Quelle  kfd_packet_manager.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0 OR MIT
/*
 * 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"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in
 * all copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
 * OTHER DEALINGS IN THE SOFTWARE.
 *
 */

#include <linux/slab.h>
#include <linux/mutex.h>
#include "kfd_device_queue_manager.h"
#include "kfd_kernel_queue.h"
#include "kfd_priv.h"

#define OVER_SUBSCRIPTION_PROCESS_COUNT (1 << 0)
#define OVER_SUBSCRIPTION_COMPUTE_QUEUE_COUNT (1 << 1)
#define OVER_SUBSCRIPTION_GWS_QUEUE_COUNT (1 << 2)
#define OVER_SUBSCRIPTION_XNACK_CONFLICT (1 << 3)

static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes,
    unsigned int buffer_size_bytes)
{
 unsigned int temp = *wptr + increment_bytes / sizeof(uint32_t);

 WARN((temp * sizeof(uint32_t)) > buffer_size_bytes,
      "Runlist IB overflow");
 *wptr = temp;
}

static void pm_calc_rlib_size(struct packet_manager *pm,
    unsigned int *rlib_size,
    int *over_subscription,
    int xnack_conflict)
{
 unsigned int process_count, queue_count, compute_queue_count, gws_queue_count;
 unsigned int map_queue_size;
 unsigned int max_proc_per_quantum = 1;
 struct kfd_node *node = pm->dqm->dev;
 struct device *dev = node->adev->dev;

 process_count = pm->dqm->processes_count;
 queue_count = pm->dqm->active_queue_count;
 compute_queue_count = pm->dqm->active_cp_queue_count;
 gws_queue_count = pm->dqm->gws_queue_count;

 /* check if there is over subscription
 * Note: the arbitration between the number of VMIDs and
 * hws_max_conc_proc has been done in
 * kgd2kfd_device_init().
 */
 *over_subscription = 0;

 if (node->max_proc_per_quantum > 1)
  max_proc_per_quantum = node->max_proc_per_quantum;

 if (process_count > max_proc_per_quantum)
  *over_subscription |= OVER_SUBSCRIPTION_PROCESS_COUNT;
 if (compute_queue_count > get_cp_queues_num(pm->dqm))
  *over_subscription |= OVER_SUBSCRIPTION_COMPUTE_QUEUE_COUNT;
 if (gws_queue_count > 1)
  *over_subscription |= OVER_SUBSCRIPTION_GWS_QUEUE_COUNT;
 if (xnack_conflict && (node->adev->gmc.xnack_flags & AMDGPU_GMC_XNACK_FLAG_CHAIN))
  *over_subscription |= OVER_SUBSCRIPTION_XNACK_CONFLICT;

 if (*over_subscription)
  dev_dbg(dev, "Over subscribed runlist\n");

 map_queue_size = pm->pmf->map_queues_size;
 /* calculate run list ib allocation size */
 *rlib_size = process_count * pm->pmf->map_process_size +
       queue_count * map_queue_size;

 /*
 * Increase the allocation size in case we need a chained run list
 * when over subscription
 */
 if (*over_subscription)
  *rlib_size += pm->pmf->runlist_size;

 dev_dbg(dev, "runlist ib size %d\n", *rlib_size);
}

static int pm_allocate_runlist_ib(struct packet_manager *pm,
    unsigned int **rl_buffer,
    uint64_t *rl_gpu_buffer,
    unsigned int *rl_buffer_size,
    int *is_over_subscription,
    int xnack_conflict)
{
 struct kfd_node *node = pm->dqm->dev;
 struct device *dev = node->adev->dev;
 int retval;

 if (WARN_ON(pm->allocated))
  return -EINVAL;

 pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription,
    xnack_conflict);

 mutex_lock(&pm->lock);

 retval = kfd_gtt_sa_allocate(node, *rl_buffer_size, &pm->ib_buffer_obj);

 if (retval) {
  dev_err(dev, "Failed to allocate runlist IB\n");
  goto out;
 }

 *(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr;
 *rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr;

 memset(*rl_buffer, 0, *rl_buffer_size);
 pm->allocated = true;

out:
 mutex_unlock(&pm->lock);
 return retval;
}

static int pm_create_runlist_ib(struct packet_manager *pm,
    struct list_head *queues,
    uint64_t *rl_gpu_addr,
    size_t *rl_size_bytes)
{
 unsigned int alloc_size_bytes;
 unsigned int *rl_buffer, rl_wptr, i;
 struct kfd_node *node = pm->dqm->dev;
 struct device *dev = node->adev->dev;
 int retval, processes_mapped;
 struct device_process_node *cur;
 struct qcm_process_device *qpd;
 struct queue *q;
 struct kernel_queue *kq;
 int is_over_subscription;
 int xnack_enabled = -1;
 bool xnack_conflict = 0;

 rl_wptr = retval = processes_mapped = 0;

 /* Check if processes set different xnack modes */
 list_for_each_entry(cur, queues, list) {
  qpd = cur->qpd;
  if (xnack_enabled < 0)
   /* First process */
   xnack_enabled = qpd->pqm->process->xnack_enabled;
  else if (qpd->pqm->process->xnack_enabled != xnack_enabled) {
   /* Found a process with a different xnack mode */
   xnack_conflict = 1;
   break;
  }
 }

 retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr,
    &alloc_size_bytes, &is_over_subscription,
    xnack_conflict);
 if (retval)
  return retval;

 *rl_size_bytes = alloc_size_bytes;
 pm->ib_size_bytes = alloc_size_bytes;

 dev_dbg(dev, "Building runlist ib process count: %d queues count %d\n",
  pm->dqm->processes_count, pm->dqm->active_queue_count);

build_runlist_ib:
 /* build the run list ib packet */
 list_for_each_entry(cur, queues, list) {
  qpd = cur->qpd;
  /* group processes with the same xnack mode together */
  if (qpd->pqm->process->xnack_enabled != xnack_enabled)
   continue;
  /* build map process packet */
  if (processes_mapped >= pm->dqm->processes_count) {
   dev_dbg(dev, "Not enough space left in runlist IB\n");
   pm_release_ib(pm);
   return -ENOMEM;
  }

  retval = pm->pmf->map_process(pm, &rl_buffer[rl_wptr], qpd);
  if (retval)
   return retval;

  processes_mapped++;
  inc_wptr(&rl_wptr, pm->pmf->map_process_size,
    alloc_size_bytes);

  list_for_each_entry(kq, &qpd->priv_queue_list, list) {
   if (!kq->queue->properties.is_active)
    continue;

   dev_dbg(dev,
    "static_queue, mapping kernel q %d, is debug status %d\n",
    kq->queue->queue, qpd->is_debug);

   retval = pm->pmf->map_queues(pm,
      &rl_buffer[rl_wptr],
      kq->queue,
      qpd->is_debug);
   if (retval)
    return retval;

   inc_wptr(&rl_wptr,
    pm->pmf->map_queues_size,
    alloc_size_bytes);
  }

  list_for_each_entry(q, &qpd->queues_list, list) {
   if (!q->properties.is_active)
    continue;

   dev_dbg(dev,
    "static_queue, mapping user queue %d, is debug status %d\n",
    q->queue, qpd->is_debug);

   retval = pm->pmf->map_queues(pm,
      &rl_buffer[rl_wptr],
      q,
      qpd->is_debug);

   if (retval)
    return retval;

   inc_wptr(&rl_wptr,
    pm->pmf->map_queues_size,
    alloc_size_bytes);
  }
 }
 if (xnack_conflict) {
  /* pick up processes with the other xnack mode */
  xnack_enabled = !xnack_enabled;
  xnack_conflict = 0;
  goto build_runlist_ib;
 }

 dev_dbg(dev, "Finished map process and queues to runlist\n");

 if (is_over_subscription) {
  if (!pm->is_over_subscription)
   dev_warn(dev, "Runlist is getting oversubscribed due to%s%s%s%s. Expect reduced ROCm performance.\n",
    is_over_subscription & OVER_SUBSCRIPTION_PROCESS_COUNT ?
    " too many processes" : "",
    is_over_subscription & OVER_SUBSCRIPTION_COMPUTE_QUEUE_COUNT ?
    " too many queues" : "",
    is_over_subscription & OVER_SUBSCRIPTION_GWS_QUEUE_COUNT ?
    " multiple processes using cooperative launch" : "",
    is_over_subscription & OVER_SUBSCRIPTION_XNACK_CONFLICT ?
    " xnack on/off processes mixed on gfx9" : "");

  retval = pm->pmf->runlist(pm, &rl_buffer[rl_wptr],
     *rl_gpu_addr,
     alloc_size_bytes / sizeof(uint32_t),
     true);
 }
 pm->is_over_subscription = !!is_over_subscription;

 for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++)
  pr_debug("0x%2X ", rl_buffer[i]);
 pr_debug("\n");

 return retval;
}

int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm)
{
 switch (dqm->dev->adev->asic_type) {
 case CHIP_KAVERI:
 case CHIP_HAWAII:
  /* PM4 packet structures on CIK are the same as on VI */
 case CHIP_CARRIZO:
 case CHIP_TONGA:
 case CHIP_FIJI:
 case CHIP_POLARIS10:
 case CHIP_POLARIS11:
 case CHIP_POLARIS12:
 case CHIP_VEGAM:
  pm->pmf = &kfd_vi_pm_funcs;
  break;
 default:
  if (KFD_GC_VERSION(dqm->dev) == IP_VERSION(9, 4, 2) ||
      KFD_GC_VERSION(dqm->dev) == IP_VERSION(9, 4, 3) ||
      KFD_GC_VERSION(dqm->dev) == IP_VERSION(9, 4, 4) ||
      KFD_GC_VERSION(dqm->dev) == IP_VERSION(9, 5, 0))
   pm->pmf = &kfd_aldebaran_pm_funcs;
  else if (KFD_GC_VERSION(dqm->dev) >= IP_VERSION(9, 0, 1))
   pm->pmf = &kfd_v9_pm_funcs;
  else {
   WARN(1, "Unexpected ASIC family %u",
        dqm->dev->adev->asic_type);
   return -EINVAL;
  }
 }

 pm->dqm = dqm;
 mutex_init(&pm->lock);
 pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ);
 if (!pm->priv_queue) {
  mutex_destroy(&pm->lock);
  return -ENOMEM;
 }
 pm->allocated = false;

 return 0;
}

void pm_uninit(struct packet_manager *pm)
{
 mutex_destroy(&pm->lock);
 kernel_queue_uninit(pm->priv_queue);
 pm->priv_queue = NULL;
}

int pm_send_set_resources(struct packet_manager *pm,
    struct scheduling_resources *res)
{
 struct kfd_node *node = pm->dqm->dev;
 struct device *dev = node->adev->dev;
 uint32_t *buffer, size;
 int retval = 0;

 size = pm->pmf->set_resources_size;
 mutex_lock(&pm->lock);
 kq_acquire_packet_buffer(pm->priv_queue,
     size / sizeof(uint32_t),
     (unsigned int **)&buffer);
 if (!buffer) {
  dev_err(dev, "Failed to allocate buffer on kernel queue\n");
  retval = -ENOMEM;
  goto out;
 }

 retval = pm->pmf->set_resources(pm, buffer, res);
 if (!retval)
  retval = kq_submit_packet(pm->priv_queue);
 else
  kq_rollback_packet(pm->priv_queue);

out:
 mutex_unlock(&pm->lock);

 return retval;
}

int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues)
{
 uint64_t rl_gpu_ib_addr;
 uint32_t *rl_buffer;
 size_t rl_ib_size, packet_size_dwords;
 int retval;

 retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr,
     &rl_ib_size);
 if (retval)
  goto fail_create_runlist_ib;

 pr_debug("runlist IB address: 0x%llX\n", rl_gpu_ib_addr);

 packet_size_dwords = pm->pmf->runlist_size / sizeof(uint32_t);
 mutex_lock(&pm->lock);

 retval = kq_acquire_packet_buffer(pm->priv_queue,
     packet_size_dwords, &rl_buffer);
 if (retval)
  goto fail_acquire_packet_buffer;

 retval = pm->pmf->runlist(pm, rl_buffer, rl_gpu_ib_addr,
     rl_ib_size / sizeof(uint32_t), false);
 if (retval)
  goto fail_create_runlist;

 retval = kq_submit_packet(pm->priv_queue);

 mutex_unlock(&pm->lock);

 return retval;

fail_create_runlist:
 kq_rollback_packet(pm->priv_queue);
fail_acquire_packet_buffer:
 mutex_unlock(&pm->lock);
fail_create_runlist_ib:
 pm_release_ib(pm);
 return retval;
}

int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
   uint64_t fence_value)
{
 struct kfd_node *node = pm->dqm->dev;
 struct device *dev = node->adev->dev;
 uint32_t *buffer, size;
 int retval = 0;

 if (WARN_ON(!fence_address))
  return -EFAULT;

 size = pm->pmf->query_status_size;
 mutex_lock(&pm->lock);
 kq_acquire_packet_buffer(pm->priv_queue,
   size / sizeof(uint32_t), (unsigned int **)&buffer);
 if (!buffer) {
  dev_err(dev, "Failed to allocate buffer on kernel queue\n");
  retval = -ENOMEM;
  goto out;
 }

 retval = pm->pmf->query_status(pm, buffer, fence_address, fence_value);
 if (!retval)
  retval = kq_submit_packet(pm->priv_queue);
 else
  kq_rollback_packet(pm->priv_queue);

out:
 mutex_unlock(&pm->lock);
 return retval;
}

/* pm_config_dequeue_wait_counts: Configure dequeue timer Wait Counts
 *  by writing to CP_IQ_WAIT_TIME2 registers.
 *
 *  @cmd: See emum kfd_config_dequeue_wait_counts_cmd definition
 *  @value: Depends on the cmd. This parameter is unused for
 *    KFD_DEQUEUE_WAIT_INIT and KFD_DEQUEUE_WAIT_RESET. For
 *    KFD_DEQUEUE_WAIT_SET_SCH_WAVE it holds value to be set
 *
 */
int pm_config_dequeue_wait_counts(struct packet_manager *pm,
  enum kfd_config_dequeue_wait_counts_cmd cmd,
  uint32_t value)
{
 struct kfd_node *node = pm->dqm->dev;
 struct device *dev = node->adev->dev;
 int retval = 0;
 uint32_t *buffer, size;

 if (!pm->pmf->config_dequeue_wait_counts ||
     !pm->pmf->config_dequeue_wait_counts_size)
  return 0;

 if (cmd == KFD_DEQUEUE_WAIT_INIT && (KFD_GC_VERSION(pm->dqm->dev) < IP_VERSION(9, 4, 1) ||
    KFD_GC_VERSION(pm->dqm->dev) >= IP_VERSION(10, 0, 0)))
  return 0;

 size = pm->pmf->config_dequeue_wait_counts_size;

 mutex_lock(&pm->lock);

 if (size) {
  kq_acquire_packet_buffer(pm->priv_queue,
   size / sizeof(uint32_t),
   (unsigned int **)&buffer);

  if (!buffer) {
   dev_err(dev,
    "Failed to allocate buffer on kernel queue\n");
   retval = -ENOMEM;
   goto out;
  }

  retval = pm->pmf->config_dequeue_wait_counts(pm, buffer,
            cmd, value);
  if (!retval) {
   retval = kq_submit_packet(pm->priv_queue);

   /* If default value is modified, cache that in dqm->wait_times */
   if (!retval && cmd == KFD_DEQUEUE_WAIT_INIT)
    update_dqm_wait_times(pm->dqm);
  } else {
   kq_rollback_packet(pm->priv_queue);
  }
 }
out:
 mutex_unlock(&pm->lock);
 return retval;
}

int pm_send_unmap_queue(struct packet_manager *pm,
   enum kfd_unmap_queues_filter filter,
   uint32_t filter_param, bool reset)
{
 struct kfd_node *node = pm->dqm->dev;
 struct device *dev = node->adev->dev;
 uint32_t *buffer, size;
 int retval = 0;

 size = pm->pmf->unmap_queues_size;
 mutex_lock(&pm->lock);
 kq_acquire_packet_buffer(pm->priv_queue,
   size / sizeof(uint32_t), (unsigned int **)&buffer);
 if (!buffer) {
  dev_err(dev, "Failed to allocate buffer on kernel queue\n");
  retval = -ENOMEM;
  goto out;
 }

 retval = pm->pmf->unmap_queues(pm, buffer, filter, filter_param, reset);
 if (!retval)
  retval = kq_submit_packet(pm->priv_queue);
 else
  kq_rollback_packet(pm->priv_queue);

out:
 mutex_unlock(&pm->lock);
 return retval;
}

void pm_release_ib(struct packet_manager *pm)
{
 mutex_lock(&pm->lock);
 if (pm->allocated) {
  kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj);
  pm->allocated = false;
 }
 mutex_unlock(&pm->lock);
}

#if defined(CONFIG_DEBUG_FS)

int pm_debugfs_runlist(struct seq_file *m, void *data)
{
 struct packet_manager *pm = data;

 mutex_lock(&pm->lock);

 if (!pm->allocated) {
  seq_puts(m, " No active runlist\n");
  goto out;
 }

 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4,
       pm->ib_buffer_obj->cpu_ptr, pm->ib_size_bytes, false);

out:
 mutex_unlock(&pm->lock);
 return 0;
}

int pm_debugfs_hang_hws(struct packet_manager *pm)
{
 struct kfd_node *node = pm->dqm->dev;
 struct device *dev = node->adev->dev;
 uint32_t *buffer, size;
 int r = 0;

 if (!pm->priv_queue)
  return -EAGAIN;

 size = pm->pmf->query_status_size;
 mutex_lock(&pm->lock);
 kq_acquire_packet_buffer(pm->priv_queue,
   size / sizeof(uint32_t), (unsigned int **)&buffer);
 if (!buffer) {
  dev_err(dev, "Failed to allocate buffer on kernel queue\n");
  r = -ENOMEM;
  goto out;
 }
 memset(buffer, 0x55, size);
 kq_submit_packet(pm->priv_queue);

 dev_info(dev, "Submitting %x %x %x %x %x %x %x to HIQ to hang the HWS.",
   buffer[0], buffer[1], buffer[2], buffer[3], buffer[4],
   buffer[5], buffer[6]);
out:
 mutex_unlock(&pm->lock);
 return r;
}


#endif