[Intel-xe] [PATCH 09/22] drm/xe: Add TDR for invalidation fence timeout cleanup
Rodrigo Vivi
rodrigo.vivi at intel.com
Fri Feb 3 20:23:56 UTC 2023
From: Matthew Brost <matthew.brost at intel.com>
Endless fences are not good, add a TDR to cleanup any invalidation
fences which have not received an invalidation message within a timeout
period.
Signed-off-by: Matthew Brost <matthew.brost at intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi at intel.com>
---
drivers/gpu/drm/xe/xe_gt_tlb_invalidation.c | 58 +++++++++++++++++--
.../gpu/drm/xe/xe_gt_tlb_invalidation_types.h | 2 +
drivers/gpu/drm/xe/xe_gt_types.h | 5 ++
drivers/gpu/drm/xe/xe_trace.h | 5 ++
4 files changed, 65 insertions(+), 5 deletions(-)
diff --git a/drivers/gpu/drm/xe/xe_gt_tlb_invalidation.c b/drivers/gpu/drm/xe/xe_gt_tlb_invalidation.c
index 4d179357ce65..9e026fd0a45d 100644
--- a/drivers/gpu/drm/xe/xe_gt_tlb_invalidation.c
+++ b/drivers/gpu/drm/xe/xe_gt_tlb_invalidation.c
@@ -9,12 +9,45 @@
#include "xe_guc_ct.h"
#include "xe_trace.h"
+#define TLB_TIMEOUT (HZ / 4)
+
static struct xe_gt *
guc_to_gt(struct xe_guc *guc)
{
return container_of(guc, struct xe_gt, uc.guc);
}
+static void xe_gt_tlb_fence_timeout(struct work_struct *work)
+{
+ struct xe_gt *gt = container_of(work, struct xe_gt,
+ tlb_invalidation.fence_tdr.work);
+ struct xe_gt_tlb_invalidation_fence *fence, *next;
+
+ mutex_lock(>->uc.guc.ct.lock);
+ list_for_each_entry_safe(fence, next,
+ >->tlb_invalidation.pending_fences, link) {
+ s64 since_inval_ms = ktime_ms_delta(ktime_get(),
+ fence->invalidation_time);
+
+ if (msecs_to_jiffies(since_inval_ms) < TLB_TIMEOUT)
+ break;
+
+ trace_xe_gt_tlb_invalidation_fence_timeout(fence);
+ drm_err(>_to_xe(gt)->drm, "TLB invalidation fence timeout, seqno=%d",
+ fence->seqno);
+
+ list_del(&fence->link);
+ fence->base.error = -ETIME;
+ dma_fence_signal(&fence->base);
+ dma_fence_put(&fence->base);
+ }
+ if (!list_empty(>->tlb_invalidation.pending_fences))
+ queue_delayed_work(system_wq,
+ >->tlb_invalidation.fence_tdr,
+ TLB_TIMEOUT);
+ mutex_unlock(>->uc.guc.ct.lock);
+}
+
/**
* xe_gt_tlb_invalidation_init - Initialize GT TLB invalidation state
* @gt: graphics tile
@@ -30,6 +63,8 @@ int xe_gt_tlb_invalidation_init(struct xe_gt *gt)
INIT_LIST_HEAD(>->tlb_invalidation.pending_fences);
spin_lock_init(>->tlb_invalidation.lock);
gt->tlb_invalidation.fence_context = dma_fence_context_alloc(1);
+ INIT_DELAYED_WORK(>->tlb_invalidation.fence_tdr,
+ xe_gt_tlb_fence_timeout);
return 0;
}
@@ -44,6 +79,8 @@ int xe_gt_tlb_invalidation_init(struct xe_gt *gt)
{
struct xe_gt_tlb_invalidation_fence *fence, *next;
+ cancel_delayed_work(>->tlb_invalidation.fence_tdr);
+
mutex_lock(>->uc.guc.ct.lock);
list_for_each_entry_safe(fence, next,
>->tlb_invalidation.pending_fences, link) {
@@ -67,6 +104,7 @@ static int send_tlb_invalidation(struct xe_guc *guc,
};
int seqno;
int ret;
+ bool queue_work;
/*
* XXX: The seqno algorithm relies on TLB invalidation being processed
@@ -76,10 +114,7 @@ static int send_tlb_invalidation(struct xe_guc *guc,
mutex_lock(&guc->ct.lock);
seqno = gt->tlb_invalidation.seqno;
if (fence) {
- /*
- * FIXME: How to deal TLB invalidation timeout, right now we
- * just have an endless fence which isn't ideal.
- */
+ queue_work = list_empty(>->tlb_invalidation.pending_fences);
fence->seqno = seqno;
list_add_tail(&fence->link,
>->tlb_invalidation.pending_fences);
@@ -92,6 +127,13 @@ static int send_tlb_invalidation(struct xe_guc *guc,
gt->tlb_invalidation.seqno = 1;
ret = xe_guc_ct_send_locked(&guc->ct, action, ARRAY_SIZE(action),
G2H_LEN_DW_TLB_INVALIDATE, 1);
+ if (!ret && fence) {
+ fence->invalidation_time = ktime_get();
+ if (queue_work)
+ queue_delayed_work(system_wq,
+ >->tlb_invalidation.fence_tdr,
+ TLB_TIMEOUT);
+ }
if (!ret)
ret = seqno;
mutex_unlock(&guc->ct.lock);
@@ -152,7 +194,7 @@ int xe_gt_tlb_invalidation_wait(struct xe_gt *gt, int seqno)
*/
ret = wait_event_timeout(guc->ct.wq,
tlb_invalidation_seqno_past(gt, seqno),
- HZ / 5);
+ TLB_TIMEOUT);
if (!ret) {
drm_err(&xe->drm, "TLB invalidation time'd out, seqno=%d, recv=%d\n",
seqno, gt->tlb_invalidation.seqno_recv);
@@ -201,6 +243,12 @@ int xe_guc_tlb_invalidation_done_handler(struct xe_guc *guc, u32 *msg, u32 len)
if (fence && tlb_invalidation_seqno_past(gt, fence->seqno)) {
trace_xe_gt_tlb_invalidation_fence_signal(fence);
list_del(&fence->link);
+ if (!list_empty(>->tlb_invalidation.pending_fences))
+ mod_delayed_work(system_wq,
+ >->tlb_invalidation.fence_tdr,
+ TLB_TIMEOUT);
+ else
+ cancel_delayed_work(>->tlb_invalidation.fence_tdr);
dma_fence_signal(&fence->base);
dma_fence_put(&fence->base);
}
diff --git a/drivers/gpu/drm/xe/xe_gt_tlb_invalidation_types.h b/drivers/gpu/drm/xe/xe_gt_tlb_invalidation_types.h
index ab57c14c6d14..934c828efe31 100644
--- a/drivers/gpu/drm/xe/xe_gt_tlb_invalidation_types.h
+++ b/drivers/gpu/drm/xe/xe_gt_tlb_invalidation_types.h
@@ -21,6 +21,8 @@ struct xe_gt_tlb_invalidation_fence {
struct list_head link;
/** @seqno: seqno of TLB invalidation to signal fence one */
int seqno;
+ /** @invalidation_time: time of TLB invalidation */
+ ktime_t invalidation_time;
};
#endif
diff --git a/drivers/gpu/drm/xe/xe_gt_types.h b/drivers/gpu/drm/xe/xe_gt_types.h
index 3b2d9842add7..a40fab262ac9 100644
--- a/drivers/gpu/drm/xe/xe_gt_types.h
+++ b/drivers/gpu/drm/xe/xe_gt_types.h
@@ -174,6 +174,11 @@ struct xe_gt {
* invaliations, protected by CT lock
*/
struct list_head pending_fences;
+ /**
+ * @fence_tdr: schedules a delayed call to
+ * xe_gt_tlb_fence_timeout after the timeut interval is over.
+ */
+ struct delayed_work fence_tdr;
/** @fence_context: context for TLB invalidation fences */
u64 fence_context;
/**
diff --git a/drivers/gpu/drm/xe/xe_trace.h b/drivers/gpu/drm/xe/xe_trace.h
index 373b0825ec79..1774658b18b7 100644
--- a/drivers/gpu/drm/xe/xe_trace.h
+++ b/drivers/gpu/drm/xe/xe_trace.h
@@ -69,6 +69,11 @@ DEFINE_EVENT(xe_gt_tlb_invalidation_fence, xe_gt_tlb_invalidation_fence_signal,
TP_ARGS(fence)
);
+DEFINE_EVENT(xe_gt_tlb_invalidation_fence, xe_gt_tlb_invalidation_fence_timeout,
+ TP_PROTO(struct xe_gt_tlb_invalidation_fence *fence),
+ TP_ARGS(fence)
+);
+
DECLARE_EVENT_CLASS(xe_bo,
TP_PROTO(struct xe_bo *bo),
TP_ARGS(bo),
--
2.39.1
More information about the Intel-xe
mailing list