在Linux Kernel 的AMDGPU驱动实现中，dma-fence扮演着重要角色，AMDGPU的Render/解码操作可能涉及到多个方面同时引用buffer的情况，以渲染/视频解码场景为例，应用将渲染/解码命令写入和GPU共享的BUFFER之后，需要将任务提交给GPU的运行队列等待执行，这个时候，应用不仅要监听buffer中任务的执行情况，还要保证在任务执行完毕之前，不能再去修改BUFFER中的数据。而AMDGPU也需要排它性地执行BUFFER命令。在GPU执行结束，应用希望及时得到执行完的信息，以便回收BUFFER重新利用，通知一般由绑定到某个BUFFER上的中断完成。这些操作，如果使用经典的共享BUFFER+锁的方式进行保护和同步，不但效率十分低下，而且各类内核机制杂糅在一起，缺乏一个统一的管理框架，使代码难以维护。 dma-fence则提供了一套简单便捷的机框架，将原子操作，休眠唤醒，同步/异步事件通知等功能嵌入到各种类型的BUFFER管理框架中，将各类机制有机的结合在一起，减少了用户态空间的忙等，使buffer使用更加智能高效。
以AMDGPU解码视频为例，利用dma-fence ring buffer隔离了应用解码请求和解码任务任务本身，提交任务和解码完成通知均通过BUFFER绑定的dma-fence进行交互：
为了便于分析，我把5.4内核中AMDGPU这部分的实现提取出来，写了一个可以独立加载执行的内核模块demo，类似于CMODEL，任务提交上下文和解码完成通知上下文在AMDGPU驱动中分别用内核线程和中断实现，在demo中则全部用内核线程实现.
工作模型如下图所示：
fence array有256个slot,每个位置代表一个fence.构成一个RING BUFFER。sync_seq为写指针序号，sync_seq只增不减，sync_seq mod 256为在array中的index，相当于RING BUFFER的写指针。
last_seq代表当前已经消耗掉（已完成处理的和fence绑定的buffer)的序号，同sync_seq一样，它也是单调递增的自然序列，last_seq mod 256 为在array中的index，相当于ring buffer的读指针。
而fence_seq则表示写位置的一个抽样，这是一个在AMDGPU中的hardware ring是所有进程共享的，每个进程随时可能会提交新的渲染任务，所以sync_seq时刻在更新，不适合处理。fence_seq则是某个时间点的快照，在last_seq和fence_seq之间的buffer(fence)，都会在下一批中一并得到signal.
当提交过快，但是消耗较慢时，写方追上读方，此时需要进行生产-消费之间的同步，读端将会反压到写端，写端调用dma_fence_wait进行同步：
当消费方完成下一个任务后，释放当前slot，并且调用dma_fence_signal唤醒写线程，写线程受到信号后，继续执行，向fence array中填充新的任务：
dma fence 的free
拔出萝卜带出泥，通过job fence的free触发schedule fence和finish fence的释放，释放是通过引用计数归零触发的回调，在回调用中利用RCU异步释放FENCE占用的空间。
以上就是fence array的使用过程中的几个corner case.
附code：
#include <linux/kernel.h>
#include <linux/init.h>
#include <linux/types.h>
#include <linux/spinlock.h>
#include <linux/module.h>
#include <linux/fs.h>
#include <linux/errno.h>
#include <linux/mm.h>
#include <linux/sched/signal.h>
#include <linux/dma-fence.h>
#include <linux/slab.h>
#include <linux/delay.h>
#include <linux/kthread.h>
#include <linux/sched.h>
#include <uapi/linux/sched/types.h>#define assert(expr) \if (!(expr)) { \printk( "Assertion failed! %s,%s,%s,line=%d\n",\#expr,__FILE__,__func__,__LINE__); \BUG(); \}#define num_hw_submission 128
struct fence_driver {uint64_t gpu_addr;volatile uint32_t *cpu_addr;uint32_t sync_seq;atomic_t last_seq;bool initialized;bool initialized_emit;bool initialized_recv;unsigned num_fences_mask;spinlock_t lock;struct dma_fence **fences;struct mutex mutex;struct dma_fence *last_fence;struct timer_list timer;struct timer_list work_timer;wait_queue_head_t job_scheduled;
};struct fence_set {struct dma_fence scheduled;struct dma_fence finished;spinlock_t lock;
};struct job_fence {struct dma_fence job;void *data;
};static uint32_t fence_seq;
static struct fence_driver *ring;
static struct task_struct *fence_emit_task;
static struct task_struct *fence_recv_task;
static struct kmem_cache *job_fence_slab;
static struct kmem_cache *sched_fence_slab;static const char *dma_fence_get_name(struct dma_fence *fence)
{return "dma-fence-drv";
}static bool dma_fence_enable_signal(struct dma_fence *fence)
{if (!timer_pending(&ring->work_timer)) {mod_timer(&ring->work_timer, jiffies + HZ / 10);}printk("%s line %d, signal fenceno %lld.\n", __func__, __LINE__, fence->seqno);return true;
}void fencedrv_fence_free(struct rcu_head *rcu)
{struct fence_set *fs;struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);struct job_fence *jb = container_of(f, struct job_fence, job);assert(jb->data != NULL);fs = (struct fence_set *)jb->data;// the dma_fence_get must be symmentry with dma_fence_put during all the alive.assert((kref_read(&fs->scheduled.refcount) == 1));assert((kref_read(&fs->finished.refcount) == 1));//dma_fence_put(&fs->scheduled);dma_fence_put(&fs->finished);kmem_cache_free(job_fence_slab, jb);// dump_stack();
}static void fencedrv_dma_fence_release(struct dma_fence *fence)
{// typically usage for dma fence release by rcu.call_rcu(&fence->rcu, fencedrv_fence_free);
}static const struct dma_fence_ops fence_ops = {.get_driver_name = dma_fence_get_name,.get_timeline_name = dma_fence_get_name,.enable_signaling = dma_fence_enable_signal,.release = fencedrv_dma_fence_release,
};static int32_t fencedrv_get_ring_avail(void)
{uint32_t read_seq, write_seq;do {read_seq = atomic_read(&ring->last_seq);write_seq = ring->sync_seq;} while (atomic_read(&ring->last_seq) != read_seq);read_seq &= ring->num_fences_mask;write_seq &= ring->num_fences_mask;pr_err("%s line %d, read_seq %d, write_seq %d.\n",__func__, __LINE__, read_seq, write_seq);if (read_seq <= write_seq) {return write_seq - read_seq;} else {return write_seq + num_hw_submission * 2 - read_seq;}
}static const char *dma_fence_get_name_scheduled(struct dma_fence *fence)
{return "dma-fence-scheduled";
}static const char *dma_fence_get_name_finished(struct dma_fence *fence)
{return "dma-fence-finished";
}static void sched_fence_free(struct rcu_head *head);
static void fencedrv_dma_fence_release_scheduled(struct dma_fence *fence)
{struct fence_set *fs = container_of(fence, struct fence_set, scheduled);// typically usage for dma fence release by rcu.call_rcu(&fs->finished.rcu, sched_fence_free);
}static void fencedrv_dma_fence_release_finished(struct dma_fence *fence)
{struct fence_set *fs = container_of(fence, struct fence_set, finished);dma_fence_put(&fs->scheduled);//while (1);
}static const struct dma_fence_ops fence_scheduled_ops = {.get_driver_name = dma_fence_get_name_scheduled,.get_timeline_name = dma_fence_get_name_scheduled,.release = fencedrv_dma_fence_release_scheduled,
};static const struct dma_fence_ops fence_finished_ops = {.get_driver_name = dma_fence_get_name_finished,.get_timeline_name = dma_fence_get_name_finished,.release = fencedrv_dma_fence_release_finished,
};static struct fence_set *to_sched_fence(struct dma_fence *f)
{if (f->ops == &fence_scheduled_ops) {return container_of(f, struct fence_set, scheduled);}if (f->ops == &fence_finished_ops) {return container_of(f, struct fence_set, finished);}return NULL;
}static void sched_fence_free(struct rcu_head *head)
{struct dma_fence *f = container_of(head, struct dma_fence, rcu);struct fence_set *fs = to_sched_fence(f);if (fs == NULL)return;assert(f == &fs->finished);kmem_cache_free(sched_fence_slab, fs);dump_stack();
}static struct fence_set *init_fence_set(void)
{struct fence_set *fs = kmem_cache_alloc(sched_fence_slab, GFP_KERNEL);if (fs == NULL) {pr_err("%s line %d, alloc fence set from fence set slab failure.\n",__func__, __LINE__);return NULL;}spin_lock_init(&fs->lock);dma_fence_init(&fs->scheduled, &fence_scheduled_ops, &fs->lock, 0, 0);dma_fence_init(&fs->finished, &fence_finished_ops, &fs->lock, 0, 0);return fs;
}// ref amdgpu_fence_process
static int fence_recv_task_thread(void *data)
{struct sched_param sparam = {.sched_priority = 1};sched_setscheduler(current, SCHED_FIFO, &sparam);//mutex_lock(&ring->mutex);while (ring->initialized == false) {set_current_state(TASK_UNINTERRUPTIBLE);if (ring->initialized == true) {break;}//mutex_unlock(&ring->mutex);schedule();//mutex_lock(&ring->mutex);}set_current_state(TASK_RUNNING);//mutex_unlock(&ring->mutex);while (!kthread_should_stop() && ring->initialized_recv == true) {uint32_t seqno_next = 0;uint32_t seq, last_seq;int r;do {// last_seq is the read pointer of fence ring buffer.last_seq = atomic_read(&ring->last_seq);seq = *ring->cpu_addr;if (kthread_should_stop())return 0;} while (atomic_cmpxchg(&ring->last_seq, last_seq, seq) != last_seq);if (del_timer(&ring->work_timer) &&seq != ring->sync_seq) {mod_timer(&ring->work_timer, jiffies + HZ / 10);}//printk("%s line %d, last_seq %d, seq %d, sync_seq %d.\n", __func__, __LINE__, last_seq, seq, ring->sync_seq);if (unlikely(seq == last_seq)) {msleep(10);continue;}assert(seq > last_seq);last_seq &= ring->num_fences_mask;seq &= ring->num_fences_mask;//printk("%s line %d, last_seq %d, seq %d, sync_seq %d.\n", __func__, __LINE__, last_seq, seq, ring->sync_seq);do {struct dma_fence *fence, **ptr;++last_seq;last_seq &= ring->num_fences_mask;ptr = &ring->fences[last_seq];fence = rcu_dereference_protected(*ptr, 1);RCU_INIT_POINTER(*ptr, NULL);if (!fence) {continue;}if (seqno_next == 0 || seqno_next == fence->seqno) {seqno_next = fence->seqno + 1;} else { /*if (seqno_next != 0 && seqno_next != fence->seqno)*/pr_err("%s line %d, seqno is not continue, exptect %d, actual %lld.\n",__func__, __LINE__, seqno_next, fence->seqno);}printk("%s line %d, last_seq/slot %d, seq %d, signal %lld.\n",__func__, __LINE__, last_seq, seq, fence->seqno);if (list_empty(&fence->cb_list)) {printk("%s line %d, fence cb list is empty.\n",__func__, __LINE__);} else {printk("%s line %d, fence cb list is not empty.\n",__func__, __LINE__);}r = dma_fence_signal(fence);if (kthread_should_stop()) {dma_fence_put(fence);return 0;}if (r) {pr_err("%s line %d, fence already signaled.\n",__func__, __LINE__);continue;//BUG();}dma_fence_put(fence);} while (last_seq != seq);wake_up(&ring->job_scheduled);}set_current_state(TASK_RUNNING);return 0;
}// ref amdgpu_fence_emit.
static int fence_emit_task_thread(void *data)
{int r;uint64_t oldwaitseqno = 0;struct sched_param sparam = {.sched_priority = 1};sched_setscheduler(current, SCHED_FIFO, &sparam);//mutex_lock(&ring->mutex);while (ring->initialized == false) {set_current_state(TASK_UNINTERRUPTIBLE);if (ring->initialized == true) {break;}//mutex_unlock(&ring->mutex);schedule();//mutex_lock(&ring->mutex);}set_current_state(TASK_RUNNING);//mutex_unlock(&ring->mutex);while (!kthread_should_stop() && ring->initialized_emit == true) {
#if 0msleep(1000);printk("%s line %d.\n", __func__, __LINE__);
#elsestruct dma_fence __rcu **ptr;struct job_fence *fence;uint32_t seq;struct fence_set *fs = init_fence_set();fence = kmem_cache_alloc(job_fence_slab, GFP_KERNEL);if (fence == NULL) {pr_err("%s line %d, alloc fence from fence slab failure.\n",__func__, __LINE__);return -1;}// ring->sync_seq is fence ring write pointer.seq = ++ring->sync_seq;dma_fence_init(&fence->job, &fence_ops, &ring->lock, 0, seq);fence->data = fs;ptr = &ring->fences[seq & ring->num_fences_mask];//printk("%s line %d, seq = %d.\n", __func__, __LINE__, seq);if (kthread_should_stop()) {// will call fence_ops.release directly to free the fence.dma_fence_put(&fence->job);continue;}if (unlikely(rcu_dereference_protected(*ptr, 1))) {struct dma_fence *old;int diff;rcu_read_lock();old = dma_fence_get_rcu_safe(ptr);rcu_read_unlock();if (old) {mutex_lock(&ring->mutex);//dma_fence_get(old);ring->last_fence = old;mutex_unlock(&ring->mutex);r = dma_fence_wait(old, false);mutex_lock(&ring->mutex);ring->last_fence = NULL;dma_fence_put(old);mutex_unlock(&ring->mutex);if (kthread_should_stop() || r) {// will call fence_ops.release directly to free the fence.dma_fence_put(&fence->job);continue;}// if overlap happened, there must be a congruences on seq and old->seqno,which means seq≡ old->seqno mod(num_hw_submission * 2)// this implies seq = q*(num_hw_submission * 2) + old->seqno. q=1 typically.diff = seq - old->seqno;printk("%s line %d, fence wokenup, wakeseqno %lld, new adding seq %d, slot %d, diff %d, waken interval %lld, latestseq %d, avail %d.\n",__func__, __LINE__, old->seqno, seq, seq & ring->num_fences_mask, diff, old->seqno - oldwaitseqno,ring->sync_seq, fencedrv_get_ring_avail());if (diff != num_hw_submission * 2) {pr_err("%s line %d, fatal error, diff not match totoal ring.\n",__func__, __LINE__);}oldwaitseqno = old->seqno;}}#if 0printk("%s line %d, fence emit, seqno %lld, seq %d, slot %d.\n",__func__, __LINE__, fence->seqno, seq, seq & ring->num_fences_mask);
#endifrcu_assign_pointer(*ptr, dma_fence_get(&fence->job));// because no outer usage of fence, so put it here for free ok.dma_fence_put(&fence->job);
#endif}set_current_state(TASK_RUNNING);return 0;
}void work_timer_fn(struct timer_list *timer)
{uint32_t seqno_next = 0;uint32_t seq, last_seq;int r;do {last_seq = atomic_read(&ring->last_seq);seq = *ring->cpu_addr;} while (atomic_cmpxchg(&ring->last_seq, last_seq, seq) != last_seq);if (unlikely(seq == last_seq)) {goto end;}assert(seq > last_seq);last_seq &= ring->num_fences_mask;seq &= ring->num_fences_mask;do {struct dma_fence *fence, **ptr;++last_seq;last_seq &= ring->num_fences_mask;ptr = &ring->fences[last_seq];fence = rcu_dereference_protected(*ptr, 1);RCU_INIT_POINTER(*ptr, NULL);if (!fence) {continue;}if (seqno_next == 0 || seqno_next == fence->seqno) {seqno_next = fence->seqno + 1;} else { /*if (seqno_next != 0 && seqno_next != fence->seqno)*/pr_err("%s line %d, seqno is not continue, exptect %d, actual %lld.\n",__func__, __LINE__, seqno_next, fence->seqno);}r = dma_fence_signal(fence);if (r) {pr_err("%s line %d, fence already signaled.\n",__func__, __LINE__);continue;//BUG();}dma_fence_put(fence);} while (last_seq != seq);
end:pr_err("%s line %d, work timer triggerd.\n", __func__, __LINE__);mod_timer(timer, jiffies + HZ / 10);
}void gpu_process_thread(struct timer_list *timer)
{uint32_t seq, oldseq;seq = ring->sync_seq;oldseq = fence_seq;// trigger a job done on device.if (fence_seq == 0) {if (seq > 6)fence_seq = seq - 4;} else if ((seq - fence_seq) > 10) {fence_seq += (seq - fence_seq) / 2;assert(fence_seq > oldseq);}printk("%s line %d, timer trigger job, latest consume fence %d.\n",__func__, __LINE__, fence_seq);mod_timer(timer, jiffies + HZ / 2);
}static int fencedrv_wait_empty(void)
{uint64_t seq = READ_ONCE(ring->sync_seq);struct dma_fence *fence, **ptr;int r;if (!seq)return 0;fence_seq = seq;ptr = &ring->fences[seq & ring->num_fences_mask];rcu_read_lock();fence = rcu_dereference(*ptr);if (!fence || !dma_fence_get_rcu(fence)) {rcu_read_unlock();return 0;}rcu_read_unlock();r = dma_fence_wait(fence, false);printk("%s line %d, wait last fence %lld, seq %lld, r %d.\n", \__func__, __LINE__, fence->seqno, seq, r);dma_fence_put(fence);return r;
}static int __init fencedrv_init(void)
{if ((num_hw_submission & (num_hw_submission - 1)) != 0) {pr_err("%s line %d, num_hw_submission must be power of two.\n",__func__, __LINE__);return -1;}ring = kzalloc(sizeof(*ring), GFP_KERNEL);if (ring == NULL) {pr_err("%s line %d, alloc fence driver failure.\n",__func__, __LINE__);return -ENOMEM;}// fence_seq is a snap shot of sync_seq for deal with fence batchlly.ring->cpu_addr = &fence_seq;ring->gpu_addr = (uint64_t)&fence_seq;ring->sync_seq = 0;atomic_set(&ring->last_seq, 0);ring->initialized = false;ring->initialized_emit = false;ring->initialized_recv = false;ring->last_fence = NULL;ring->num_fences_mask = num_hw_submission * 2 - 1;init_waitqueue_head(&ring->job_scheduled);spin_lock_init(&ring->lock);ring->fences = kcalloc(num_hw_submission * 2, sizeof(void *), GFP_KERNEL);if (!ring->fences) {pr_err("%s line %d, alloc fence buffer failure.\n",__func__, __LINE__);return -ENOMEM;}printk("%s line %d, fence mask 0x%x, num_hw_submission 0x%x.\n",__func__, __LINE__, ring->num_fences_mask, num_hw_submission);job_fence_slab = kmem_cache_create("job_fence_slab", sizeof(struct job_fence), 0,SLAB_HWCACHE_ALIGN, NULL);if (!job_fence_slab) {pr_err("%s line %d, alloc job_fence_slab falure.\n",__func__, __LINE__);return -ENOMEM;}sched_fence_slab = kmem_cache_create("sched_fence_slab", sizeof(struct fence_set), 0,SLAB_HWCACHE_ALIGN, NULL);if (!sched_fence_slab) {pr_err("%s line %d, alloc sched_fence_slab falure.\n",__func__, __LINE__);return -ENOMEM;}mutex_init(&ring->mutex);fence_emit_task = kthread_run(fence_emit_task_thread, NULL, "fence_emit");if (IS_ERR(fence_emit_task)) {pr_err("%s line %d, create fence emit tsk failure.\n",__func__, __LINE__);return -1;}fence_recv_task = kthread_run(fence_recv_task_thread, NULL, "fence_recv");if (IS_ERR(fence_recv_task)) {pr_err("%s line %d, create fence recv tsk failure.\n",__func__, __LINE__);return -1;}timer_setup(&ring->timer, gpu_process_thread, TIMER_IRQSAFE);add_timer(&ring->timer);mod_timer(&ring->timer, jiffies + HZ / 2);timer_setup(&ring->work_timer, work_timer_fn, TIMER_IRQSAFE);add_timer(&ring->work_timer);mod_timer(&ring->work_timer, jiffies + HZ / 10);printk("%s line %d, module init.\n", __func__, __LINE__);ring->initialized = true;ring->initialized_emit = true;ring->initialized_recv = true;wake_up_process(fence_emit_task);wake_up_process(fence_recv_task);return 0;
}static void __exit fencedrv_exit(void)
{printk("%s line %d, module unload task begin.\n", __func__, __LINE__);del_timer_sync(&ring->work_timer);mutex_lock(&ring->mutex);if ((ring->last_fence != NULL) &&(!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &ring->last_fence->flags))) {ring->initialized_emit = false;dma_fence_signal(ring->last_fence);dma_fence_put(ring->last_fence);}mutex_unlock(&ring->mutex);kthread_stop(fence_emit_task);printk("%s line %d, module unload task mid.\n", __func__, __LINE__);del_timer_sync(&ring->timer);fencedrv_wait_empty();printk("%s line %d, sync wait avail %d.\n", __func__, __LINE__, fencedrv_get_ring_avail());wait_event_killable(ring->job_scheduled, fencedrv_get_ring_avail() <= 1);ring->initialized_recv = false;kthread_stop(fence_recv_task);printk("%s line %d, module unload task end.\n", __func__, __LINE__);ring->initialized = false;rcu_barrier();kmem_cache_destroy(job_fence_slab);kmem_cache_destroy(sched_fence_slab);kfree(ring->fences);kfree(ring);fence_emit_task = NULL;fence_recv_task = NULL;printk("%s line %d, module unload.\n", __func__, __LINE__);
}module_init(fencedrv_init);
module_exit(fencedrv_exit);
MODULE_LICENSE("GPL");
MODULE_AUTHOR("czl");参考博客
Fence_dma fence_笔落梦昙的博客-CSDN博客
