Skip to contentSkip to conversation

drm/i915/guc: Don't make assumptions while getting the lrca offset

PATCH In review Michel Thierry 3 participants 10 messages Jul 11, 2017 – Jul 12, 2017

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/i915/guc: Don't make assumptions while getting the lrca offset

Using the HWSP ggtt_offset to get the lrca offset is only correct if the HWSP happens to be before it (when we reuse the PPHWSP of the kernel context as the engine HWSP). Instead of making this assumption, get the lrca offset from the kernel_context engine state.

And while looking at this part of the GuC interaction, it was also noticed that the firmware expects the size of only the engine context (context minus the execlist part, i.e. don't include the first 80 dwords), so pass the right size.

Cc Daniele Ceraolo Spurio <[email protected]>
Cc Michal Wajdeczko <[email protected]>
Cc Oscar Mateo <[email protected]>
---
 drivers/gpu/drm/i915/i915_guc_submission.c | 15 ++++++++++++---
 1 file changed, 12 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/i915_guc_submission.c +12 -3
index 48a1e9349a2c..97ec11a7cc9a 100644
@@ -1018,6 +1018,12 @@ static void guc_policies_init(struct guc_policies *policies)
 	policies->is_valid = 1;
 }
 
+/*
+ * The first 80 dwords of the register state context, containing the
+ * execlists and ppgtt registers.
+ */
+#define LR_HW_CONTEXT_SIZE	(80 * sizeof(u32))
+
 static int guc_ads_create(struct intel_guc *guc)
 {
 	struct drm_i915_private *dev_priv = guc_to_i915(guc);
@@ -1033,6 +1039,7 @@ static int guc_ads_create(struct intel_guc *guc)
 	struct intel_engine_cs *engine;
 	enum intel_engine_id id;
 	u32 base;
+	u32 lrca_offset = LRC_PPHWSP_PN * PAGE_SIZE;
 
 	GEM_BUG_ON(guc->ads_vma);
 
@@ -1062,13 +1069,15 @@ static int guc_ads_create(struct intel_guc *guc)
 	 * engines after a reset. Here we use the Render ring default
 	 * context, which must already exist and be pinned in the GGTT,
 	 * so its address won't change after we've told the GuC where
-	 * to find it.
+	 * to find it. GuC will skip the PPHWSP and 'Execlist Context',
+	 * thus reduce the size by 1 page (PPHWSP) and 80 dwords.
 	 */
 	blob->ads.golden_context_lrca =
-		dev_priv->engine[RCS]->status_page.ggtt_offset;
+		guc_ggtt_offset(dev_priv->kernel_context->engine[RCS].state) + lrca_offset;
 
+	/* context_size - PPHWSP - first 80 dwords */
 	for_each_engine(engine, dev_priv, id)
-		blob->ads.eng_state_size[engine->guc_id] = engine->context_size;
+		blob->ads.eng_state_size[engine->guc_id] = engine->context_size - lrca_offset - LR_HW_CONTEXT_SIZE;
 
 	base = guc_ggtt_offset(vma);
 	blob->ads.scheduler_policies = base + ptr_offset(blob, policies);
Signature · 6 lines
2.13.0

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#1 lore raw

✓ Fi.CI.BAT: success for drm/i915/guc: Don't make assumptions while getting the lrca offset

== Series Details ==

== Series Details ==

Series: drm/i915/guc: Don't make assumptions while getting the lrca offset
URL : https://patchwork.freedesktop.org/series/27136/
State : success

== Summary ==

Series 27136v1 drm/i915/guc: Don't make assumptions while getting the lrca offset
https://patchwork.freedesktop.org/api/1.0/series/27136/revisions/1/mbox/

Test gem_exec_suspend:

        Subgroup basic-s4-devices:
                pass       -> DMESG-WARN (fi-kbl-7560u) fdo#100125
Test kms_flip:
        Subgroup basic-flip-vs-wf_vblank:
                pass       -> FAIL       (fi-skl-6770hq) fdo#99739

fdo#100125 https://bugs.freedesktop.org/show_bug.cgi?id=100125
fdo#99739 https://bugs.freedesktop.org/show_bug.cgi?id=99739

fi-bdw-5557u     total:279  pass:268  dwarn:0   dfail:0   fail:0   skip:11  time:436s
fi-bdw-gvtdvm    total:279  pass:265  dwarn:0   dfail:0   fail:0   skip:14  time:433s
fi-blb-e6850     total:279  pass:224  dwarn:1   dfail:0   fail:0   skip:54  time:361s
fi-bsw-n3050     total:279  pass:243  dwarn:0   dfail:0   fail:0   skip:36  time:525s
fi-bxt-j4205     total:279  pass:260  dwarn:0   dfail:0   fail:0   skip:19  time:508s
fi-byt-j1900     total:279  pass:254  dwarn:1   dfail:0   fail:0   skip:24  time:487s
fi-byt-n2820     total:279  pass:250  dwarn:1   dfail:0   fail:0   skip:28  time:487s
fi-glk-2a        total:279  pass:260  dwarn:0   dfail:0   fail:0   skip:19  time:598s
fi-hsw-4770      total:279  pass:263  dwarn:0   dfail:0   fail:0   skip:16  time:443s
fi-hsw-4770r     total:279  pass:263  dwarn:0   dfail:0   fail:0   skip:16  time:412s
fi-ilk-650       total:279  pass:229  dwarn:0   dfail:0   fail:0   skip:50  time:421s
fi-ivb-3520m     total:279  pass:261  dwarn:0   dfail:0   fail:0   skip:18  time:494s
fi-ivb-3770      total:279  pass:261  dwarn:0   dfail:0   fail:0   skip:18  time:471s
fi-kbl-7500u     total:279  pass:261  dwarn:0   dfail:0   fail:0   skip:18  time:461s
fi-kbl-7560u     total:279  pass:268  dwarn:1   dfail:0   fail:0   skip:10  time:572s
fi-kbl-r         total:279  pass:260  dwarn:1   dfail:0   fail:0   skip:18  time:576s
fi-pnv-d510      total:279  pass:222  dwarn:2   dfail:0   fail:0   skip:55  time:560s
fi-skl-6260u     total:279  pass:269  dwarn:0   dfail:0   fail:0   skip:10  time:467s
fi-skl-6700hq    total:279  pass:262  dwarn:0   dfail:0   fail:0   skip:17  time:585s
fi-skl-6700k     total:279  pass:257  dwarn:4   dfail:0   fail:0   skip:18  time:468s
fi-skl-6770hq    total:279  pass:268  dwarn:0   dfail:0   fail:1   skip:10  time:470s
fi-skl-gvtdvm    total:279  pass:266  dwarn:0   dfail:0   fail:0   skip:13  time:436s
fi-skl-x1585l    total:279  pass:269  dwarn:0   dfail:0   fail:0   skip:10  time:505s
fi-snb-2520m     total:279  pass:251  dwarn:0   dfail:0   fail:0   skip:28  time:550s
fi-snb-2600      total:279  pass:250  dwarn:0   dfail:0   fail:0   skip:29  time:406s

8ad9e19aafea47c272163c2cbf554e06ff7f9857 drm-tip: 2017y-07m-11d-19h-08m-20s UTC integration manifest 84116ad drm/i915/guc: Don't make assumptions while getting the lrca offset

== Logs ==

For more details see: https://intel-gfx-ci.01.org/CI/Patchwork_5166/
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#2 lore raw
In reply to Michel Thierry

Quoting Michel Thierry (2017-07-11 22:29:39)

Quoted 37 lines

Using the HWSP ggtt_offset to get the lrca offset is only correct if the HWSP happens to be before it (when we reuse the PPHWSP of the kernel context as the engine HWSP). Instead of making this assumption, get the lrca offset from the kernel_context engine state.

And while looking at this part of the GuC interaction, it was also noticed that the firmware expects the size of only the engine context (context minus the execlist part, i.e. don't include the first 80 dwords), so pass the right size.

Cc Daniele Ceraolo Spurio <[email protected]>
Cc Michal Wajdeczko <[email protected]>
Cc Oscar Mateo <[email protected]>
---
 drivers/gpu/drm/i915/i915_guc_submission.c | 15 ++++++++++++---
 1 file changed, 12 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/i915_guc_submission.c +6 -0
index 48a1e9349a2c..97ec11a7cc9a 100644
@@ -1018,6 +1018,12 @@ static void guc_policies_init(struct guc_policies *policies)
        policies->is_valid = 1;
 }
 
+/*
+ * The first 80 dwords of the register state context, containing the
+ * execlists and ppgtt registers.
+ */
+#define LR_HW_CONTEXT_SIZE     (80 * sizeof(u32))
+
 static int guc_ads_create(struct intel_guc *guc)
 {
        struct drm_i915_private *dev_priv = guc_to_i915(guc);

drivers/gpu/drm/i915/i915_guc_submission.c static int guc_ads_create(struct intel_guc *guc) · line 1042

Quoted 4 lines
drivers/gpu/drm/i915/i915_guc_submission.c +1 -0
@@ -1033,6 +1039,7 @@ static int guc_ads_create(struct intel_guc *guc)
        struct intel_engine_cs *engine;
        enum intel_engine_id id;
        u32 base;
+       u32 lrca_offset = LRC_PPHWSP_PN * PAGE_SIZE;

Make it const and try to avoid the visual breaks in line length? (Whilst also trying not to gratuitously mix types.)

Quoted 9 lines
drivers/gpu/drm/i915/i915_guc_submission.c +1 -1
        GEM_BUG_ON(guc->ads_vma);
 
@@ -1062,13 +1069,15 @@ static int guc_ads_create(struct intel_guc *guc)
         * engines after a reset. Here we use the Render ring default
         * context, which must already exist and be pinned in the GGTT,
         * so its address won't change after we've told the GuC where
-        * to find it.
+        * to find it. GuC will skip the PPHWSP and 'Execlist Context',

drivers/gpu/drm/i915/i915_guc_submission.c static int guc_ads_create(struct intel_guc *guc) · line 1080

Quoted 10 lines
drivers/gpu/drm/i915/i915_guc_submission.c +4 -2
+        * thus reduce the size by 1 page (PPHWSP) and 80 dwords.
         */
        blob->ads.golden_context_lrca =
-               dev_priv->engine[RCS]->status_page.ggtt_offset;
+               guc_ggtt_offset(dev_priv->kernel_context->engine[RCS].state) + lrca_offset;
 
+       /* context_size - PPHWSP - first 80 dwords */
        for_each_engine(engine, dev_priv, id)
-               blob->ads.eng_state_size[engine->guc_id] = engine->context_size;
+               blob->ads.eng_state_size[engine->guc_id] = engine->context_size - lrca_offset - LR_HW_CONTEXT_SIZE;

The first LRC_PPHWSP_PN pages are not included in engine->context_size, they are added on by execlist_context_deferred_alloc().

I'm finding it odd that we pass in a pointer to the start of the context image, but that the guc wants to be told the size of the context minus the bytes that *it* is choosing to skip at the start. (May be correct, but looks weird and merits a warning.) The alternative explanation is that it is not skipping the first 80 dwords, but the last. -Chris
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#3 lore raw
Quoted 37 lines · Chris Wilson

Quoting Michel Thierry (2017-07-11 22:29:39)

Quoted 36 lines

Using the HWSP ggtt_offset to get the lrca offset is only correct if the HWSP happens to be before it (when we reuse the PPHWSP of the kernel context as the engine HWSP). Instead of making this assumption, get the lrca offset from the kernel_context engine state.

And while looking at this part of the GuC interaction, it was also noticed that the firmware expects the size of only the engine context (context minus the execlist part, i.e. don't include the first 80 dwords), so pass the right size.

Cc Daniele Ceraolo Spurio <[email protected]>
Cc Michal Wajdeczko <[email protected]>
Cc Oscar Mateo <[email protected]>
---
  drivers/gpu/drm/i915/i915_guc_submission.c | 15 ++++++++++++---
  1 file changed, 12 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/i915_guc_submission.c +6 -0
index 48a1e9349a2c..97ec11a7cc9a 100644
@@ -1018,6 +1018,12 @@ static void guc_policies_init(struct guc_policies *policies)
         policies->is_valid = 1;
  }
 
+/*
+ * The first 80 dwords of the register state context, containing the
+ * execlists and ppgtt registers.
+ */
+#define LR_HW_CONTEXT_SIZE     (80 * sizeof(u32))
+
  static int guc_ads_create(struct intel_guc *guc)
  {
         struct drm_i915_private *dev_priv = guc_to_i915(guc);
Quoted 8 lines

drivers/gpu/drm/i915/i915_guc_submission.c static int guc_ads_create(struct intel_guc *guc) · line 1042

Quoted 5 lines
drivers/gpu/drm/i915/i915_guc_submission.c +1 -0
@@ -1033,6 +1039,7 @@ static int guc_ads_create(struct intel_guc *guc)
         struct intel_engine_cs *engine;
         enum intel_engine_id id;
         u32 base;
+       u32 lrca_offset = LRC_PPHWSP_PN * PAGE_SIZE;

Make it const and try to avoid the visual breaks in line length? (Whilst also trying not to gratuitously mix types.)

ok

Quoted 15 lines
Quoted 15 lines
drivers/gpu/drm/i915/i915_guc_submission.c +3 -2
         GEM_BUG_ON(guc->ads_vma);
 
@@ -1062,13 +1069,15 @@ static int guc_ads_create(struct intel_guc *guc)
          * engines after a reset. Here we use the Render ring default
          * context, which must already exist and be pinned in the GGTT,
          * so its address won't change after we've told the GuC where
-        * to find it.
+        * to find it. GuC will skip the PPHWSP and 'Execlist Context',
+        * thus reduce the size by 1 page (PPHWSP) and 80 dwords.
          */
         blob->ads.golden_context_lrca =
-               dev_priv->engine[RCS]->status_page.ggtt_offset;
+               guc_ggtt_offset(dev_priv->kernel_context->engine[RCS].state) + lrca_offset;
 
Quoted 8 lines

drivers/gpu/drm/i915/i915_guc_submission.c static int guc_ads_create(struct intel_guc *guc) · line 1080

Quoted 4 lines
drivers/gpu/drm/i915/i915_guc_submission.c +2 -1
+       /* context_size - PPHWSP - first 80 dwords */
         for_each_engine(engine, dev_priv, id)
-               blob->ads.eng_state_size[engine->guc_id] = engine->context_size;
+               blob->ads.eng_state_size[engine->guc_id] = engine->context_size - lrca_offset - LR_HW_CONTEXT_SIZE;

The first LRC_PPHWSP_PN pages are not included in engine->context_size, they are added on by execlist_context_deferred_alloc().

I think there is only one extra page added in execlists_context_deferred_alloc, and it is the "GuC shared data". And yes, we don't include that one in in the context_size. (also __intel_engine_context_size has a comment saying the size includes the HWSP).

Quoted 5 lines

I'm finding it odd that we pass in a pointer to the start of the context image, but that the guc wants to be told the size of the context minus the bytes that *it* is choosing to skip at the start. (May be correct, but looks weird and merits a warning.) The alternative explanation is that it is not skipping the first 80 dwords, but the last.

Yes, it's really odd style. Daniele & I looked at it, and guc takes the pointer and uses "size of the PPHWSP + the size of the execlist_context (those 80 dwords)" as the starting point of the engine-state... that's why ads.eng_state_size needs to be decreased.

Anyway, since we always round up the ctx size to the next page, I imagine we've been lucky and there are enough 0's at the end of the context.

Quoted 2 lines

-Chris

Thanks!
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#4 lore raw

Quoting Michel Thierry (2017-07-12 00:01:04)

Quoted 73 lines · Chris Wilson
Quoted 37 lines

Quoting Michel Thierry (2017-07-11 22:29:39)

Quoted 36 lines

Using the HWSP ggtt_offset to get the lrca offset is only correct if the HWSP happens to be before it (when we reuse the PPHWSP of the kernel context as the engine HWSP). Instead of making this assumption, get the lrca offset from the kernel_context engine state.

And while looking at this part of the GuC interaction, it was also noticed that the firmware expects the size of only the engine context (context minus the execlist part, i.e. don't include the first 80 dwords), so pass the right size.

Cc Daniele Ceraolo Spurio <[email protected]>
Cc Michal Wajdeczko <[email protected]>
Cc Oscar Mateo <[email protected]>
---
  drivers/gpu/drm/i915/i915_guc_submission.c | 15 ++++++++++++---
  1 file changed, 12 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/i915_guc_submission.c +6 -0
index 48a1e9349a2c..97ec11a7cc9a 100644
@@ -1018,6 +1018,12 @@ static void guc_policies_init(struct guc_policies *policies)
         policies->is_valid = 1;
  }
 
+/*
+ * The first 80 dwords of the register state context, containing the
+ * execlists and ppgtt registers.
+ */
+#define LR_HW_CONTEXT_SIZE     (80 * sizeof(u32))
+
  static int guc_ads_create(struct intel_guc *guc)
  {
         struct drm_i915_private *dev_priv = guc_to_i915(guc);
Quoted 8 lines

drivers/gpu/drm/i915/i915_guc_submission.c static int guc_ads_create(struct intel_guc *guc) · line 1042

Quoted 5 lines
drivers/gpu/drm/i915/i915_guc_submission.c +1 -0
@@ -1033,6 +1039,7 @@ static int guc_ads_create(struct intel_guc *guc)
         struct intel_engine_cs *engine;
         enum intel_engine_id id;
         u32 base;
+       u32 lrca_offset = LRC_PPHWSP_PN * PAGE_SIZE;

Make it const and try to avoid the visual breaks in line length? (Whilst also trying not to gratuitously mix types.)

ok

Quoted 15 lines
Quoted 15 lines
drivers/gpu/drm/i915/i915_guc_submission.c +3 -2
         GEM_BUG_ON(guc->ads_vma);
 
@@ -1062,13 +1069,15 @@ static int guc_ads_create(struct intel_guc *guc)
          * engines after a reset. Here we use the Render ring default
          * context, which must already exist and be pinned in the GGTT,
          * so its address won't change after we've told the GuC where
-        * to find it.
+        * to find it. GuC will skip the PPHWSP and 'Execlist Context',
+        * thus reduce the size by 1 page (PPHWSP) and 80 dwords.
          */
         blob->ads.golden_context_lrca =
-               dev_priv->engine[RCS]->status_page.ggtt_offset;
+               guc_ggtt_offset(dev_priv->kernel_context->engine[RCS].state) + lrca_offset;
 
Quoted 8 lines

drivers/gpu/drm/i915/i915_guc_submission.c static int guc_ads_create(struct intel_guc *guc) · line 1080

Quoted 4 lines
drivers/gpu/drm/i915/i915_guc_submission.c +2 -1
+       /* context_size - PPHWSP - first 80 dwords */
         for_each_engine(engine, dev_priv, id)
-               blob->ads.eng_state_size[engine->guc_id] = engine->context_size;
+               blob->ads.eng_state_size[engine->guc_id] = engine->context_size - lrca_offset - LR_HW_CONTEXT_SIZE;

The first LRC_PPHWSP_PN pages are not included in engine->context_size, they are added on by execlist_context_deferred_alloc().

Quoted 5 lines

I think there is only one extra page added in execlists_context_deferred_alloc, and it is the "GuC shared data". And yes, we don't include that one in in the context_size.
(also __intel_engine_context_size has a comment saying the size includes the HWSP).

Oh, I see what my confusion is. You are using lrca_offset for two very different meanings just because they are the same value. (And also the use of LRC_PPHWSP_PN for execlists_context_deferred_alloc() has the same dubious semantics, which lead me to think that PPHWSP couldn't possibly mean per-process hwsp anymore!) The first is an offset from the start of the allocation to the HWSP, which is the correct meaning of LRC_PPHWSP_PN (that is, its page index). In second case, you just mean PAGE_SIZE to remove the single page being skipped. Please, I am easily confused! That does strongly suggest the comment needs an actual sentence and not just rehashing of the code:

/*
* The GuC expects us to exclude the portion of the context image that
* it skips from the size it is to read. It starts reading from after
* the execlist context (so skipping the first page [PPHWSP] and 80
* dwords). Weird guc is weird.
*/

for_each_engine(engine) {
	const u32 skipped = PAGE_SIZE + LR_HW_CONTEXT_SIZE;
	blob->ads.eng_state_size[engine->guc_id] = engine->context_size - skipped;
}

Also feel free to try and fixup the other confusion in intel_lrc.c. Perhaps something like:

drivers/gpu/drm/i915/intel_lrc.c +3 -2
index b0738d2..f498aa6 100644
@@ -2031,8 +2031,9 @@ static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
 
        context_size = round_up(engine->context_size, I915_GTT_PAGE_SIZE);
 
-       /* One extra page as the sharing data between driver and GuC */
-       context_size += PAGE_SIZE * LRC_PPHWSP_PN;
+       /* Add an extra page at the start for sharing data with the GuC */
+       BUILD_BUG_ON(LRC_PPHWSP_PN != 1);
+       context_size += PAGE_SIZE;
 
        ctx_obj = i915_gem_object_create(ctx->i915, context_size);
        if (IS_ERR(ctx_obj)) {

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#5 lore raw

Quoting Chris Wilson (2017-07-12 01:00:02)

Quoted 9 lines

Also feel free to try and fixup the other confusion in intel_lrc.c. Perhaps something like:

drivers/gpu/drm/i915/intel_lrc.c +0 -0
index b0738d2..f498aa6 100644
@@ -2031,8 +2031,9 @@ static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
 

drivers/gpu/drm/i915/intel_lrc.c static int execlists_context_deferred_alloc(struct i915_gem_context *ctx, · line 2039

Quoted 10 lines
drivers/gpu/drm/i915/intel_lrc.c +3 -2
        context_size = round_up(engine->context_size, I915_GTT_PAGE_SIZE);
 
-       /* One extra page as the sharing data between driver and GuC */
-       context_size += PAGE_SIZE * LRC_PPHWSP_PN;
+       /* Add an extra page at the start for sharing data with the GuC */
+       BUILD_BUG_ON(LRC_PPHWSP_PN != 1);
+       context_size += PAGE_SIZE;
 
        ctx_obj = i915_gem_object_create(ctx->i915, context_size);
        if (IS_ERR(ctx_obj)) {

Or better

drivers/gpu/drm/i915/intel_lrc.c +2 -2
index b0738d2..0b7fb38 100644
@@ -2031,8 +2031,8 @@ static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
 
        context_size = round_up(engine->context_size, I915_GTT_PAGE_SIZE);
 
-       /* One extra page as the sharing data between driver and GuC */
-       context_size += PAGE_SIZE * LRC_PPHWSP_PN;
+       /* Add an extra page at the start for sharing data with the GuC */
+       context_size += LRC_GUCSHR_PAGES * PAGE_SIZE;
 
        ctx_obj = i915_gem_object_create(ctx->i915, context_size);
        if (IS_ERR(ctx_obj)) {
drivers/gpu/drm/i915/intel_lrc.h +2 -1
index 52b3a1f..848ca21 100644
@@ -72,7 +72,8 @@ int logical_xcs_ring_init(struct intel_engine_cs *engine);
 
 /* One extra page is added before LRC for GuC as shared data */
 #define LRC_GUCSHR_PN  (0)
-#define LRC_PPHWSP_PN  (LRC_GUCSHR_PN + 1)
+#define LRC_GUCSHR_PAGES 1
+#define LRC_PPHWSP_PN  (LRC_GUCSHR_PN + LRC_GUCSHR_PAGES)
 #define LRC_STATE_PN   (LRC_PPHWSP_PN + 1)
 
 struct drm_i915_private;

Nowhere do we use LRC_GUCSHR_PN. Having it just gives us the mirage of flexibility...
-Chris
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#6 lore raw
Quoted 53 lines · Chris Wilson

Quoting Chris Wilson (2017-07-12 01:00:02)

Quoted 9 lines

Also feel free to try and fixup the other confusion in intel_lrc.c. Perhaps something like:

drivers/gpu/drm/i915/intel_lrc.c +0 -0
index b0738d2..f498aa6 100644
@@ -2031,8 +2031,9 @@ static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
 

drivers/gpu/drm/i915/intel_lrc.c static int execlists_context_deferred_alloc(struct i915_gem_context *ctx, · line 2039

Quoted 10 lines
drivers/gpu/drm/i915/intel_lrc.c +3 -2
         context_size = round_up(engine->context_size, I915_GTT_PAGE_SIZE);
 
-       /* One extra page as the sharing data between driver and GuC */
-       context_size += PAGE_SIZE * LRC_PPHWSP_PN;
+       /* Add an extra page at the start for sharing data with the GuC */
+       BUILD_BUG_ON(LRC_PPHWSP_PN != 1);
+       context_size += PAGE_SIZE;
 
         ctx_obj = i915_gem_object_create(ctx->i915, context_size);
         if (IS_ERR(ctx_obj)) {

Or better

drivers/gpu/drm/i915/intel_lrc.c +2 -2
index b0738d2..0b7fb38 100644
@@ -2031,8 +2031,8 @@ static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
 
         context_size = round_up(engine->context_size, I915_GTT_PAGE_SIZE);
 
-       /* One extra page as the sharing data between driver and GuC */
-       context_size += PAGE_SIZE * LRC_PPHWSP_PN;
+       /* Add an extra page at the start for sharing data with the GuC */
+       context_size += LRC_GUCSHR_PAGES * PAGE_SIZE;
 
         ctx_obj = i915_gem_object_create(ctx->i915, context_size);
         if (IS_ERR(ctx_obj)) {
drivers/gpu/drm/i915/intel_lrc.h +2 -1
index 52b3a1f..848ca21 100644
@@ -72,7 +72,8 @@ int logical_xcs_ring_init(struct intel_engine_cs *engine);
 
  /* One extra page is added before LRC for GuC as shared data */
  #define LRC_GUCSHR_PN  (0)
-#define LRC_PPHWSP_PN  (LRC_GUCSHR_PN + 1)
+#define LRC_GUCSHR_PAGES 1
+#define LRC_PPHWSP_PN  (LRC_GUCSHR_PN + LRC_GUCSHR_PAGES)
  #define LRC_STATE_PN   (LRC_PPHWSP_PN + 1)
 
  struct drm_i915_private;
Quoted 2 lines

Nowhere do we use LRC_GUCSHR_PN. Having it just gives us the mirage of flexibility...

LRC_GUCSHR_PN is not used because people assumed it is always page 0 of the kernel ctx :S

One more thing to fix, use LRC_GUCSHR_PN in all the places where this _shared_ GuC page is used (guc_ggtt_offset(ctx->engine[RCS].state))

Quoted 2 lines

-Chris

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#7 lore raw

Quoting Michel Thierry (2017-07-12 01:14:46)

Quoted 30 lines · Chris Wilson
Quoted 29 lines

Quoting Chris Wilson (2017-07-12 01:00:02)

Quoted 9 lines

Also feel free to try and fixup the other confusion in intel_lrc.c. Perhaps something like:

drivers/gpu/drm/i915/intel_lrc.c +0 -0
index b0738d2..f498aa6 100644
@@ -2031,8 +2031,9 @@ static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
 

drivers/gpu/drm/i915/intel_lrc.c static int execlists_context_deferred_alloc(struct i915_gem_context *ctx, · line 2039

Quoted 10 lines
drivers/gpu/drm/i915/intel_lrc.c +3 -2
         context_size = round_up(engine->context_size, I915_GTT_PAGE_SIZE);
 
-       /* One extra page as the sharing data between driver and GuC */
-       context_size += PAGE_SIZE * LRC_PPHWSP_PN;
+       /* Add an extra page at the start for sharing data with the GuC */
+       BUILD_BUG_ON(LRC_PPHWSP_PN != 1);
+       context_size += PAGE_SIZE;
 
         ctx_obj = i915_gem_object_create(ctx->i915, context_size);
         if (IS_ERR(ctx_obj)) {

Or better

drivers/gpu/drm/i915/intel_lrc.c +0 -0
index b0738d2..0b7fb38 100644
@@ -2031,8 +2031,8 @@ static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
 
Quoted 6 lines

drivers/gpu/drm/i915/intel_lrc.c static int execlists_context_deferred_alloc(struct i915_gem_context *ctx, · line 2035

Quoted 6 lines
drivers/gpu/drm/i915/intel_lrc.c +2 -2
         context_size = round_up(engine->context_size, I915_GTT_PAGE_SIZE);
 
-       /* One extra page as the sharing data between driver and GuC */
-       context_size += PAGE_SIZE * LRC_PPHWSP_PN;
+       /* Add an extra page at the start for sharing data with the GuC */
+       context_size += LRC_GUCSHR_PAGES * PAGE_SIZE;

One more thought about this, and with a little coaxing LRC_PPHWSP_PN may be the most appropriate constant to use here.

/* Before the actual start of the context image, we insert a few pages
* for our own use and for sharing with the GuC.
*/
context_size += LRC_HEADER_PAGES * PAGE_SIZE;

Quoted 8 lines
Quoted 8 lines
drivers/gpu/drm/i915/intel_lrc.c +0 -0
         ctx_obj = i915_gem_object_create(ctx->i915, context_size);
         if (IS_ERR(ctx_obj)) {
drivers/gpu/drm/i915/intel_lrc.h +0 -0
index 52b3a1f..848ca21 100644
@@ -72,7 +72,8 @@ int logical_xcs_ring_init(struct intel_engine_cs *engine);
 
Quoted 6 lines

drivers/gpu/drm/i915/intel_lrc.h int logical_xcs_ring_init(struct intel_engine_cs *engine); · line 77

Quoted 6 lines
drivers/gpu/drm/i915/intel_lrc.h +2 -1
  /* One extra page is added before LRC for GuC as shared data */
  #define LRC_GUCSHR_PN  (0)
-#define LRC_PPHWSP_PN  (LRC_GUCSHR_PN + 1)
+#define LRC_GUCSHR_PAGES 1
+#define LRC_PPHWSP_PN  (LRC_GUCSHR_PN + LRC_GUCSHR_PAGES)
  #define LRC_STATE_PN   (LRC_PPHWSP_PN + 1)

but

/* We allocate a header at the start of the context image for our own
* use, therefore the actual location of the logical state is offset
* from the start of the VMA. The layout is
*

 * | [guc]          | [hwsp] [logical state] |
 * |<- our header ->|<- context image      ->|

*
*/

/* One extra page is used for sharing with the GuC */ #define LRC_GUCSHR_PN (0)
#define LRC_GUCSHR_PAGES 1

/* At the start of the context image is its per-process HWS page */ #define LRC_PPHWSP_PN (LRC_GUCSHR_PN + LRC_GUCSHR_PAGES)

/* Finally we have the logical state for the context */
#define LRC_STATE_PN (LRC_PPHWSP_PN + 1)

/* Currently we include the PPHWSP in __intel_engine_context_size() so
* the size of the header is synonymous with the start of the PPHWSP. */
#define LRC_HEADER_PAGES LRC_PPHWSP_PN

Quoted 11 lines
Quoted 5 lines
drivers/gpu/drm/i915/intel_lrc.h +0 -0
  struct drm_i915_private;
 

Nowhere do we use LRC_GUCSHR_PN. Having it just gives us the mirage of flexibility...

LRC_GUCSHR_PN is not used because people assumed it is always page 0 of the kernel ctx :S

One more thing to fix, use LRC_GUCSHR_PN in all the places where this _shared_ GuC page is used (guc_ggtt_offset(ctx->engine[RCS].state))

That may prove very handy.
-Chris
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#8 lore raw
Quoted 79 lines · Chris Wilson

Quoting Michel Thierry (2017-07-12 01:14:46)

Quoted 30 lines · Chris Wilson
Quoted 29 lines

Quoting Chris Wilson (2017-07-12 01:00:02)

Quoted 9 lines

Also feel free to try and fixup the other confusion in intel_lrc.c. Perhaps something like:

drivers/gpu/drm/i915/intel_lrc.c +0 -0
index b0738d2..f498aa6 100644
@@ -2031,8 +2031,9 @@ static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
 

drivers/gpu/drm/i915/intel_lrc.c static int execlists_context_deferred_alloc(struct i915_gem_context *ctx, · line 2039

Quoted 10 lines
drivers/gpu/drm/i915/intel_lrc.c +3 -2
          context_size = round_up(engine->context_size, I915_GTT_PAGE_SIZE);
 
-       /* One extra page as the sharing data between driver and GuC */
-       context_size += PAGE_SIZE * LRC_PPHWSP_PN;
+       /* Add an extra page at the start for sharing data with the GuC */
+       BUILD_BUG_ON(LRC_PPHWSP_PN != 1);
+       context_size += PAGE_SIZE;
 
          ctx_obj = i915_gem_object_create(ctx->i915, context_size);
          if (IS_ERR(ctx_obj)) {

Or better

drivers/gpu/drm/i915/intel_lrc.c +0 -0
index b0738d2..0b7fb38 100644
@@ -2031,8 +2031,8 @@ static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
 
Quoted 6 lines

drivers/gpu/drm/i915/intel_lrc.c static int execlists_context_deferred_alloc(struct i915_gem_context *ctx, · line 2035

Quoted 6 lines
drivers/gpu/drm/i915/intel_lrc.c +2 -2
          context_size = round_up(engine->context_size, I915_GTT_PAGE_SIZE);
 
-       /* One extra page as the sharing data between driver and GuC */
-       context_size += PAGE_SIZE * LRC_PPHWSP_PN;
+       /* Add an extra page at the start for sharing data with the GuC */
+       context_size += LRC_GUCSHR_PAGES * PAGE_SIZE;

One more thought about this, and with a little coaxing LRC_PPHWSP_PN may be the most appropriate constant to use here.

/* Before the actual start of the context image, we insert a few pages
* for our own use and for sharing with the GuC.
*/
context_size += LRC_HEADER_PAGES * PAGE_SIZE;

Quoted 8 lines
Quoted 8 lines
drivers/gpu/drm/i915/intel_lrc.c +0 -0
          ctx_obj = i915_gem_object_create(ctx->i915, context_size);
          if (IS_ERR(ctx_obj)) {
drivers/gpu/drm/i915/intel_lrc.h +0 -0
index 52b3a1f..848ca21 100644
@@ -72,7 +72,8 @@ int logical_xcs_ring_init(struct intel_engine_cs *engine);
 
Quoted 6 lines

drivers/gpu/drm/i915/intel_lrc.h int logical_xcs_ring_init(struct intel_engine_cs *engine); · line 77

Quoted 6 lines
drivers/gpu/drm/i915/intel_lrc.h +2 -1
   /* One extra page is added before LRC for GuC as shared data */
   #define LRC_GUCSHR_PN  (0)
-#define LRC_PPHWSP_PN  (LRC_GUCSHR_PN + 1)
+#define LRC_GUCSHR_PAGES 1
+#define LRC_PPHWSP_PN  (LRC_GUCSHR_PN + LRC_GUCSHR_PAGES)
   #define LRC_STATE_PN   (LRC_PPHWSP_PN + 1)

but

/* We allocate a header at the start of the context image for our own
* use, therefore the actual location of the logical state is offset
* from the start of the VMA. The layout is
*

  * | [guc]          | [hwsp] [logical state] |
  * |<- our header ->|<- context image      ->|

*
*/

/* One extra page is used for sharing with the GuC */ #define LRC_GUCSHR_PN (0)
#define LRC_GUCSHR_PAGES 1

/* At the start of the context image is its per-process HWS page */ #define LRC_PPHWSP_PN (LRC_GUCSHR_PN + LRC_GUCSHR_PAGES)

Quoted 8 lines

/* Finally we have the logical state for the context */
#define LRC_STATE_PN (LRC_PPHWSP_PN + 1)

/* Currently we include the PPHWSP in __intel_engine_context_size() so
* the size of the header is synonymous with the start of the PPHWSP. */
#define LRC_HEADER_PAGES LRC_PPHWSP_PN

It cannot be clearer now.

The only extra thing I was thinking about was also defining the 'size' (length?) of the guc and pphwsp sections, in case those "+ 1" were not clear:

 #define LRC_GUCSHR_PN  (0)
-#define LRC_PPHWSP_PN  (LRC_GUCSHR_PN + 1)
-#define LRC_STATE_PN   (LRC_PPHWSP_PN + 1)
+#define LRC_GUCSHR_SZ  (1)
+#define LRC_PPHWSP_PN  (LRC_GUCSHR_PN + LRC_GUCSHR_SZ)
+#define LRC_PPHWSP_SZ  (1)
+#define LRC_STATE_PN   (LRC_PPHWSP_PN + LRC_PPHWSP_SZ)
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines
drivers/gpu/drm/i915/intel_lrc.h +0 -0
   struct drm_i915_private;
 

Nowhere do we use LRC_GUCSHR_PN. Having it just gives us the mirage of flexibility...

Quoted 9 lines
Quoted 5 lines

LRC_GUCSHR_PN is not used because people assumed it is always page 0 of the kernel ctx :S

One more thing to fix, use LRC_GUCSHR_PN in all the places where this _shared_ GuC page is used (guc_ggtt_offset(ctx->engine[RCS].state))

That may prove very handy.
-Chris

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#9 lore raw

Quoting Michel Thierry (2017-07-12 02:01:29)

Quoted 34 lines · Chris Wilson
Quoted 16 lines

/* We allocate a header at the start of the context image for our own
* use, therefore the actual location of the logical state is offset
* from the start of the VMA. The layout is
*

  * | [guc]          | [hwsp] [logical state] |
  * |<- our header ->|<- context image      ->|

*
*/

/* One extra page is used for sharing with the GuC */ #define LRC_GUCSHR_PN (0)
#define LRC_GUCSHR_PAGES 1

/* At the start of the context image is its per-process HWS page */ #define LRC_PPHWSP_PN (LRC_GUCSHR_PN + LRC_GUCSHR_PAGES)

Quoted 8 lines

/* Finally we have the logical state for the context */
#define LRC_STATE_PN (LRC_PPHWSP_PN + 1)

/* Currently we include the PPHWSP in __intel_engine_context_size() so
* the size of the header is synonymous with the start of the PPHWSP. */
#define LRC_HEADER_PAGES LRC_PPHWSP_PN

It cannot be clearer now.

The only extra thing I was thinking about was also defining the 'size' (length?) of the guc and pphwsp sections, in case those "+ 1" were not clear:

  #define LRC_GUCSHR_PN  (0)

Review

Quoted 6 lines
(quoted diff) +4 -2
-#define LRC_PPHWSP_PN  (LRC_GUCSHR_PN + 1)
-#define LRC_STATE_PN   (LRC_PPHWSP_PN + 1)
+#define LRC_GUCSHR_SZ  (1)
+#define LRC_PPHWSP_PN  (LRC_GUCSHR_PN + LRC_GUCSHR_SZ)
+#define LRC_PPHWSP_SZ  (1)
+#define LRC_STATE_PN   (LRC_PPHWSP_PN + LRC_PPHWSP_SZ)

Yes, I erred on the side of laziness thinking that this is common enough that there should be a generator somewhere, but was to lazy to find/write one.
-Chris
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx