[PATCH 2/2] drm/xe/guc: Add support for G2G communications

Daniele Ceraolo Spurio daniele.ceraolospurio at intel.com
Fri Nov 15 21:22:43 UTC 2024




On 11/8/2024 12:22 PM, John.C.Harrison at Intel.com wrote:
> From: John Harrison <John.C.Harrison at Intel.com>
>
> Some features require inter-GuC communication channels on multi-tile
> devices. So allocate and enable such.
>
> Signed-off-by: John Harrison <John.C.Harrison at Intel.com>
> ---
>   drivers/gpu/drm/xe/abi/guc_actions_abi.h |  20 ++
>   drivers/gpu/drm/xe/xe_guc.c              | 276 ++++++++++++++++++++++-
>   drivers/gpu/drm/xe/xe_guc_types.h        |  10 +
>   3 files changed, 305 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/xe/abi/guc_actions_abi.h b/drivers/gpu/drm/xe/abi/guc_actions_abi.h
> index b54fe40fc5a9..fee385532fb0 100644
> --- a/drivers/gpu/drm/xe/abi/guc_actions_abi.h
> +++ b/drivers/gpu/drm/xe/abi/guc_actions_abi.h
> @@ -134,6 +134,8 @@ enum xe_guc_action {
>   	XE_GUC_ACTION_DEREGISTER_CONTEXT = 0x4503,
>   	XE_GUC_ACTION_REGISTER_COMMAND_TRANSPORT_BUFFER = 0x4505,
>   	XE_GUC_ACTION_DEREGISTER_COMMAND_TRANSPORT_BUFFER = 0x4506,
> +	XE_GUC_ACTION_REGISTER_G2G = 0x4507,
> +	XE_GUC_ACTION_DEREGISTER_G2G = 0x4508,
>   	XE_GUC_ACTION_DEREGISTER_CONTEXT_DONE = 0x4600,
>   	XE_GUC_ACTION_REGISTER_CONTEXT_MULTI_LRC = 0x4601,
>   	XE_GUC_ACTION_CLIENT_SOFT_RESET = 0x5507,
> @@ -218,4 +220,22 @@ enum xe_guc_tlb_inval_mode {
>   	XE_GUC_TLB_INVAL_MODE_LITE = 0x1,
>   };
>   
> +/*
> + * GuC to GuC communication (de-)registration fields:
> + */
> +enum xe_guc_g2g_type {
> +	XE_G2G_TYPE_IN = 0x0,
> +	XE_G2G_TYPE_OUT,
> +	XE_G2G_TYPE_LIMIT,
> +};
> +
> +#define XE_G2G_REGISTER_DEVICE	REG_GENMASK(16, 16)

I'd call this DEVICE_ID like in the GuC specs, or GT_ID if you want a 
shorter define. Just DEVICE sounds like it's going to a different card 
instead of a different GuC device. Not a blocker.

> +#define XE_G2G_REGISTER_TILE	REG_GENMASK(15, 12)
> +#define XE_G2G_REGISTER_TYPE	REG_GENMASK(11, 8)
> +#define XE_G2G_REGISTER_SIZE	REG_GENMASK(7, 0)
> +
> +#define XE_G2G_DEREGISTER_DEVICE	REG_GENMASK(16, 16)
> +#define XE_G2G_DEREGISTER_TILE	REG_GENMASK(15, 12)
> +#define XE_G2G_DEREGISTER_TYPE	REG_GENMASK(11, 8)
> +
>   #endif
> diff --git a/drivers/gpu/drm/xe/xe_guc.c b/drivers/gpu/drm/xe/xe_guc.c
> index 1cbad8a91798..899094423291 100644
> --- a/drivers/gpu/drm/xe/xe_guc.c
> +++ b/drivers/gpu/drm/xe/xe_guc.c
> @@ -44,7 +44,15 @@ static u32 guc_bo_ggtt_addr(struct xe_guc *guc,
>   			    struct xe_bo *bo)
>   {
>   	struct xe_device *xe = guc_to_xe(guc);
> -	u32 addr = xe_bo_ggtt_addr(bo);
> +	u32 addr;
> +
> +	/*
> +	 * For most BOs, the address on the allocating tile is fine. However for
> +	 * some, e.g. G2G CTB, the address on a specific tile is required as it
> +	 * might be different for each tile. So, just always ask for the address
> +	 * on the target GuC.
> +	 */
> +	addr = __xe_bo_ggtt_addr(bo, gt_to_tile(guc_to_gt(guc))->id);
>   
>   	/* GuC addresses above GUC_GGTT_TOP don't map through the GTT */
>   	xe_assert(xe, addr >= xe_wopcm_size(guc_to_xe(guc)));
> @@ -244,6 +252,263 @@ static void guc_write_params(struct xe_guc *guc)
>   		xe_mmio_write32(&gt->mmio, SOFT_SCRATCH(1 + i), guc->params[i]);
>   }
>   
> +static int guc_action_register_g2g_buffer(struct xe_guc *guc, u32 type, u32 dst_tile, u32 dst_dev,
> +					  u32 desc_addr, u32 buff_addr, u32 size)
> +{
> +	struct xe_gt *gt = guc_to_gt(guc);
> +	struct xe_device *xe = gt_to_xe(gt);
> +	u32 action[] = {
> +		XE_GUC_ACTION_REGISTER_G2G,
> +		FIELD_PREP(XE_G2G_REGISTER_SIZE, size / SZ_4K - 1) |

This needs a comment somewhere to explain that the GuC expects the 
number of pages minus 1, because otherwise that -1 is confusing.

> +		FIELD_PREP(XE_G2G_REGISTER_TYPE, type) |
> +		FIELD_PREP(XE_G2G_REGISTER_TILE, dst_tile) |
> +		FIELD_PREP(XE_G2G_REGISTER_DEVICE, dst_dev),
> +		desc_addr,
> +		buff_addr,
> +	};
> +
> +	xe_assert(xe, (type == XE_G2G_TYPE_IN) || (type == XE_G2G_TYPE_OUT));
> +	xe_assert(xe, !(size % SZ_4K));

Is it worth having asserts for max size (1MB) and for dst_dev (0 or 1)?

> +
> +	return xe_guc_ct_send_block(&guc->ct, action, ARRAY_SIZE(action));
> +}
> +
> +static int guc_action_deregister_g2g_buffer(struct xe_guc *guc, u32 type, u32 dst_tile, u32 dst_dev)
> +{
> +	struct xe_gt *gt = guc_to_gt(guc);
> +	struct xe_device *xe = gt_to_xe(gt);
> +	u32 action[] = {
> +		XE_GUC_ACTION_DEREGISTER_G2G,
> +		FIELD_PREP(XE_G2G_DEREGISTER_TYPE, type) |
> +		FIELD_PREP(XE_G2G_DEREGISTER_TILE, dst_tile) |
> +		FIELD_PREP(XE_G2G_DEREGISTER_DEVICE, dst_dev),
> +	};
> +
> +	xe_assert(xe, (type == XE_G2G_TYPE_IN) || (type == XE_G2G_TYPE_OUT));
> +
> +	return xe_guc_ct_send_block(&guc->ct, action, ARRAY_SIZE(action));
> +}
> +
> +#define G2G_DEV(gt)	(((gt)->info.type == XE_GT_TYPE_MAIN) ? 0 : 1)
> +
> +#define G2G_BUFFER_SIZE (SZ_4K)
> +#define G2G_DESC_SIZE (64)
> +#define G2G_DESC_AREA_SIZE (SZ_4K)
> +
> +/*
> + * Generate a unique id for each bi-directional CTB for each pair of
> + * near and far tiles/devices. The id can then be used as an index into
> + * a single allocation that is sub-divided into multiple CTBs.
> + *
> + * For example, with two devices per tile and two tiles, the table should
> + * look like:
> + *           Far <tile>.<dev>
> + *         0.0   0.1   1.0   1.1
> + * N 0.0  --/-- 00/01 02/03 04/05
> + * e 0.1  01/00 --/-- 06/07 08/09
> + * a 1.0  03/02 07/06 --/-- 10/11
> + * r 1.1  05/04 09/08 11/10 --/--
> + *
> + * Where each entry is Rx/Tx channel id.
> + *
> + * So GuC #3 (tile 1, dev 1) talking to GuC #2 (tile 1, dev 0) would
> + * be reading from channel #11 and writing to channel #10. Whereas,
> + * GuC #2 talking to GuC #3 would be read on #10 and write to #11.
> + */
> +static unsigned int g2g_slot(u32 near_tile, u32 near_dev, u32 far_tile, u32 far_dev,
> +			     u32 type, u32 max_inst, bool have_dev)
> +{
> +	u32 near = near_tile, far = far_tile;
> +	u32 idx = 0;
> +	int i;
> +
> +	if (have_dev) {
> +		near = (near << 1) | near_dev;
> +		far = (far << 1) | far_dev;
> +	}
> +
> +	if (far > near) {
> +		for (i = near; i > 0; i--)
> +			idx += max_inst - i;

The loop can be replaced by a single line with:

idx += near * (2 * max_inst  - near - 1) / 2

But the loop is probably easier to understand.
(I've spent way too much time trying to find a better formula with 
little luck :P)

> +		idx += (far - 1 - near);
> +		idx *= 2;
> +		idx += type;
> +		return idx;
> +	}

This needs some comments, because otherwise it's very hard to parse. 
Maybe stick with to the table example? something like

/* Count all BO pairs from the rows above the one where the pair we want 
is */
for (i = near; i > 0; i--)
         idx += max_inst - i;

/* Count all BO pairs on the same row and to the left of the pair we want */
idx += (far - 1 - near);

/* Switch from BO pairs to individual BOs */
idx *= 2;

/* Select which of the 2 BOs in the pair we want */
idx += type;
     return idx;

Same thing for the other one, just swap rows with columns and left with 
above.

I'm also ok with a different comments as along as there are some.

> +
> +	if (far < near) {
> +		for (i = far; i > 0; i--)
> +			idx += max_inst - i;
> +		idx += (near - 1 - far);
> +		idx *= 2;
> +		idx += (1 - type);
> +		return idx;
> +	}
> +
> +	return -1;
> +}
> +
> +static int guc_g2g_register(struct xe_guc *near_guc, struct xe_gt *far_gt, u32 type, bool have_dev)

you don't really need the have_dev passed in from outside, you can just 
check for tile->media_gt being not null.

> +{
> +	struct xe_gt *near_gt = guc_to_gt(near_guc);
> +	struct xe_device *xe = gt_to_xe(near_gt);
> +	struct xe_bo *g2g_bo;
> +	u32 near_tile = gt_to_tile(near_gt)->id;
> +	u32 near_dev = G2G_DEV(near_gt);
> +	u32 far_tile = gt_to_tile(far_gt)->id;
> +	u32 far_dev = G2G_DEV(far_gt);
> +	u32 max = xe->info.gt_count;
> +	u32 base, desc, buf;
> +	int slot;
> +
> +	xe_assert(xe, xe == gt_to_xe(far_gt));
> +
> +	g2g_bo = near_guc->g2g.bo;
> +	xe_assert(xe, g2g_bo);
> +
> +	slot = g2g_slot(near_tile, near_dev, far_tile, far_dev, type, max, have_dev);
> +	xe_assert(xe, slot >= 0);
> +
> +	base = guc_bo_ggtt_addr(near_guc, g2g_bo);
> +	desc = base + slot * G2G_DESC_SIZE;
> +	buf = base + slot * G2G_BUFFER_SIZE + G2G_DESC_AREA_SIZE;

nit: I'd reorder this to have G2G_DESC_AREA_SIZE before the slot, so 
it's in order to how the memory is actually filled.

> +
> +	xe_assert(xe, (desc - base + G2G_DESC_SIZE) <= G2G_DESC_AREA_SIZE);

nit: We should be able to assert this only once at alloc time (see below).

> +	xe_assert(xe, (buf - base + G2G_BUFFER_SIZE) <= g2g_bo->size);
> +
> +	return guc_action_register_g2g_buffer(near_guc, type, far_tile, far_dev,
> +					      desc, buf, G2G_BUFFER_SIZE);
> +}
> +
> +static void guc_g2g_deregister(struct xe_guc *guc, u32 far_tile, u32 far_dev, u32 type)
> +{
> +	guc_action_deregister_g2g_buffer(guc, type, far_tile, far_dev);
> +}
> +
> +static u32 guc_g2g_size(struct xe_guc *guc)
> +{
> +	struct xe_gt *gt = guc_to_gt(guc);
> +	struct xe_device *xe = gt_to_xe(gt);
> +	unsigned int count = xe->info.gt_count;
> +	u32 num_channels = (count * (count - 1)) / 2;
> +
> +	return num_channels * XE_G2G_TYPE_LIMIT * G2G_BUFFER_SIZE + G2G_DESC_AREA_SIZE;

Here we can assert:
num_channels * XE_G2G_TYPE_LIMIT * G2G_DESC_SIZE <= G2G_DESC_AREA_SIZE

to cover all the descriptors in one go

> +}
> +
> +static bool xe_guc_g2g_wanted(struct xe_device *xe)
> +{
> +	/* Can't do GuC to GuC communication if there is only one GuC */
> +	if (xe->info.gt_count <= 1)
> +		return false;
> +
> +	/* No current user */
> +	return false;
> +}
> +
> +static int guc_g2g_alloc(struct xe_guc *guc)
> +{
> +	struct xe_gt *gt = guc_to_gt(guc);
> +	struct xe_device *xe = gt_to_xe(gt);
> +	struct xe_tile *tile = gt_to_tile(gt);
> +	struct xe_bo *bo;
> +	u32 g2g_size;
> +
> +	if (guc->g2g.bo)
> +		return 0;
> +
> +	if (gt->info.id != 0) {
> +		struct xe_gt *root_gt = xe_device_get_gt(xe, 0);
> +		struct xe_guc *root_guc = &root_gt->uc.guc;
> +		struct xe_bo *bo;
> +
> +		bo = xe_bo_get(root_guc->g2g.bo);
> +		if (IS_ERR(bo))
> +			return PTR_ERR(bo);
> +
> +		guc->g2g.bo = bo;
> +		guc->g2g.owned = false;

It doesn't look like you're actually using the "owned" variable (which 
makes sense, because the BO is refcounted so it doesn't matter which GuC 
it was originally allocated for)

Daniele

> +		return 0;
> +	}
> +
> +	g2g_size = guc_g2g_size(guc);
> +	bo = xe_managed_bo_create_pin_map(xe, tile, g2g_size,
> +					  XE_BO_FLAG_VRAM_IF_DGFX(tile) |
> +					  XE_BO_FLAG_GGTT |
> +					  XE_BO_FLAG_GGTT_ALL |
> +					  XE_BO_FLAG_GGTT_INVALIDATE);
> +	if (IS_ERR(bo))
> +		return PTR_ERR(bo);
> +
> +	xe_map_memset(xe, &bo->vmap, 0, 0, g2g_size);
> +	guc->g2g.bo = bo;
> +	guc->g2g.owned = true;
> +
> +	return 0;
> +}
> +
> +static int guc_g2g_start(struct xe_guc *guc)
> +{
> +	struct xe_gt *far_gt, *gt = guc_to_gt(guc);
> +	struct xe_device *xe = gt_to_xe(gt);
> +	unsigned int i, j;
> +	int t, err;
> +	bool have_dev;
> +
> +	if (!guc->g2g.bo) {
> +		int ret;
> +
> +		ret = guc_g2g_alloc(guc);
> +		if (ret)
> +			return ret;
> +	}
> +
> +	/* GuC interface will need extending if more GT device types are ever created. */
> +	xe_gt_assert(gt, (gt->info.type == XE_GT_TYPE_MAIN) || (gt->info.type == XE_GT_TYPE_MEDIA));
> +
> +	/* Channel numbering depends on whether there are multiple GTs per tile */
> +	have_dev = xe->info.gt_count > xe->info.tile_count;
> +
> +	for_each_gt(far_gt, xe, i) {
> +		u32 far_tile, far_dev;
> +
> +		if (far_gt->info.id == gt->info.id)
> +			continue;
> +
> +		far_tile = gt_to_tile(far_gt)->id;
> +		far_dev = G2G_DEV(far_gt);
> +
> +		for (t = 0; t < XE_G2G_TYPE_LIMIT; t++) {
> +			err = guc_g2g_register(guc, far_gt, t, have_dev);
> +			if (err) {
> +				while (--t >= 0)
> +					guc_g2g_deregister(guc, far_tile, far_dev, t);
> +				goto err_deregister;
> +			}
> +		}
> +	}
> +
> +	return 0;
> +
> +err_deregister:
> +	for_each_gt(far_gt, xe, j) {
> +		u32 tile, dev;
> +
> +		if (far_gt->info.id == gt->info.id)
> +			continue;
> +
> +		if (j >= i)
> +			break;
> +
> +		tile = gt_to_tile(far_gt)->id;
> +		dev = G2G_DEV(far_gt);
> +
> +		for (t = 0; t < XE_G2G_TYPE_LIMIT; t++)
> +			guc_g2g_deregister(guc, tile, dev, t);
> +	}
> +
> +	return err;
> +}
> +
>   static void guc_fini_hw(void *arg)
>   {
>   	struct xe_guc *guc = arg;
> @@ -423,7 +688,16 @@ int xe_guc_init_post_hwconfig(struct xe_guc *guc)
>   
>   int xe_guc_post_load_init(struct xe_guc *guc)
>   {
> +	int ret;
> +
>   	xe_guc_ads_populate_post_load(&guc->ads);
> +
> +	if (xe_guc_g2g_wanted(guc_to_xe(guc))) {
> +		ret = guc_g2g_start(guc);
> +		if (ret)
> +			return ret;
> +	}
> +
>   	guc->submission_state.enabled = true;
>   
>   	return 0;
> diff --git a/drivers/gpu/drm/xe/xe_guc_types.h b/drivers/gpu/drm/xe/xe_guc_types.h
> index fa75f57bf5da..83a41ebcdc91 100644
> --- a/drivers/gpu/drm/xe/xe_guc_types.h
> +++ b/drivers/gpu/drm/xe/xe_guc_types.h
> @@ -64,6 +64,15 @@ struct xe_guc {
>   	struct xe_guc_pc pc;
>   	/** @dbm: GuC Doorbell Manager */
>   	struct xe_guc_db_mgr dbm;
> +
> +	/** @g2g: GuC to GuC communication state */
> +	struct {
> +		/** @g2g.bo: Storage for GuC to GuC communication channels */
> +		struct xe_bo *bo;
> +		/** @g2g.owned: Is the BO owned by this GT or just mapped in */
> +		bool owned;
> +	} g2g;
> +
>   	/** @submission_state: GuC submission state */
>   	struct {
>   		/** @submission_state.idm: GuC context ID Manager */
> @@ -79,6 +88,7 @@ struct xe_guc {
>   		/** @submission_state.fini_wq: submit fini wait queue */
>   		wait_queue_head_t fini_wq;
>   	} submission_state;
> +
>   	/** @hwconfig: Hardware config state */
>   	struct {
>   		/** @hwconfig.bo: buffer object of the hardware config */



More information about the Intel-xe mailing list