diff mbox series

drm/i915/gt: Poison GTT scratch pages

Message ID 20200123085124.1256919-1-chris@chris-wilson.co.uk (mailing list archive)
State New, archived
Headers show
Series drm/i915/gt: Poison GTT scratch pages | expand

Commit Message

Chris Wilson Jan. 23, 2020, 8:51 a.m. UTC
Using a clear page for scratch means that we have relatively benign
errors in case it is accidentally used, but that can be rather too
benign for debugging. If we poison the scratch, ideally it quickly
results in an obvious error.

Suggested-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
---
 .../drm/i915/gem/selftests/i915_gem_context.c | 43 ++++++++++++++++---
 drivers/gpu/drm/i915/gt/intel_gtt.c           | 19 ++++++++
 2 files changed, 56 insertions(+), 6 deletions(-)

Comments

Matthew Auld Jan. 23, 2020, 9:38 a.m. UTC | #1
On Thu, 23 Jan 2020 at 08:51, Chris Wilson <chris@chris-wilson.co.uk> wrote:
>
> Using a clear page for scratch means that we have relatively benign
> errors in case it is accidentally used, but that can be rather too
> benign for debugging. If we poison the scratch, ideally it quickly
> results in an obvious error.
>
> Suggested-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>

[snip]

>
> +static void poison_scratch_page(struct page *page, unsigned long size)
> +{
> +       if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
> +               return;
> +
> +       kunmap_atomic(memset(kmap_atomic(page), POISON_FREE, size));

kmap works for !order-zero?

> +}
> +
>  int setup_scratch_page(struct i915_address_space *vm, gfp_t gfp)
>  {
>         unsigned long size;
> @@ -331,6 +339,17 @@ int setup_scratch_page(struct i915_address_space *vm, gfp_t gfp)
>                 if (unlikely(!page))
>                         goto skip;
>
> +               /*
> +                * Use a non-zero scratch page for debugging.
> +                *
> +                * We want a value that should be reasonably obvious
> +                * to spot in the error state, while also causing a GPU hang
> +                * if executed. We prefer using a clear page in production, so
> +                * should it ever be accidentally used, the effect should be
> +                * fairly benign.
> +                */
> +               poison_scratch_page(page, size);
> +
>                 addr = dma_map_page_attrs(vm->dma,
>                                           page, 0, size,
>                                           PCI_DMA_BIDIRECTIONAL,
> --
> 2.25.0
>
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Chris Wilson Jan. 23, 2020, 9:55 a.m. UTC | #2
Quoting Matthew Auld (2020-01-23 09:38:02)
> On Thu, 23 Jan 2020 at 08:51, Chris Wilson <chris@chris-wilson.co.uk> wrote:
> >
> > Using a clear page for scratch means that we have relatively benign
> > errors in case it is accidentally used, but that can be rather too
> > benign for debugging. If we poison the scratch, ideally it quickly
> > results in an obvious error.
> >
> > Suggested-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
> > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> > Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
> 
> [snip]
> 
> >
> > +static void poison_scratch_page(struct page *page, unsigned long size)
> > +{
> > +       if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
> > +               return;
> > +
> > +       kunmap_atomic(memset(kmap_atomic(page), POISON_FREE, size));
> 
> kmap works for !order-zero?

Definitely works if not highmem :)
But needs a page by page approach just in case, fine.
-Chris
diff mbox series

Patch

diff --git a/drivers/gpu/drm/i915/gem/selftests/i915_gem_context.c b/drivers/gpu/drm/i915/gem/selftests/i915_gem_context.c
index 7fc46861a54d..c93fb0d8c597 100644
--- a/drivers/gpu/drm/i915/gem/selftests/i915_gem_context.c
+++ b/drivers/gpu/drm/i915/gem/selftests/i915_gem_context.c
@@ -1575,7 +1575,7 @@  static int read_from_scratch(struct i915_gem_context *ctx,
 	struct drm_i915_private *i915 = ctx->i915;
 	struct drm_i915_gem_object *obj;
 	struct i915_address_space *vm;
-	const u32 RCS_GPR0 = 0x2600; /* not all engines have their own GPR! */
+	const u32 GPR0 = engine->mmio_base + 0x600;
 	const u32 result = 0x100;
 	struct i915_request *rq;
 	struct i915_vma *vma;
@@ -1597,19 +1597,19 @@  static int read_from_scratch(struct i915_gem_context *ctx,
 	memset(cmd, POISON_INUSE, PAGE_SIZE);
 	if (INTEL_GEN(i915) >= 8) {
 		*cmd++ = MI_LOAD_REGISTER_MEM_GEN8;
-		*cmd++ = RCS_GPR0;
+		*cmd++ = GPR0;
 		*cmd++ = lower_32_bits(offset);
 		*cmd++ = upper_32_bits(offset);
 		*cmd++ = MI_STORE_REGISTER_MEM_GEN8;
-		*cmd++ = RCS_GPR0;
+		*cmd++ = GPR0;
 		*cmd++ = result;
 		*cmd++ = 0;
 	} else {
 		*cmd++ = MI_LOAD_REGISTER_MEM;
-		*cmd++ = RCS_GPR0;
+		*cmd++ = GPR0;
 		*cmd++ = offset;
 		*cmd++ = MI_STORE_REGISTER_MEM;
-		*cmd++ = RCS_GPR0;
+		*cmd++ = GPR0;
 		*cmd++ = result;
 	}
 	*cmd = MI_BATCH_BUFFER_END;
@@ -1686,6 +1686,23 @@  static int read_from_scratch(struct i915_gem_context *ctx,
 	return err;
 }
 
+static int check_scratch_page(struct i915_gem_context *ctx, u32 *out)
+{
+	struct page *page = ctx_vm(ctx)->scratch[0].base.page;
+	u32 *vaddr;
+	int err = 0;
+
+	vaddr = kmap(page);
+	memcpy(out, vaddr, sizeof(*out));
+	if (memchr_inv(vaddr, *out, PAGE_SIZE)) {
+		pr_err("Inconsistent initial state of scratch page!\n");
+		err = -EINVAL;
+	}
+	kunmap(page);
+
+	return err;
+}
+
 static int igt_vm_isolation(void *arg)
 {
 	struct drm_i915_private *i915 = arg;
@@ -1696,6 +1713,7 @@  static int igt_vm_isolation(void *arg)
 	I915_RND_STATE(prng);
 	struct file *file;
 	u64 vm_total;
+	u32 expected;
 	int err;
 
 	if (INTEL_GEN(i915) < 7)
@@ -1720,12 +1738,21 @@  static int igt_vm_isolation(void *arg)
 		goto out_file;
 	}
 
+	/* Read the initial state of the scratch page */
+	err = check_scratch_page(ctx_a, &expected);
+	if (err)
+		goto out_file;
+
 	ctx_b = live_context(i915, file);
 	if (IS_ERR(ctx_b)) {
 		err = PTR_ERR(ctx_b);
 		goto out_file;
 	}
 
+	err = check_scratch_page(ctx_b, &expected);
+	if (err)
+		goto out_file;
+
 	/* We can only test vm isolation, if the vm are distinct */
 	if (ctx_vm(ctx_a) == ctx_vm(ctx_b))
 		goto out_file;
@@ -1743,6 +1770,10 @@  static int igt_vm_isolation(void *arg)
 		if (!intel_engine_can_store_dword(engine))
 			continue;
 
+		/* not all engines have their own GPR! */
+		if (INTEL_GEN(i915) < 9 && engine->class != RENDER_CLASS)
+			continue;
+
 		while (!__igt_timeout(end_time, NULL)) {
 			u32 value = 0xc5c5c5c5;
 			u64 offset;
@@ -1760,7 +1791,7 @@  static int igt_vm_isolation(void *arg)
 			if (err)
 				goto out_file;
 
-			if (value) {
+			if (value != expected) {
 				pr_err("%s: Read %08x from scratch (offset 0x%08x_%08x), after %lu reads!\n",
 				       engine->name, value,
 				       upper_32_bits(offset),
diff --git a/drivers/gpu/drm/i915/gt/intel_gtt.c b/drivers/gpu/drm/i915/gt/intel_gtt.c
index 45d8e0019a8e..804da478d8f9 100644
--- a/drivers/gpu/drm/i915/gt/intel_gtt.c
+++ b/drivers/gpu/drm/i915/gt/intel_gtt.c
@@ -299,6 +299,14 @@  fill_page_dma(const struct i915_page_dma *p, const u64 val, unsigned int count)
 	kunmap_atomic(memset64(kmap_atomic(p->page), val, count));
 }
 
+static void poison_scratch_page(struct page *page, unsigned long size)
+{
+	if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
+		return;
+
+	kunmap_atomic(memset(kmap_atomic(page), POISON_FREE, size));
+}
+
 int setup_scratch_page(struct i915_address_space *vm, gfp_t gfp)
 {
 	unsigned long size;
@@ -331,6 +339,17 @@  int setup_scratch_page(struct i915_address_space *vm, gfp_t gfp)
 		if (unlikely(!page))
 			goto skip;
 
+		/*
+		 * Use a non-zero scratch page for debugging.
+		 *
+		 * We want a value that should be reasonably obvious
+		 * to spot in the error state, while also causing a GPU hang
+		 * if executed. We prefer using a clear page in production, so
+		 * should it ever be accidentally used, the effect should be
+		 * fairly benign.
+		 */
+		poison_scratch_page(page, size);
+
 		addr = dma_map_page_attrs(vm->dma,
 					  page, 0, size,
 					  PCI_DMA_BIDIRECTIONAL,