* [PATCH] habanalabs: add CS completion and timeout properties
@ 2021-01-26 7:00 Oded Gabbay
0 siblings, 0 replies; only message in thread
From: Oded Gabbay @ 2021-01-26 7:00 UTC (permalink / raw)
To: linux-kernel; +Cc: Ofir Bitton
From: Ofir Bitton <obitton@habana.ai>
In order to support staged submission feature, we need to
distinguish on which command submission we want to receive
timeout and for which we want to receive completion.
Signed-off-by: Ofir Bitton <obitton@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
---
.../habanalabs/common/command_submission.c | 104 ++++++++++++++----
drivers/misc/habanalabs/common/habanalabs.h | 15 ++-
drivers/misc/habanalabs/common/hw_queue.c | 24 +++-
3 files changed, 117 insertions(+), 26 deletions(-)
diff --git a/drivers/misc/habanalabs/common/command_submission.c b/drivers/misc/habanalabs/common/command_submission.c
index a5e9bb0a4855..57daff0e59ae 100644
--- a/drivers/misc/habanalabs/common/command_submission.c
+++ b/drivers/misc/habanalabs/common/command_submission.c
@@ -149,9 +149,10 @@ void hl_fence_get(struct hl_fence *fence)
kref_get(&fence->refcount);
}
-static void hl_fence_init(struct hl_fence *fence)
+static void hl_fence_init(struct hl_fence *fence, u64 sequence)
{
kref_init(&fence->refcount);
+ fence->cs_sequence = sequence;
fence->error = 0;
fence->timestamp = ktime_set(0, 0);
init_completion(&fence->completion);
@@ -184,6 +185,28 @@ static void cs_job_put(struct hl_cs_job *job)
kref_put(&job->refcount, cs_job_do_release);
}
+bool cs_needs_completion(struct hl_cs *cs)
+{
+ /* In case this is a staged CS, only the last CS in sequence should
+ * get a completion, any non staged CS will always get a completion
+ */
+ if (cs->staged_cs && !cs->staged_last)
+ return false;
+
+ return true;
+}
+
+bool cs_needs_timeout(struct hl_cs *cs)
+{
+ /* In case this is a staged CS, only the first CS in sequence should
+ * get a timeout, any non staged CS will always get a timeout
+ */
+ if (cs->staged_cs && !cs->staged_first)
+ return false;
+
+ return true;
+}
+
static bool is_cb_patched(struct hl_device *hdev, struct hl_cs_job *job)
{
/*
@@ -225,7 +248,7 @@ static int cs_parser(struct hl_fpriv *hpriv, struct hl_cs_job *job)
parser.queue_type = job->queue_type;
parser.is_kernel_allocated_cb = job->is_kernel_allocated_cb;
job->patched_cb = NULL;
- parser.completion = true;
+ parser.completion = cs_needs_completion(job->cs);
rc = hdev->asic_funcs->cs_parser(hdev, &parser);
@@ -291,8 +314,21 @@ static void complete_job(struct hl_device *hdev, struct hl_cs_job *job)
hl_debugfs_remove_job(hdev, job);
- if (job->queue_type == QUEUE_TYPE_EXT ||
- job->queue_type == QUEUE_TYPE_HW)
+ /* We decrement reference only for a CS that gets completion
+ * because the reference was incremented only for this kind of CS
+ * right before it was scheduled.
+ *
+ * In staged submission, only the last CS marked as 'staged_last'
+ * gets completion, hence its release function will be called from here.
+ * As for all the rest CS's in the staged submission which do not get
+ * completion, their CS reference will be decremented by the
+ * 'staged_last' CS during the CS release flow.
+ * All relevant PQ CI counters will be incremented during the CS release
+ * flow by calling 'hl_hw_queue_update_ci'.
+ */
+ if (cs_needs_completion(cs) &&
+ (job->queue_type == QUEUE_TYPE_EXT ||
+ job->queue_type == QUEUE_TYPE_HW))
cs_put(cs);
cs_job_put(job);
@@ -347,8 +383,8 @@ static void cs_do_release(struct kref *ref)
hdev->asic_funcs->hw_queues_unlock(hdev);
- /* Need to update CI for internal queues */
- hl_int_hw_queue_update_ci(cs);
+ /* Need to update CI for all queue jobs that does not get completion */
+ hl_hw_queue_update_ci(cs);
/* remove CS from CS mirror list */
spin_lock(&hdev->cs_mirror_lock);
@@ -359,6 +395,7 @@ static void cs_do_release(struct kref *ref)
* running from the TDR context
*/
if (!cs->timedout && hdev->timeout_jiffies != MAX_SCHEDULE_TIMEOUT) {
+ bool next_entry_found = false;
struct hl_cs *next;
if (cs->tdr_active)
@@ -367,10 +404,13 @@ static void cs_do_release(struct kref *ref)
spin_lock(&hdev->cs_mirror_lock);
/* queue TDR for next CS */
- next = list_first_entry_or_null(&hdev->cs_mirror_list,
- struct hl_cs, mirror_node);
+ list_for_each_entry(next, &hdev->cs_mirror_list, mirror_node)
+ if (cs_needs_timeout(next)) {
+ next_entry_found = true;
+ break;
+ }
- if (next && !next->tdr_active) {
+ if (next_entry_found && !next->tdr_active) {
next->tdr_active = true;
schedule_delayed_work(&next->work_tdr,
hdev->timeout_jiffies);
@@ -462,7 +502,8 @@ 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)
+ enum hl_cs_type cs_type, u64 user_sequence,
+ struct hl_cs **cs_new)
{
struct hl_cs_counters_atomic *cntr;
struct hl_fence *other = NULL;
@@ -511,6 +552,18 @@ static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx,
(hdev->asic_prop.max_pending_cs - 1)];
if (other && !completion_done(&other->completion)) {
+ /* If the following statement is true, it means we have reached
+ * a point in which only part of the staged submission was
+ * submitted and we don't have enough room in the 'cs_pending'
+ * array for the rest of the submission.
+ * This causes a deadlock because this CS will never be
+ * completed as it depends on future CS's for completion.
+ */
+ if (other->cs_sequence == user_sequence)
+ dev_crit_ratelimited(hdev->dev,
+ "Staged CS %llu deadlock due to lack of resources",
+ user_sequence);
+
dev_dbg_ratelimited(hdev->dev,
"Rejecting CS because of too many in-flights CS\n");
atomic64_inc(&ctx->cs_counters.max_cs_in_flight_drop_cnt);
@@ -529,7 +582,7 @@ static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx,
}
/* init hl_fence */
- hl_fence_init(&cs_cmpl->base_fence);
+ hl_fence_init(&cs_cmpl->base_fence, cs_cmpl->cs_seq);
cs->sequence = cs_cmpl->cs_seq;
@@ -825,7 +878,7 @@ static int hl_cs_copy_chunk_array(struct hl_device *hdev,
static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks,
u32 num_chunks, u64 *cs_seq, u32 flags)
{
- bool int_queues_only = true;
+ bool staged_mid, int_queues_only = true;
struct hl_device *hdev = hpriv->hdev;
struct hl_cs_chunk *cs_chunk_array;
struct hl_cs_counters_atomic *cntr;
@@ -833,9 +886,11 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks,
struct hl_cs_job *job;
struct hl_cs *cs;
struct hl_cb *cb;
+ u64 user_sequence;
int rc, i;
cntr = &hdev->aggregated_cs_counters;
+ user_sequence = *cs_seq;
*cs_seq = ULLONG_MAX;
rc = hl_cs_copy_chunk_array(hdev, &cs_chunk_array, chunks, num_chunks,
@@ -843,7 +898,14 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks,
if (rc)
goto out;
- rc = allocate_cs(hdev, hpriv->ctx, CS_TYPE_DEFAULT, &cs);
+ if ((flags & HL_CS_FLAGS_STAGED_SUBMISSION) &&
+ !(flags & HL_CS_FLAGS_STAGED_SUBMISSION_FIRST))
+ staged_mid = true;
+ else
+ staged_mid = false;
+
+ rc = allocate_cs(hdev, hpriv->ctx, CS_TYPE_DEFAULT,
+ staged_mid ? user_sequence : ULLONG_MAX, &cs);
if (rc)
goto free_cs_chunk_array;
@@ -911,8 +973,9 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks,
* Only increment for JOB on external or H/W queues, because
* only for those JOBs we get completion
*/
- if (job->queue_type == QUEUE_TYPE_EXT ||
- job->queue_type == QUEUE_TYPE_HW)
+ if (cs_needs_completion(cs) &&
+ (job->queue_type == QUEUE_TYPE_EXT ||
+ job->queue_type == QUEUE_TYPE_HW))
cs_get(cs);
hl_debugfs_add_job(hdev, job);
@@ -928,11 +991,14 @@ static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks,
}
}
- if (int_queues_only) {
+ /* We allow a CS with any queue type combination as long as it does
+ * not get a completion
+ */
+ if (int_queues_only && cs_needs_completion(cs)) {
atomic64_inc(&ctx->cs_counters.validation_drop_cnt);
atomic64_inc(&cntr->validation_drop_cnt);
dev_err(hdev->dev,
- "Reject CS %d.%llu because only internal queues jobs are present\n",
+ "Reject CS %d.%llu since it contains only internal queues jobs and needs completion\n",
cs->ctx->asid, cs->sequence);
rc = -EINVAL;
goto free_cs_object;
@@ -1037,7 +1103,7 @@ static int hl_submit_pending_cb(struct hl_fpriv *hpriv)
list_move_tail(&pending_cb->cb_node, &local_cb_list);
spin_unlock(&ctx->pending_cb_lock);
- rc = allocate_cs(hdev, ctx, CS_TYPE_DEFAULT, &cs);
+ rc = allocate_cs(hdev, ctx, CS_TYPE_DEFAULT, ULLONG_MAX, &cs);
if (rc)
goto add_list_elements;
@@ -1410,7 +1476,7 @@ static int cs_ioctl_signal_wait(struct hl_fpriv *hpriv, enum hl_cs_type cs_type,
}
}
- rc = allocate_cs(hdev, ctx, cs_type, &cs);
+ rc = allocate_cs(hdev, ctx, cs_type, ULLONG_MAX, &cs);
if (rc) {
if (cs_type == CS_TYPE_WAIT ||
cs_type == CS_TYPE_COLLECTIVE_WAIT)
diff --git a/drivers/misc/habanalabs/common/habanalabs.h b/drivers/misc/habanalabs/common/habanalabs.h
index e105612ed577..fd2fffd20ba1 100644
--- a/drivers/misc/habanalabs/common/habanalabs.h
+++ b/drivers/misc/habanalabs/common/habanalabs.h
@@ -486,6 +486,7 @@ struct asic_fixed_properties {
* struct hl_fence - software synchronization primitive
* @completion: fence is implemented using completion
* @refcount: refcount for this fence
+ * @cs_sequence: sequence of the corresponding command submission
* @error: mark this fence with error
* @timestamp: timestamp upon completion
*
@@ -493,6 +494,7 @@ struct asic_fixed_properties {
struct hl_fence {
struct completion completion;
struct kref refcount;
+ u64 cs_sequence;
int error;
ktime_t timestamp;
};
@@ -1176,7 +1178,11 @@ struct hl_userptr {
* @tdr_active: true if TDR was activated for this CS (to prevent
* double TDR activation).
* @aborted: true if CS was aborted due to some device error.
- * @timestamp: true if a timestmap must be captured upon completion
+ * @timestamp: true if a timestmap must be captured upon completion.
+ * @staged_last: true if this is the last staged CS and needs completion.
+ * @staged_first: true if this is the first staged CS and we need to receive
+ * timeout for this CS.
+ * @staged_cs: true if this CS is part of a staged submission.
*/
struct hl_cs {
u16 *jobs_in_queue_cnt;
@@ -1198,6 +1204,9 @@ struct hl_cs {
u8 tdr_active;
u8 aborted;
u8 timestamp;
+ u8 staged_last;
+ u8 staged_first;
+ u8 staged_cs;
};
/**
@@ -2118,7 +2127,7 @@ int hl_hw_queue_send_cb_no_cmpl(struct hl_device *hdev, u32 hw_queue_id,
int hl_hw_queue_schedule_cs(struct hl_cs *cs);
u32 hl_hw_queue_add_ptr(u32 ptr, u16 val);
void hl_hw_queue_inc_ci_kernel(struct hl_device *hdev, u32 hw_queue_id);
-void hl_int_hw_queue_update_ci(struct hl_cs *cs);
+void hl_hw_queue_update_ci(struct hl_cs *cs);
void hl_hw_queue_reset(struct hl_device *hdev, bool hard_reset);
#define hl_queue_inc_ptr(p) hl_hw_queue_add_ptr(p, 1)
@@ -2196,6 +2205,8 @@ int hl_gen_sob_mask(u16 sob_base, u8 sob_mask, u8 *mask);
void hl_fence_put(struct hl_fence *fence);
void hl_fence_get(struct hl_fence *fence);
void cs_get(struct hl_cs *cs);
+bool cs_needs_completion(struct hl_cs *cs);
+bool cs_needs_timeout(struct hl_cs *cs);
void goya_set_asic_funcs(struct hl_device *hdev);
void gaudi_set_asic_funcs(struct hl_device *hdev);
diff --git a/drivers/misc/habanalabs/common/hw_queue.c b/drivers/misc/habanalabs/common/hw_queue.c
index 76217258780a..ad440ae785a3 100644
--- a/drivers/misc/habanalabs/common/hw_queue.c
+++ b/drivers/misc/habanalabs/common/hw_queue.c
@@ -38,7 +38,7 @@ static inline int queue_free_slots(struct hl_hw_queue *q, u32 queue_len)
return (abs(delta) - queue_len);
}
-void hl_int_hw_queue_update_ci(struct hl_cs *cs)
+void hl_hw_queue_update_ci(struct hl_cs *cs)
{
struct hl_device *hdev = cs->ctx->hdev;
struct hl_hw_queue *q;
@@ -53,8 +53,13 @@ void hl_int_hw_queue_update_ci(struct hl_cs *cs)
if (!hdev->asic_prop.max_queues || q->queue_type == QUEUE_TYPE_HW)
return;
+ /* We must increment CI for every queue that will never get a
+ * completion, there are 2 scenarios this can happen:
+ * 1. All queues of a non completion CS will never get a completion.
+ * 2. Internal queues never gets completion.
+ */
for (i = 0 ; i < hdev->asic_prop.max_queues ; i++, q++) {
- if (q->queue_type == QUEUE_TYPE_INT)
+ if (!cs_needs_completion(cs) || q->queue_type == QUEUE_TYPE_INT)
atomic_add(cs->jobs_in_queue_cnt[i], &q->ci);
}
}
@@ -292,6 +297,10 @@ static void ext_queue_schedule_job(struct hl_cs_job *job)
len = job->job_cb_size;
ptr = cb->bus_address;
+ /* Skip completion flow in case this is a non completion CS */
+ if (!cs_needs_completion(job->cs))
+ goto submit_bd;
+
cq_pkt.data = cpu_to_le32(
((q->pi << CQ_ENTRY_SHADOW_INDEX_SHIFT)
& CQ_ENTRY_SHADOW_INDEX_MASK) |
@@ -318,6 +327,7 @@ static void ext_queue_schedule_job(struct hl_cs_job *job)
cq->pi = hl_cq_inc_ptr(cq->pi);
+submit_bd:
ext_and_hw_queue_submit_bd(hdev, q, ctl, len, ptr);
}
@@ -525,6 +535,7 @@ int hl_hw_queue_schedule_cs(struct hl_cs *cs)
struct hl_cs_job *job, *tmp;
struct hl_hw_queue *q;
int rc = 0, i, cq_cnt;
+ bool first_entry;
u32 max_queues;
cntr = &hdev->aggregated_cs_counters;
@@ -548,7 +559,9 @@ int hl_hw_queue_schedule_cs(struct hl_cs *cs)
switch (q->queue_type) {
case QUEUE_TYPE_EXT:
rc = ext_queue_sanity_checks(hdev, q,
- cs->jobs_in_queue_cnt[i], true);
+ cs->jobs_in_queue_cnt[i],
+ cs_needs_completion(cs) ?
+ true : false);
break;
case QUEUE_TYPE_INT:
rc = int_queue_sanity_checks(hdev, q,
@@ -586,9 +599,10 @@ int hl_hw_queue_schedule_cs(struct hl_cs *cs)
list_add_tail(&cs->mirror_node, &hdev->cs_mirror_list);
/* Queue TDR if the CS is the first entry and if timeout is wanted */
+ first_entry = list_first_entry(&hdev->cs_mirror_list,
+ struct hl_cs, mirror_node) == cs;
if ((hdev->timeout_jiffies != MAX_SCHEDULE_TIMEOUT) &&
- (list_first_entry(&hdev->cs_mirror_list,
- struct hl_cs, mirror_node) == cs)) {
+ first_entry && cs_needs_timeout(cs)) {
cs->tdr_active = true;
schedule_delayed_work(&cs->work_tdr, hdev->timeout_jiffies);
--
2.25.1
^ permalink raw reply related [flat|nested] only message in thread
only message in thread, other threads:[~2021-01-26 11:12 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-01-26 7:00 [PATCH] habanalabs: add CS completion and timeout properties Oded Gabbay
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).