With the current deferred-submission model, if a problem arises part-way
through the insertion of instructions into the ringbuffer (e.g. due to
one of the begin() calls finding there's not enough space), we avoid
sending the incomplete sequence to the h/w; but currently have no means
of undoing the work so far, which will lead to undefined behaviour when
the next batch is submitted (probably TDR will trigger a reset first,
though, and clean up the ring state).
A future idea is to move to an atomic-submission model, where all the
space required for a batch submission is reserved up front, and in the
event of failure partway through, the work can be abandoned without
side-effects. This will be required for the forthcoming GPU scheduler
(specifically, for preemption).
To support this, we allow nested begin/advance pairs. Specifically,
the outermost pair defines the total space reservation; inner pairs
can be nested ad lib, but all inner reservations at any level must
fit entirely within the outermost one. Thus, this is permitted:
begin(128) - guarantees that up to 128 dwords can now be
emitted without waiting for more freespace
begin(6)
advance
begin(10)
advance
begin(8)
advance
etc, as long as the total is no more than 128 dwords
advance-and-submit
The execbuffer code will later be enhanced to use this approach. In the
mean time, the traditional single-level begin/advance mechanism remains
fully supported.
This commit changes only the begin/advance checking code, to permit (but
not require) nested begin/advance pairs.
Signed-off-by: Dave Gordon <[email protected]>
---
drivers/gpu/drm/i915/intel_ringbuffer.h | 15 ++++++++++++---
1 file changed, 12 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.h
b/drivers/gpu/drm/i915/intel_ringbuffer.h
index a10d271..71cb3ef 100644
--- a/drivers/gpu/drm/i915/intel_ringbuffer.h
+++ b/drivers/gpu/drm/i915/intel_ringbuffer.h
@@ -416,8 +416,17 @@ static inline void __intel_ringbuffer_begin(struct
intel_ringbuffer *ringbuf,
WARN_ON(nbytes <= 0);
if (ringbuf->rsv_level++) {
- /* begin() called twice or more without advance() */
- WARN_ON(1);
+ /*
+ * A nested reservation; check that it falls entirely
+ * within the outer block. Don't adjust remaining space.
+ */
+ WARN_ON(ringbuf->rsv_start < 0);
+ WARN_ON(ringbuf->rsv_start & 7);
+ WARN_ON(ringbuf->tail & 7);
+ WARN_ON(ringbuf->tail > ringbuf->effective_size);
+ WARN_ON(ringbuf->tail > ringbuf->rsv_start + ringbuf->rsv_size);
+ WARN_ON(ringbuf->tail + nbytes > ringbuf->effective_size);
+ WARN_ON(ringbuf->tail + nbytes > ringbuf->rsv_start +
ringbuf->rsv_size);
} else {
/*
* A new reservation; validate and record the start and
@@ -437,7 +446,7 @@ static inline void __intel_ringbuffer_begin(struct
intel_ringbuffer *ringbuf,
static inline void __intel_ringbuffer_check(struct intel_ringbuffer *ringbuf)
{
#if 1
- WARN_ON(ringbuf->rsv_level-- != 1);
+ WARN_ON(ringbuf->rsv_level-- <= 0);
WARN_ON(ringbuf->rsv_start < 0 || ringbuf->rsv_size < 0);
WARN_ON(ringbuf->tail & 7);
WARN_ON(ringbuf->tail > ringbuf->rsv_start + ringbuf->rsv_size);
--
1.7.9.5
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
