xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c (revision b49024d79fb7304f646003fcd8846ef26dea7e92)

// SPDX-License-Identifier: GPL-2.0 OR MIT
/*
 * Copyright 2020-2022 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/poll.h>
#include <linux/wait.h>
#include <linux/anon_inodes.h>
#include <uapi/linux/kfd_ioctl.h>
#include "amdgpu.h"
#include "amdgpu_vm.h"
#include "kfd_priv.h"
#include "kfd_smi_events.h"
#include "amdgpu_reset.h"

struct kfd_smi_client {
	struct list_head list;
	struct kfifo fifo;
	wait_queue_head_t wait_queue;
	/* events enabled */
	uint64_t events;
	struct kfd_node *dev;
	spinlock_t lock;
	struct rcu_head rcu;
	pid_t pid;
	bool suser;
};

#define KFD_MAX_KFIFO_SIZE	8192

static __poll_t kfd_smi_ev_poll(struct file *, struct poll_table_struct *);
static ssize_t kfd_smi_ev_read(struct file *, char __user *, size_t, loff_t *);
static ssize_t kfd_smi_ev_write(struct file *, const char __user *, size_t,
				loff_t *);
static int kfd_smi_ev_release(struct inode *, struct file *);

static const char kfd_smi_name[] = "kfd_smi_ev";

static const struct file_operations kfd_smi_ev_fops = {
	.owner = THIS_MODULE,
	.poll = kfd_smi_ev_poll,
	.read = kfd_smi_ev_read,
	.write = kfd_smi_ev_write,
	.release = kfd_smi_ev_release
};

static __poll_t kfd_smi_ev_poll(struct file *filep,
				struct poll_table_struct *wait)
{
	struct kfd_smi_client *client = filep->private_data;
	__poll_t mask = 0;

	poll_wait(filep, &client->wait_queue, wait);

	spin_lock(&client->lock);
	if (!kfifo_is_empty(&client->fifo))
		mask = EPOLLIN | EPOLLRDNORM;
	spin_unlock(&client->lock);

	return mask;
}

static ssize_t kfd_smi_ev_read(struct file *filep, char __user *user,
			       size_t size, loff_t *offset)
{
	int ret;
	size_t to_copy;
	struct kfd_smi_client *client = filep->private_data;
	unsigned char *buf;

	size = min_t(size_t, size, KFD_MAX_KFIFO_SIZE);
	buf = kmalloc(size, GFP_KERNEL);
	if (!buf)
		return -ENOMEM;

	/* kfifo_to_user can sleep so we can't use spinlock protection around
	 * it. Instead, we kfifo out as spinlocked then copy them to the user.
	 */
	spin_lock(&client->lock);
	to_copy = kfifo_len(&client->fifo);
	if (!to_copy) {
		spin_unlock(&client->lock);
		ret = -EAGAIN;
		goto ret_err;
	}
	to_copy = min(size, to_copy);
	ret = kfifo_out(&client->fifo, buf, to_copy);
	spin_unlock(&client->lock);
	if (ret <= 0) {
		ret = -EAGAIN;
		goto ret_err;
	}

	ret = copy_to_user(user, buf, to_copy);
	if (ret) {
		ret = -EFAULT;
		goto ret_err;
	}

	kfree(buf);
	return to_copy;

ret_err:
	kfree(buf);
	return ret;
}

static ssize_t kfd_smi_ev_write(struct file *filep, const char __user *user,
				size_t size, loff_t *offset)
{
	struct kfd_smi_client *client = filep->private_data;
	uint64_t events;

	if (!access_ok(user, size) || size < sizeof(events))
		return -EFAULT;
	if (copy_from_user(&events, user, sizeof(events)))
		return -EFAULT;

	WRITE_ONCE(client->events, events);

	return sizeof(events);
}

static void kfd_smi_ev_client_free(struct rcu_head *p)
{
	struct kfd_smi_client *ev = container_of(p, struct kfd_smi_client, rcu);

	kfifo_free(&ev->fifo);
	kfree(ev);
}

static int kfd_smi_ev_release(struct inode *inode, struct file *filep)
{
	struct kfd_smi_client *client = filep->private_data;
	struct kfd_node *dev = client->dev;

	spin_lock(&dev->smi_lock);
	list_del_rcu(&client->list);
	spin_unlock(&dev->smi_lock);

	call_rcu(&client->rcu, kfd_smi_ev_client_free);
	return 0;
}

static bool kfd_smi_ev_enabled(pid_t pid, struct kfd_smi_client *client,
			       unsigned int event)
{
	uint64_t events = READ_ONCE(client->events);

	if (pid && client->pid != pid && !client->suser)
		return false;

	return events & KFD_SMI_EVENT_MASK_FROM_INDEX(event);
}

static void add_event_to_kfifo(pid_t pid, struct kfd_node *dev,
			       unsigned int smi_event, char *event_msg, int len)
{
	struct kfd_smi_client *client;

	rcu_read_lock();

	list_for_each_entry_rcu(client, &dev->smi_clients, list) {
		if (!kfd_smi_ev_enabled(pid, client, smi_event))
			continue;
		spin_lock(&client->lock);
		if (kfifo_avail(&client->fifo) >= len) {
			kfifo_in(&client->fifo, event_msg, len);
			wake_up_all(&client->wait_queue);
		} else {
			pr_debug("smi_event(EventID: %u): no space left\n",
					smi_event);
		}
		spin_unlock(&client->lock);
	}

	rcu_read_unlock();
}

/**
 * kfd_smi_task_to_pid - Convert task to namespace-aware PID
 * @task: task_struct pointer (typically p->lead_thread)
 *
 * Returns the PID as it appears in the task's own PID namespace.
 * For containerized processes, this returns the container-local PID
 * (what getpid() returns), not the global host PID.
 *
 * Returns 0 if task is NULL.
 */
static inline pid_t kfd_smi_task_to_pid(struct task_struct *task)
{
	return task ? task_tgid_nr_ns(task, task_active_pid_ns(task)) : 0;
}

__printf(4, 5)
static void kfd_smi_event_add(struct task_struct *task, struct kfd_node *dev,
			      unsigned int event, char *fmt, ...)
{
	char fifo_in[KFD_SMI_EVENT_MSG_SIZE];
	int len;
	va_list args;
	pid_t pid;

	if (list_empty(&dev->smi_clients))
		return;

	pid = kfd_smi_task_to_pid(task);

	len = scnprintf(fifo_in, sizeof(fifo_in), "%x ", event);

	va_start(args, fmt);
	len += vscnprintf(fifo_in + len, sizeof(fifo_in) - len, fmt, args);
	va_end(args);

	add_event_to_kfifo(pid, dev, event, fifo_in, len);
}

void kfd_smi_event_update_gpu_reset(struct kfd_node *dev, bool post_reset,
				    struct amdgpu_reset_context *reset_context)
{
	unsigned int event;
	char reset_cause[64];

	if (post_reset) {
		event = KFD_SMI_EVENT_GPU_POST_RESET;
	} else {
		event = KFD_SMI_EVENT_GPU_PRE_RESET;
		++(dev->reset_seq_num);
	}

	memset(reset_cause, 0, sizeof(reset_cause));

	if (reset_context)
		amdgpu_reset_get_desc(reset_context, reset_cause,
				      sizeof(reset_cause));

	kfd_smi_event_add(NULL, dev, event, KFD_EVENT_FMT_UPDATE_GPU_RESET(
			  dev->reset_seq_num, reset_cause));
}

void kfd_smi_event_update_thermal_throttling(struct kfd_node *dev,
					     uint64_t throttle_bitmask)
{
	kfd_smi_event_add(NULL, dev, KFD_SMI_EVENT_THERMAL_THROTTLE,
			  KFD_EVENT_FMT_THERMAL_THROTTLING(
			  throttle_bitmask,
			  amdgpu_dpm_get_thermal_throttling_counter(dev->adev)));
}

void kfd_smi_event_update_vmfault(struct kfd_node *dev, uint16_t pasid)
{
	struct amdgpu_task_info *task_info;

	task_info = amdgpu_vm_get_task_info_pasid(dev->adev, pasid);
	if (task_info) {
		/* Report VM faults from user applications, not retry from kernel */
		if (task_info->task.pid)
			kfd_smi_event_add(NULL, dev, KFD_SMI_EVENT_VMFAULT, KFD_EVENT_FMT_VMFAULT(
					  task_info->task.pid, task_info->task.comm));
		amdgpu_vm_put_task_info(task_info);
	}
}

void kfd_smi_event_page_fault_start(struct kfd_node *node, struct task_struct *task,
				    unsigned long address, bool write_fault,
				    ktime_t ts)
{
	kfd_smi_event_add(task, node, KFD_SMI_EVENT_PAGE_FAULT_START,
			  KFD_EVENT_FMT_PAGEFAULT_START(ktime_to_ns(ts),
			  kfd_smi_task_to_pid(task), address, node->id,
			  write_fault ? 'W' : 'R'));
}

void kfd_smi_event_page_fault_end(struct kfd_node *node, struct task_struct *task,
				  unsigned long address, bool migration)
{
	kfd_smi_event_add(task, node, KFD_SMI_EVENT_PAGE_FAULT_END,
			  KFD_EVENT_FMT_PAGEFAULT_END(ktime_get_boottime_ns(),
			  kfd_smi_task_to_pid(task), address, node->id,
			  migration ? 'M' : 'U'));
}

void kfd_smi_event_migration_start(struct kfd_node *node, struct task_struct *task,
				   unsigned long start, unsigned long end,
				   uint32_t from, uint32_t to,
				   uint32_t prefetch_loc, uint32_t preferred_loc,
				   uint32_t trigger)
{
	kfd_smi_event_add(task, node, KFD_SMI_EVENT_MIGRATE_START,
			  KFD_EVENT_FMT_MIGRATE_START(ktime_get_boottime_ns(),
			  kfd_smi_task_to_pid(task), start, end - start, from,
			  to, prefetch_loc, preferred_loc, trigger));
}

void kfd_smi_event_migration_end(struct kfd_node *node, struct task_struct *task,
				 unsigned long start, unsigned long end,
				 uint32_t from, uint32_t to, uint32_t trigger,
				 int error_code)
{
	kfd_smi_event_add(task, node, KFD_SMI_EVENT_MIGRATE_END,
			  KFD_EVENT_FMT_MIGRATE_END(ktime_get_boottime_ns(),
			  kfd_smi_task_to_pid(task), start, end - start, from,
			  to, trigger, error_code));
}

void kfd_smi_event_queue_eviction(struct kfd_node *node, struct task_struct *task,
				  uint32_t trigger)
{
	kfd_smi_event_add(task, node, KFD_SMI_EVENT_QUEUE_EVICTION,
			  KFD_EVENT_FMT_QUEUE_EVICTION(ktime_get_boottime_ns(),
			  kfd_smi_task_to_pid(task), node->id, trigger));
}

void kfd_smi_event_queue_restore(struct kfd_node *node, struct task_struct *task)
{
	kfd_smi_event_add(task, node, KFD_SMI_EVENT_QUEUE_RESTORE,
			  KFD_EVENT_FMT_QUEUE_RESTORE(ktime_get_boottime_ns(),
			  kfd_smi_task_to_pid(task), node->id, '0'));
}

void kfd_smi_event_queue_restore_rescheduled(struct mm_struct *mm)
{
	struct kfd_process *p;
	int i;

	p = kfd_lookup_process_by_mm(mm);
	if (!p)
		return;

	for (i = 0; i < p->n_pdds; i++) {
		struct kfd_process_device *pdd = p->pdds[i];

		kfd_smi_event_add(p->lead_thread, pdd->dev,
				  KFD_SMI_EVENT_QUEUE_RESTORE,
				  KFD_EVENT_FMT_QUEUE_RESTORE(ktime_get_boottime_ns(),
				  kfd_smi_task_to_pid(p->lead_thread),
				  pdd->dev->id, 'R'));
	}
	kfd_unref_process(p);
}

void kfd_smi_event_unmap_from_gpu(struct kfd_node *node, struct task_struct *task,
				  unsigned long address, unsigned long last,
				  uint32_t trigger)
{
	kfd_smi_event_add(task, node, KFD_SMI_EVENT_UNMAP_FROM_GPU,
			  KFD_EVENT_FMT_UNMAP_FROM_GPU(ktime_get_boottime_ns(),
			  kfd_smi_task_to_pid(task), address,
			  last - address + 1, node->id, trigger));
}

void kfd_smi_event_process(struct kfd_process_device *pdd, bool start)
{
	struct amdgpu_task_info *task_info;
	struct amdgpu_vm *avm;

	if (!pdd->drm_priv)
		return;

	avm = drm_priv_to_vm(pdd->drm_priv);
	task_info = amdgpu_vm_get_task_info_vm(avm);

	if (task_info) {
		kfd_smi_event_add(NULL, pdd->dev,
				  start ? KFD_SMI_EVENT_PROCESS_START :
				  KFD_SMI_EVENT_PROCESS_END,
				  KFD_EVENT_FMT_PROCESS(task_info->task.pid,
				  task_info->task.comm));
		amdgpu_vm_put_task_info(task_info);
	}
}

int kfd_smi_event_open(struct kfd_node *dev, uint32_t *fd)
{
	struct kfd_smi_client *client;
	int ret;

	client = kzalloc_obj(struct kfd_smi_client);
	if (!client)
		return -ENOMEM;
	INIT_LIST_HEAD(&client->list);

	ret = kfifo_alloc(&client->fifo, KFD_MAX_KFIFO_SIZE, GFP_KERNEL);
	if (ret) {
		kfree(client);
		return ret;
	}

	init_waitqueue_head(&client->wait_queue);
	spin_lock_init(&client->lock);
	client->events = 0;
	client->dev = dev;
	client->pid = kfd_smi_task_to_pid(current);
	client->suser = capable(CAP_SYS_ADMIN);

	spin_lock(&dev->smi_lock);
	list_add_rcu(&client->list, &dev->smi_clients);
	spin_unlock(&dev->smi_lock);

	ret = anon_inode_getfd(kfd_smi_name, &kfd_smi_ev_fops, (void *)client,
			       O_RDWR);
	if (ret < 0) {
		spin_lock(&dev->smi_lock);
		list_del_rcu(&client->list);
		spin_unlock(&dev->smi_lock);

		synchronize_rcu();

		kfifo_free(&client->fifo);
		kfree(client);
		return ret;
	}
	*fd = ret;

	return 0;
}