_of_cache_types; ct++) { + cu_processor_id = gpu_processor_id; + if (pcache_info[ct].cache_level == 1) { + for (xcc = start; xcc < end; xcc++) { + for (i = 0; i < gfx_info->max_shader_engines; i++) { + for (j = 0; j < gfx_info->max_sh_per_se; j++) { + for (k = 0; k < gfx_info->max_cu_per_sh; k += pcache_info[ct].num_cu_shared) { + + ret = fill_in_l1_pcache(&props_ext, pcache_info, + cu_info->bitmap[xcc][i % 4][j + i / 4], ct, + cu_processor_id, k); + + if (ret < 0) + break; + + if (!ret) { + num_of_entries++; + list_add_tail(&props_ext->list, &dev->cache_props); + } + + /* Move to next CU block */ + num_cu_shared = ((k + pcache_info[ct].num_cu_shared) <= + gfx_info->max_cu_per_sh) ? + pcache_info[ct].num_cu_shared : + (gfx_info->max_cu_per_sh - k); + cu_processor_id += num_cu_shared; + } + } + } + } + } else { + ret = fill_in_l2_l3_pcache(&props_ext, pcache_info, + cu_info, gfx_info, ct, cu_processor_id, kdev); + + if (ret < 0) + break; + + if (!ret) { + num_of_entries++; + list_add_tail(&props_ext->list, &dev->cache_props); + } + } + } + dev->node_props.caches_count += num_of_entries; + pr_debug("Added [%d] GPU cache entries\n", num_of_entries); +} + +static int kfd_topology_add_device_locked(struct kfd_node *gpu, + struct kfd_topology_device **dev) +{ + int proximity_domain = ++topology_crat_proximity_domain; + struct list_head temp_topology_device_list; + void *crat_image = NULL; + size_t image_size = 0; + int res; + + res = kfd_create_crat_image_virtual(&crat_image, &image_size, + COMPUTE_UNIT_GPU, gpu, + proximity_domain); + if (res) { + dev_err(gpu->adev->dev, "Error creating VCRAT\n"); + topology_crat_proximity_domain--; + goto err; + } + + INIT_LIST_HEAD(&temp_topology_device_list); + + res = kfd_parse_crat_table(crat_image, + &temp_topology_device_list, + proximity_domain); + if (res) { + dev_err(gpu->adev->dev, "Error parsing VCRAT\n"); + topology_crat_proximity_domain--; + goto err; + } + + kfd_topology_update_device_list(&temp_topology_device_list, + &topology_device_list); + + *dev = kfd_assign_gpu(gpu); + if (WARN_ON(!*dev)) { + res = -ENODEV; + goto err; + } + + /* Fill the cache affinity information here for the GPUs + * using VCRAT + */ + kfd_fill_cache_non_crat_info(*dev, gpu); + + /* Update the SYSFS tree, since we added another topology + * device + */ + res = kfd_topology_update_sysfs(); + if (!res) + sys_props.generation_count++; + else + dev_err(gpu->adev->dev, "Failed to update GPU to sysfs topology. res=%d\n", + res); + +err: + kfd_destroy_crat_image(crat_image); + return res; +} + +static void kfd_topology_set_dbg_firmware_support(struct kfd_topology_device *dev) +{ + bool firmware_supported = true; + + if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(11, 0, 0) && + KFD_GC_VERSION(dev->gpu) < IP_VERSION(12, 0, 0)) { + uint32_t mes_api_rev = (dev->gpu->adev->mes.sched_version & + AMDGPU_MES_API_VERSION_MASK) >> + AMDGPU_MES_API_VERSION_SHIFT; + uint32_t mes_rev = dev->gpu->adev->mes.sched_version & + AMDGPU_MES_VERSION_MASK; + + firmware_supported = (mes_api_rev >= 14) && (mes_rev >= 64); + goto out; + } + + /* + * Note: Any unlisted devices here are assumed to support exception handling. + * Add additional checks here as needed. + */ + switch (KFD_GC_VERSION(dev->gpu)) { + case IP_VERSION(9, 0, 1): + firmware_supported = dev->gpu->kfd->mec_fw_version >= 459 + 32768; + break; + case IP_VERSION(9, 1, 0): + case IP_VERSION(9, 2, 1): + case IP_VERSION(9, 2, 2): + case IP_VERSION(9, 3, 0): + case IP_VERSION(9, 4, 0): + firmware_supported = dev->gpu->kfd->mec_fw_version >= 459; + break; + case IP_VERSION(9, 4, 1): + firmware_supported = dev->gpu->kfd->mec_fw_version >= 60; + break; + case IP_VERSION(9, 4, 2): + firmware_supported = dev->gpu->kfd->mec_fw_version >= 51; + break; + case IP_VERSION(10, 1, 10): + case IP_VERSION(10, 1, 2): + case IP_VERSION(10, 1, 1): + firmware_supported = dev->gpu->kfd->mec_fw_version >= 144; + break; + case IP_VERSION(10, 3, 0): + case IP_VERSION(10, 3, 2): + case IP_VERSION(10, 3, 1): + case IP_VERSION(10, 3, 4): + case IP_VERSION(10, 3, 5): + firmware_supported = dev->gpu->kfd->mec_fw_version >= 89; + break; + case IP_VERSION(10, 1, 3): + case IP_VERSION(10, 3, 3): + firmware_supported = false; + break; + default: + break; + } + +out: + if (firmware_supported) + dev->node_props.capability |= HSA_CAP_TRAP_DEBUG_FIRMWARE_SUPPORTED; +} + +static void kfd_topology_set_capabilities(struct kfd_topology_device *dev) +{ + dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 << + HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & + HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); + + dev->node_props.capability |= HSA_CAP_TRAP_DEBUG_SUPPORT | + HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_TRAP_OVERRIDE_SUPPORTED | + HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_MODE_SUPPORTED; + + if (kfd_dbg_has_ttmps_always_setup(dev->gpu)) + dev->node_props.debug_prop |= HSA_DBG_DISPATCH_INFO_ALWAYS_VALID; + + if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(10, 0, 0)) { + if (KFD_GC_VERSION(dev->gpu) == IP_VERSION(9, 4, 3) || + KFD_GC_VERSION(dev->gpu) == IP_VERSION(9, 4, 4)) + dev->node_props.debug_prop |= + HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX9_4_3 | + HSA_DBG_WATCH_ADDR_MASK_HI_BIT_GFX9_4_3; + else + dev->node_props.debug_prop |= + HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX9 | + HSA_DBG_WATCH_ADDR_MASK_HI_BIT; + + if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(9, 4, 2)) + dev->node_props.capability |= + HSA_CAP_TRAP_DEBUG_PRECISE_MEMORY_OPERATIONS_SUPPORTED; + + if (!amdgpu_sriov_vf(dev->gpu->adev)) + dev->node_props.capability |= HSA_CAP_PER_QUEUE_RESET_SUPPORTED; + + } else { + dev->node_props.debug_prop |= HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX10 | + HSA_DBG_WATCH_ADDR_MASK_HI_BIT; + + if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(12, 0, 0)) + dev->node_props.capability |= + HSA_CAP_TRAP_DEBUG_PRECISE_ALU_OPERATIONS_SUPPORTED; + } + + kfd_topology_set_dbg_firmware_support(dev); +} + +int kfd_topology_add_device(struct kfd_node *gpu) +{ + uint32_t gpu_id; + struct kfd_topology_device *dev; + int res = 0; + int i; + const char *asic_name = amdgpu_asic_name[gpu->adev->asic_type]; + struct amdgpu_gfx_config *gfx_info = &gpu->adev->gfx.config; + struct amdgpu_cu_info *cu_info = &gpu->adev->gfx.cu_info; + + if (gpu->xcp && !gpu->xcp->ddev) { + dev_warn(gpu->adev->dev, + "Won't add GPU to topology since it has no drm node assigned."); + return 0; + } else { + dev_dbg(gpu->adev->dev, "Adding new GPU to topology\n"); + } + + /* Check to see if this gpu device exists in the topology_device_list. + * If so, assign the gpu to that device, + * else create a Virtual CRAT for this gpu device and then parse that + * CRAT to create a new topology device. Once created assign the gpu to + * that topology device + */ + down_write(&topology_lock); + dev = kfd_assign_gpu(gpu); + if (!dev) + res = kfd_topology_add_device_locked(gpu, &dev); + up_write(&topology_lock); + if (res) + return res; + + gpu_id = kfd_generate_gpu_id(gpu); 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); + + kfd_dev_create_p2p_links(); + + /* TODO: Move the following lines to function + * kfd_add_non_crat_information + */ + + /* Fill-in additional information that is not available in CRAT but + * needed for the topology + */ + for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1; i++) { + dev->node_props.name[i] = __tolower(asic_name[i]); + if (asic_name[i] == '\0') + break; + } + dev->node_props.name[i] = '\0'; + + dev->node_props.simd_arrays_per_engine = + gfx_info->max_sh_per_se; + + dev->node_props.gfx_target_version = + gpu->kfd->device_info.gfx_target_version; + dev->node_props.vendor_id = gpu->adev->pdev->vendor; + dev->node_props.device_id = gpu->adev->pdev->device; + dev->node_props.capability |= + ((dev->gpu->adev->rev_id << HSA_CAP_ASIC_REVISION_SHIFT) & + HSA_CAP_ASIC_REVISION_MASK); + + dev->node_props.location_id = pci_dev_id(gpu->adev->pdev); + if (gpu->kfd->num_nodes > 1) + dev->node_props.location_id |= dev->gpu->node_id; + + dev->node_props.domain = pci_domain_nr(gpu->adev->pdev->bus); + dev->node_props.max_engine_clk_fcompute = + amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->adev); + dev->node_props.max_engine_clk_ccompute = + cpufreq_quick_get_max(0) / 1000; + + if (gpu->xcp) + dev->node_props.drm_render_minor = gpu->xcp->ddev->render->index; + else + dev->node_props.drm_render_minor = + gpu->kfd->shared_resources.drm_render_minor; + + dev->node_props.hive_id = gpu->kfd->hive_id; + dev->node_props.num_sdma_engines = kfd_get_num_sdma_engines(gpu); + dev->node_props.num_sdma_xgmi_engines = + kfd_get_num_xgmi_sdma_engines(gpu); + dev->node_props.num_sdma_queues_per_engine = + gpu->kfd->device_info.num_sdma_queues_per_engine - + gpu->kfd->device_info.num_reserved_sdma_queues_per_engine; + dev->node_props.num_gws = (dev->gpu->gws && + dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ? + dev->gpu->adev->gds.gws_size : 0; + dev->node_props.num_cp_queues = get_cp_queues_num(dev->gpu->dqm); + + kfd_fill_mem_clk_max_info(dev); + kfd_fill_iolink_non_crat_info(dev); + + switch (dev->gpu->adev->asic_type) { + case CHIP_KAVERI: + case CHIP_HAWAII: + case CHIP_TONGA: + dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 << + HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & + HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); + break; + case CHIP_CARRIZO: + case CHIP_FIJI: + case CHIP_POLARIS10: + case CHIP_POLARIS11: + case CHIP_POLARIS12: + case CHIP_VEGAM: + pr_debug("Adding doorbell packet type capability\n"); + dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 << + HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & + HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); + break; + default: + if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(9, 0, 1)) + WARN(1, "Unexpected ASIC family %u", + dev->gpu->adev->asic_type); + else + kfd_topology_set_capabilities(dev); + } + /* - * TODO: Retrieve max engine clock values from KGD + * Overwrite ATS capability according to needs_iommu_device to fix + * potential missing corresponding bit in CRAT of BIOS. */ + dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT; - if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) { - dev->node_props.capability |= HSA_CAP_DOORBELL_PACKET_TYPE; - pr_info("amdkfd: adding doorbell packet type capability\n"); + /* Fix errors in CZ CRAT. + * simd_count: Carrizo CRAT reports wrong simd_count, probably + * because it doesn't consider masked out CUs + * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd + */ + if (dev->gpu->adev->asic_type == CHIP_CARRIZO) { + dev->node_props.simd_count = + cu_info->simd_per_cu * cu_info->number; + dev->node_props.max_waves_per_simd = 10; } - res = 0; + /* kfd only concerns sram ecc on GFX and HBM ecc on UMC */ + dev->node_props.capability |= + ((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0) ? + HSA_CAP_SRAM_EDCSUPPORTED : 0; + dev->node_props.capability |= + ((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ? + HSA_CAP_MEM_EDCSUPPORTED : 0; -err: - up_write(&topology_lock); + if (KFD_GC_VERSION(dev->gpu) != IP_VERSION(9, 0, 1)) + dev->node_props.capability |= (dev->gpu->adev->ras_enabled != 0) ? + HSA_CAP_RASEVENTNOTIFY : 0; - if (res == 0) - kfd_notify_gpu_change(gpu_id, 1); + if (KFD_IS_SVM_API_SUPPORTED(dev->gpu->adev)) + dev->node_props.capability |= HSA_CAP_SVMAPI_SUPPORTED; - return res; + if (dev->gpu->adev->gmc.is_app_apu || + dev->gpu->adev->gmc.xgmi.connected_to_cpu) + dev->node_props.capability |= HSA_CAP_FLAGS_COHERENTHOSTACCESS; + + kfd_queue_ctx_save_restore_size(dev); + + kfd_debug_print_topology(); + + kfd_notify_gpu_change(gpu_id, 1); + + return 0; } -int kfd_topology_remove_device(struct kfd_dev *gpu) +/** + * kfd_topology_update_io_links() - Update IO links after device removal. + * @proximity_domain: Proximity domain value of the dev being removed. + * + * The topology list currently is arranged in increasing order of + * proximity domain. + * + * Two things need to be done when a device is removed: + * 1. All the IO links to this device need to be removed. + * 2. All nodes after the current device node need to move + * up once this device node is removed from the topology + * list. As a result, the proximity domain values for + * all nodes after the node being deleted reduce by 1. + * This would also cause the proximity domain values for + * io links to be updated based on new proximity domain + * values. + * + * Context: The caller must hold write topology_lock. + */ +static void kfd_topology_update_io_links(int proximity_domain) { struct kfd_topology_device *dev; + struct kfd_iolink_properties *iolink, *p2plink, *tmp; + + list_for_each_entry(dev, &topology_device_list, list) { + if (dev->proximity_domain > proximity_domain) + dev->proximity_domain--; + + list_for_each_entry_safe(iolink, tmp, &dev->io_link_props, list) { + /* + * If there is an io link to the dev being deleted + * then remove that IO link also. + */ + if (iolink->node_to == proximity_domain) { + list_del(&iolink->list); + dev->node_props.io_links_count--; + } else { + if (iolink->node_from > proximity_domain) + iolink->node_from--; + if (iolink->node_to > proximity_domain) + iolink->node_to--; + } + } + + list_for_each_entry_safe(p2plink, tmp, &dev->p2p_link_props, list) { + /* + * If there is a p2p link to the dev being deleted + * then remove that p2p link also. + */ + if (p2plink->node_to == proximity_domain) { + list_del(&p2plink->list); + dev->node_props.p2p_links_count--; + } else { + if (p2plink->node_from > proximity_domain) + p2plink->node_from--; + if (p2plink->node_to > proximity_domain) + p2plink->node_to--; + } + } + } +} + +int kfd_topology_remove_device(struct kfd_node *gpu) +{ + struct kfd_topology_device *dev, *tmp; uint32_t gpu_id; int res = -ENODEV; - - BUG_ON(!gpu); + int i = 0; down_write(&topology_lock); - list_for_each_entry(dev, &topology_device_list, list) + list_for_each_entry_safe(dev, tmp, &topology_device_list, list) { if (dev->gpu == gpu) { gpu_id = dev->gpu_id; kfd_remove_sysfs_node_entry(dev); kfd_release_topology_device(dev); + sys_props.num_devices--; + kfd_topology_update_io_links(i); + topology_crat_proximity_domain = sys_props.num_devices-1; + sys_props.generation_count++; res = 0; if (kfd_topology_update_sysfs() < 0) kfd_topology_release_sysfs(); break; } + i++; + } up_write(&topology_lock); - if (res == 0) + if (!res) kfd_notify_gpu_change(gpu_id, 0); return res; } -/* - * When idx is out of bounds, the function will return NULL +/* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD + * topology. If GPU device is found @idx, then valid kfd_dev pointer is + * returned through @kdev + * Return - 0: On success (@kdev will be NULL for non GPU nodes) + * -1: If end of list */ -struct kfd_dev *kfd_topology_enum_kfd_devices(uint8_t idx) +int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev) { struct kfd_topology_device *top_dev; - struct kfd_dev *device = NULL; uint8_t device_idx = 0; + *kdev = NULL; down_read(&topology_lock); list_for_each_entry(top_dev, &topology_device_list, list) { if (device_idx == idx) { - device = top_dev->gpu; - break; + *kdev = top_dev->gpu; + up_read(&topology_lock); + return 0; } device_idx++; @@ -1256,6 +2318,89 @@ struct kfd_dev *kfd_topology_enum_kfd_devices(uint8_t idx) up_read(&topology_lock); - return device; + return -1; } + +static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask) +{ + int first_cpu_of_numa_node; + + if (!cpumask || cpumask == cpu_none_mask) + return -1; + first_cpu_of_numa_node = cpumask_first(cpumask); + if (first_cpu_of_numa_node >= nr_cpu_ids) + return -1; +#ifdef CONFIG_X86_64 + return cpu_data(first_cpu_of_numa_node).topo.apicid; +#else + return first_cpu_of_numa_node; +#endif +} + +/* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor + * of the given NUMA node (numa_node_id) + * Return -1 on failure + */ +int kfd_numa_node_to_apic_id(int numa_node_id) +{ + if (numa_node_id == -1) { + pr_warn("Invalid NUMA Node. Use online CPU mask\n"); + return kfd_cpumask_to_apic_id(cpu_online_mask); + } + return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id)); +} + +#if defined(CONFIG_DEBUG_FS) + +int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data) +{ + struct kfd_topology_device *dev; + unsigned int i = 0; + int r = 0; + + down_read(&topology_lock); + + list_for_each_entry(dev, &topology_device_list, list) { + if (!dev->gpu) { + i++; + continue; + } + + seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id); + r = dqm_debugfs_hqds(m, dev->gpu->dqm); + if (r) + break; + } + + up_read(&topology_lock); + + return r; +} + +int kfd_debugfs_rls_by_device(struct seq_file *m, void *data) +{ + struct kfd_topology_device *dev; + unsigned int i = 0; + int r = 0; + + down_read(&topology_lock); + + list_for_each_entry(dev, &topology_device_list, list) { + if (!dev->gpu) { + i++; + continue; + } + + seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id); + r = pm_debugfs_runlist(m, &dev->gpu->dqm->packet_mgr); + if (r) + break; + } + + up_read(&topology_lock); + + return r; +} + +#endif |