UbiquitousOS
/
OpenCloudOS-Kernel
mirror of https://gitee.com/OpenCloudOS/OpenCloudOS-Kernel.git
11
0
Fork 0
Code Issues Projects Releases Wiki Activity

habanalabs: change aggregate cs counters to atomic 

In case we will have multiple contexts/processes, we can't just
increment aggregated counters. We need to make them atomic as they can
be incremented by multiple processes

Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
This commit is contained in:
Oded Gabbay 2020-09-30 14:25:55 +03:00
parent 6c38e7c60f
commit 23c15ae615
4 changed files with 60 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

42
drivers/misc/habanalabs/common/command_submission.c
View File

@ -242,20 +242,6 @@ static void free_job(struct hl_device *hdev, struct hl_cs_job *job)
kfree(job);
}
static void cs_counters_aggregate(struct hl_device *hdev, struct hl_ctx *ctx)
{
hdev->aggregated_cs_counters.device_in_reset_drop_cnt +=
ctx->cs_counters.device_in_reset_drop_cnt;
hdev->aggregated_cs_counters.out_of_mem_drop_cnt +=
ctx->cs_counters.out_of_mem_drop_cnt;
hdev->aggregated_cs_counters.parsing_drop_cnt +=
ctx->cs_counters.parsing_drop_cnt;
hdev->aggregated_cs_counters.queue_full_drop_cnt +=
ctx->cs_counters.queue_full_drop_cnt;
hdev->aggregated_cs_counters.max_cs_in_flight_drop_cnt +=
ctx->cs_counters.max_cs_in_flight_drop_cnt;
}
static void cs_do_release(struct kref *ref)
{
struct hl_cs *cs = container_of(ref, struct hl_cs,
@ -358,7 +344,6 @@ static void cs_do_release(struct kref *ref)
complete_all(&cs->fence->completion);
hl_fence_put(cs->fence);
cs_counters_aggregate(hdev, cs->ctx);
kfree(cs->jobs_in_queue_cnt);
kfree(cs);
@ -397,11 +382,14 @@ static void cs_timedout(struct work_struct *work)
static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx,
enum hl_cs_type cs_type, struct hl_cs **cs_new)
{
struct hl_cs_compl *cs_cmpl;
struct hl_cs_counters_atomic *cntr;
struct hl_fence *other = NULL;
struct hl_cs_compl *cs_cmpl;
struct hl_cs *cs;
int rc;
cntr = &hdev->aggregated_cs_counters;
cs = kzalloc(sizeof(*cs), GFP_ATOMIC);
if (!cs)
return -ENOMEM;
@ -436,6 +424,7 @@ static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx,
dev_dbg_ratelimited(hdev->dev,
"Rejecting CS because of too many in-flights CS\n");
ctx->cs_counters.max_cs_in_flight_drop_cnt++;
atomic64_inc(&cntr->max_cs_in_flight_drop_cnt);
rc = -EAGAIN;
goto free_fence;
}
@ -610,6 +599,7 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks,
{
struct hl_device *hdev = hpriv->hdev;
struct hl_cs_chunk *cs_chunk_array;
struct hl_cs_counters_atomic *cntr;
struct hl_cs_job *job;
struct hl_cs *cs;
struct hl_cb *cb;
@ -617,6 +607,7 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks,
u32 size_to_copy;
int rc, i;
cntr = &hdev->aggregated_cs_counters;
*cs_seq = ULLONG_MAX;
if (num_chunks > HL_MAX_JOBS_PER_CS) {
@ -664,6 +655,7 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks,
&is_kernel_allocated_cb);
if (rc) {
hpriv->ctx->cs_counters.parsing_drop_cnt++;
atomic64_inc(&cntr->parsing_drop_cnt);
goto free_cs_object;
}
@ -671,6 +663,7 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks,
cb = get_cb_from_cs_chunk(hdev, &hpriv->cb_mgr, chunk);
if (!cb) {
hpriv->ctx->cs_counters.parsing_drop_cnt++;
atomic64_inc(&cntr->parsing_drop_cnt);
rc = -EINVAL;
goto free_cs_object;
}
@ -685,6 +678,7 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks,
is_kernel_allocated_cb);
if (!job) {
hpriv->ctx->cs_counters.out_of_mem_drop_cnt++;
atomic64_inc(&cntr->out_of_mem_drop_cnt);
dev_err(hdev->dev, "Failed to allocate a new job\n");
rc = -ENOMEM;
if (is_kernel_allocated_cb)
@ -718,6 +712,7 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks,
rc = cs_parser(hpriv, job);
if (rc) {
hpriv->ctx->cs_counters.parsing_drop_cnt++;
atomic64_inc(&cntr->parsing_drop_cnt);
dev_err(hdev->dev,
"Failed to parse JOB %d.%llu.%d, err %d, rejecting the CS\n",
cs->ctx->asid, cs->sequence, job->id, rc);
@ -727,6 +722,7 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks,
if (int_queues_only) {
hpriv->ctx->cs_counters.parsing_drop_cnt++;
atomic64_inc(&cntr->parsing_drop_cnt);
dev_err(hdev->dev,
"Reject CS %d.%llu because only internal queues jobs are present\n",
cs->ctx->asid, cs->sequence);
@ -768,20 +764,22 @@ static int cs_ioctl_signal_wait(struct hl_fpriv *hpriv, enum hl_cs_type cs_type,
void __user *chunks, u32 num_chunks,
u64 *cs_seq)
{
struct hl_device *hdev = hpriv->hdev;
struct hl_ctx *ctx = hpriv->ctx;
u32 size_to_copy, q_idx, signal_seq_arr_len, cb_size;
struct hl_cs_chunk *cs_chunk_array, *chunk;
struct hw_queue_properties *hw_queue_prop;
u64 *signal_seq_arr = NULL, signal_seq;
struct hl_device *hdev = hpriv->hdev;
struct hl_cs_counters_atomic *cntr;
struct hl_fence *sig_fence = NULL;
struct hl_ctx *ctx = hpriv->ctx;
enum hl_queue_type q_type;
struct hl_cs_job *job;
struct hl_cs *cs;
struct hl_cb *cb;
enum hl_queue_type q_type;
u64 *signal_seq_arr = NULL, signal_seq;
u32 size_to_copy, q_idx, signal_seq_arr_len, cb_size;
int rc;
*cs_seq = ULLONG_MAX;
cntr = &hdev->aggregated_cs_counters;
if (num_chunks > HL_MAX_JOBS_PER_CS) {
dev_err(hdev->dev,
@ -920,6 +918,7 @@ static int cs_ioctl_signal_wait(struct hl_fpriv *hpriv, enum hl_cs_type cs_type,
job = hl_cs_allocate_job(hdev, q_type, true);
if (!job) {
ctx->cs_counters.out_of_mem_drop_cnt++;
atomic64_inc(&cntr->out_of_mem_drop_cnt);
dev_err(hdev->dev, "Failed to allocate a new job\n");
rc = -ENOMEM;
goto put_cs;
@ -934,6 +933,7 @@ static int cs_ioctl_signal_wait(struct hl_fpriv *hpriv, enum hl_cs_type cs_type,
q_type == QUEUE_TYPE_HW && hdev->mmu_enable);
if (!cb) {
ctx->cs_counters.out_of_mem_drop_cnt++;
atomic64_inc(&cntr->out_of_mem_drop_cnt);
kfree(job);
rc = -EFAULT;
goto put_cs;

17
drivers/misc/habanalabs/common/habanalabs.h
View File

@ -1067,6 +1067,21 @@ struct hl_cs_parser {
u8 contains_dma_pkt;
};
/**
* struct hl_info_cs_counters - command submission counters
* @out_of_mem_drop_cnt: dropped due to memory allocation issue
* @parsing_drop_cnt: dropped due to error in packet parsing
* @queue_full_drop_cnt: dropped due to queue full
* @device_in_reset_drop_cnt: dropped due to device in reset
* @max_cs_in_flight_drop_cnt: dropped due to maximum CS in-flight
*/
struct hl_cs_counters_atomic {
atomic64_t out_of_mem_drop_cnt;
atomic64_t parsing_drop_cnt;
atomic64_t queue_full_drop_cnt;
atomic64_t device_in_reset_drop_cnt;
atomic64_t max_cs_in_flight_drop_cnt;
};
/*
* MEMORY STRUCTURE
@ -1649,7 +1664,7 @@ struct hl_device {
struct hl_device_idle_busy_ts *idle_busy_ts_arr;
struct hl_cs_counters aggregated_cs_counters;
struct hl_cs_counters_atomic aggregated_cs_counters;
struct hl_mmu_priv mmu_priv;
struct hl_mmu_funcs mmu_func;

20
drivers/misc/habanalabs/common/habanalabs_ioctl.c
View File

@ -314,10 +314,13 @@ static int clk_throttle_info(struct hl_fpriv *hpriv, struct hl_info_args *args)
static int cs_counters_info(struct hl_fpriv *hpriv, struct hl_info_args *args)
{
struct hl_device *hdev = hpriv->hdev;
struct hl_info_cs_counters cs_counters = { {0} };
u32 max_size = args->return_size;
void __user *out = (void __user *) (uintptr_t) args->return_pointer;
struct hl_info_cs_counters cs_counters = { {0} };
struct hl_device *hdev = hpriv->hdev;
struct hl_cs_counters_atomic *cntr;
u32 max_size = args->return_size;
cntr = &hdev->aggregated_cs_counters;
if ((!max_size) || (!out))
return -EINVAL;
@ -325,6 +328,17 @@ static int cs_counters_info(struct hl_fpriv *hpriv, struct hl_info_args *args)
memcpy(&cs_counters.cs_counters, &hdev->aggregated_cs_counters,
sizeof(struct hl_cs_counters));
cs_counters.cs_counters.out_of_mem_drop_cnt =
atomic64_read(&cntr->out_of_mem_drop_cnt);
cs_counters.cs_counters.parsing_drop_cnt =
atomic64_read(&cntr->parsing_drop_cnt);
cs_counters.cs_counters.queue_full_drop_cnt =
atomic64_read(&cntr->queue_full_drop_cnt);
cs_counters.cs_counters.device_in_reset_drop_cnt =
atomic64_read(&cntr->device_in_reset_drop_cnt);
cs_counters.cs_counters.max_cs_in_flight_drop_cnt =
atomic64_read(&cntr->max_cs_in_flight_drop_cnt);
if (hpriv->ctx)
memcpy(&cs_counters.ctx_cs_counters, &hpriv->ctx->cs_counters,
sizeof(struct hl_cs_counters));

7
drivers/misc/habanalabs/common/hw_queue.c
View File

@ -484,17 +484,21 @@ static void init_signal_wait_cs(struct hl_cs *cs)
*/
int hl_hw_queue_schedule_cs(struct hl_cs *cs)
{
struct hl_cs_counters_atomic *cntr;
struct hl_ctx *ctx = cs->ctx;
struct hl_device *hdev = ctx->hdev;
struct hl_cs_job *job, *tmp;
struct hl_hw_queue *q;
u32 max_queues;
int rc = 0, i, cq_cnt;
u32 max_queues;
cntr = &hdev->aggregated_cs_counters;
hdev->asic_funcs->hw_queues_lock(hdev);
if (hl_device_disabled_or_in_reset(hdev)) {
ctx->cs_counters.device_in_reset_drop_cnt++;
atomic64_inc(&cntr->device_in_reset_drop_cnt);
dev_err(hdev->dev,
"device is disabled or in reset, CS rejected!\n");
rc = -EPERM;
@ -528,6 +532,7 @@ int hl_hw_queue_schedule_cs(struct hl_cs *cs)
if (rc) {
ctx->cs_counters.queue_full_drop_cnt++;
atomic64_inc(&cntr->queue_full_drop_cnt);
goto unroll_cq_resv;
}