[PATCH v3 1/2] drm/xe/vf: Refactor CCS save/restore to use default migration context
Summers, Stuart
stuart.summers at intel.com
Wed Aug 6 16:08:01 UTC 2025
On Wed, 2025-08-06 at 13:59 +0530, Satyanarayana K V P wrote:
> Previously, CCS save/restore operations created separate migration
> contexts with new VM memory allocations, resulting in significant
> overhead.
>
> This commit eliminates redundant context creation reusing the default
> migration context by registering new execution queues for CCS save
> and
> restore on the existing migrate VM.
>
> Signed-off-by: Satyanarayana K V P <satyanarayana.k.v.p at intel.com>
> Suggested-by: Matthew Brost <matthew.brost at intel.com>
> Cc: Michal Wajdeczko <michal.wajdeczko at intel.com>
>
> ---
> V2 -> V3:
> - Fixed review comments (Matthew Brost).
>
> V1 -> V2:
> - Fixed kernel-doc issues reported by patchworks.
> ---
> drivers/gpu/drm/xe/xe_migrate.c | 18 ++++----
> drivers/gpu/drm/xe/xe_migrate.h | 2 +-
> drivers/gpu/drm/xe/xe_pm.c | 3 ++
> drivers/gpu/drm/xe/xe_sriov_vf_ccs.c | 49 ++++++++++++--------
> --
> drivers/gpu/drm/xe/xe_sriov_vf_ccs_types.h | 6 +--
> 5 files changed, 43 insertions(+), 35 deletions(-)
>
> diff --git a/drivers/gpu/drm/xe/xe_migrate.c
> b/drivers/gpu/drm/xe/xe_migrate.c
> index 3a276e2348a2..7d3c7c4e2185 100644
> --- a/drivers/gpu/drm/xe/xe_migrate.c
> +++ b/drivers/gpu/drm/xe/xe_migrate.c
> @@ -960,14 +960,14 @@ struct xe_lrc *xe_migrate_lrc(struct xe_migrate
> *migrate)
> return migrate->q->lrc[0];
> }
>
> -static int emit_flush_invalidate(struct xe_migrate *m, u32 *dw, int
> i,
> +static int emit_flush_invalidate(struct xe_exec_queue *q, u32 *dw,
> int i,
> u32 flags)
> {
> dw[i++] = MI_FLUSH_DW | MI_INVALIDATE_TLB |
> MI_FLUSH_DW_OP_STOREDW |
> MI_FLUSH_IMM_DW | flags;
> - dw[i++] =
> lower_32_bits(xe_lrc_start_seqno_ggtt_addr(xe_migrate_lrc(m))) |
> + dw[i++] = lower_32_bits(xe_lrc_start_seqno_ggtt_addr(q-
> >lrc[0])) |
> MI_FLUSH_DW_USE_GTT;
> - dw[i++] =
> upper_32_bits(xe_lrc_start_seqno_ggtt_addr(xe_migrate_lrc(m)));
> + dw[i++] = upper_32_bits(xe_lrc_start_seqno_ggtt_addr(q-
> >lrc[0]));
> dw[i++] = MI_NOOP;
> dw[i++] = MI_NOOP;
>
> @@ -976,7 +976,8 @@ static int emit_flush_invalidate(struct
> xe_migrate *m, u32 *dw, int i,
>
> /**
> * xe_migrate_ccs_rw_copy() - Copy content of TTM resources.
> - * @m: The migration context.
> + * @tile: Tile whose migration context to be used.
> + * @q : Execution to be used along with migration context.
> * @src_bo: The buffer object @src is currently bound to.
> * @read_write : Creates BB commands for CCS read/write.
> *
> @@ -987,7 +988,7 @@ static int emit_flush_invalidate(struct
> xe_migrate *m, u32 *dw, int i,
> *
> * Return: 0 if successful, negative error code on failure.
> */
> -int xe_migrate_ccs_rw_copy(struct xe_migrate *m,
> +int xe_migrate_ccs_rw_copy(struct xe_tile *tile, struct
> xe_exec_queue *q,
> struct xe_bo *src_bo,
> enum xe_sriov_vf_ccs_rw_ctxs read_write)
>
> @@ -995,7 +996,8 @@ int xe_migrate_ccs_rw_copy(struct xe_migrate *m,
> bool src_is_pltt = read_write == XE_SRIOV_VF_CCS_READ_CTX;
> bool dst_is_pltt = read_write == XE_SRIOV_VF_CCS_WRITE_CTX;
> struct ttm_resource *src = src_bo->ttm.resource;
> - struct xe_gt *gt = m->tile->primary_gt;
> + struct xe_migrate *m = tile->migrate;
> + struct xe_gt *gt = tile->primary_gt;
> u32 batch_size, batch_size_allocated;
> struct xe_device *xe = gt_to_xe(gt);
> struct xe_res_cursor src_it, ccs_it;
> @@ -1078,11 +1080,11 @@ int xe_migrate_ccs_rw_copy(struct xe_migrate
> *m,
>
> emit_pte(m, bb, ccs_pt, false, false, &ccs_it,
> ccs_size, src);
>
> - bb->len = emit_flush_invalidate(m, bb->cs, bb->len,
> flush_flags);
> + bb->len = emit_flush_invalidate(q, bb->cs, bb->len,
> flush_flags);
> flush_flags = xe_migrate_ccs_copy(m, bb, src_L0_ofs,
> src_is_pltt,
> src_L0_ofs,
> dst_is_pltt,
> src_L0, ccs_ofs,
> true);
> - bb->len = emit_flush_invalidate(m, bb->cs, bb->len,
> flush_flags);
> + bb->len = emit_flush_invalidate(q, bb->cs, bb->len,
> flush_flags);
>
> size -= src_L0;
> }
> diff --git a/drivers/gpu/drm/xe/xe_migrate.h
> b/drivers/gpu/drm/xe/xe_migrate.h
> index e81ea6b27fb5..9e20da6d58c2 100644
> --- a/drivers/gpu/drm/xe/xe_migrate.h
> +++ b/drivers/gpu/drm/xe/xe_migrate.h
> @@ -124,7 +124,7 @@ struct dma_fence *xe_migrate_copy(struct
> xe_migrate *m,
> struct ttm_resource *dst,
> bool copy_only_ccs);
>
> -int xe_migrate_ccs_rw_copy(struct xe_migrate *m,
> +int xe_migrate_ccs_rw_copy(struct xe_tile *tile, struct
> xe_exec_queue *q,
> struct xe_bo *src_bo,
> enum xe_sriov_vf_ccs_rw_ctxs read_write);
>
> diff --git a/drivers/gpu/drm/xe/xe_pm.c b/drivers/gpu/drm/xe/xe_pm.c
> index 44aaf154ddf7..5e8126ca8e27 100644
> --- a/drivers/gpu/drm/xe/xe_pm.c
> +++ b/drivers/gpu/drm/xe/xe_pm.c
> @@ -209,6 +209,9 @@ int xe_pm_resume(struct xe_device *xe)
>
> xe_pxp_pm_resume(xe->pxp);
>
> + if (IS_SRIOV_VF(xe))
> + xe_sriov_vf_ccs_register_context(xe);
> +
> drm_dbg(&xe->drm, "Device resumed\n");
> return 0;
> err:
> diff --git a/drivers/gpu/drm/xe/xe_sriov_vf_ccs.c
> b/drivers/gpu/drm/xe/xe_sriov_vf_ccs.c
> index f0ca2a9b2bb7..a87f39eae4dc 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_vf_ccs.c
> +++ b/drivers/gpu/drm/xe/xe_sriov_vf_ccs.c
> @@ -8,6 +8,7 @@
> #include "xe_bb.h"
> #include "xe_bo.h"
> #include "xe_device.h"
> +#include "xe_exec_queue.h"
> #include "xe_exec_queue_types.h"
> #include "xe_guc_submit.h"
> #include "xe_lrc.h"
> @@ -168,7 +169,7 @@ static int alloc_bb_pool(struct xe_tile *tile,
> struct xe_tile_vf_ccs *ctx)
>
> static void ccs_rw_update_ring(struct xe_tile_vf_ccs *ctx)
> {
> - struct xe_lrc *lrc = xe_migrate_lrc(ctx->migrate);
> + struct xe_lrc *lrc = ctx->q->lrc[0];
> u64 addr = xe_sa_manager_gpu_addr(ctx->mem.ccs_bb_pool);
> u32 dw[10], i = 0;
>
> @@ -183,13 +184,12 @@ static void ccs_rw_update_ring(struct
> xe_tile_vf_ccs *ctx)
> xe_lrc_set_ring_tail(lrc, lrc->ring.tail);
> }
>
> -static int register_save_restore_context(struct xe_migrate *m,
> - enum xe_sriov_vf_ccs_rw_ctxs
> ctx_id)
> +static int register_save_restore_context(struct xe_tile_vf_ccs *ctx)
> {
> int err = -EINVAL;
> int ctx_type;
>
> - switch (ctx_id) {
> + switch (ctx->ctx_id) {
> case XE_SRIOV_VF_CCS_READ_CTX:
> ctx_type = GUC_CONTEXT_COMPRESSION_SAVE;
> break;
> @@ -200,7 +200,7 @@ static int register_save_restore_context(struct
> xe_migrate *m,
> return err;
> }
>
> - xe_guc_register_exec_queue(xe_migrate_exec_queue(m),
> ctx_type);
> + xe_guc_register_exec_queue(ctx->q, ctx_type);
> return 0;
> }
>
> @@ -225,7 +225,7 @@ int xe_sriov_vf_ccs_register_context(struct
> xe_device *xe)
>
> for_each_ccs_rw_ctx(ctx_id) {
> ctx = &tile->sriov.vf.ccs[ctx_id];
> - err = register_save_restore_context(ctx->migrate,
> ctx_id);
> + err = register_save_restore_context(ctx);
> if (err)
> return err;
> }
> @@ -236,13 +236,14 @@ int xe_sriov_vf_ccs_register_context(struct
> xe_device *xe)
> static void xe_sriov_vf_ccs_fini(void *arg)
> {
> struct xe_tile_vf_ccs *ctx = arg;
> - struct xe_lrc *lrc = xe_migrate_lrc(ctx->migrate);
> + struct xe_lrc *lrc = ctx->q->lrc[0];
>
> /*
> * Make TAIL = HEAD in the ring so that no issues are seen if
> Guc
> * submits this context to HW on VF pause after unbinding
> device.
> */
> xe_lrc_set_ring_tail(lrc, xe_lrc_ring_head(lrc));
> + xe_exec_queue_put(ctx->q);
> }
>
> /**
> @@ -258,8 +259,9 @@ int xe_sriov_vf_ccs_init(struct xe_device *xe)
> {
> struct xe_tile *tile = xe_device_get_root_tile(xe);
> enum xe_sriov_vf_ccs_rw_ctxs ctx_id;
> - struct xe_migrate *migrate;
> struct xe_tile_vf_ccs *ctx;
> + struct xe_exec_queue *q;
> + u32 flags;
> int err;
>
> xe_assert(xe, IS_SRIOV_VF(xe));
> @@ -270,27 +272,25 @@ int xe_sriov_vf_ccs_init(struct xe_device *xe)
> ctx = &tile->sriov.vf.ccs[ctx_id];
> ctx->ctx_id = ctx_id;
>
> - migrate = xe_migrate_alloc(tile);
> - if (!migrate) {
> - err = -ENOMEM;
> + flags = EXEC_QUEUE_FLAG_KERNEL |
> + EXEC_QUEUE_FLAG_PERMANENT |
> + EXEC_QUEUE_FLAG_MIGRATE;
When USM is supported, we also set EXEC_QUEUE_FLAG_HIGH_PRIORITY for
the migration context. Is there a reason not to do that here?
Otherwise the change makes sense to me. We still have a unique queue
here, just not the VM creation per VF which I agree seems like a good
idea.
Thanks,
Stuart
> + q = xe_exec_queue_create_bind(xe, tile, flags, 0);
> + if (IS_ERR(q)) {
> + err = PTR_ERR(q);
> goto err_ret;
> }
> -
> - err = xe_migrate_init(migrate);
> - if (err)
> - goto err_ret;
> -
> - ctx->migrate = migrate;
> + ctx->q = q;
>
> err = alloc_bb_pool(tile, ctx);
> if (err)
> - goto err_ret;
> + goto err_free_queue;
>
> ccs_rw_update_ring(ctx);
>
> - err = register_save_restore_context(ctx->migrate,
> ctx_id);
> + err = register_save_restore_context(ctx);
> if (err)
> - goto err_ret;
> + goto err_free_queue;
>
> err = devm_add_action_or_reset(xe->drm.dev,
> xe_sriov_vf_ccs_fini,
> @@ -301,6 +301,9 @@ int xe_sriov_vf_ccs_init(struct xe_device *xe)
>
> return 0;
>
> +err_free_queue:
> + xe_exec_queue_put(q);
> +
> err_ret:
> return err;
> }
> @@ -319,7 +322,7 @@ int xe_sriov_vf_ccs_attach_bo(struct xe_bo *bo)
> {
> struct xe_device *xe = xe_bo_device(bo);
> enum xe_sriov_vf_ccs_rw_ctxs ctx_id;
> - struct xe_migrate *migrate;
> + struct xe_tile_vf_ccs *ctx;
> struct xe_tile *tile;
> struct xe_bb *bb;
> int err = 0;
> @@ -334,8 +337,8 @@ int xe_sriov_vf_ccs_attach_bo(struct xe_bo *bo)
> /* bb should be NULL here. Assert if not NULL */
> xe_assert(xe, !bb);
>
> - migrate = tile->sriov.vf.ccs[ctx_id].migrate;
> - err = xe_migrate_ccs_rw_copy(migrate, bo, ctx_id);
> + ctx = &tile->sriov.vf.ccs[ctx_id];
> + err = xe_migrate_ccs_rw_copy(tile, ctx->q, bo,
> ctx_id);
> }
> return err;
> }
> diff --git a/drivers/gpu/drm/xe/xe_sriov_vf_ccs_types.h
> b/drivers/gpu/drm/xe/xe_sriov_vf_ccs_types.h
> index e240f3fd18af..1add0541aed8 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_vf_ccs_types.h
> +++ b/drivers/gpu/drm/xe/xe_sriov_vf_ccs_types.h
> @@ -41,11 +41,11 @@ struct xe_sa_manager;
> struct xe_tile_vf_ccs {
> /** @id: Id to which context it belongs to */
> enum xe_sriov_vf_ccs_rw_ctxs ctx_id;
> - /** @migrate: Migration helper for save/restore of CCS data
> */
> - struct xe_migrate *migrate;
> + /** @q: exec queues used for migration */
> + struct xe_exec_queue *q;
>
> struct {
> - /** @ccs_rw_bb_pool: Pool from which batch buffers
> are allocated. */
> + /** @ccs_bb_pool: Pool from which batch buffers are
> allocated. */
> struct xe_sa_manager *ccs_bb_pool;
> } mem;
> };
More information about the Intel-xe
mailing list