On Tue, Dec 12, 2023 at 8:10 AM Fabiano Rosas <faro...@suse.de> wrote:
>
> Hao Xiang <hao.xi...@bytedance.com> writes:
>
> > * Use a safe thread queue for DSA task enqueue/dequeue.
> > * Implement DSA task submission.
> > * Implement DSA batch task submission.
> >
> > Signed-off-by: Hao Xiang <hao.xi...@bytedance.com>
> > ---
> > include/qemu/dsa.h | 35 ++++++++
> > util/dsa.c | 196 +++++++++++++++++++++++++++++++++++++++++++++
> > 2 files changed, 231 insertions(+)
> >
> > diff --git a/include/qemu/dsa.h b/include/qemu/dsa.h
> > index 30246b507e..23f55185be 100644
> > --- a/include/qemu/dsa.h
> > +++ b/include/qemu/dsa.h
> > @@ -12,6 +12,41 @@
> > #include <linux/idxd.h>
> > #include "x86intrin.h"
> >
> > +enum dsa_task_type {
>
> Our coding style requires CamelCase for enums and typedef'ed structures.
When I wrote this, I found numerous instances where snake case and no
typedef enum are used. But I do see the camel case and typedef'ed
instances now. Converted to that.
>
> > + DSA_TASK = 0,
> > + DSA_BATCH_TASK
> > +};
> > +
> > +enum dsa_task_status {
> > + DSA_TASK_READY = 0,
> > + DSA_TASK_PROCESSING,
> > + DSA_TASK_COMPLETION
> > +};
> > +
> > +typedef void (*buffer_zero_dsa_completion_fn)(void *);
>
> We don't really need the "buffer_zero" mention in any of this
> code. Simply dsa_batch_task or batch_task would suffice.
I removed "buffer_zero" prefix in some of the places.
>
> > +
> > +typedef struct buffer_zero_batch_task {
> > + struct dsa_hw_desc batch_descriptor;
> > + struct dsa_hw_desc *descriptors;
> > + struct dsa_completion_record batch_completion
> > __attribute__((aligned(32)));
> > + struct dsa_completion_record *completions;
> > + struct dsa_device_group *group;
> > + struct dsa_device *device;
> > + buffer_zero_dsa_completion_fn completion_callback;
> > + QemuSemaphore sem_task_complete;
> > + enum dsa_task_type task_type;
> > + enum dsa_task_status status;
> > + bool *results;
> > + int batch_size;
> > + QSIMPLEQ_ENTRY(buffer_zero_batch_task) entry;
> > +} buffer_zero_batch_task;
>
> I see data specific to this implementation and data coming from the
> library, maybe these would be better organized in two separate
> structures with the qemu-specific having a pointer to the generic
> one. Looking ahead in the series, there seems to be migration data
> coming into this as well.
I refactored to create a generic structure batch_task and a DSA
specific version dsa_batch_task. batch_task has a pointer to
dsa_batch_task if DSA compilation option is enabled.
>
> > +
> > +#else
> > +
> > +struct buffer_zero_batch_task {
> > + bool *results;
> > +};
> > +
> > #endif
> >
> > /**
> > diff --git a/util/dsa.c b/util/dsa.c
> > index 8edaa892ec..f82282ce99 100644
> > --- a/util/dsa.c
> > +++ b/util/dsa.c
> > @@ -245,6 +245,200 @@ dsa_device_group_get_next_device(struct
> > dsa_device_group *group)
> > return &group->dsa_devices[current];
> > }
> >
> > +/**
> > + * @brief Empties out the DSA task queue.
> > + *
> > + * @param group A pointer to the DSA device group.
> > + */
> > +static void
> > +dsa_empty_task_queue(struct dsa_device_group *group)
> > +{
> > + qemu_mutex_lock(&group->task_queue_lock);
> > + dsa_task_queue *task_queue = &group->task_queue;
> > + while (!QSIMPLEQ_EMPTY(task_queue)) {
> > + QSIMPLEQ_REMOVE_HEAD(task_queue, entry);
> > + }
> > + qemu_mutex_unlock(&group->task_queue_lock);
> > +}
> > +
> > +/**
> > + * @brief Adds a task to the DSA task queue.
> > + *
> > + * @param group A pointer to the DSA device group.
> > + * @param context A pointer to the DSA task to enqueue.
> > + *
> > + * @return int Zero if successful, otherwise a proper error code.
> > + */
> > +static int
> > +dsa_task_enqueue(struct dsa_device_group *group,
> > + struct buffer_zero_batch_task *task)
> > +{
> > + dsa_task_queue *task_queue = &group->task_queue;
> > + QemuMutex *task_queue_lock = &group->task_queue_lock;
> > + QemuCond *task_queue_cond = &group->task_queue_cond;
> > +
> > + bool notify = false;
> > +
> > + qemu_mutex_lock(task_queue_lock);
> > +
> > + if (!group->running) {
> > + fprintf(stderr, "DSA: Tried to queue task to stopped device
> > queue\n");
> > + qemu_mutex_unlock(task_queue_lock);
> > + return -1;
> > + }
> > +
> > + // The queue is empty. This enqueue operation is a 0->1 transition.
> > + if (QSIMPLEQ_EMPTY(task_queue))
> > + notify = true;
> > +
> > + QSIMPLEQ_INSERT_TAIL(task_queue, task, entry);
> > +
> > + // We need to notify the waiter for 0->1 transitions.
> > + if (notify)
> > + qemu_cond_signal(task_queue_cond);
> > +
> > + qemu_mutex_unlock(task_queue_lock);
> > +
> > + return 0;
> > +}
> > +
> > +/**
> > + * @brief Takes a DSA task out of the task queue.
> > + *
> > + * @param group A pointer to the DSA device group.
> > + * @return buffer_zero_batch_task* The DSA task being dequeued.
> > + */
> > +__attribute__((unused))
> > +static struct buffer_zero_batch_task *
> > +dsa_task_dequeue(struct dsa_device_group *group)
> > +{
> > + struct buffer_zero_batch_task *task = NULL;
> > + dsa_task_queue *task_queue = &group->task_queue;
> > + QemuMutex *task_queue_lock = &group->task_queue_lock;
> > + QemuCond *task_queue_cond = &group->task_queue_cond;
> > +
> > + qemu_mutex_lock(task_queue_lock);
> > +
> > + while (true) {
> > + if (!group->running)
> > + goto exit;
> > + task = QSIMPLEQ_FIRST(task_queue);
> > + if (task != NULL) {
> > + break;
> > + }
> > + qemu_cond_wait(task_queue_cond, task_queue_lock);
> > + }
> > +
> > + QSIMPLEQ_REMOVE_HEAD(task_queue, entry);
> > +
> > +exit:
> > + qemu_mutex_unlock(task_queue_lock);
> > + return task;
> > +}
> > +
> > +/**
> > + * @brief Submits a DSA work item to the device work queue.
> > + *
> > + * @param wq A pointer to the DSA work queue's device memory.
> > + * @param descriptor A pointer to the DSA work item descriptor.
> > + *
> > + * @return Zero if successful, non-zero otherwise.
> > + */
> > +static int
> > +submit_wi_int(void *wq, struct dsa_hw_desc *descriptor)
> > +{
> > + uint64_t retry = 0;
> > +
> > + _mm_sfence();
> > +
> > + while (true) {
> > + if (_enqcmd(wq, descriptor) == 0) {
> > + break;
> > + }
> > + retry++;
> > + if (retry > max_retry_count) {
>
> 'max_retry_count' is UINT64_MAX so 'retry' will wrap around.
>
> > + fprintf(stderr, "Submit work retry %lu times.\n", retry);
> > + exit(1);
>
> Is this not the case where we'd fallback to the CPU?
"retry" here means _enqcmd returned a failure because the shared DSA
queue is full. When we run out of retry counts, we definitely have a
bug that prevents any DSA task from completing. So this situation is
really not expected and we don't want to fallback to use CPU.
>
> You should not exit() here, but return non-zero as the documentation
> mentions and the callers expect.
I will propagate this error all the way up to multifd_send_thread.
>
> > + }
> > + }
> > +
> > + return 0;
> > +}
> > +
> > +/**
> > + * @brief Synchronously submits a DSA work item to the
> > + * device work queue.
> > + *
> > + * @param wq A pointer to the DSA worjk queue's device memory.
> > + * @param descriptor A pointer to the DSA work item descriptor.
> > + *
> > + * @return int Zero if successful, non-zero otherwise.
> > + */
> > +__attribute__((unused))
> > +static int
> > +submit_wi(void *wq, struct dsa_hw_desc *descriptor)
> > +{
> > + return submit_wi_int(wq, descriptor);
> > +}
> > +
> > +/**
> > + * @brief Asynchronously submits a DSA work item to the
> > + * device work queue.
> > + *
> > + * @param task A pointer to the buffer zero task.
> > + *
> > + * @return int Zero if successful, non-zero otherwise.
> > + */
> > +__attribute__((unused))
> > +static int
> > +submit_wi_async(struct buffer_zero_batch_task *task)
> > +{
> > + struct dsa_device_group *device_group = task->group;
> > + struct dsa_device *device_instance = task->device;
> > + int ret;
> > +
> > + assert(task->task_type == DSA_TASK);
> > +
> > + task->status = DSA_TASK_PROCESSING;
> > +
> > + ret = submit_wi_int(device_instance->work_queue,
> > + &task->descriptors[0]);
> > + if (ret != 0)
> > + return ret;
> > +
> > + return dsa_task_enqueue(device_group, task);
> > +}
> > +
> > +/**
> > + * @brief Asynchronously submits a DSA batch work item to the
> > + * device work queue.
> > + *
> > + * @param batch_task A pointer to the batch buffer zero task.
> > + *
> > + * @return int Zero if successful, non-zero otherwise.
> > + */
> > +__attribute__((unused))
> > +static int
> > +submit_batch_wi_async(struct buffer_zero_batch_task *batch_task)
> > +{
> > + struct dsa_device_group *device_group = batch_task->group;
> > + struct dsa_device *device_instance = batch_task->device;
> > + int ret;
> > +
> > + assert(batch_task->task_type == DSA_BATCH_TASK);
> > + assert(batch_task->batch_descriptor.desc_count <=
> > batch_task->batch_size);
> > + assert(batch_task->status == DSA_TASK_READY);
> > +
> > + batch_task->status = DSA_TASK_PROCESSING;
> > +
> > + ret = submit_wi_int(device_instance->work_queue,
> > + &batch_task->batch_descriptor);
> > + if (ret != 0)
> > + return ret;
> > +
> > + return dsa_task_enqueue(device_group, batch_task);
> > +}
>
> At this point in the series submit_wi_async() and
> submit_batch_wi_async() look the same to me without the asserts. Can't
> we consolidate them?
>
> There's also the fact that both functions receive a _batch_ task but one
> is supposed to work in batches and the other is not. That could be
> solved by renaming the structure I guess.
So we do need to have two functions to handle a single task and a
batch task respectively. This is due to how DSA is designed at the
lower level. When we submit a task to DSA hardware, the task
description can be an individual task or a batch task containing a
pointer to an array of individual tasks. The workflow tries to
aggregate a lot of individual tasks and put them into a batch task.
However, there are times when only 1 task is available but DSA doesn't
accept a batch task description with only 1 individual task in it so
we always need a path to submit an individual task. I used to have two
data structures representing an individual task and a batch task but I
converged them into the batch task right now. The two functions are
just using different fields of the same structure to process
individual task vs batch task. submit_wi_async and
submit_batch_wi_async are different on the actual descriptor passed
into the submit_wi_int call. Yes, the two functions look similar but
they are not completely the same and because its implementation is so
simple it doesn't worth adding a unified helper layer to have them
both calling that helper layer. I went back and forth between the
current implementation and the solution you suggested but ended up
using the current implementation. But let me know if you still prefer
a converged helper function.
>
> > +
> > /**
> > * @brief Check if DSA is running.
> > *
> > @@ -301,6 +495,8 @@ void dsa_stop(void)
> > if (!group->running) {
> > return;
> > }
> > +
> > + dsa_empty_task_queue(group);
> > }
> >
> > /**
