Quelle  kfd_packet_manager_v9.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0 OR MIT
/*
 * Copyright 2016-2022 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"),
 * 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 "kfd_kernel_queue.h"
#include "kfd_device_queue_manager.h"
#include "kfd_pm4_headers_ai.h"
#include "kfd_pm4_headers_aldebaran.h"
#include "kfd_pm4_opcodes.h"
#include "gc/gc_10_1_0_sh_mask.h"

static int pm_map_process_v9(struct packet_manager *pm,
  uint32_t *buffer, struct qcm_process_device *qpd)
{
 struct pm4_mes_map_process *packet;
 uint64_t vm_page_table_base_addr = qpd->page_table_base;
 struct kfd_node *kfd = pm->dqm->dev;
 struct kfd_process_device *pdd =
   container_of(qpd, struct kfd_process_device, qpd);
 struct amdgpu_device *adev = kfd->adev;

 packet = (struct pm4_mes_map_process *)buffer;
 memset(buffer, 0, sizeof(struct pm4_mes_map_process));
 packet->header.u32All = pm_build_pm4_header(IT_MAP_PROCESS,
     sizeof(struct pm4_mes_map_process));
 if (adev->enforce_isolation[kfd->node_id] == AMDGPU_ENFORCE_ISOLATION_ENABLE)
  packet->bitfields2.exec_cleaner_shader = 1;
 packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0;
 packet->bitfields2.process_quantum = 10;
 packet->bitfields2.pasid = pdd->pasid;
 packet->bitfields14.gds_size = qpd->gds_size & 0x3F;
 packet->bitfields14.gds_size_hi = (qpd->gds_size >> 6) & 0xF;
 packet->bitfields14.num_gws = (qpd->mapped_gws_queue) ? qpd->num_gws : 0;
 packet->bitfields14.num_oac = qpd->num_oac;
 packet->bitfields14.sdma_enable = 1;
 packet->bitfields14.num_queues = (qpd->is_debug) ? 0 : qpd->queue_count;

 if (kfd->dqm->trap_debug_vmid && pdd->process->debug_trap_enabled &&
   pdd->process->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED) {
  packet->bitfields2.debug_vmid = kfd->dqm->trap_debug_vmid;
  packet->bitfields2.new_debug = 1;
 }

 packet->sh_mem_config = qpd->sh_mem_config;
 packet->sh_mem_bases = qpd->sh_mem_bases;
 if (qpd->tba_addr) {
  packet->sq_shader_tba_lo = lower_32_bits(qpd->tba_addr >> 8);
  /* On GFX9, unlike GFX10, bit TRAP_EN of SQ_SHADER_TBA_HI is
   * not defined, so setting it won't do any harm.
 */
  packet->sq_shader_tba_hi = upper_32_bits(qpd->tba_addr >> 8)
    | 1 << SQ_SHADER_TBA_HI__TRAP_EN__SHIFT;

  packet->sq_shader_tma_lo = lower_32_bits(qpd->tma_addr >> 8);
  packet->sq_shader_tma_hi = upper_32_bits(qpd->tma_addr >> 8);
 }

 packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area);
 packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area);

 packet->vm_context_page_table_base_addr_lo32 =
   lower_32_bits(vm_page_table_base_addr);
 packet->vm_context_page_table_base_addr_hi32 =
   upper_32_bits(vm_page_table_base_addr);

 return 0;
}

static int pm_map_process_aldebaran(struct packet_manager *pm,
  uint32_t *buffer, struct qcm_process_device *qpd)
{
 struct pm4_mes_map_process_aldebaran *packet;
 uint64_t vm_page_table_base_addr = qpd->page_table_base;
 struct kfd_dev *kfd = pm->dqm->dev->kfd;
 struct kfd_node *knode = pm->dqm->dev;
 struct kfd_process_device *pdd =
   container_of(qpd, struct kfd_process_device, qpd);
 int i;
 struct amdgpu_device *adev = kfd->adev;

 packet = (struct pm4_mes_map_process_aldebaran *)buffer;
 memset(buffer, 0, sizeof(struct pm4_mes_map_process_aldebaran));
 packet->header.u32All = pm_build_pm4_header(IT_MAP_PROCESS,
   sizeof(struct pm4_mes_map_process_aldebaran));
 if (adev->enforce_isolation[knode->node_id] ==
     AMDGPU_ENFORCE_ISOLATION_ENABLE)
  packet->bitfields2.exec_cleaner_shader = 1;
 packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0;
 packet->bitfields2.process_quantum = 10;
 packet->bitfields2.pasid = pdd->pasid;
 packet->bitfields14.gds_size = qpd->gds_size & 0x3F;
 packet->bitfields14.gds_size_hi = (qpd->gds_size >> 6) & 0xF;
 packet->bitfields14.num_gws = (qpd->mapped_gws_queue) ? qpd->num_gws : 0;
 packet->bitfields14.num_oac = qpd->num_oac;
 packet->bitfields14.sdma_enable = 1;
 packet->bitfields14.num_queues = (qpd->is_debug) ? 0 : qpd->queue_count;
 packet->spi_gdbg_per_vmid_cntl = pdd->spi_dbg_override |
      pdd->spi_dbg_launch_mode;

 if (pdd->process->debug_trap_enabled) {
  for (i = 0; i < kfd->device_info.num_of_watch_points; i++)
   packet->tcp_watch_cntl[i] = pdd->watch_points[i];

  packet->bitfields2.single_memops =
    !!(pdd->process->dbg_flags & KFD_DBG_TRAP_FLAG_SINGLE_MEM_OP);
 }

 packet->sh_mem_config = qpd->sh_mem_config;
 packet->sh_mem_bases = qpd->sh_mem_bases;
 if (qpd->tba_addr) {
  packet->sq_shader_tba_lo = lower_32_bits(qpd->tba_addr >> 8);
  packet->sq_shader_tba_hi = upper_32_bits(qpd->tba_addr >> 8);
  packet->sq_shader_tma_lo = lower_32_bits(qpd->tma_addr >> 8);
  packet->sq_shader_tma_hi = upper_32_bits(qpd->tma_addr >> 8);
 }

 packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area);
 packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area);

 packet->vm_context_page_table_base_addr_lo32 =
   lower_32_bits(vm_page_table_base_addr);
 packet->vm_context_page_table_base_addr_hi32 =
   upper_32_bits(vm_page_table_base_addr);

 return 0;
}

static int pm_runlist_v9(struct packet_manager *pm, uint32_t *buffer,
   uint64_t ib, size_t ib_size_in_dwords, bool chain)
{
 struct pm4_mes_runlist *packet;

 int concurrent_proc_cnt = 0;
 struct kfd_node *kfd = pm->dqm->dev;
 struct amdgpu_device *adev = kfd->adev;

 /* Determine the number of processes to map together to HW:
  * it can not exceed the number of VMIDs available to the
  * scheduler, and it is determined by the smaller of the number
  * of processes in the runlist and kfd module parameter
  * hws_max_conc_proc.
  * However, if enforce_isolation is set (toggle LDS/VGPRs/SGPRs
  * cleaner between process switch), enable single-process mode
  * in HWS.
  * Note: the arbitration between the number of VMIDs and
  * hws_max_conc_proc has been done in
  * kgd2kfd_device_init().
 */
 concurrent_proc_cnt = (adev->enforce_isolation[kfd->node_id] ==
          AMDGPU_ENFORCE_ISOLATION_ENABLE) ?
  1 : min(pm->dqm->processes_count, kfd->max_proc_per_quantum);

 packet = (struct pm4_mes_runlist *)buffer;

 memset(buffer, 0, sizeof(struct pm4_mes_runlist));
 packet->header.u32All = pm_build_pm4_header(IT_RUN_LIST,
      sizeof(struct pm4_mes_runlist));

 packet->bitfields4.ib_size = ib_size_in_dwords;
 packet->bitfields4.chain = chain ? 1 : 0;
 packet->bitfields4.offload_polling = 0;
 packet->bitfields4.chained_runlist_idle_disable = chain ? 1 : 0;
 packet->bitfields4.valid = 1;
 packet->bitfields4.process_cnt = concurrent_proc_cnt;
 packet->ordinal2 = lower_32_bits(ib);
 packet->ib_base_hi = upper_32_bits(ib);

 return 0;
}

static int pm_set_resources_v9(struct packet_manager *pm, uint32_t *buffer,
    struct scheduling_resources *res)
{
 struct pm4_mes_set_resources *packet;

 packet = (struct pm4_mes_set_resources *)buffer;
 memset(buffer, 0, sizeof(struct pm4_mes_set_resources));

 packet->header.u32All = pm_build_pm4_header(IT_SET_RESOURCES,
     sizeof(struct pm4_mes_set_resources));

 packet->bitfields2.queue_type =
   queue_type__mes_set_resources__hsa_interface_queue_hiq;
 packet->bitfields2.vmid_mask = res->vmid_mask;
 packet->bitfields2.unmap_latency = KFD_UNMAP_LATENCY_MS / 100;
 if (pm->dqm->dev->adev->gmc.xnack_flags & AMDGPU_GMC_XNACK_FLAG_CHAIN)
  packet->bitfields2.enb_xnack_retry_disable_check = 1;
 packet->bitfields7.oac_mask = res->oac_mask;
 packet->bitfields8.gds_heap_base = res->gds_heap_base;
 packet->bitfields8.gds_heap_size = res->gds_heap_size;

 packet->gws_mask_lo = lower_32_bits(res->gws_mask);
 packet->gws_mask_hi = upper_32_bits(res->gws_mask);

 packet->queue_mask_lo = lower_32_bits(res->queue_mask);
 packet->queue_mask_hi = upper_32_bits(res->queue_mask);

 return 0;
}

static inline bool pm_use_ext_eng(struct kfd_dev *dev)
{
 return amdgpu_ip_version(dev->adev, SDMA0_HWIP, 0) >=
        IP_VERSION(5, 2, 0);
}

static int pm_map_queues_v9(struct packet_manager *pm, uint32_t *buffer,
  struct queue *q, bool is_static)
{
 struct pm4_mes_map_queues *packet;

 packet = (struct pm4_mes_map_queues *)buffer;
 memset(buffer, 0, sizeof(struct pm4_mes_map_queues));

 packet->header.u32All = pm_build_pm4_header(IT_MAP_QUEUES,
     sizeof(struct pm4_mes_map_queues));
 packet->bitfields2.num_queues = 1;
 packet->bitfields2.queue_sel =
  queue_sel__mes_map_queues__map_to_hws_determined_queue_slots_vi;

 packet->bitfields2.engine_sel =
  engine_sel__mes_map_queues__compute_vi;
 packet->bitfields2.gws_control_queue = q->properties.is_gws ? 1 : 0;
 packet->bitfields2.extended_engine_sel =
  extended_engine_sel__mes_map_queues__legacy_engine_sel;
 packet->bitfields2.queue_type =
  queue_type__mes_map_queues__normal_compute_vi;

 switch (q->properties.type) {
 case KFD_QUEUE_TYPE_COMPUTE:
  if (is_static)
   packet->bitfields2.queue_type =
  queue_type__mes_map_queues__normal_latency_static_queue_vi;
  break;
 case KFD_QUEUE_TYPE_DIQ:
  packet->bitfields2.queue_type =
   queue_type__mes_map_queues__debug_interface_queue_vi;
  break;
 case KFD_QUEUE_TYPE_SDMA:
 case KFD_QUEUE_TYPE_SDMA_XGMI:
  if (q->properties.sdma_engine_id < 2 &&
      !pm_use_ext_eng(q->device->kfd))
   packet->bitfields2.engine_sel = q->properties.sdma_engine_id +
    engine_sel__mes_map_queues__sdma0_vi;
  else {
   /*
    * For GFX9.4.3, SDMA engine id can be greater than 8.
    * For such cases, set extended_engine_sel to 2 and
    * ensure engine_sel lies between 0-7.
 */
   if (q->properties.sdma_engine_id >= 8)
    packet->bitfields2.extended_engine_sel =
     extended_engine_sel__mes_map_queues__sdma8_to_15_sel;
   else
    packet->bitfields2.extended_engine_sel =
     extended_engine_sel__mes_map_queues__sdma0_to_7_sel;

   packet->bitfields2.engine_sel = q->properties.sdma_engine_id % 8;
  }
  break;
 default:
  WARN(1, "queue type %d", q->properties.type);
  return -EINVAL;
 }
 packet->bitfields3.doorbell_offset =
   q->properties.doorbell_off;

 packet->mqd_addr_lo =
   lower_32_bits(q->gart_mqd_addr);

 packet->mqd_addr_hi =
   upper_32_bits(q->gart_mqd_addr);

 packet->wptr_addr_lo =
   lower_32_bits((uint64_t)q->properties.write_ptr);

 packet->wptr_addr_hi =
   upper_32_bits((uint64_t)q->properties.write_ptr);

 return 0;
}

static inline void pm_build_dequeue_wait_counts_packet_info(struct packet_manager *pm,
   uint32_t sch_value, uint32_t que_sleep, uint32_t *reg_offset,
   uint32_t *reg_data)
{
 pm->dqm->dev->kfd2kgd->build_dequeue_wait_counts_packet_info(
  pm->dqm->dev->adev,
  pm->dqm->wait_times,
  sch_value,
  que_sleep,
  reg_offset,
  reg_data);
}

/* pm_config_dequeue_wait_counts_v9: Builds WRITE_DATA packet with
 *    register/value for configuring dequeue wait counts
 *
 * @return: -ve for failure and 0 for success and buffer is
 *  filled in with packet
 *
 **/
static int pm_config_dequeue_wait_counts_v9(struct packet_manager *pm,
  uint32_t *buffer,
  enum kfd_config_dequeue_wait_counts_cmd cmd,
  uint32_t value)
{
 struct pm4_mec_write_data_mmio *packet;
 uint32_t reg_offset = 0;
 uint32_t reg_data = 0;

 switch (cmd) {
 case KFD_DEQUEUE_WAIT_INIT: {
  uint32_t sch_wave = 0, que_sleep = 1;

  /* For all gfx9 ASICs > gfx941,
   * Reduce CP_IQ_WAIT_TIME2.QUE_SLEEP to 0x1 from default 0x40.
   * On a 1GHz machine this is roughly 1 microsecond, which is
   * about how long it takes to load data out of memory during
   * queue connect
   * QUE_SLEEP: Wait Count for Dequeue Retry.
   *
   * Set CWSR grace period to 1x1000 cycle for GFX9.4.3 APU
 */
  if (KFD_GC_VERSION(pm->dqm->dev) < IP_VERSION(9, 4, 1) ||
      KFD_GC_VERSION(pm->dqm->dev) >= IP_VERSION(10, 0, 0))
   return -EPERM;

  if (amdgpu_emu_mode == 0 && pm->dqm->dev->adev->gmc.is_app_apu &&
      (KFD_GC_VERSION(pm->dqm->dev) == IP_VERSION(9, 4, 3)))
   sch_wave = 1;

  pm_build_dequeue_wait_counts_packet_info(pm, sch_wave, que_sleep,
   ®_offset, ®_data);

  break;
 }
 case KFD_DEQUEUE_WAIT_RESET:
  /* reg_data would be set to dqm->wait_times */
  pm_build_dequeue_wait_counts_packet_info(pm, 0, 0, ®_offset, ®_data);
  break;

 case KFD_DEQUEUE_WAIT_SET_SCH_WAVE:
  /* The CP cannot handle value 0 and it will result in
   * an infinite grace period being set so set to 1 to prevent this. Also
   * avoid debugger API breakage as it sets 0 and expects a low value.
 */
  if (!value)
   value = 1;
  pm_build_dequeue_wait_counts_packet_info(pm, value, 0, ®_offset, ®_data);
  break;
 default:
  pr_err("Invalid dequeue wait cmd\n");
  return -EINVAL;
 }

 packet = (struct pm4_mec_write_data_mmio *)buffer;
 memset(buffer, 0, sizeof(struct pm4_mec_write_data_mmio));

 packet->header.u32All = pm_build_pm4_header(IT_WRITE_DATA,
     sizeof(struct pm4_mec_write_data_mmio));

 packet->bitfields2.dst_sel  = dst_sel___write_data__mem_mapped_register;
 packet->bitfields2.addr_incr =
   addr_incr___write_data__do_not_increment_address;

 packet->bitfields3.dst_mmreg_addr = reg_offset;

 packet->data = reg_data;

 return 0;
}

static int pm_unmap_queues_v9(struct packet_manager *pm, uint32_t *buffer,
   enum kfd_unmap_queues_filter filter,
   uint32_t filter_param, bool reset)
{
 struct pm4_mes_unmap_queues *packet;

 packet = (struct pm4_mes_unmap_queues *)buffer;
 memset(buffer, 0, sizeof(struct pm4_mes_unmap_queues));

 packet->header.u32All = pm_build_pm4_header(IT_UNMAP_QUEUES,
     sizeof(struct pm4_mes_unmap_queues));

 packet->bitfields2.extended_engine_sel =
    pm_use_ext_eng(pm->dqm->dev->kfd) ?
  extended_engine_sel__mes_unmap_queues__sdma0_to_7_sel :
  extended_engine_sel__mes_unmap_queues__legacy_engine_sel;

 packet->bitfields2.engine_sel =
  engine_sel__mes_unmap_queues__compute;

 if (reset)
  packet->bitfields2.action =
   action__mes_unmap_queues__reset_queues;
 else
  packet->bitfields2.action =
   action__mes_unmap_queues__preempt_queues;

 switch (filter) {
 case KFD_UNMAP_QUEUES_FILTER_BY_PASID:
  packet->bitfields2.queue_sel =
   queue_sel__mes_unmap_queues__perform_request_on_pasid_queues;
  packet->bitfields3a.pasid = filter_param;
  break;
 case KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES:
  packet->bitfields2.queue_sel =
   queue_sel__mes_unmap_queues__unmap_all_queues;
  break;
 case KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES:
  /* in this case, we do not preempt static queues */
  packet->bitfields2.queue_sel =
   queue_sel__mes_unmap_queues__unmap_all_non_static_queues;
  break;
 default:
  WARN(1, "filter %d", filter);
  return -EINVAL;
 }

 return 0;

}

static int pm_query_status_v9(struct packet_manager *pm, uint32_t *buffer,
   uint64_t fence_address, uint64_t fence_value)
{
 struct pm4_mes_query_status *packet;

 packet = (struct pm4_mes_query_status *)buffer;
 memset(buffer, 0, sizeof(struct pm4_mes_query_status));


 packet->header.u32All = pm_build_pm4_header(IT_QUERY_STATUS,
     sizeof(struct pm4_mes_query_status));

 packet->bitfields2.context_id = 0;
 packet->bitfields2.interrupt_sel =
   interrupt_sel__mes_query_status__completion_status;
 packet->bitfields2.command =
   command__mes_query_status__fence_only_after_write_ack;

 packet->addr_hi = upper_32_bits((uint64_t)fence_address);
 packet->addr_lo = lower_32_bits((uint64_t)fence_address);
 packet->data_hi = upper_32_bits((uint64_t)fence_value);
 packet->data_lo = lower_32_bits((uint64_t)fence_value);

 return 0;
}

const struct packet_manager_funcs kfd_v9_pm_funcs = {
 .map_process  = pm_map_process_v9,
 .runlist  = pm_runlist_v9,
 .set_resources  = pm_set_resources_v9,
 .map_queues  = pm_map_queues_v9,
 .unmap_queues  = pm_unmap_queues_v9,
 .config_dequeue_wait_counts = pm_config_dequeue_wait_counts_v9,
 .query_status  = pm_query_status_v9,
 .release_mem  = NULL,
 .map_process_size = sizeof(struct pm4_mes_map_process),
 .runlist_size  = sizeof(struct pm4_mes_runlist),
 .set_resources_size = sizeof(struct pm4_mes_set_resources),
 .map_queues_size = sizeof(struct pm4_mes_map_queues),
 .unmap_queues_size = sizeof(struct pm4_mes_unmap_queues),
 .config_dequeue_wait_counts_size  = sizeof(struct pm4_mec_write_data_mmio),
 .query_status_size = sizeof(struct pm4_mes_query_status),
 .release_mem_size = 0,
};

const struct packet_manager_funcs kfd_aldebaran_pm_funcs = {
 .map_process  = pm_map_process_aldebaran,
 .runlist  = pm_runlist_v9,
 .set_resources  = pm_set_resources_v9,
 .map_queues  = pm_map_queues_v9,
 .unmap_queues  = pm_unmap_queues_v9,
 .config_dequeue_wait_counts = pm_config_dequeue_wait_counts_v9,
 .query_status  = pm_query_status_v9,
 .release_mem  = NULL,
 .map_process_size = sizeof(struct pm4_mes_map_process_aldebaran),
 .runlist_size  = sizeof(struct pm4_mes_runlist),
 .set_resources_size = sizeof(struct pm4_mes_set_resources),
 .map_queues_size = sizeof(struct pm4_mes_map_queues),
 .unmap_queues_size = sizeof(struct pm4_mes_unmap_queues),
 .config_dequeue_wait_counts_size  = sizeof(struct pm4_mec_write_data_mmio),
 .query_status_size = sizeof(struct pm4_mes_query_status),
 .release_mem_size = 0,
};