[PATCH 2/3] drm/amdkfd: Add process eviction counters to sysfs
Felix Kuehling
felix.kuehling at amd.com
Thu Sep 10 22:43:34 UTC 2020
Am 2020-09-10 um 2:54 p.m. schrieb Philip Cox:
> Add per-process eviction counters to sysfs to keep track of
> how many eviction events have happened for each process.
>
> Signed-off-by: Philip Cox <Philip.Cox at amd.com>
> ---
> drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 15 ++-
> drivers/gpu/drm/amd/amdkfd/kfd_process.c | 117 +++++++++++++++++++++++
> 2 files changed, 130 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> index 023629f28495..f6902e9c6077 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> @@ -631,7 +631,7 @@ enum kfd_pdd_bound {
> PDD_BOUND_SUSPENDED,
> };
>
> -#define MAX_SYSFS_FILENAME_LEN 11
> +#define MAX_SYSFS_FILENAME_LEN 15
>
> /*
> * SDMA counter runs at 100MHz frequency.
> @@ -692,10 +692,19 @@ struct kfd_process_device {
> uint64_t sdma_past_activity_counter;
> struct attribute attr_sdma;
> char sdma_filename[MAX_SYSFS_FILENAME_LEN];
> +
> + /* Eviction activity tracking */
> + atomic64_t evict_duration_counter;
> + struct attribute attr_evict;
> + char evict_filename[MAX_SYSFS_FILENAME_LEN];
I don't think this name needs to be in a per-pdd variable, because it's
the same for all pdds. see below.
> };
>
> #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
>
> +struct kobj_counters_list {
> + struct list_head counters_list;
> + struct kobject *kobj;
> +};
> /* Process data */
> struct kfd_process {
> /*
> @@ -766,13 +775,15 @@ struct kfd_process {
> /* seqno of the last scheduled eviction */
> unsigned int last_eviction_seqno;
> /* Approx. the last timestamp (in jiffies) when the process was
> - * restored after an eviction
> + * restored or evicted.
> */
> unsigned long last_restore_timestamp;
> + unsigned long last_evict_timestamp;
>
> /* Kobj for our procfs */
> struct kobject *kobj;
> struct kobject *kobj_queues;
> + struct kobj_counters_list counters;
> struct attribute attr_pasid;
> };
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> index 1e15aa7d8ae8..2a81d5a790a0 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> @@ -344,6 +344,26 @@ static ssize_t kfd_procfs_queue_show(struct kobject *kobj,
>
> return 0;
> }
> +static ssize_t kfd_procfs_counters_show(struct kobject *kobj,
> + struct attribute *attr, char *buffer)
> +{
> + if (strcmp(attr->name, "evictions") == 0) {
> + struct kfd_process_device *pdd = container_of(attr,
> + struct kfd_process_device,
> + attr_evict);
> + uint64_t evict_jiffies;
> +
> + evict_jiffies = atomic64_read(&pdd->evict_duration_counter);
> +
> + return snprintf(buffer,
> + PAGE_SIZE,
> + "%llu\n",
> + jiffies64_to_msecs(evict_jiffies));
> + } else
> + pr_err("Invalid attribute");
> +
> + return 0;
> +}
>
> static struct attribute attr_queue_size = {
> .name = "size",
> @@ -376,6 +396,19 @@ static struct kobj_type procfs_queue_type = {
> .default_attrs = procfs_queue_attrs,
> };
>
> +static const struct sysfs_ops procfs_counters_ops = {
> + .show = kfd_procfs_counters_show,
> +};
> +
> +static struct attribute *procfs_counters_attrs[] = {
> + NULL
> +};
> +
> +static struct kobj_type procfs_counters_type = {
> + .sysfs_ops = &procfs_counters_ops,
> + .default_attrs = procfs_counters_attrs,
> +};
> +
> int kfd_procfs_add_queue(struct queue *q)
> {
> struct kfd_process *proc;
> @@ -417,6 +450,70 @@ static int kfd_sysfs_create_file(struct kfd_process *p, struct attribute *attr,
> return ret;
> }
>
> +static int kfd_procfs_add_sysfs_counters(struct kfd_process *p)
> +{
> + int ret = 0;
> + struct kfd_process_device *pdd;
> + char counter_dir_filename[MAX_SYSFS_FILENAME_LEN];
> +
> + if (!p)
> + return -EINVAL;
> +
> + if (!p->kobj)
> + return -EFAULT;
> +
> + INIT_LIST_HEAD(&p->counters.counters_list);
> + /*
> + * Create sysfs files for each GPU:
> + * - proc/<pid>/counters_<gpuid>/
> + * - proc/<pid>/counters_<gpuid>/evictions
> + */
> + list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
> + struct kobj_counters_list *kobj_counter;
> +
> + kobj_counter = kzalloc(sizeof(*kobj_counter),
> + GFP_KERNEL);
> + if (!kobj_counter)
> + return -ENOMEM;
> +
> + snprintf(counter_dir_filename, MAX_SYSFS_FILENAME_LEN,
> + "counters_%u", pdd->dev->id);
As discussed on another email thread, let's rename this to stats_%u, as
the eviction file isn't technically a counter and we're thinking of
adding other stats that aren't counters either (e.g. CU occupancy).
> + kobj_counter->kobj = kfd_alloc_struct(kobj_counter->kobj);
> + if (!kobj_counter->kobj) {
> + kfree(kobj_counter);
> + return -ENOMEM;
> + }
> +
> + ret = kobject_init_and_add(kobj_counter->kobj,
> + &procfs_counters_type,
> + p->kobj,
> + counter_dir_filename);
> +
> + if (ret) {
> + pr_warn("Creating KFD proc/counters_%s folder failed",
> + counter_dir_filename);
> + kobject_put(kobj_counter->kobj);
> + kfree(kobj_counter);
> + goto err;
> + }
> +
> + list_add(&kobj_counter->counters_list,
> + &p->counters.counters_list);
> + snprintf(pdd->evict_filename,
> + MAX_SYSFS_FILENAME_LEN,
> + "evictions");
> + pdd->attr_evict.name = pdd->evict_filename;
Let's rename this to "evicted_ms" to better describe what is counted
here. Also, you don't need this in a per-pdd variable, because the name
is the same for every pdd. I think you can just assign the string
literal directly:
pdd->attr_evict.name = "evicted_ms";
> + pdd->attr_evict.mode = KFD_SYSFS_FILE_MODE;
> + sysfs_attr_init(&pdd->attr_evict);
> + ret = sysfs_create_file(kobj_counter->kobj, &pdd->attr_evict);
> + if (ret)
> + pr_warn("Creating eviction counter for gpuid %d failed",
> + (int)pdd->dev->id);
> + }
> +err:
> + return ret;
> +}
> +
> static int kfd_procfs_add_sysfs_files(struct kfd_process *p)
> {
> int ret = 0;
> @@ -660,6 +757,11 @@ struct kfd_process *kfd_create_process(struct file *filep)
> if (!process->kobj_queues)
> pr_warn("Creating KFD proc/queues folder failed");
>
> + ret = kfd_procfs_add_sysfs_counters(process);
> + if (ret)
> + pr_warn("Creating sysfs counter dir for pid %d failed",
> + (int)process->lead_thread->pid);
> +
> ret = kfd_procfs_add_sysfs_files(process);
> if (ret)
> pr_warn("Creating sysfs usage file for pid %d failed",
> @@ -806,6 +908,7 @@ static void kfd_process_wq_release(struct work_struct *work)
> release_work);
> struct kfd_process_device *pdd;
>
> + struct kobj_counters_list *counters;
> /* Remove the procfs files */
> if (p->kobj) {
> sysfs_remove_file(p->kobj, &p->attr_pasid);
> @@ -818,6 +921,14 @@ static void kfd_process_wq_release(struct work_struct *work)
> sysfs_remove_file(p->kobj, &pdd->attr_sdma);
> }
>
> + list_for_each_entry(counters,
> + &p->counters.counters_list,
> + counters_list) {
> + sysfs_remove_file(p->kobj, &pdd->attr_evict);
> + kobject_del(counters->kobj);
> + kobject_put(counters->kobj);
> + }
> +
> kobject_del(p->kobj);
> kobject_put(p->kobj);
> p->kobj = NULL;
> @@ -1125,6 +1236,7 @@ struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
> pdd->runtime_inuse = false;
> pdd->vram_usage = 0;
> pdd->sdma_past_activity_counter = 0;
> + atomic64_set(&pdd->evict_duration_counter, 0);
> list_add(&pdd->per_device_list, &p->per_device_data);
>
> /* Init idr used for memory handle translation */
> @@ -1388,7 +1500,9 @@ int kfd_process_restore_queues(struct kfd_process *p)
> {
> struct kfd_process_device *pdd;
> int r, ret = 0;
> + uint64_t eviction_duration;
>
> + eviction_duration = p->last_restore_timestamp - p->last_evict_timestamp;
These timestamps only count the time that a process was evicted due to a
TTM buffer eviction. They don't include evicted time from
userptr-related MMU notifiers.
To get a reliable measure of the time that a process spends evicted, you
should get timestamps in evict_process_queues_nocpsch and
evict_process_queues_cpsch and the corresponding restore functions in
kfd_device_queue_manager.c. That's where it does the reference counting
for all the possible eviction triggers and does the actual per-pdd
eviction. You'll need to maintain a last_evicted timestamp and a
evicted_duration counter in the pdd structure, separate from the per
process eviction timestamps.
Regards,
Felix
> list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
> r = pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
> &pdd->qpd);
> @@ -1397,6 +1511,7 @@ int kfd_process_restore_queues(struct kfd_process *p)
> if (!ret)
> ret = r;
> }
> + atomic64_add(eviction_duration, &pdd->evict_duration_counter);
> }
>
> return ret;
> @@ -1425,6 +1540,8 @@ static void evict_process_worker(struct work_struct *work)
> */
> flush_delayed_work(&p->restore_work);
>
> + p->last_evict_timestamp = get_jiffies_64();
> +
> pr_debug("Started evicting pasid 0x%x\n", p->pasid);
> ret = kfd_process_evict_queues(p);
> if (!ret) {
More information about the amd-gfx
mailing list