On 10/4/2021 15:06, Matthew Brost wrote:
For some users of multi-lrc, e.g. split frame, it isn't safe to preempt
mid BB. To safely enable preemption at the BB boundary, a handshake
between to parent and child is needed. This is implemented via custom
between to parent -> between parent
emit_bb_start & emit_fini_breadcrumb functions and enabled via by
via by -> by
I'm also not seeing any mention of the forced re-group behavioural
change in either the comments or commit description.
default if a context is configured by set parallel extension.
v2:
(John Harrison)
- Fix a few comments wording
- Add struture for parent page layout
Signed-off-by: Matthew Brost <matthew.brost@xxxxxxxxx>
---
drivers/gpu/drm/i915/gt/intel_context.c | 2 +-
drivers/gpu/drm/i915/gt/intel_context_types.h | 2 +
drivers/gpu/drm/i915/gt/uc/intel_guc_fwif.h | 2 +-
.../gpu/drm/i915/gt/uc/intel_guc_submission.c | 330 +++++++++++++++++-
4 files changed, 324 insertions(+), 12 deletions(-)
diff --git a/drivers/gpu/drm/i915/gt/intel_context.c b/drivers/gpu/drm/i915/gt/intel_context.c
index 3b340eb59ada..ee84259959d0 100644
--- a/drivers/gpu/drm/i915/gt/intel_context.c
+++ b/drivers/gpu/drm/i915/gt/intel_context.c
@@ -569,7 +569,7 @@ void intel_context_bind_parent_child(struct intel_context *parent,
GEM_BUG_ON(intel_context_is_child(child));
GEM_BUG_ON(intel_context_is_parent(child));
- parent->parallel.number_children++;
+ parent->parallel.child_index = parent->parallel.number_children++;
list_add_tail(&child->parallel.child_link,
&parent->parallel.child_list);
child->parallel.parent = parent;
diff --git a/drivers/gpu/drm/i915/gt/intel_context_types.h b/drivers/gpu/drm/i915/gt/intel_context_types.h
index 1d880303a7e4..95a5b94b4ece 100644
--- a/drivers/gpu/drm/i915/gt/intel_context_types.h
+++ b/drivers/gpu/drm/i915/gt/intel_context_types.h
@@ -250,6 +250,8 @@ struct intel_context {
struct i915_request *last_rq;
/** @number_children: number of children if parent */
u8 number_children;
+ /** @child_index: index into child_list if child */
+ u8 child_index;
/** @guc: GuC specific members for parallel submission */
struct {
/** @wqi_head: head pointer in work queue */
diff --git a/drivers/gpu/drm/i915/gt/uc/intel_guc_fwif.h b/drivers/gpu/drm/i915/gt/uc/intel_guc_fwif.h
index a00eeddc1449..663950d3badc 100644
--- a/drivers/gpu/drm/i915/gt/uc/intel_guc_fwif.h
+++ b/drivers/gpu/drm/i915/gt/uc/intel_guc_fwif.h
@@ -181,7 +181,7 @@ struct guc_process_desc {
u32 wq_status;
u32 engine_presence;
u32 priority;
- u32 reserved[30];
+ u32 reserved[36];
Not seeing the promised explanation of this bug fix.
} __packed;
#define CONTEXT_REGISTRATION_FLAG_KMD BIT(0)
diff --git a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
index 12ee8ca76249..f28e36aa77c2 100644
--- a/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
+++ b/drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c
@@ -11,6 +11,7 @@
#include "gt/intel_context.h"
#include "gt/intel_engine_pm.h"
#include "gt/intel_engine_heartbeat.h"
+#include "gt/intel_gpu_commands.h"
#include "gt/intel_gt.h"
#include "gt/intel_gt_irq.h"
#include "gt/intel_gt_pm.h"
@@ -368,10 +369,16 @@ static inline struct i915_priolist *to_priolist(struct rb_node *rb)
/*
* When using multi-lrc submission an extra page in the context state is
- * reserved for the process descriptor and work queue.
+ * reserved for the process descriptor, work queue, and handshake between the
+ * parent + childlren contexts to insert safe preemption points between each set
+ * of BBs.
*
* The layout of this page is below:
* 0 guc_process_desc
+ * + sizeof(struct guc_process_desc) child go
+ * + CACHELINE_BYTES child join[0]
+ * ...
+ * + CACHELINE_BYTES child join[n - 1]
* ... unused
* PAGE_SIZE / 2 work queue start
* ... work queue
@@ -379,7 +386,25 @@ static inline struct i915_priolist *to_priolist(struct rb_node *rb)
*/
#define WQ_SIZE (PAGE_SIZE / 2)
#define WQ_OFFSET (PAGE_SIZE - WQ_SIZE)
-static u32 __get_process_desc_offset(struct intel_context *ce)
+
+struct parent_page {
+ struct guc_process_desc pdesc;
+
+ u32 child_go_memory;
+ u8 unused0[CACHELINE_BYTES - sizeof(u32)];
+
+ struct {
+ u32 child_join_memory;
+ u8 unused1[CACHELINE_BYTES - sizeof(u32)];
+ } join[MAX_ENGINE_INSTANCE + 1];
Could have a common structure for these. Call the u32 'semaphore_memory'
or something then just have:
struct sync_semaphore go;
struct sync_semaphore go[MAX + 1];
+
+ u8 unused2[(WQ_OFFSET - sizeof(struct guc_process_desc) -
+ CACHELINE_BYTES * (MAX_ENGINE_INSTANCE + 2))];
And this bit could be 'sizeof(struct sync_semaphore) * MAX + 2' to be
clearer what it refers to.
And to be totally paranoid about it, could also add
'BUILD_BUG_ON(sizeof(struct sync_semaphore) != CACHELINE_BYTES'.
And BUILD_BUG_ON(sizeof(parent_page) == PARENT_PAGE_SIZE)'.
+
+ u32 wq[WQ_SIZE / sizeof(u32)];
+};
+
+static u32 __get_parent_page_offset(struct intel_context *ce)
{
GEM_BUG_ON(!ce->parallel.guc.parent_page);
@@ -388,23 +413,35 @@ static u32 __get_process_desc_offset(struct intel_context *ce)
static u32 __get_wq_offset(struct intel_context *ce)
{
- return __get_process_desc_offset(ce) + WQ_OFFSET;
+ BUILD_BUG_ON(offsetof(struct parent_page, wq) != WQ_OFFSET);
+
+ return __get_parent_page_offset(ce) + WQ_OFFSET;
}
-static struct guc_process_desc *
-__get_process_desc(struct intel_context *ce)
+static struct parent_page *
+__get_parent_page(struct intel_context *ce)
{
+ BUILD_BUG_ON(sizeof(struct parent_page) != PAGE_SIZE);
+
/*
* Need to subtract LRC_STATE_OFFSET here as the
* parallel.guc.parent_page is the offset into ce->state while
* ce->lrc_reg_reg is ce->state + LRC_STATE_OFFSET.
*/
- return (struct guc_process_desc *)
+ return (struct parent_page *)
(ce->lrc_reg_state +
- ((__get_process_desc_offset(ce) -
+ ((__get_parent_page_offset(ce) -
LRC_STATE_OFFSET) / sizeof(u32)));
}
+static struct guc_process_desc *
+__get_process_desc(struct intel_context *ce)
+{
+ struct parent_page *pp = __get_parent_page(ce);
+
+ return &pp->pdesc;
+}
+
static u32 *get_wq_pointer(struct guc_process_desc *desc,
struct intel_context *ce,
u32 wqi_size)
@@ -424,8 +461,7 @@ static u32 *get_wq_pointer(struct guc_process_desc *desc,
}
#undef AVAILABLE_SPACE
- return ((u32 *)__get_process_desc(ce)) +
- ((WQ_OFFSET + ce->parallel.guc.wqi_tail) / sizeof(u32));
+ return &__get_parent_page(ce)->wq[ce->parallel.guc.wqi_tail / sizeof(u32)];
}
static struct guc_lrc_desc *__get_lrc_desc(struct intel_guc *guc, u32 index)
@@ -1829,6 +1865,26 @@ static int deregister_context(struct intel_context *ce, u32 guc_id)
return __guc_action_deregister_context(guc, guc_id);
}
+static inline void clear_children_join_go_memory(struct intel_context *ce)
+{
+ u32 *mem = (u32 *)(&__get_parent_page(ce)->child_go_memory);
+ u8 i;
+
+ for (i = 0; i < ce->parallel.number_children + 1; ++i)
+ mem[i * (CACHELINE_BYTES / sizeof(u32))] = 0;
Can't this be written as:
pp->child_go_memory = 0;
for(i = 0 to number_children)
pp->child_join_memory = 0;
Seems like that would be much clearer than this magic casting and
offsetting. I mean, that was the whole point of creating the parent_page
structure.
+}
+
+static inline u32 get_children_go_value(struct intel_context *ce)
+{
+ return __get_parent_page(ce)->child_go_memory;
+}
+
+static inline u32 get_children_join_value(struct intel_context *ce,
+ u8 child_index)
+{
+ return __get_parent_page(ce)->join[child_index].child_join_memory;
+}
+
static void guc_context_policy_init(struct intel_engine_cs *engine,
struct guc_lrc_desc *desc)
{
@@ -1888,7 +1944,7 @@ static int guc_lrc_desc_pin(struct intel_context *ce, bool loop)
ce->parallel.guc.wqi_head = 0;
desc->process_desc = i915_ggtt_offset(ce->state) +
- __get_process_desc_offset(ce);
+ __get_parent_page_offset(ce);
desc->wq_addr = i915_ggtt_offset(ce->state) +
__get_wq_offset(ce);
desc->wq_size = WQ_SIZE;
@@ -1910,6 +1966,8 @@ static int guc_lrc_desc_pin(struct intel_context *ce, bool loop)
desc->context_flags = CONTEXT_REGISTRATION_FLAG_KMD;
guc_context_policy_init(engine, desc);
}
+
+ clear_children_join_go_memory(ce);
}
/*
@@ -2976,6 +3034,31 @@ static const struct intel_context_ops virtual_child_context_ops = {
.get_sibling = guc_virtual_get_sibling,
};
+/*
+ * The below override of the breadcrumbs is enabled when the user configures a
+ * context for parallel submission (multi-lrc, parent-child).
+ *
+ * The overridden breadcrumbs implements an algorithm which allows the GuC to
+ * safely preempt all the hw contexts configured for parallel submission
+ * between each BB. The contract between the i915 and GuC is if the parent
+ * context can be preempted, all the children can be preempted, and the GuC will
+ * always try to preempt the parent before the children. A handshake between the
+ * parent / children breadcrumbs ensures the i915 holds up its end of the deal
+ * creating a window to preempt between each set of BBs.
+ */
+static int emit_bb_start_parent_no_preempt_mid_batch(struct i915_request *rq,
+ u64 offset, u32 len,
+ const unsigned int flags);
+static int emit_bb_start_child_no_preempt_mid_batch(struct i915_request *rq,
+ u64 offset, u32 len,
+ const unsigned int flags);
+static u32 *
+emit_fini_breadcrumb_parent_no_preempt_mid_batch(struct i915_request *rq,
+ u32 *cs);
+static u32 *
+emit_fini_breadcrumb_child_no_preempt_mid_batch(struct i915_request *rq,
+ u32 *cs);
+
static struct intel_context *
guc_create_parallel(struct intel_engine_cs **engines,
unsigned int num_siblings,
@@ -3011,6 +3094,20 @@ guc_create_parallel(struct intel_engine_cs **engines,
}
}
+ parent->engine->emit_bb_start =
+ emit_bb_start_parent_no_preempt_mid_batch;
+ parent->engine->emit_fini_breadcrumb =
+ emit_fini_breadcrumb_parent_no_preempt_mid_batch;
+ parent->engine->emit_fini_breadcrumb_dw =
+ 12 + 4 * parent->parallel.number_children;
+ for_each_child(parent, ce) {
+ ce->engine->emit_bb_start =
+ emit_bb_start_child_no_preempt_mid_batch;
+ ce->engine->emit_fini_breadcrumb =
+ emit_fini_breadcrumb_child_no_preempt_mid_batch;
+ ce->engine->emit_fini_breadcrumb_dw = 16;
+ }
+
kfree(siblings);
return parent;
@@ -3840,6 +3937,17 @@ void intel_guc_submission_print_context_info(struct intel_guc *guc,
drm_printf(p, "\t\tWQI Status: %u\n\n",
READ_ONCE(desc->wq_status));
+ if (ce->engine->emit_bb_start ==
+ emit_bb_start_parent_no_preempt_mid_batch) {
+ u8 i;
+
+ drm_printf(p, "\t\tChildren Go: %u\n\n",
+ get_children_go_value(ce));
+ for (i = 0; i < ce->parallel.number_children; ++i)
+ drm_printf(p, "\t\tChildren Join: %u\n",
+ get_children_join_value(ce, i));
+ }
+
for_each_child(ce, child)
guc_log_context(p, child);
}
@@ -3847,6 +3955,208 @@ void intel_guc_submission_print_context_info(struct intel_guc *guc,
xa_unlock_irqrestore(&guc->context_lookup, flags);
}
+static inline u32 get_children_go_addr(struct intel_context *ce)
+{
+ GEM_BUG_ON(!intel_context_is_parent(ce));
+ BUILD_BUG_ON(offsetof(struct parent_page, child_go_memory) !=
+ sizeof(struct guc_process_desc));
+
+ return i915_ggtt_offset(ce->state) +
+ __get_parent_page_offset(ce) +
+ sizeof(struct guc_process_desc);
Rather than relying on the BUILD_BUG to make sure that the magic
calculation matches the structure definition, can't this just say
"ggtt_offset + pp_offset + offsetof(pp, child_go)"?
+}
+
+static inline u32 get_children_join_addr(struct intel_context *ce,
+ u8 child_index)
+{
+ GEM_BUG_ON(!intel_context_is_parent(ce));
+
+ return get_children_go_addr(ce) + (child_index + 1) * CACHELINE_BYTES;
"ggtt_offset + pp_offset + offsetof(pp, child_join[i])"?
+}
+
+#define PARENT_GO_BB 1
+#define PARENT_GO_FINI_BREADCRUMB 0
+#define CHILD_GO_BB 1
+#define CHILD_GO_FINI_BREADCRUMB 0
+static int emit_bb_start_parent_no_preempt_mid_batch(struct i915_request *rq,
+ u64 offset, u32 len,
+ const unsigned int flags)
+{
+ struct intel_context *ce = rq->context;
+ u32 *cs;
+ u8 i;
+
+ GEM_BUG_ON(!intel_context_is_parent(ce));
+
+ cs = intel_ring_begin(rq, 10 + 4 * ce->parallel.number_children);
+ if (IS_ERR(cs))
+ return PTR_ERR(cs);
+
+ /* Wait on children */
+ for (i = 0; i < ce->parallel.number_children; ++i) {
+ *cs++ = (MI_SEMAPHORE_WAIT |
+ MI_SEMAPHORE_GLOBAL_GTT |
+ MI_SEMAPHORE_POLL |
+ MI_SEMAPHORE_SAD_EQ_SDD);
+ *cs++ = PARENT_GO_BB;
+ *cs++ = get_children_join_addr(ce, i);
+ *cs++ = 0;
+ }
+
+ /* Turn off preemption */
+ *cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
+ *cs++ = MI_NOOP;
+
+ /* Tell children go */
+ cs = gen8_emit_ggtt_write(cs,
+ CHILD_GO_BB,
+ get_children_go_addr(ce),
+ 0);
+
+ /* Jump to batch */
+ *cs++ = MI_BATCH_BUFFER_START_GEN8 |
+ (flags & I915_DISPATCH_SECURE ? 0 : BIT(8));
+ *cs++ = lower_32_bits(offset);
+ *cs++ = upper_32_bits(offset);
+ *cs++ = MI_NOOP;
+
+ intel_ring_advance(rq, cs);
+
+ return 0;
+}
+
+static int emit_bb_start_child_no_preempt_mid_batch(struct i915_request *rq,
+ u64 offset, u32 len,
+ const unsigned int flags)
+{
+ struct intel_context *ce = rq->context;
+ struct intel_context *parent = intel_context_to_parent(ce);
+ u32 *cs;
+
+ GEM_BUG_ON(!intel_context_is_child(ce));
+
+ cs = intel_ring_begin(rq, 12);
+ if (IS_ERR(cs))
+ return PTR_ERR(cs);
+
+ /* Signal parent */
+ cs = gen8_emit_ggtt_write(cs,
+ PARENT_GO_BB,
+ get_children_join_addr(parent,
+ ce->parallel.child_index),
+ 0);
+
+ /* Wait on parent for go */
+ *cs++ = (MI_SEMAPHORE_WAIT |
+ MI_SEMAPHORE_GLOBAL_GTT |
+ MI_SEMAPHORE_POLL |
+ MI_SEMAPHORE_SAD_EQ_SDD);
+ *cs++ = CHILD_GO_BB;
+ *cs++ = get_children_go_addr(parent);
+ *cs++ = 0;
+
+ /* Turn off preemption */
+ *cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
+
+ /* Jump to batch */
+ *cs++ = MI_BATCH_BUFFER_START_GEN8 |
+ (flags & I915_DISPATCH_SECURE ? 0 : BIT(8));
+ *cs++ = lower_32_bits(offset);
+ *cs++ = upper_32_bits(offset);
+
+ intel_ring_advance(rq, cs);
+
+ return 0;
+}
+
+static u32 *
+emit_fini_breadcrumb_parent_no_preempt_mid_batch(struct i915_request *rq,
+ u32 *cs)
+{
+ struct intel_context *ce = rq->context;
+ u8 i;
+
+ GEM_BUG_ON(!intel_context_is_parent(ce));
+
+ /* Wait on children */
+ for (i = 0; i < ce->parallel.number_children; ++i) {
+ *cs++ = (MI_SEMAPHORE_WAIT |
+ MI_SEMAPHORE_GLOBAL_GTT |
+ MI_SEMAPHORE_POLL |
+ MI_SEMAPHORE_SAD_EQ_SDD);
+ *cs++ = PARENT_GO_FINI_BREADCRUMB;
+ *cs++ = get_children_join_addr(ce, i);
+ *cs++ = 0;
+ }
+
+ /* Turn on preemption */
+ *cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
+ *cs++ = MI_NOOP;
+
You mentioned possibly needing to add an MI_ARB_CHECK in here but I'm
not seeing it. Did the testing happen? I don't see that it should be
necessary. Once you execute the MI_ARB_ENABLE, the CS can preempt
anywhere, I thought? Even if it can't there should be an MI_ARB_CHECK
added at the next level up after the breadcrumb code. Or do we not have
those in between batches any more?
John.
+ /* Tell children go */
+ cs = gen8_emit_ggtt_write(cs,
+ CHILD_GO_FINI_BREADCRUMB,
+ get_children_go_addr(ce),
+ 0);
+
+ /* Emit fini breadcrumb */
+ cs = gen8_emit_ggtt_write(cs,
+ rq->fence.seqno,
+ i915_request_active_timeline(rq)->hwsp_offset,
+ 0);
+
+ /* User interrupt */
+ *cs++ = MI_USER_INTERRUPT;
+ *cs++ = MI_NOOP;
+
+ rq->tail = intel_ring_offset(rq, cs);
+
+ return cs;
+}
+
+static u32 *
+emit_fini_breadcrumb_child_no_preempt_mid_batch(struct i915_request *rq, u32 *cs)
+{
+ struct intel_context *ce = rq->context;
+ struct intel_context *parent = intel_context_to_parent(ce);
+
+ GEM_BUG_ON(!intel_context_is_child(ce));
+
+ /* Turn on preemption */
+ *cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
+ *cs++ = MI_NOOP;
+
+ /* Signal parent */
+ cs = gen8_emit_ggtt_write(cs,
+ PARENT_GO_FINI_BREADCRUMB,
+ get_children_join_addr(parent,
+ ce->parallel.child_index),
+ 0);
+
+ /* Wait parent on for go */
+ *cs++ = (MI_SEMAPHORE_WAIT |
+ MI_SEMAPHORE_GLOBAL_GTT |
+ MI_SEMAPHORE_POLL |
+ MI_SEMAPHORE_SAD_EQ_SDD);
+ *cs++ = CHILD_GO_FINI_BREADCRUMB;
+ *cs++ = get_children_go_addr(parent);
+ *cs++ = 0;
+
+ /* Emit fini breadcrumb */
+ cs = gen8_emit_ggtt_write(cs,
+ rq->fence.seqno,
+ i915_request_active_timeline(rq)->hwsp_offset,
+ 0);
+
+ /* User interrupt */
+ *cs++ = MI_USER_INTERRUPT;
+ *cs++ = MI_NOOP;
+
+ rq->tail = intel_ring_offset(rq, cs);
+
+ return cs;
+}
+
static struct intel_context *
guc_create_virtual(struct intel_engine_cs **siblings, unsigned int count,
unsigned long flags)