From 2725d86da22e2d2e46d970f0b2f2193a0b8e653b Mon Sep 17 00:00:00 2001 From: Jose Maria Casanova Crespo Date: Tue, 7 Feb 2023 13:54:02 +0100 Subject: [PATCH] drm/v3d: New debugfs end-points to query GPU usage stats. Two new debugfs interfaces are implemented: - gpu_usage: exposes the total runtime since boot of each of the 5 scheduling queues available at V3D (BIN, RENDER, CSD, TFU, CACHE_CLEAN). So if the interface is queried at two different points of time the usage percentage of each of the queues can be calculated. - gpu_pid_usage: exposes the same information but to the level of detail of each process using the V3D driver. The runtime for process using the driver is stored. So the percentages of usage by PID can be calculated with measures at different timestamps. The storage of gpu_pid_usage stats is only done if the debugfs interface is polled during the last 70 seconds. If a process does not submit a GPU job during last 70 seconds its stats will also be purged. Signed-off-by: Jose Maria Casanova Crespo --- drivers/gpu/drm/v3d/v3d_debugfs.c | 79 +++++++++++++++++ drivers/gpu/drm/v3d/v3d_drv.h | 59 +++++++++++++ drivers/gpu/drm/v3d/v3d_gem.c | 1 + drivers/gpu/drm/v3d/v3d_irq.c | 5 ++ drivers/gpu/drm/v3d/v3d_sched.c | 139 +++++++++++++++++++++++++++++- 5 files changed, 282 insertions(+), 1 deletion(-) --- a/drivers/gpu/drm/v3d/v3d_debugfs.c +++ b/drivers/gpu/drm/v3d/v3d_debugfs.c @@ -6,6 +6,7 @@ #include #include #include +#include #include @@ -202,6 +203,82 @@ static int v3d_debugfs_bo_stats(struct s return 0; } +static int v3d_debugfs_gpu_usage(struct seq_file *m, void *unused) +{ + struct drm_info_node *node = (struct drm_info_node *)m->private; + struct drm_device *dev = node->minor->dev; + struct v3d_dev *v3d = to_v3d_dev(dev); + struct v3d_queue_stats *queue_stats; + enum v3d_queue queue; + u64 timestamp = local_clock(); + u64 active_runtime; + + seq_printf(m, "timestamp;%llu;\n", local_clock()); + seq_printf(m, "\"QUEUE\";\"JOBS\";\"RUNTIME\";\"ACTIVE\";\n"); + for (queue = 0; queue < V3D_MAX_QUEUES; queue++) { + if (!v3d->queue[queue].sched.ready) + continue; + + queue_stats = &v3d->gpu_queue_stats[queue]; + mutex_lock(&queue_stats->lock); + v3d_sched_stats_update(queue_stats); + if (queue_stats->last_pid) + active_runtime = timestamp - queue_stats->last_exec_start; + else + active_runtime = 0; + + seq_printf(m, "%s;%d;%llu;%c;\n", + v3d_queue_to_string(queue), + queue_stats->jobs_sent, + queue_stats->runtime + active_runtime, + queue_stats->last_pid?'1':'0'); + mutex_unlock(&queue_stats->lock); + } + + return 0; +} + +static int v3d_debugfs_gpu_pid_usage(struct seq_file *m, void *unused) +{ + struct drm_info_node *node = (struct drm_info_node *)m->private; + struct drm_device *dev = node->minor->dev; + struct v3d_dev *v3d = to_v3d_dev(dev); + struct v3d_queue_stats *queue_stats; + struct v3d_queue_pid_stats *cur; + enum v3d_queue queue; + u64 active_runtime; + u64 timestamp = local_clock(); + + seq_printf(m, "timestamp;%llu;\n", timestamp); + seq_printf(m, "\"QUEUE\";\"PID\",\"JOBS\";\"RUNTIME\";\"ACTIVE\";\n"); + for (queue = 0; queue < V3D_MAX_QUEUES; queue++) { + + if (!v3d->queue[queue].sched.ready) + continue; + + queue_stats = &v3d->gpu_queue_stats[queue]; + mutex_lock(&queue_stats->lock); + queue_stats->gpu_pid_stats_timeout = jiffies + V3D_QUEUE_STATS_TIMEOUT; + v3d_sched_stats_update(queue_stats); + list_for_each_entry(cur, &queue_stats->pid_stats_list, list) { + + if (cur->pid == queue_stats->last_pid) + active_runtime = timestamp - queue_stats->last_exec_start; + else + active_runtime = 0; + + seq_printf(m, "%s;%d;%d;%llu;%c;\n", + v3d_queue_to_string(queue), + cur->pid, cur->jobs_sent, + cur->runtime + active_runtime, + cur->pid == queue_stats->last_pid ? '1' : '0'); + } + mutex_unlock(&queue_stats->lock); + } + + return 0; +} + static int v3d_measure_clock(struct seq_file *m, void *unused) { struct drm_info_node *node = (struct drm_info_node *)m->private; @@ -241,6 +318,8 @@ static const struct drm_info_list v3d_de {"v3d_regs", v3d_v3d_debugfs_regs, 0}, {"measure_clock", v3d_measure_clock, 0}, {"bo_stats", v3d_debugfs_bo_stats, 0}, + {"gpu_usage", v3d_debugfs_gpu_usage, 0}, + {"gpu_pid_usage", v3d_debugfs_gpu_pid_usage, 0}, }; void --- a/drivers/gpu/drm/v3d/v3d_drv.h +++ b/drivers/gpu/drm/v3d/v3d_drv.h @@ -21,6 +21,19 @@ struct reset_control; #define V3D_MAX_QUEUES (V3D_CACHE_CLEAN + 1) +static inline char * +v3d_queue_to_string(enum v3d_queue queue) +{ + switch (queue) { + case V3D_BIN: return "v3d_bin"; + case V3D_RENDER: return "v3d_render"; + case V3D_TFU: return "v3d_tfu"; + case V3D_CSD: return "v3d_csd"; + case V3D_CACHE_CLEAN: return "v3d_cache_clean"; + } + return "UNKNOWN"; +} + struct v3d_queue_state { struct drm_gpu_scheduler sched; @@ -28,6 +41,44 @@ struct v3d_queue_state { u64 emit_seqno; }; +struct v3d_queue_pid_stats { + struct list_head list; + u64 runtime; + /* Time in jiffes.to purge the stats of this process. Every time a + * process sends a new job to the queue, this timeout is delayed by + * V3D_QUEUE_STATS_TIMEOUT while the gpu_pid_stats_timeout of the + * queue is not reached. + */ + unsigned long timeout_purge; + u32 jobs_sent; + pid_t pid; +}; + +struct v3d_queue_stats { + struct mutex lock; + u64 last_exec_start; + u64 last_exec_end; + u64 runtime; + u32 jobs_sent; + /* Time in jiffes to stop collecting gpu stats by process. This is + * increased by every access to*the debugfs interface gpu_pid_usage. + * If the debugfs is not used stats are not collected. + */ + unsigned long gpu_pid_stats_timeout; + pid_t last_pid; + struct list_head pid_stats_list; +}; + +/* pid_stats by process (v3d_queue_pid_stats) are recorded if there is an + * access to the gpu_pid_usageare debugfs interface for the last + * V3D_QUEUE_STATS_TIMEOUT (70s). + * + * The same timeout is used to purge the stats by process for those process + * that have not sent jobs this period. + */ +#define V3D_QUEUE_STATS_TIMEOUT (70 * HZ) + + /* Performance monitor object. The perform lifetime is controlled by userspace * using perfmon related ioctls. A perfmon can be attached to a submit_cl * request, and when this is the case, HW perf counters will be activated just @@ -147,6 +198,8 @@ struct v3d_dev { u32 num_allocated; u32 pages_allocated; } bo_stats; + + struct v3d_queue_stats gpu_queue_stats[V3D_MAX_QUEUES]; }; static inline struct v3d_dev * @@ -244,6 +297,11 @@ struct v3d_job { */ struct v3d_perfmon *perfmon; + /* PID of the process that submitted the job that could be used to + * for collecting stats by process of gpu usage. + */ + pid_t client_pid; + /* Callback for the freeing of the job on refcount going to 0. */ void (*free)(struct kref *ref); }; @@ -408,6 +466,7 @@ void v3d_mmu_remove_ptes(struct v3d_bo * /* v3d_sched.c */ int v3d_sched_init(struct v3d_dev *v3d); void v3d_sched_fini(struct v3d_dev *v3d); +void v3d_sched_stats_update(struct v3d_queue_stats *queue_stats); /* v3d_perfmon.c */ void v3d_perfmon_get(struct v3d_perfmon *perfmon); --- a/drivers/gpu/drm/v3d/v3d_gem.c +++ b/drivers/gpu/drm/v3d/v3d_gem.c @@ -516,6 +516,7 @@ v3d_job_init(struct v3d_dev *v3d, struct job = *container; job->v3d = v3d; job->free = free; + job->client_pid = current->pid; ret = drm_sched_job_init(&job->base, &v3d_priv->sched_entity[queue], v3d_priv); --- a/drivers/gpu/drm/v3d/v3d_irq.c +++ b/drivers/gpu/drm/v3d/v3d_irq.c @@ -14,6 +14,7 @@ */ #include +#include #include "v3d_drv.h" #include "v3d_regs.h" @@ -100,6 +101,7 @@ v3d_irq(int irq, void *arg) if (intsts & V3D_INT_FLDONE) { struct v3d_fence *fence = to_v3d_fence(v3d->bin_job->base.irq_fence); + v3d->gpu_queue_stats[V3D_BIN].last_exec_end = local_clock(); trace_v3d_bcl_irq(&v3d->drm, fence->seqno); dma_fence_signal(&fence->base); @@ -109,6 +111,7 @@ v3d_irq(int irq, void *arg) if (intsts & V3D_INT_FRDONE) { struct v3d_fence *fence = to_v3d_fence(v3d->render_job->base.irq_fence); + v3d->gpu_queue_stats[V3D_RENDER].last_exec_end = local_clock(); trace_v3d_rcl_irq(&v3d->drm, fence->seqno); dma_fence_signal(&fence->base); @@ -118,6 +121,7 @@ v3d_irq(int irq, void *arg) if (intsts & V3D_INT_CSDDONE) { struct v3d_fence *fence = to_v3d_fence(v3d->csd_job->base.irq_fence); + v3d->gpu_queue_stats[V3D_CSD].last_exec_end = local_clock(); trace_v3d_csd_irq(&v3d->drm, fence->seqno); dma_fence_signal(&fence->base); @@ -154,6 +158,7 @@ v3d_hub_irq(int irq, void *arg) if (intsts & V3D_HUB_INT_TFUC) { struct v3d_fence *fence = to_v3d_fence(v3d->tfu_job->base.irq_fence); + v3d->gpu_queue_stats[V3D_TFU].last_exec_end = local_clock(); trace_v3d_tfu_irq(&v3d->drm, fence->seqno); dma_fence_signal(&fence->base); --- a/drivers/gpu/drm/v3d/v3d_sched.c +++ b/drivers/gpu/drm/v3d/v3d_sched.c @@ -19,6 +19,7 @@ */ #include +#include #include "v3d_drv.h" #include "v3d_regs.h" @@ -72,6 +73,114 @@ v3d_switch_perfmon(struct v3d_dev *v3d, v3d_perfmon_start(v3d, job->perfmon); } +/* + * Updates the scheduling stats of the gpu queues runtime for completed jobs. + * + * It should be called before any new job submission to the queue or before + * accessing the stats from the debugfs interface. + * + * It is expected that calls to this function are done with queue_stats->lock + * locked. + */ +void +v3d_sched_stats_update(struct v3d_queue_stats *queue_stats) +{ + struct list_head *pid_stats_list = &queue_stats->pid_stats_list; + struct v3d_queue_pid_stats *cur, *tmp; + u64 runtime = 0; + bool store_pid_stats = + time_is_after_jiffies(queue_stats->gpu_pid_stats_timeout); + + /* If debugfs stats gpu_pid_usage has not been polled for a period, + * the pid stats collection is stopped and we purge any existing + * pid_stats. + * + * pid_stats are also purged for clients that have reached the + * timeout_purge because the process probably does not exist anymore. + */ + list_for_each_entry_safe_reverse(cur, tmp, pid_stats_list, list) { + if (!store_pid_stats || time_is_before_jiffies(cur->timeout_purge)) { + list_del(&cur->list); + kfree(cur); + } else { + break; + } + } + /* If a job has finished its stats are updated. */ + if (queue_stats->last_pid && queue_stats->last_exec_end) { + runtime = queue_stats->last_exec_end - + queue_stats->last_exec_start; + queue_stats->runtime += runtime; + + if (store_pid_stats) { + struct v3d_queue_pid_stats *pid_stats; + /* Last job info is always at the head of the list */ + pid_stats = list_first_entry_or_null(pid_stats_list, + struct v3d_queue_pid_stats, list); + if (pid_stats && + pid_stats->pid == queue_stats->last_pid) { + pid_stats->runtime += runtime; + } + } + queue_stats->last_pid = 0; + } +} + +/* + * Updates the queue usage adding the information of a new job that is + * about to be sent to the GPU to be executed. + */ +int +v3d_sched_stats_add_job(struct v3d_queue_stats *queue_stats, + struct drm_sched_job *sched_job) +{ + + struct v3d_queue_pid_stats *pid_stats = NULL; + struct v3d_job *job = sched_job?to_v3d_job(sched_job):NULL; + struct v3d_queue_pid_stats *cur; + struct list_head *pid_stats_list = &queue_stats->pid_stats_list; + int ret = 0; + + mutex_lock(&queue_stats->lock); + + /* Completion of previous job requires an update of its runtime stats */ + v3d_sched_stats_update(queue_stats); + + queue_stats->last_exec_start = local_clock(); + queue_stats->last_exec_end = 0; + queue_stats->jobs_sent++; + queue_stats->last_pid = job->client_pid; + + /* gpu usage stats by process are being collected */ + if (time_is_after_jiffies(queue_stats->gpu_pid_stats_timeout)) { + list_for_each_entry(cur, pid_stats_list, list) { + if (cur->pid == job->client_pid) { + pid_stats = cur; + break; + } + } + /* pid_stats of this client is moved to the head of the list. */ + if (pid_stats) { + list_move(&pid_stats->list, pid_stats_list); + } else { + pid_stats = kzalloc(sizeof(struct v3d_queue_pid_stats), + GFP_KERNEL); + if (!pid_stats) { + ret = -ENOMEM; + goto err_mem; + } + pid_stats->pid = job->client_pid; + list_add(&pid_stats->list, pid_stats_list); + } + pid_stats->jobs_sent++; + pid_stats->timeout_purge = jiffies + V3D_QUEUE_STATS_TIMEOUT; + } + +err_mem: + mutex_unlock(&queue_stats->lock); + return ret; +} + static struct dma_fence *v3d_bin_job_run(struct drm_sched_job *sched_job) { struct v3d_bin_job *job = to_bin_job(sched_job); @@ -107,6 +216,7 @@ static struct dma_fence *v3d_bin_job_run trace_v3d_submit_cl(dev, false, to_v3d_fence(fence)->seqno, job->start, job->end); + v3d_sched_stats_add_job(&v3d->gpu_queue_stats[V3D_BIN], sched_job); v3d_switch_perfmon(v3d, &job->base); /* Set the current and end address of the control list. @@ -158,6 +268,7 @@ static struct dma_fence *v3d_render_job_ trace_v3d_submit_cl(dev, true, to_v3d_fence(fence)->seqno, job->start, job->end); + v3d_sched_stats_add_job(&v3d->gpu_queue_stats[V3D_RENDER], sched_job); v3d_switch_perfmon(v3d, &job->base); /* XXX: Set the QCFG */ @@ -190,6 +301,7 @@ v3d_tfu_job_run(struct drm_sched_job *sc trace_v3d_submit_tfu(dev, to_v3d_fence(fence)->seqno); + v3d_sched_stats_add_job(&v3d->gpu_queue_stats[V3D_TFU], sched_job); V3D_WRITE(V3D_TFU_IIA, job->args.iia); V3D_WRITE(V3D_TFU_IIS, job->args.iis); V3D_WRITE(V3D_TFU_ICA, job->args.ica); @@ -231,6 +343,7 @@ v3d_csd_job_run(struct drm_sched_job *sc trace_v3d_submit_csd(dev, to_v3d_fence(fence)->seqno); + v3d_sched_stats_add_job(&v3d->gpu_queue_stats[V3D_CSD], sched_job); v3d_switch_perfmon(v3d, &job->base); for (i = 1; i <= 6; i++) @@ -247,7 +360,10 @@ v3d_cache_clean_job_run(struct drm_sched struct v3d_job *job = to_v3d_job(sched_job); struct v3d_dev *v3d = job->v3d; + v3d_sched_stats_add_job(&v3d->gpu_queue_stats[V3D_CACHE_CLEAN], + sched_job); v3d_clean_caches(v3d); + v3d->gpu_queue_stats[V3D_CACHE_CLEAN].last_exec_end = local_clock(); return NULL; } @@ -385,8 +501,18 @@ v3d_sched_init(struct v3d_dev *v3d) int hw_jobs_limit = 1; int job_hang_limit = 0; int hang_limit_ms = 500; + enum v3d_queue q; int ret; + for (q = 0; q < V3D_MAX_QUEUES; q++) { + INIT_LIST_HEAD(&v3d->gpu_queue_stats[q].pid_stats_list); + /* Setting timeout before current jiffies disables collecting + * pid_stats on scheduling init. + */ + v3d->gpu_queue_stats[q].gpu_pid_stats_timeout = jiffies - 1; + mutex_init(&v3d->gpu_queue_stats[q].lock); + } + ret = drm_sched_init(&v3d->queue[V3D_BIN].sched, &v3d_bin_sched_ops, hw_jobs_limit, job_hang_limit, @@ -440,9 +566,20 @@ void v3d_sched_fini(struct v3d_dev *v3d) { enum v3d_queue q; + struct v3d_queue_stats *queue_stats; for (q = 0; q < V3D_MAX_QUEUES; q++) { - if (v3d->queue[q].sched.ready) + if (v3d->queue[q].sched.ready) { + queue_stats = &v3d->gpu_queue_stats[q]; + mutex_lock(&queue_stats->lock); + /* Setting gpu_pid_stats_timeout to jiffies-1 will + * make v3d_sched_stats_update to purge all + * allocated pid_stats. + */ + queue_stats->gpu_pid_stats_timeout = jiffies - 1; + v3d_sched_stats_update(queue_stats); + mutex_unlock(&queue_stats->lock); drm_sched_fini(&v3d->queue[q].sched); + } } }