[Intel-gfx] [PATCH] drm/i915: Support to enable TRTT on GEN9

[Chris Wilson]

On Mon, Jan 11, 2016 at 01:09:50PM +0530, Goel, Akash wrote:
> 
> 
> On 1/10/2016 11:09 PM, Chris Wilson wrote:
> >On Sat, Jan 09, 2016 at 05:00:21PM +0530, akash.goel at intel.com wrote:
> >>From: Akash Goel <akash.goel at intel.com>
> >>
> >>Gen9 has an additional address translation hardware support in form of
> >>Tiled Resource Translation Table (TR-TT) which provides an extra level
> >>of abstraction over PPGTT.
> >>This is useful for mapping Sparse/Tiled texture resources.
> >>Sparse resources are created as virtual-only allocations. Regions of the
> >>resource that the application intends to use is bound to the physical memory
> >>on the fly and can be re-bound to different memory allocations over the
> >>lifetime of the resource.
> >>
> >>TR-TT is tightly coupled with PPGTT, a new instance of TR-TT will be required
> >>for a new PPGTT instance, but TR-TT may not enabled for every context.
> >>1/16th of the 48bit PPGTT space is earmarked for the translation by TR-TT,
> >>which such chunk to use is conveyed to HW through a register.
> >>Any GFX address, which lies in that reserved 44 bit range will be translated
> >>through TR-TT first and then through PPGTT to get the actual physical address,
> >>so the output of translation from TR-TT will be a PPGTT offset.
> >>
> >>TRTT is constructed as a 3 level tile Table. Each tile is 64KB is size which
> >>leaves behind 44-16=28 address bits. 28bits are partitioned as 9+9+10, and
> >>each level is contained within a 4KB page hence L3 and L2 is composed of
> >>512 64b entries and L1 is composed of 1024 32b entries.
> >>
> >>There is a provision to keep TR-TT Tables in virtual space, where the pages of
> >>TRTT tables will be mapped to PPGTT.
> >>Currently this is the supported mode, in this mode UMD will have a full control
> >>on TR-TT management, with bare minimum support from KMD.
> >>So the entries of L3 table will contain the PPGTT offset of L2 Table pages,
> >>similarly entries of L2 table will contain the PPGTT offset of L1 Table pages.
> >>The entries of L1 table will contain the PPGTT offset of BOs actually backing
> >>the Sparse resources.
> >
> >>The assumption here is that UMD only will do the complete PPGTT address space
> >>management and use the Soft Pin API for all the buffer objects associated with
> >>a given Context.
> >
> >That is a poor assumption, and not one required for this to work.
> >
> This is not a strict requirement.
> But I thought that conflicts will be minimized if UMD itself can do
> the full address space management.
> At least UMD has to ensure that PPGTT offset of L3 table remains
> same throughout.

Yes, userspace must control that object, and that would require softpin
to preserve it across execbuffer calls. The kernel does not require that
all addresses be handled in userspace afterwards, that's the language I
wish to avoid. (Hence I don't like using "assumption" as that just
invites userspace to break the kernel.)
 
> >>So UMD will also have to allocate the L3/L2/L1 table pages
> >>as a regular GEM BO only & assign them a PPGTT address through the Soft Pin API.
> >>UMD would have to emit the MI_STORE_DATA_IMM commands in the batch buffer to
> >>program the relevant entries of L3/L2/L1 tables.
> >
> >This only applies to te TR-TT L1-L3 cache, right?
> >
> Yes applies only to the TR-TT L1-L3 tables.
> The backing pages of L3/L2/L1 tables shall be allocated as a BO,
> which should be assigned a PPGTT address.
> The table entries could be written directly also by UMD by mmapping
> the table BOs, but adding MI_STORE_DATA_IMM commands in the batch
> buffer itself would help to achieve serialization (implicitly).

Can you tighten up the phrasing here? My first read was that you indeed
for all PTE writes to be in userspace, which is scary.

"UMD will then allocate the L3/L32/L1 page tables for TR-TT as a regular
bo, and will use softpin to assign it to the l3_table_address when used.
UMD will also maintain the entries in the TR-TT page tables using
regular batch commands (MI_STORE_DATA_IMM), or via mmapping of the page
table bo."

> >>autonomously and KMD will be oblivious of it.
> >>The BOs must not be assigned an address from TR-TT segment, they will be mapped
> >
> >s/The BOs/Any object/
> >
> Ok will use 'Any object'
> >>to PPGTT in a regular way by KMD
> >
> >s/using the Soft Pin offset provided by UMD// as this is irrelevant.
> >
> You mean to say that it is needless or inappropriate to state that
> KMD will use the Soft PIN offset provided by UMD, it doesn't matter
> that whether the Soft PIN offset is used or KMD itself assigns an
> address.

I just want to avoid implying that userspace must use softpin on every
single bo for this to work. (Mainly because I don't really want
userspace to have to do full address space management, as we will always
have to do the double check inside the kernel. Unless there is a real
need (e.g. svm), I'd rather improve the kernel allocator/verification, rather
than try and circumvent it.)

> >>@@ -172,6 +172,9 @@ static int i915_getparam(struct drm_device *dev, void *data,
> >>  	case I915_PARAM_HAS_EXEC_SOFTPIN:
> >>  		value = 1;
> >>  		break;
> >>+	case I915_PARAM_HAS_TRTT:
> >>+		value = HAS_TRTT(dev);
> >>+		break;
> >
> >Should we do this here, or just query the context? In fact you are
> >missing the context getparam path any way.
> >
> Sorry, do you mean to say that with -ENODEV error also, on context
> setparam, User can make out the TR-TT support, so no need to have an
> explicit getparam case.
> 
> Would the context getparam path be really useful for TR-TT?.
> If its needed, then would be better to rename
> I915_CONTEXT_PARAM_ENABLE_TRTT to I915_CONTEXT_PARAM_TRTT_INFO ?

The question I have is do we want:

GETPARAM + CONTEXT_SETPARAM

or

CONTEXT_GETPARAM + CONTEXT_SETPARAM

the latter seems more symmetric and flexible, and we can use as a double
check later on that we set the right address etc.

Indeed, hindsight says ENABLE_TRTT is a bad name :)

I915_CONTEXT_PARAM_TRTT (let's assume for now  this will be the last,
any future PARAM_TRTT can think of a good name for its extension).

> >>  	default:
> >>  		DRM_DEBUG("Unknown parameter %d\n", param->param);
> >>  		return -EINVAL;
> >>diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> >>index c6dd4db..12c612e 100644
> >>--- a/drivers/gpu/drm/i915/i915_drv.h
> >>+++ b/drivers/gpu/drm/i915/i915_drv.h
> >>@@ -839,6 +839,7 @@ struct i915_ctx_hang_stats {
> >>  #define DEFAULT_CONTEXT_HANDLE 0
> >>
> >>  #define CONTEXT_NO_ZEROMAP (1<<0)
> >>+#define CONTEXT_USE_TRTT   (1<<1)
> >
> >Make flags unsigned whilst you are here, and fix the holes!
> >
> 
> Ok will change the type of 'flags' field inside 'intel_context' to unsigned.
> Sorry, but apart from this anything else required here ?

No, it's just belated anger about the silly "int flags".

> 
> >>  /**
> >>   * struct intel_context - as the name implies, represents a context.
> >>   * @ref: reference count.
> >>@@ -881,6 +882,15 @@ struct intel_context {
> >>  		int pin_count;
> >>  	} engine[I915_NUM_RINGS];
> >>
> >>+	/* TRTT info */
> >>+	struct {
> >
> >Give this a name now, we will be thankful in the future.
> >
> Would ctx_trtt_params be fine ?

struct intel_context_trtt

(Avoid using params for internals, let's keep those for uAPI - that
helps us distinguish pieces of code / context.)

> >>
> >>  void i915_gem_context_free(struct kref *ctx_ref)
> >>@@ -512,6 +515,35 @@ i915_gem_context_get(struct drm_i915_file_private *file_priv, u32 id)
> >>  	return ctx;
> >>  }
> >>
> >>+static int
> >>+i915_setup_trtt_ctx(struct intel_context *ctx,
> >>+		    struct drm_i915_gem_context_trtt_param *trtt_params)
> >>+{
> >>+	if (ctx->flags & CONTEXT_USE_TRTT)
> >>+		return -EEXIST;
> >>+
> >>+	/* basic sanity checks for the l3 table pointer */
> >>+	if ((ctx->trtt_info.l3_table_address >= GEN9_TRTT_SEGMENT_START) &&
> >>+	    (ctx->trtt_info.l3_table_address <
> >>+			(GEN9_TRTT_SEGMENT_START + GEN9_TRTT_SEGMENT_SIZE)))
> >
> >Presumably l3_table has an actual size and you want to do a range
> >overlap test, not just the start address.
> >
> Yes intend to do a range overlap test only. But since L3 table size
> is fixed as 4KB, thought there is no real need to also include the
> size in the range check, considering the allocations are always in
> multiple of 4KB.

Ok. You have a choice of writting that up as a comment, or just doing
the page overlap test :) Honestly, I would just go for the range test
since this is a one-off init path and the reader then doesn't even have
to think.

> >>+		return -EINVAL;
> >>+
> >>+	if (ctx->trtt_info.l3_table_address & ~GEN9_TRTT_L3_GFXADDR_MASK)
> >>+		return -EINVAL;
> >
> >These are worth adding DRM_DEBUG() or even better start using dev_debug()
> >so that we can debug userspace startup issues.
> >
> Fine, I think DRM_DEBUG_DRIVER will be more appropriate compared to
> DRM_DEBUG.

No, these are userspace errors for which we use DRM_DEBUG.
DRM_DEBUG_DRIVER is for a driver error :)

> Or
> 	dev_dbg(dev->dev, "invalid l3 table address\n");

Include as much info as you can (without giving away kernel internals),
since the user gave use the l3_address that we reject, report it. That
makes it easier to spot if it is the same address as they expected.

dev_dbg() would be my preference.

#define i915_dbg(DEV, args...) dev_dbg(__I915__(DEV)->dev->dev, ##args)
(not the prettiest yet, the pointer dancing is in the wrong direction!)

and let's get the ball rolling.

> >>  int i915_ppgtt_init_hw(struct drm_device *dev)
> >>  {
> >>+	if (HAS_TRTT(dev) && USES_FULL_48BIT_PPGTT(dev)) {
> >>+		struct drm_i915_private *dev_priv = dev->dev_private;
> >>+
> >>+		I915_WRITE(GEN9_TR_CHICKEN_BIT_VECTOR,
> >>+			   GEN9_TRTT_BYPASS_DISABLE);
> >
> >Shouldn't this be a context specific register? In which case you need to
> >set it in the context image instead.
> >
> >Hmm. given you already do the context image tweaks, how does work with
> >non-trtt contexts?
> >
> 
> GEN9_TR_CHICKEN_BIT_VECTOR is not a context specific register.
> It globally enables TR-TT support in Hw. Still TR-TT enabling on per
> context basis is required.
> Non-trtt contexts are not affected by this setting.

Please add that as a comment here. What are the downsides, potential
regressions? It's behind a chicken bit after all...

> Ok so need to define a new wrapper function,
> 	i915_vm_reserve_node(vm, START, SIZE, &vma).
> 
> After looking at the other callsites of drm_mm_reserve_node,
> including i915_vgpu, I think it would be better to have the
> prototype as,
>    i915_vm_reserve_node(vm, &node);
> 
> However this should be done as a separate patch ?

Yes, I was just recognising the code duplication and found 3 places
where we could use it - 3 being the magic number to refactor.
 
> >>+struct drm_i915_gem_context_trtt_param {
> >>+	__u64 l3_table_address;
> >>+	__u32 invd_tile_val;
> >>+	__u32 null_tile_val;
> >>+};
> >
> >Passes the ABI structure sanity checks.
> 
> Should we allow User to also choose the location of TR-TT segment
> (size is anyways fixed as 1<<44).

The kernel is much more agnostic with your approach than I anticipated,
so from our pov, allowing the user to shoot themselves in the foot is
ok.

There is only one sensible location, but that one location may be
sensible for a few things.

i.e. it shouldn't be below 1<<40 so that you can do full aliasing
between CPU and GPU addresses, and you want to avoid cutting your
address space in two, so it has to go at the ends, ergo it should be at
the very top.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre