diff mbox series

[v6,3/5] drm/i915: Introduce guard pages to i915_vma

Message ID 20221201203912.346110-1-andi.shyti@linux.intel.com (mailing list archive)
State New, archived
Headers show
Series None | expand

Commit Message

Andi Shyti Dec. 1, 2022, 8:39 p.m. UTC
From: Chris Wilson <chris@chris-wilson.co.uk>

Introduce the concept of padding the i915_vma with guard pages before
and after. The major consequence is that all ordinary uses of i915_vma
must use i915_vma_offset/i915_vma_size and not i915_vma.node.start/size
directly, as the drm_mm_node will include the guard pages that surround
our object.

The biggest connundrum is how exactly to mix requesting a fixed address
with guard pages, particularly through the existing uABI. The user does
not know about guard pages, so such must be transparent to the user, and
so the execobj.offset must be that of the object itself excluding the
guard. So a PIN_OFFSET_FIXED must then be exclusive of the guard pages.
The caveat is that some placements will be impossible with guard pages,
as wrap arounds need to be avoided, and the vma itself will require a
larger node. We must not report EINVAL but ENOSPC as these are unavailable
locations within the GTT rather than conflicting user requirements.

In the next patch, we start using guard pages for scanout objects. While
these are limited to GGTT vma, on a few platforms these vma (or at least
an alias of the vma) is shared with userspace, so we may leak the
existence of such guards if we are not careful to ensure that the
execobj.offset is transparent and excludes the guards. (On such platforms
like ivb, without full-ppgtt, userspace has to use relocations so the
presence of more untouchable regions within its GTT such be of no further
issue.)

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Tejas Upadhyay <tejaskumarx.surendrakumar.upadhyay@intel.com>
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: Andi Shyti <andi.shyti@linux.intel.com>
---
Hi Tvrtko,

I removed your r-b in this version because I restored the original value
of the guard being aligned with the vma size alignment. Turns out that
CI failed with the latest version because the guard was becoming too big
(we would have hit the GEM_BUG_ON)[*].

The reason why now the guard is aligned with the vma alignment is that
the area is already aligned and if we use as a starting address start +
guard, guard needs to be aligned, otherwise we screw up all the memory
alignment.

Let me know if it makes sense to you.

Thanks,
Andi

Changelog
=========
v5 -> v6:
 - restore the original alignment of guard so that it's aligned
   coherently with the vma area's alignment.
v4 -> v5:
 - remove again the GEM_BUG_ON()
 - fix an oversight where the rounding was called without assigning the
   value to the guard.
v1 -> v4:
 - refer to the cover letter.

[*] https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_110720v5/fi-rkl-11600/igt@i915_module_load@load.html#dmesg-warnings300

 drivers/gpu/drm/i915/gt/intel_ggtt.c     | 14 +++++---
 drivers/gpu/drm/i915/i915_gem_gtt.h      |  3 +-
 drivers/gpu/drm/i915/i915_vma.c          | 43 ++++++++++++++++++++----
 drivers/gpu/drm/i915/i915_vma.h          |  5 +--
 drivers/gpu/drm/i915/i915_vma_resource.c |  4 +--
 drivers/gpu/drm/i915/i915_vma_resource.h |  7 +++-
 drivers/gpu/drm/i915/i915_vma_types.h    |  1 +
 7 files changed, 60 insertions(+), 17 deletions(-)

Comments

Tvrtko Ursulin Dec. 2, 2022, 10:20 a.m. UTC | #1
On 01/12/2022 20:39, Andi Shyti wrote:
> From: Chris Wilson <chris@chris-wilson.co.uk>
> 
> Introduce the concept of padding the i915_vma with guard pages before
> and after. The major consequence is that all ordinary uses of i915_vma
> must use i915_vma_offset/i915_vma_size and not i915_vma.node.start/size
> directly, as the drm_mm_node will include the guard pages that surround
> our object.
> 
> The biggest connundrum is how exactly to mix requesting a fixed address
> with guard pages, particularly through the existing uABI. The user does
> not know about guard pages, so such must be transparent to the user, and
> so the execobj.offset must be that of the object itself excluding the
> guard. So a PIN_OFFSET_FIXED must then be exclusive of the guard pages.
> The caveat is that some placements will be impossible with guard pages,
> as wrap arounds need to be avoided, and the vma itself will require a
> larger node. We must not report EINVAL but ENOSPC as these are unavailable
> locations within the GTT rather than conflicting user requirements.
> 
> In the next patch, we start using guard pages for scanout objects. While
> these are limited to GGTT vma, on a few platforms these vma (or at least
> an alias of the vma) is shared with userspace, so we may leak the
> existence of such guards if we are not careful to ensure that the
> execobj.offset is transparent and excludes the guards. (On such platforms
> like ivb, without full-ppgtt, userspace has to use relocations so the
> presence of more untouchable regions within its GTT such be of no further
> issue.)
> 
> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> Signed-off-by: Tejas Upadhyay <tejaskumarx.surendrakumar.upadhyay@intel.com>
> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
> Signed-off-by: Andi Shyti <andi.shyti@linux.intel.com>
> ---
> Hi Tvrtko,
> 
> I removed your r-b in this version because I restored the original value
> of the guard being aligned with the vma size alignment. Turns out that
> CI failed with the latest version because the guard was becoming too big
> (we would have hit the GEM_BUG_ON)[*].
> 
> The reason why now the guard is aligned with the vma alignment is that
> the area is already aligned and if we use as a starting address start +
> guard, guard needs to be aligned, otherwise we screw up all the memory
> alignment.
> 
> Let me know if it makes sense to you.

Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>

Conditional to promise of a prioritised follow up improvement, if it 
turns out GGTT wastage due a bit over zealous guard size comes to bite.

Regards,

Tvrtko

> 
> Thanks,
> Andi
> 
> Changelog
> =========
> v5 -> v6:
>   - restore the original alignment of guard so that it's aligned
>     coherently with the vma area's alignment.
> v4 -> v5:
>   - remove again the GEM_BUG_ON()
>   - fix an oversight where the rounding was called without assigning the
>     value to the guard.
> v1 -> v4:
>   - refer to the cover letter.
> 
> [*] https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_110720v5/fi-rkl-11600/igt@i915_module_load@load.html#dmesg-warnings300
> 
>   drivers/gpu/drm/i915/gt/intel_ggtt.c     | 14 +++++---
>   drivers/gpu/drm/i915/i915_gem_gtt.h      |  3 +-
>   drivers/gpu/drm/i915/i915_vma.c          | 43 ++++++++++++++++++++----
>   drivers/gpu/drm/i915/i915_vma.h          |  5 +--
>   drivers/gpu/drm/i915/i915_vma_resource.c |  4 +--
>   drivers/gpu/drm/i915/i915_vma_resource.h |  7 +++-
>   drivers/gpu/drm/i915/i915_vma_types.h    |  1 +
>   7 files changed, 60 insertions(+), 17 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/gt/intel_ggtt.c b/drivers/gpu/drm/i915/gt/intel_ggtt.c
> index 7644738b9cdbe..784d4a8c43ba9 100644
> --- a/drivers/gpu/drm/i915/gt/intel_ggtt.c
> +++ b/drivers/gpu/drm/i915/gt/intel_ggtt.c
> @@ -296,8 +296,11 @@ static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
>   	 */
>   
>   	gte = (gen8_pte_t __iomem *)ggtt->gsm;
> -	gte += vma_res->start / I915_GTT_PAGE_SIZE;
> -	end = gte + vma_res->node_size / I915_GTT_PAGE_SIZE;
> +	gte += (vma_res->start - vma_res->guard) / I915_GTT_PAGE_SIZE;
> +	end = gte + vma_res->guard / I915_GTT_PAGE_SIZE;
> +	while (gte < end)
> +		gen8_set_pte(gte++, vm->scratch[0]->encode);
> +	end += (vma_res->node_size + vma_res->guard) / I915_GTT_PAGE_SIZE;
>   
>   	for_each_sgt_daddr(addr, iter, vma_res->bi.pages)
>   		gen8_set_pte(gte++, pte_encode | addr);
> @@ -347,9 +350,12 @@ static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
>   	dma_addr_t addr;
>   
>   	gte = (gen6_pte_t __iomem *)ggtt->gsm;
> -	gte += vma_res->start / I915_GTT_PAGE_SIZE;
> -	end = gte + vma_res->node_size / I915_GTT_PAGE_SIZE;
> +	gte += (vma_res->start - vma_res->guard) / I915_GTT_PAGE_SIZE;
>   
> +	end = gte + vma_res->guard / I915_GTT_PAGE_SIZE;
> +	while (gte < end)
> +		iowrite32(vm->scratch[0]->encode, gte++);
> +	end += (vma_res->node_size + vma_res->guard) / I915_GTT_PAGE_SIZE;
>   	for_each_sgt_daddr(addr, iter, vma_res->bi.pages)
>   		iowrite32(vm->pte_encode(addr, level, flags), gte++);
>   	GEM_BUG_ON(gte > end);
> diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.h b/drivers/gpu/drm/i915/i915_gem_gtt.h
> index 8c2f57eb5ddaa..2434197830523 100644
> --- a/drivers/gpu/drm/i915/i915_gem_gtt.h
> +++ b/drivers/gpu/drm/i915/i915_gem_gtt.h
> @@ -44,7 +44,8 @@ int i915_gem_gtt_insert(struct i915_address_space *vm,
>   #define PIN_HIGH		BIT_ULL(5)
>   #define PIN_OFFSET_BIAS		BIT_ULL(6)
>   #define PIN_OFFSET_FIXED	BIT_ULL(7)
> -#define PIN_VALIDATE		BIT_ULL(8) /* validate placement only, no need to call unpin() */
> +#define PIN_OFFSET_GUARD	BIT_ULL(8)
> +#define PIN_VALIDATE		BIT_ULL(9) /* validate placement only, no need to call unpin() */
>   
>   #define PIN_GLOBAL		BIT_ULL(10) /* I915_VMA_GLOBAL_BIND */
>   #define PIN_USER		BIT_ULL(11) /* I915_VMA_LOCAL_BIND */
> diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c
> index fefee5fef38d3..34f0e6c923c26 100644
> --- a/drivers/gpu/drm/i915/i915_vma.c
> +++ b/drivers/gpu/drm/i915/i915_vma.c
> @@ -419,7 +419,7 @@ i915_vma_resource_init_from_vma(struct i915_vma_resource *vma_res,
>   			       obj->mm.rsgt, i915_gem_object_is_readonly(obj),
>   			       i915_gem_object_is_lmem(obj), obj->mm.region,
>   			       vma->ops, vma->private, __i915_vma_offset(vma),
> -			       __i915_vma_size(vma), vma->size);
> +			       __i915_vma_size(vma), vma->size, vma->guard);
>   }
>   
>   /**
> @@ -677,6 +677,10 @@ bool i915_vma_misplaced(const struct i915_vma *vma,
>   	    i915_vma_offset(vma) != (flags & PIN_OFFSET_MASK))
>   		return true;
>   
> +	if (flags & PIN_OFFSET_GUARD &&
> +	    vma->guard < (flags & PIN_OFFSET_MASK))
> +		return true;
> +
>   	return false;
>   }
>   
> @@ -749,15 +753,16 @@ static int
>   i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
>   		u64 size, u64 alignment, u64 flags)
>   {
> -	unsigned long color;
> +	unsigned long color, guard;
>   	u64 start, end;
>   	int ret;
>   
>   	GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
>   	GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
> +	GEM_BUG_ON(hweight64(flags & (PIN_OFFSET_GUARD | PIN_OFFSET_FIXED | PIN_OFFSET_BIAS)) > 1);
>   
>   	size = max(size, vma->size);
> -	alignment = max(alignment, vma->display_alignment);
> +	alignment = max_t(typeof(alignment), alignment, vma->display_alignment);
>   	if (flags & PIN_MAPPABLE) {
>   		size = max_t(typeof(size), size, vma->fence_size);
>   		alignment = max_t(typeof(alignment),
> @@ -768,6 +773,18 @@ i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
>   	GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
>   	GEM_BUG_ON(!is_power_of_2(alignment));
>   
> +	guard = vma->guard; /* retain guard across rebinds */
> +	if (flags & PIN_OFFSET_GUARD) {
> +		GEM_BUG_ON(overflows_type(flags & PIN_OFFSET_MASK, u32));
> +		guard = max_t(u32, guard, flags & PIN_OFFSET_MASK);
> +	}
> +	/*
> +	 * As we align the node upon insertion, but the hardware gets
> +	 * node.start + guard, the easiest way to make that work is
> +	 * to make the guard a multiple of the alignment size.
> +	 */
> +	guard = ALIGN(guard, alignment);
> +
>   	start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
>   	GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
>   
> @@ -780,11 +797,12 @@ i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
>   
>   	alignment = max(alignment, i915_vm_obj_min_alignment(vma->vm, vma->obj));
>   
> -	/* If binding the object/GGTT view requires more space than the entire
> +	/*
> +	 * If binding the object/GGTT view requires more space than the entire
>   	 * aperture has, reject it early before evicting everything in a vain
>   	 * attempt to find space.
>   	 */
> -	if (size > end) {
> +	if (size > end - 2 * guard) {
>   		drm_dbg(&to_i915(vma->obj->base.dev)->drm,
>   			"Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
>   			size, flags & PIN_MAPPABLE ? "mappable" : "total", end);
> @@ -801,13 +819,23 @@ i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
>   		if (!IS_ALIGNED(offset, alignment) ||
>   		    range_overflows(offset, size, end))
>   			return -EINVAL;
> +		/*
> +		 * The caller knows not of the guard added by others and
> +		 * requests for the offset of the start of its buffer
> +		 * to be fixed, which may not be the same as the position
> +		 * of the vma->node due to the guard pages.
> +		 */
> +		if (offset < guard || offset + size > end - guard)
> +			return -ENOSPC;
>   
>   		ret = i915_gem_gtt_reserve(vma->vm, ww, &vma->node,
> -					   size, offset, color,
> -					   flags);
> +					   size + 2 * guard,
> +					   offset - guard,
> +					   color, flags);
>   		if (ret)
>   			return ret;
>   	} else {
> +		size += 2 * guard;
>   		/*
>   		 * We only support huge gtt pages through the 48b PPGTT,
>   		 * however we also don't want to force any alignment for
> @@ -855,6 +883,7 @@ i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
>   	GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
>   
>   	list_move_tail(&vma->vm_link, &vma->vm->bound_list);
> +	vma->guard = guard;
>   
>   	return 0;
>   }
> diff --git a/drivers/gpu/drm/i915/i915_vma.h b/drivers/gpu/drm/i915/i915_vma.h
> index 3fd4512b1f65f..ed5c9d682a1b2 100644
> --- a/drivers/gpu/drm/i915/i915_vma.h
> +++ b/drivers/gpu/drm/i915/i915_vma.h
> @@ -128,7 +128,7 @@ static inline bool i915_vma_is_closed(const struct i915_vma *vma)
>   /* Internal use only. */
>   static inline u64 __i915_vma_size(const struct i915_vma *vma)
>   {
> -	return vma->node.size;
> +	return vma->node.size - 2 * vma->guard;
>   }
>   
>   /**
> @@ -150,7 +150,8 @@ static inline u64 i915_vma_size(const struct i915_vma *vma)
>   /* Internal use only. */
>   static inline u64 __i915_vma_offset(const struct i915_vma *vma)
>   {
> -	return vma->node.start;
> +	/* The actual start of the vma->pages is after the guard pages. */
> +	return vma->node.start + vma->guard;
>   }
>   
>   /**
> diff --git a/drivers/gpu/drm/i915/i915_vma_resource.c b/drivers/gpu/drm/i915/i915_vma_resource.c
> index de1342dbfa128..6ba7a7feceba1 100644
> --- a/drivers/gpu/drm/i915/i915_vma_resource.c
> +++ b/drivers/gpu/drm/i915/i915_vma_resource.c
> @@ -34,8 +34,8 @@ static struct kmem_cache *slab_vma_resources;
>    * and removal of fences increases as O(ln(pending_unbinds)) instead of
>    * O(1) for a single fence without interval tree.
>    */
> -#define VMA_RES_START(_node) ((_node)->start)
> -#define VMA_RES_LAST(_node) ((_node)->start + (_node)->node_size - 1)
> +#define VMA_RES_START(_node) ((_node)->start - (_node)->guard)
> +#define VMA_RES_LAST(_node) ((_node)->start + (_node)->node_size + (_node)->guard - 1)
>   INTERVAL_TREE_DEFINE(struct i915_vma_resource, rb,
>   		     u64, __subtree_last,
>   		     VMA_RES_START, VMA_RES_LAST, static, vma_res_itree);
> diff --git a/drivers/gpu/drm/i915/i915_vma_resource.h b/drivers/gpu/drm/i915/i915_vma_resource.h
> index 54edf3739ca0b..c1864e3d0b43e 100644
> --- a/drivers/gpu/drm/i915/i915_vma_resource.h
> +++ b/drivers/gpu/drm/i915/i915_vma_resource.h
> @@ -57,6 +57,7 @@ struct i915_page_sizes {
>    * @node_size: Size of the allocated range manager node with padding
>    * subtracted.
>    * @vma_size: Bind size.
> + * @guard: The size of guard area preceding and trailing the bind.
>    * @page_sizes_gtt: Resulting page sizes from the bind operation.
>    * @bound_flags: Flags indicating binding status.
>    * @allocated: Backend private data. TODO: Should move into @private.
> @@ -115,6 +116,7 @@ struct i915_vma_resource {
>   	u64 start;
>   	u64 node_size;
>   	u64 vma_size;
> +	u32 guard;
>   	u32 page_sizes_gtt;
>   
>   	u32 bound_flags;
> @@ -179,6 +181,7 @@ static inline void i915_vma_resource_put(struct i915_vma_resource *vma_res)
>    * @start: Offset into the address space of bind range start after padding.
>    * @node_size: Size of the allocated range manager node minus padding.
>    * @size: Bind size.
> + * @guard: The size of the guard area preceding and trailing the bind.
>    *
>    * Initializes a vma resource allocated using i915_vma_resource_alloc().
>    * The reason for having separate allocate and initialize function is that
> @@ -197,7 +200,8 @@ static inline void i915_vma_resource_init(struct i915_vma_resource *vma_res,
>   					  void *private,
>   					  u64 start,
>   					  u64 node_size,
> -					  u64 size)
> +					  u64 size,
> +					  u32 guard)
>   {
>   	__i915_vma_resource_init(vma_res);
>   	vma_res->vm = vm;
> @@ -215,6 +219,7 @@ static inline void i915_vma_resource_init(struct i915_vma_resource *vma_res,
>   	vma_res->start = start;
>   	vma_res->node_size = node_size;
>   	vma_res->vma_size = size;
> +	vma_res->guard = guard;
>   }
>   
>   static inline void i915_vma_resource_fini(struct i915_vma_resource *vma_res)
> diff --git a/drivers/gpu/drm/i915/i915_vma_types.h b/drivers/gpu/drm/i915/i915_vma_types.h
> index 0375812792b9c..77fda2244d161 100644
> --- a/drivers/gpu/drm/i915/i915_vma_types.h
> +++ b/drivers/gpu/drm/i915/i915_vma_types.h
> @@ -202,6 +202,7 @@ struct i915_vma {
>   	/* mmap-offset associated with fencing for this vma */
>   	struct i915_mmap_offset	*mmo;
>   
> +	u32 guard; /* padding allocated around vma->pages within the node */
>   	u32 fence_size;
>   	u32 fence_alignment;
>   	u32 display_alignment;
Andi Shyti Dec. 2, 2022, 11:11 a.m. UTC | #2
Hi Tvrtko,

On Fri, Dec 02, 2022 at 10:20:11AM +0000, Tvrtko Ursulin wrote:
> 
> On 01/12/2022 20:39, Andi Shyti wrote:
> > From: Chris Wilson <chris@chris-wilson.co.uk>
> > 
> > Introduce the concept of padding the i915_vma with guard pages before
> > and after. The major consequence is that all ordinary uses of i915_vma
> > must use i915_vma_offset/i915_vma_size and not i915_vma.node.start/size
> > directly, as the drm_mm_node will include the guard pages that surround
> > our object.
> > 
> > The biggest connundrum is how exactly to mix requesting a fixed address
> > with guard pages, particularly through the existing uABI. The user does
> > not know about guard pages, so such must be transparent to the user, and
> > so the execobj.offset must be that of the object itself excluding the
> > guard. So a PIN_OFFSET_FIXED must then be exclusive of the guard pages.
> > The caveat is that some placements will be impossible with guard pages,
> > as wrap arounds need to be avoided, and the vma itself will require a
> > larger node. We must not report EINVAL but ENOSPC as these are unavailable
> > locations within the GTT rather than conflicting user requirements.
> > 
> > In the next patch, we start using guard pages for scanout objects. While
> > these are limited to GGTT vma, on a few platforms these vma (or at least
> > an alias of the vma) is shared with userspace, so we may leak the
> > existence of such guards if we are not careful to ensure that the
> > execobj.offset is transparent and excludes the guards. (On such platforms
> > like ivb, without full-ppgtt, userspace has to use relocations so the
> > presence of more untouchable regions within its GTT such be of no further
> > issue.)
> > 
> > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> > Signed-off-by: Tejas Upadhyay <tejaskumarx.surendrakumar.upadhyay@intel.com>
> > Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
> > Signed-off-by: Andi Shyti <andi.shyti@linux.intel.com>
> > ---
> > Hi Tvrtko,
> > 
> > I removed your r-b in this version because I restored the original value
> > of the guard being aligned with the vma size alignment. Turns out that
> > CI failed with the latest version because the guard was becoming too big
> > (we would have hit the GEM_BUG_ON)[*].
> > 
> > The reason why now the guard is aligned with the vma alignment is that
> > the area is already aligned and if we use as a starting address start +
> > guard, guard needs to be aligned, otherwise we screw up all the memory
> > alignment.
> > 
> > Let me know if it makes sense to you.
> 
> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
> 
> Conditional to promise of a prioritised follow up improvement, if it turns
> out GGTT wastage due a bit over zealous guard size comes to bite.

Sure! I'll be alert!

There are some unrelated failures from CI, just to be sure I sent
last night a trybot run.

Thanks, Tvrtko!

Andi
Tvrtko Ursulin Dec. 2, 2022, 12:38 p.m. UTC | #3
On 02/12/2022 11:11, Andi Shyti wrote:
> Hi Tvrtko,
> 
> On Fri, Dec 02, 2022 at 10:20:11AM +0000, Tvrtko Ursulin wrote:
>>
>> On 01/12/2022 20:39, Andi Shyti wrote:
>>> From: Chris Wilson <chris@chris-wilson.co.uk>
>>>
>>> Introduce the concept of padding the i915_vma with guard pages before
>>> and after. The major consequence is that all ordinary uses of i915_vma
>>> must use i915_vma_offset/i915_vma_size and not i915_vma.node.start/size
>>> directly, as the drm_mm_node will include the guard pages that surround
>>> our object.
>>>
>>> The biggest connundrum is how exactly to mix requesting a fixed address
>>> with guard pages, particularly through the existing uABI. The user does
>>> not know about guard pages, so such must be transparent to the user, and
>>> so the execobj.offset must be that of the object itself excluding the
>>> guard. So a PIN_OFFSET_FIXED must then be exclusive of the guard pages.
>>> The caveat is that some placements will be impossible with guard pages,
>>> as wrap arounds need to be avoided, and the vma itself will require a
>>> larger node. We must not report EINVAL but ENOSPC as these are unavailable
>>> locations within the GTT rather than conflicting user requirements.
>>>
>>> In the next patch, we start using guard pages for scanout objects. While
>>> these are limited to GGTT vma, on a few platforms these vma (or at least
>>> an alias of the vma) is shared with userspace, so we may leak the
>>> existence of such guards if we are not careful to ensure that the
>>> execobj.offset is transparent and excludes the guards. (On such platforms
>>> like ivb, without full-ppgtt, userspace has to use relocations so the
>>> presence of more untouchable regions within its GTT such be of no further
>>> issue.)
>>>
>>> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
>>> Signed-off-by: Tejas Upadhyay <tejaskumarx.surendrakumar.upadhyay@intel.com>
>>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
>>> Signed-off-by: Andi Shyti <andi.shyti@linux.intel.com>
>>> ---
>>> Hi Tvrtko,
>>>
>>> I removed your r-b in this version because I restored the original value
>>> of the guard being aligned with the vma size alignment. Turns out that
>>> CI failed with the latest version because the guard was becoming too big
>>> (we would have hit the GEM_BUG_ON)[*].
>>>
>>> The reason why now the guard is aligned with the vma alignment is that
>>> the area is already aligned and if we use as a starting address start +
>>> guard, guard needs to be aligned, otherwise we screw up all the memory
>>> alignment.
>>>
>>> Let me know if it makes sense to you.
>>
>> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
>>
>> Conditional to promise of a prioritised follow up improvement, if it turns
>> out GGTT wastage due a bit over zealous guard size comes to bite.
> 
> Sure! I'll be alert!
> 
> There are some unrelated failures from CI, just to be sure I sent
> last night a trybot run.

Trybot looked okay, and I just pressed re-test for the intel-gfx series 
so lets see that too.

Regards,

Tvrtko
diff mbox series

Patch

diff --git a/drivers/gpu/drm/i915/gt/intel_ggtt.c b/drivers/gpu/drm/i915/gt/intel_ggtt.c
index 7644738b9cdbe..784d4a8c43ba9 100644
--- a/drivers/gpu/drm/i915/gt/intel_ggtt.c
+++ b/drivers/gpu/drm/i915/gt/intel_ggtt.c
@@ -296,8 +296,11 @@  static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
 	 */
 
 	gte = (gen8_pte_t __iomem *)ggtt->gsm;
-	gte += vma_res->start / I915_GTT_PAGE_SIZE;
-	end = gte + vma_res->node_size / I915_GTT_PAGE_SIZE;
+	gte += (vma_res->start - vma_res->guard) / I915_GTT_PAGE_SIZE;
+	end = gte + vma_res->guard / I915_GTT_PAGE_SIZE;
+	while (gte < end)
+		gen8_set_pte(gte++, vm->scratch[0]->encode);
+	end += (vma_res->node_size + vma_res->guard) / I915_GTT_PAGE_SIZE;
 
 	for_each_sgt_daddr(addr, iter, vma_res->bi.pages)
 		gen8_set_pte(gte++, pte_encode | addr);
@@ -347,9 +350,12 @@  static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
 	dma_addr_t addr;
 
 	gte = (gen6_pte_t __iomem *)ggtt->gsm;
-	gte += vma_res->start / I915_GTT_PAGE_SIZE;
-	end = gte + vma_res->node_size / I915_GTT_PAGE_SIZE;
+	gte += (vma_res->start - vma_res->guard) / I915_GTT_PAGE_SIZE;
 
+	end = gte + vma_res->guard / I915_GTT_PAGE_SIZE;
+	while (gte < end)
+		iowrite32(vm->scratch[0]->encode, gte++);
+	end += (vma_res->node_size + vma_res->guard) / I915_GTT_PAGE_SIZE;
 	for_each_sgt_daddr(addr, iter, vma_res->bi.pages)
 		iowrite32(vm->pte_encode(addr, level, flags), gte++);
 	GEM_BUG_ON(gte > end);
diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.h b/drivers/gpu/drm/i915/i915_gem_gtt.h
index 8c2f57eb5ddaa..2434197830523 100644
--- a/drivers/gpu/drm/i915/i915_gem_gtt.h
+++ b/drivers/gpu/drm/i915/i915_gem_gtt.h
@@ -44,7 +44,8 @@  int i915_gem_gtt_insert(struct i915_address_space *vm,
 #define PIN_HIGH		BIT_ULL(5)
 #define PIN_OFFSET_BIAS		BIT_ULL(6)
 #define PIN_OFFSET_FIXED	BIT_ULL(7)
-#define PIN_VALIDATE		BIT_ULL(8) /* validate placement only, no need to call unpin() */
+#define PIN_OFFSET_GUARD	BIT_ULL(8)
+#define PIN_VALIDATE		BIT_ULL(9) /* validate placement only, no need to call unpin() */
 
 #define PIN_GLOBAL		BIT_ULL(10) /* I915_VMA_GLOBAL_BIND */
 #define PIN_USER		BIT_ULL(11) /* I915_VMA_LOCAL_BIND */
diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c
index fefee5fef38d3..34f0e6c923c26 100644
--- a/drivers/gpu/drm/i915/i915_vma.c
+++ b/drivers/gpu/drm/i915/i915_vma.c
@@ -419,7 +419,7 @@  i915_vma_resource_init_from_vma(struct i915_vma_resource *vma_res,
 			       obj->mm.rsgt, i915_gem_object_is_readonly(obj),
 			       i915_gem_object_is_lmem(obj), obj->mm.region,
 			       vma->ops, vma->private, __i915_vma_offset(vma),
-			       __i915_vma_size(vma), vma->size);
+			       __i915_vma_size(vma), vma->size, vma->guard);
 }
 
 /**
@@ -677,6 +677,10 @@  bool i915_vma_misplaced(const struct i915_vma *vma,
 	    i915_vma_offset(vma) != (flags & PIN_OFFSET_MASK))
 		return true;
 
+	if (flags & PIN_OFFSET_GUARD &&
+	    vma->guard < (flags & PIN_OFFSET_MASK))
+		return true;
+
 	return false;
 }
 
@@ -749,15 +753,16 @@  static int
 i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
 		u64 size, u64 alignment, u64 flags)
 {
-	unsigned long color;
+	unsigned long color, guard;
 	u64 start, end;
 	int ret;
 
 	GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 	GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
+	GEM_BUG_ON(hweight64(flags & (PIN_OFFSET_GUARD | PIN_OFFSET_FIXED | PIN_OFFSET_BIAS)) > 1);
 
 	size = max(size, vma->size);
-	alignment = max(alignment, vma->display_alignment);
+	alignment = max_t(typeof(alignment), alignment, vma->display_alignment);
 	if (flags & PIN_MAPPABLE) {
 		size = max_t(typeof(size), size, vma->fence_size);
 		alignment = max_t(typeof(alignment),
@@ -768,6 +773,18 @@  i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
 	GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
 	GEM_BUG_ON(!is_power_of_2(alignment));
 
+	guard = vma->guard; /* retain guard across rebinds */
+	if (flags & PIN_OFFSET_GUARD) {
+		GEM_BUG_ON(overflows_type(flags & PIN_OFFSET_MASK, u32));
+		guard = max_t(u32, guard, flags & PIN_OFFSET_MASK);
+	}
+	/*
+	 * As we align the node upon insertion, but the hardware gets
+	 * node.start + guard, the easiest way to make that work is
+	 * to make the guard a multiple of the alignment size.
+	 */
+	guard = ALIGN(guard, alignment);
+
 	start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
 	GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
 
@@ -780,11 +797,12 @@  i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
 
 	alignment = max(alignment, i915_vm_obj_min_alignment(vma->vm, vma->obj));
 
-	/* If binding the object/GGTT view requires more space than the entire
+	/*
+	 * If binding the object/GGTT view requires more space than the entire
 	 * aperture has, reject it early before evicting everything in a vain
 	 * attempt to find space.
 	 */
-	if (size > end) {
+	if (size > end - 2 * guard) {
 		drm_dbg(&to_i915(vma->obj->base.dev)->drm,
 			"Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
 			size, flags & PIN_MAPPABLE ? "mappable" : "total", end);
@@ -801,13 +819,23 @@  i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
 		if (!IS_ALIGNED(offset, alignment) ||
 		    range_overflows(offset, size, end))
 			return -EINVAL;
+		/*
+		 * The caller knows not of the guard added by others and
+		 * requests for the offset of the start of its buffer
+		 * to be fixed, which may not be the same as the position
+		 * of the vma->node due to the guard pages.
+		 */
+		if (offset < guard || offset + size > end - guard)
+			return -ENOSPC;
 
 		ret = i915_gem_gtt_reserve(vma->vm, ww, &vma->node,
-					   size, offset, color,
-					   flags);
+					   size + 2 * guard,
+					   offset - guard,
+					   color, flags);
 		if (ret)
 			return ret;
 	} else {
+		size += 2 * guard;
 		/*
 		 * We only support huge gtt pages through the 48b PPGTT,
 		 * however we also don't want to force any alignment for
@@ -855,6 +883,7 @@  i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
 	GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
 
 	list_move_tail(&vma->vm_link, &vma->vm->bound_list);
+	vma->guard = guard;
 
 	return 0;
 }
diff --git a/drivers/gpu/drm/i915/i915_vma.h b/drivers/gpu/drm/i915/i915_vma.h
index 3fd4512b1f65f..ed5c9d682a1b2 100644
--- a/drivers/gpu/drm/i915/i915_vma.h
+++ b/drivers/gpu/drm/i915/i915_vma.h
@@ -128,7 +128,7 @@  static inline bool i915_vma_is_closed(const struct i915_vma *vma)
 /* Internal use only. */
 static inline u64 __i915_vma_size(const struct i915_vma *vma)
 {
-	return vma->node.size;
+	return vma->node.size - 2 * vma->guard;
 }
 
 /**
@@ -150,7 +150,8 @@  static inline u64 i915_vma_size(const struct i915_vma *vma)
 /* Internal use only. */
 static inline u64 __i915_vma_offset(const struct i915_vma *vma)
 {
-	return vma->node.start;
+	/* The actual start of the vma->pages is after the guard pages. */
+	return vma->node.start + vma->guard;
 }
 
 /**
diff --git a/drivers/gpu/drm/i915/i915_vma_resource.c b/drivers/gpu/drm/i915/i915_vma_resource.c
index de1342dbfa128..6ba7a7feceba1 100644
--- a/drivers/gpu/drm/i915/i915_vma_resource.c
+++ b/drivers/gpu/drm/i915/i915_vma_resource.c
@@ -34,8 +34,8 @@  static struct kmem_cache *slab_vma_resources;
  * and removal of fences increases as O(ln(pending_unbinds)) instead of
  * O(1) for a single fence without interval tree.
  */
-#define VMA_RES_START(_node) ((_node)->start)
-#define VMA_RES_LAST(_node) ((_node)->start + (_node)->node_size - 1)
+#define VMA_RES_START(_node) ((_node)->start - (_node)->guard)
+#define VMA_RES_LAST(_node) ((_node)->start + (_node)->node_size + (_node)->guard - 1)
 INTERVAL_TREE_DEFINE(struct i915_vma_resource, rb,
 		     u64, __subtree_last,
 		     VMA_RES_START, VMA_RES_LAST, static, vma_res_itree);
diff --git a/drivers/gpu/drm/i915/i915_vma_resource.h b/drivers/gpu/drm/i915/i915_vma_resource.h
index 54edf3739ca0b..c1864e3d0b43e 100644
--- a/drivers/gpu/drm/i915/i915_vma_resource.h
+++ b/drivers/gpu/drm/i915/i915_vma_resource.h
@@ -57,6 +57,7 @@  struct i915_page_sizes {
  * @node_size: Size of the allocated range manager node with padding
  * subtracted.
  * @vma_size: Bind size.
+ * @guard: The size of guard area preceding and trailing the bind.
  * @page_sizes_gtt: Resulting page sizes from the bind operation.
  * @bound_flags: Flags indicating binding status.
  * @allocated: Backend private data. TODO: Should move into @private.
@@ -115,6 +116,7 @@  struct i915_vma_resource {
 	u64 start;
 	u64 node_size;
 	u64 vma_size;
+	u32 guard;
 	u32 page_sizes_gtt;
 
 	u32 bound_flags;
@@ -179,6 +181,7 @@  static inline void i915_vma_resource_put(struct i915_vma_resource *vma_res)
  * @start: Offset into the address space of bind range start after padding.
  * @node_size: Size of the allocated range manager node minus padding.
  * @size: Bind size.
+ * @guard: The size of the guard area preceding and trailing the bind.
  *
  * Initializes a vma resource allocated using i915_vma_resource_alloc().
  * The reason for having separate allocate and initialize function is that
@@ -197,7 +200,8 @@  static inline void i915_vma_resource_init(struct i915_vma_resource *vma_res,
 					  void *private,
 					  u64 start,
 					  u64 node_size,
-					  u64 size)
+					  u64 size,
+					  u32 guard)
 {
 	__i915_vma_resource_init(vma_res);
 	vma_res->vm = vm;
@@ -215,6 +219,7 @@  static inline void i915_vma_resource_init(struct i915_vma_resource *vma_res,
 	vma_res->start = start;
 	vma_res->node_size = node_size;
 	vma_res->vma_size = size;
+	vma_res->guard = guard;
 }
 
 static inline void i915_vma_resource_fini(struct i915_vma_resource *vma_res)
diff --git a/drivers/gpu/drm/i915/i915_vma_types.h b/drivers/gpu/drm/i915/i915_vma_types.h
index 0375812792b9c..77fda2244d161 100644
--- a/drivers/gpu/drm/i915/i915_vma_types.h
+++ b/drivers/gpu/drm/i915/i915_vma_types.h
@@ -202,6 +202,7 @@  struct i915_vma {
 	/* mmap-offset associated with fencing for this vma */
 	struct i915_mmap_offset	*mmo;
 
+	u32 guard; /* padding allocated around vma->pages within the node */
 	u32 fence_size;
 	u32 fence_alignment;
 	u32 display_alignment;