From 70c70f091b1ffd16b3e1a439bd595f7d539b1d5d Mon Sep 17 00:00:00 2001 From: Rob Clark Date: Fri, 30 May 2014 14:49:43 -0400 Subject: [PATCH] drm/msm: add perf logging debugfs Signed-off-by: Rob Clark --- drivers/gpu/drm/msm/Makefile | 1 + drivers/gpu/drm/msm/adreno/a3xx_gpu.c | 20 ++- drivers/gpu/drm/msm/msm_drv.c | 7 + drivers/gpu/drm/msm/msm_drv.h | 4 + drivers/gpu/drm/msm/msm_gpu.c | 103 +++++++++++++ drivers/gpu/drm/msm/msm_gpu.h | 31 ++++ drivers/gpu/drm/msm/msm_perf.c | 275 ++++++++++++++++++++++++++++++++++ 7 files changed, 436 insertions(+), 5 deletions(-) create mode 100644 drivers/gpu/drm/msm/msm_perf.c diff --git a/drivers/gpu/drm/msm/Makefile b/drivers/gpu/drm/msm/Makefile index 9a506b5..93ca49c 100644 --- a/drivers/gpu/drm/msm/Makefile +++ b/drivers/gpu/drm/msm/Makefile @@ -34,6 +34,7 @@ msm-y := \ msm_gem_submit.o \ msm_gpu.o \ msm_iommu.o \ + msm_perf.o \ msm_rd.o \ msm_ringbuffer.o diff --git a/drivers/gpu/drm/msm/adreno/a3xx_gpu.c b/drivers/gpu/drm/msm/adreno/a3xx_gpu.c index f20fbde5..942e09d 100644 --- a/drivers/gpu/drm/msm/adreno/a3xx_gpu.c +++ b/drivers/gpu/drm/msm/adreno/a3xx_gpu.c @@ -207,11 +207,11 @@ static int a3xx_hw_init(struct msm_gpu *gpu) /* Turn on performance counters: */ gpu_write(gpu, REG_A3XX_RBBM_PERFCTR_CTL, 0x01); - /* Set SP perfcounter 7 to count SP_FS_FULL_ALU_INSTRUCTIONS - * we will use this to augment our hang detection: - */ - gpu_write(gpu, REG_A3XX_SP_PERFCOUNTER7_SELECT, - SP_FS_FULL_ALU_INSTRUCTIONS); + /* Enable the perfcntrs that we use.. */ + for (i = 0; i < gpu->num_perfcntrs; i++) { + const struct msm_gpu_perfcntr *perfcntr = &gpu->perfcntrs[i]; + gpu_write(gpu, perfcntr->select_reg, perfcntr->select_val); + } gpu_write(gpu, REG_A3XX_RBBM_INT_0_MASK, A3XX_INT0_MASK); @@ -465,6 +465,13 @@ static const struct adreno_gpu_funcs funcs = { }, }; +static const struct msm_gpu_perfcntr perfcntrs[] = { + { REG_A3XX_SP_PERFCOUNTER6_SELECT, REG_A3XX_RBBM_PERFCTR_SP_6_LO, + SP_ALU_ACTIVE_CYCLES, "ALUACTIVE" }, + { REG_A3XX_SP_PERFCOUNTER7_SELECT, REG_A3XX_RBBM_PERFCTR_SP_7_LO, + SP_FS_FULL_ALU_INSTRUCTIONS, "ALUFULL" }, +}; + struct msm_gpu *a3xx_gpu_init(struct drm_device *dev) { struct a3xx_gpu *a3xx_gpu = NULL; @@ -504,6 +511,9 @@ struct msm_gpu *a3xx_gpu_init(struct drm_device *dev) DBG("fast_rate=%u, slow_rate=%u, bus_freq=%u", gpu->fast_rate, gpu->slow_rate, gpu->bus_freq); + gpu->perfcntrs = perfcntrs; + gpu->num_perfcntrs = ARRAY_SIZE(perfcntrs); + ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, config->rev); if (ret) goto fail; diff --git a/drivers/gpu/drm/msm/msm_drv.c b/drivers/gpu/drm/msm/msm_drv.c index 929f573..1d5151a 100644 --- a/drivers/gpu/drm/msm/msm_drv.c +++ b/drivers/gpu/drm/msm/msm_drv.c @@ -548,6 +548,12 @@ static int late_init_minor(struct drm_minor *minor) return ret; } + ret = msm_perf_debugfs_init(minor); + if (ret) { + dev_err(minor->dev->dev, "could not install perf debugfs\n"); + return ret; + } + return 0; } @@ -588,6 +594,7 @@ static void msm_debugfs_cleanup(struct drm_minor *minor) if (!minor->dev->dev_private) return; msm_rd_debugfs_cleanup(minor); + msm_perf_debugfs_cleanup(minor); } #endif diff --git a/drivers/gpu/drm/msm/msm_drv.h b/drivers/gpu/drm/msm/msm_drv.h index c9c71a0..8a2c5fd 100644 --- a/drivers/gpu/drm/msm/msm_drv.h +++ b/drivers/gpu/drm/msm/msm_drv.h @@ -56,6 +56,7 @@ struct msm_kms; struct msm_gpu; struct msm_mmu; struct msm_rd_state; +struct msm_perf_state; struct msm_gem_submit; #define NUM_DOMAINS 2 /* one for KMS, then one per gpu core (?) */ @@ -85,6 +86,7 @@ struct msm_drm_private { wait_queue_head_t fence_event; struct msm_rd_state *rd; + struct msm_perf_state *perf; /* list of GEM objects: */ struct list_head inactive_list; @@ -212,6 +214,8 @@ int msm_debugfs_late_init(struct drm_device *dev); int msm_rd_debugfs_init(struct drm_minor *minor); void msm_rd_debugfs_cleanup(struct drm_minor *minor); void msm_rd_dump_submit(struct msm_gem_submit *submit); +int msm_perf_debugfs_init(struct drm_minor *minor); +void msm_perf_debugfs_cleanup(struct drm_minor *minor); #else static inline int msm_debugfs_late_init(struct drm_device *dev) { return 0; } static inline void msm_rd_dump_submit(struct msm_gem_submit *submit) {} diff --git a/drivers/gpu/drm/msm/msm_gpu.c b/drivers/gpu/drm/msm/msm_gpu.c index 67371f3..c632219 100644 --- a/drivers/gpu/drm/msm/msm_gpu.c +++ b/drivers/gpu/drm/msm/msm_gpu.c @@ -320,6 +320,101 @@ static void hangcheck_handler(unsigned long data) } /* + * Performance Counters: + */ + +/* called under perf_lock */ +static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs) +{ + uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)]; + int i, n = min(ncntrs, gpu->num_perfcntrs); + + /* read current values: */ + for (i = 0; i < gpu->num_perfcntrs; i++) + current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg); + + /* update cntrs: */ + for (i = 0; i < n; i++) + cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i]; + + /* save current values: */ + for (i = 0; i < gpu->num_perfcntrs; i++) + gpu->last_cntrs[i] = current_cntrs[i]; + + return n; +} + +static void update_sw_cntrs(struct msm_gpu *gpu) +{ + ktime_t time; + uint32_t elapsed; + unsigned long flags; + + spin_lock_irqsave(&gpu->perf_lock, flags); + if (!gpu->perfcntr_active) + goto out; + + time = ktime_get(); + elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time)); + + gpu->totaltime += elapsed; + if (gpu->last_sample.active) + gpu->activetime += elapsed; + + gpu->last_sample.active = msm_gpu_active(gpu); + gpu->last_sample.time = time; + +out: + spin_unlock_irqrestore(&gpu->perf_lock, flags); +} + +void msm_gpu_perfcntr_start(struct msm_gpu *gpu) +{ + unsigned long flags; + + spin_lock_irqsave(&gpu->perf_lock, flags); + /* we could dynamically enable/disable perfcntr registers too.. */ + gpu->last_sample.active = msm_gpu_active(gpu); + gpu->last_sample.time = ktime_get(); + gpu->activetime = gpu->totaltime = 0; + gpu->perfcntr_active = true; + update_hw_cntrs(gpu, 0, NULL); + spin_unlock_irqrestore(&gpu->perf_lock, flags); +} + +void msm_gpu_perfcntr_stop(struct msm_gpu *gpu) +{ + gpu->perfcntr_active = false; +} + +/* returns -errno or # of cntrs sampled */ +int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime, + uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs) +{ + unsigned long flags; + int ret; + + spin_lock_irqsave(&gpu->perf_lock, flags); + + if (!gpu->perfcntr_active) { + ret = -EINVAL; + goto out; + } + + *activetime = gpu->activetime; + *totaltime = gpu->totaltime; + + gpu->activetime = gpu->totaltime = 0; + + ret = update_hw_cntrs(gpu, ncntrs, cntrs); + +out: + spin_unlock_irqrestore(&gpu->perf_lock, flags); + + return ret; +} + +/* * Cmdstream submission/retirement: */ @@ -361,6 +456,7 @@ void msm_gpu_retire(struct msm_gpu *gpu) { struct msm_drm_private *priv = gpu->dev->dev_private; queue_work(priv->wq, &gpu->retire_work); + update_sw_cntrs(gpu); } /* add bo's to gpu's ring, and kick gpu: */ @@ -381,6 +477,8 @@ int msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit, gpu->submitted_fence = submit->fence; + update_sw_cntrs(gpu); + ret = gpu->funcs->submit(gpu, submit, ctx); priv->lastctx = ctx; @@ -433,6 +531,9 @@ int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev, struct iommu_domain *iommu; int i, ret; + if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs))) + gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs); + gpu->dev = drm; gpu->funcs = funcs; gpu->name = name; @@ -448,6 +549,8 @@ int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev, setup_timer(&gpu->hangcheck_timer, hangcheck_handler, (unsigned long)gpu); + spin_lock_init(&gpu->perf_lock); + BUG_ON(ARRAY_SIZE(clk_names) != ARRAY_SIZE(gpu->grp_clks)); /* Map registers: */ diff --git a/drivers/gpu/drm/msm/msm_gpu.h b/drivers/gpu/drm/msm/msm_gpu.h index fad2700..9b579b7 100644 --- a/drivers/gpu/drm/msm/msm_gpu.h +++ b/drivers/gpu/drm/msm/msm_gpu.h @@ -25,6 +25,7 @@ #include "msm_ringbuffer.h" struct msm_gem_submit; +struct msm_gpu_perfcntr; /* So far, with hardware that I've seen to date, we can have: * + zero, one, or two z180 2d cores @@ -64,6 +65,18 @@ struct msm_gpu { struct drm_device *dev; const struct msm_gpu_funcs *funcs; + /* performance counters (hw & sw): */ + spinlock_t perf_lock; + bool perfcntr_active; + struct { + bool active; + ktime_t time; + } last_sample; + uint32_t totaltime, activetime; /* sw counters */ + uint32_t last_cntrs[5]; /* hw counters */ + const struct msm_gpu_perfcntr *perfcntrs; + uint32_t num_perfcntrs; + struct msm_ringbuffer *rb; uint32_t rb_iova; @@ -113,6 +126,19 @@ static inline bool msm_gpu_active(struct msm_gpu *gpu) return gpu->submitted_fence > gpu->funcs->last_fence(gpu); } +/* Perf-Counters: + * The select_reg and select_val are just there for the benefit of the child + * class that actually enables the perf counter.. but msm_gpu base class + * will handle sampling/displaying the counters. + */ + +struct msm_gpu_perfcntr { + uint32_t select_reg; + uint32_t sample_reg; + uint32_t select_val; + const char *name; +}; + static inline void gpu_write(struct msm_gpu *gpu, u32 reg, u32 data) { msm_writel(data, gpu->mmio + (reg << 2)); @@ -126,6 +152,11 @@ static inline u32 gpu_read(struct msm_gpu *gpu, u32 reg) int msm_gpu_pm_suspend(struct msm_gpu *gpu); int msm_gpu_pm_resume(struct msm_gpu *gpu); +void msm_gpu_perfcntr_start(struct msm_gpu *gpu); +void msm_gpu_perfcntr_stop(struct msm_gpu *gpu); +int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime, + uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs); + void msm_gpu_retire(struct msm_gpu *gpu); int msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit, struct msm_file_private *ctx); diff --git a/drivers/gpu/drm/msm/msm_perf.c b/drivers/gpu/drm/msm/msm_perf.c new file mode 100644 index 0000000..830857c --- /dev/null +++ b/drivers/gpu/drm/msm/msm_perf.c @@ -0,0 +1,275 @@ +/* + * Copyright (C) 2013 Red Hat + * Author: Rob Clark + * + * This program is free software; you can redistribute it and/or modify it + * under the terms of the GNU General Public License version 2 as published by + * the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, but WITHOUT + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for + * more details. + * + * You should have received a copy of the GNU General Public License along with + * this program. If not, see . + */ + +/* For profiling, userspace can: + * + * tail -f /sys/kernel/debug/dri//gpu + * + * This will enable performance counters/profiling to track the busy time + * and any gpu specific performance counters that are supported. + */ + +#ifdef CONFIG_DEBUG_FS + +#include + +#include "msm_drv.h" +#include "msm_gpu.h" + +struct msm_perf_state { + struct drm_device *dev; + + bool open; + int cnt; + struct mutex read_lock; + + char buf[256]; + int buftot, bufpos; + + unsigned long next_jiffies; + + struct dentry *ent; + struct drm_info_node *node; +}; + +#define SAMPLE_TIME (HZ/4) + +/* wait for next sample time: */ +static int wait_sample(struct msm_perf_state *perf) +{ + unsigned long start_jiffies = jiffies; + + if (time_after(perf->next_jiffies, start_jiffies)) { + unsigned long remaining_jiffies = + perf->next_jiffies - start_jiffies; + int ret = schedule_timeout_interruptible(remaining_jiffies); + if (ret > 0) { + /* interrupted */ + return -ERESTARTSYS; + } + } + perf->next_jiffies += SAMPLE_TIME; + return 0; +} + +static int refill_buf(struct msm_perf_state *perf) +{ + struct msm_drm_private *priv = perf->dev->dev_private; + struct msm_gpu *gpu = priv->gpu; + char *ptr = perf->buf; + int rem = sizeof(perf->buf); + int i, n; + + if ((perf->cnt++ % 32) == 0) { + /* Header line: */ + n = snprintf(ptr, rem, "%%BUSY"); + ptr += n; + rem -= n; + + for (i = 0; i < gpu->num_perfcntrs; i++) { + const struct msm_gpu_perfcntr *perfcntr = &gpu->perfcntrs[i]; + n = snprintf(ptr, rem, "\t%s", perfcntr->name); + ptr += n; + rem -= n; + } + } else { + /* Sample line: */ + uint32_t activetime = 0, totaltime = 0; + uint32_t cntrs[5]; + uint32_t val; + int ret; + + /* sleep until next sample time: */ + ret = wait_sample(perf); + if (ret) + return ret; + + ret = msm_gpu_perfcntr_sample(gpu, &activetime, &totaltime, + ARRAY_SIZE(cntrs), cntrs); + if (ret < 0) + return ret; + + val = totaltime ? 1000 * activetime / totaltime : 0; + n = snprintf(ptr, rem, "%3d.%d%%", val / 10, val % 10); + ptr += n; + rem -= n; + + for (i = 0; i < ret; i++) { + /* cycle counters (I think).. convert to MHz.. */ + val = cntrs[i] / 10000; + n = snprintf(ptr, rem, "\t%5d.%02d", + val / 100, val % 100); + ptr += n; + rem -= n; + } + } + + n = snprintf(ptr, rem, "\n"); + ptr += n; + rem -= n; + + perf->bufpos = 0; + perf->buftot = ptr - perf->buf; + + return 0; +} + +static ssize_t perf_read(struct file *file, char __user *buf, + size_t sz, loff_t *ppos) +{ + struct msm_perf_state *perf = file->private_data; + int n = 0, ret; + + mutex_lock(&perf->read_lock); + + if (perf->bufpos >= perf->buftot) { + ret = refill_buf(perf); + if (ret) + goto out; + } + + n = min((int)sz, perf->buftot - perf->bufpos); + ret = copy_to_user(buf, &perf->buf[perf->bufpos], n); + if (ret) + goto out; + + perf->bufpos += n; + *ppos += n; + +out: + mutex_unlock(&perf->read_lock); + if (ret) + return ret; + return n; +} + +static int perf_open(struct inode *inode, struct file *file) +{ + struct msm_perf_state *perf = inode->i_private; + struct drm_device *dev = perf->dev; + struct msm_drm_private *priv = dev->dev_private; + struct msm_gpu *gpu = priv->gpu; + int ret = 0; + + mutex_lock(&dev->struct_mutex); + + if (perf->open || !gpu) { + ret = -EBUSY; + goto out; + } + + file->private_data = perf; + perf->open = true; + perf->cnt = 0; + perf->buftot = 0; + perf->bufpos = 0; + msm_gpu_perfcntr_start(gpu); + perf->next_jiffies = jiffies + SAMPLE_TIME; + +out: + mutex_unlock(&dev->struct_mutex); + return ret; +} + +static int perf_release(struct inode *inode, struct file *file) +{ + struct msm_perf_state *perf = inode->i_private; + struct msm_drm_private *priv = perf->dev->dev_private; + msm_gpu_perfcntr_stop(priv->gpu); + perf->open = false; + return 0; +} + + +static const struct file_operations perf_debugfs_fops = { + .owner = THIS_MODULE, + .open = perf_open, + .read = perf_read, + .llseek = no_llseek, + .release = perf_release, +}; + +int msm_perf_debugfs_init(struct drm_minor *minor) +{ + struct msm_drm_private *priv = minor->dev->dev_private; + struct msm_perf_state *perf; + + /* only create on first minor: */ + if (priv->perf) + return 0; + + perf = kzalloc(sizeof(*perf), GFP_KERNEL); + if (!perf) + return -ENOMEM; + + perf->dev = minor->dev; + + mutex_init(&perf->read_lock); + priv->perf = perf; + + perf->node = kzalloc(sizeof(*perf->node), GFP_KERNEL); + if (!perf->node) + goto fail; + + perf->ent = debugfs_create_file("perf", S_IFREG | S_IRUGO, + minor->debugfs_root, perf, &perf_debugfs_fops); + if (!perf->ent) { + DRM_ERROR("Cannot create /sys/kernel/debug/dri/%s/perf\n", + minor->debugfs_root->d_name.name); + goto fail; + } + + perf->node->minor = minor; + perf->node->dent = perf->ent; + perf->node->info_ent = NULL; + + mutex_lock(&minor->debugfs_lock); + list_add(&perf->node->list, &minor->debugfs_list); + mutex_unlock(&minor->debugfs_lock); + + return 0; + +fail: + msm_perf_debugfs_cleanup(minor); + return -1; +} + +void msm_perf_debugfs_cleanup(struct drm_minor *minor) +{ + struct msm_drm_private *priv = minor->dev->dev_private; + struct msm_perf_state *perf = priv->perf; + + if (!perf) + return; + + priv->perf = NULL; + + debugfs_remove(perf->ent); + + if (perf->node) { + mutex_lock(&minor->debugfs_lock); + list_del(&perf->node->list); + mutex_unlock(&minor->debugfs_lock); + kfree(perf->node); + } + + mutex_destroy(&perf->read_lock); + + kfree(perf); +} + +#endif -- 2.7.4