drm/xe/pf: Handle VRAM migration data as part of PF control

Connect the helpers to allow save and restore of VRAM migration data in
stop_copy / resume device state.

Co-developed-by: Lukasz Laguna <lukasz.laguna@intel.com>
Signed-off-by: Lukasz Laguna <lukasz.laguna@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Link: https://patch.msgid.link/20251112132220.516975-23-michal.winiarski@intel.com
Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>

Author: mwiniars

drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c

@@ -892,6 +892,20 @@ static int pf_handle_vf_save_data(struct xe_gt *gt, unsigned int vfid)
 		return -EAGAIN;
 	}
 
+	if (xe_gt_sriov_pf_migration_save_data_pending(gt, vfid,
+						       XE_SRIOV_PACKET_TYPE_VRAM)) {
+		ret = xe_gt_sriov_pf_migration_vram_save(gt, vfid);
+		if (ret == -EAGAIN)
+			return -EAGAIN;
+		else if (ret)
+			return ret;
+
+		xe_gt_sriov_pf_migration_save_data_complete(gt, vfid,
+							    XE_SRIOV_PACKET_TYPE_VRAM);
+
+		return -EAGAIN;
+	}
+
 	return 0;
 }
 
@@ -1130,6 +1144,9 @@ static int pf_handle_vf_restore_data(struct xe_gt *gt, unsigned int vfid)
 	case XE_SRIOV_PACKET_TYPE_GUC:
 		ret = xe_gt_sriov_pf_migration_guc_restore(gt, vfid, data);
 		break;
+	case XE_SRIOV_PACKET_TYPE_VRAM:
+		ret = xe_gt_sriov_pf_migration_vram_restore(gt, vfid, data);
+		break;
 	default:
 		xe_gt_sriov_notice(gt, "Skipping VF%u unknown data type: %d\n",
 				   vfid, data->hdr.type);

drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c

@@ -19,6 +19,7 @@
 #include "xe_gt_sriov_printk.h"
 #include "xe_guc_buf.h"
 #include "xe_guc_ct.h"
+#include "xe_migrate.h"
 #include "xe_mmio.h"
 #include "xe_sriov.h"
 #include "xe_sriov_packet.h"
@@ -505,6 +506,206 @@ int xe_gt_sriov_pf_migration_mmio_restore(struct xe_gt *gt, unsigned int vfid,
 	return pf_restore_vf_mmio_mig_data(gt, vfid, data);
 }
 
+static ssize_t pf_migration_vram_size(struct xe_gt *gt, unsigned int vfid)
+{
+	if (!xe_gt_is_main_type(gt))
+		return 0;
+
+	return xe_gt_sriov_pf_config_get_lmem(gt, vfid);
+}
+
+static struct dma_fence *__pf_save_restore_vram(struct xe_gt *gt, unsigned int vfid,
+						struct xe_bo *vram, u64 vram_offset,
+						struct xe_bo *sysmem, u64 sysmem_offset,
+						size_t size, bool save)
+{
+	struct dma_fence *ret = NULL;
+	struct drm_exec exec;
+	int err;
+
+	drm_exec_init(&exec, 0, 0);
+	drm_exec_until_all_locked(&exec) {
+		err = drm_exec_lock_obj(&exec, &vram->ttm.base);
+		drm_exec_retry_on_contention(&exec);
+		if (err) {
+			ret = ERR_PTR(err);
+			goto err;
+		}
+
+		err = drm_exec_lock_obj(&exec, &sysmem->ttm.base);
+		drm_exec_retry_on_contention(&exec);
+		if (err) {
+			ret = ERR_PTR(err);
+			goto err;
+		}
+	}
+
+	ret = xe_migrate_vram_copy_chunk(vram, vram_offset, sysmem, sysmem_offset, size,
+					 save ? XE_MIGRATE_COPY_TO_SRAM : XE_MIGRATE_COPY_TO_VRAM);
+
+err:
+	drm_exec_fini(&exec);
+
+	return ret;
+}
+
+#define PF_VRAM_SAVE_RESTORE_TIMEOUT (5 * HZ)
+static int pf_save_vram_chunk(struct xe_gt *gt, unsigned int vfid,
+			      struct xe_bo *src_vram, u64 src_vram_offset,
+			      size_t size)
+{
+	struct xe_sriov_packet *data;
+	struct dma_fence *fence;
+	int ret;
+
+	data = xe_sriov_packet_alloc(gt_to_xe(gt));
+	if (!data)
+		return -ENOMEM;
+
+	ret = xe_sriov_packet_init(data, gt->tile->id, gt->info.id,
+				   XE_SRIOV_PACKET_TYPE_VRAM, src_vram_offset,
+				   size);
+	if (ret)
+		goto fail;
+
+	fence = __pf_save_restore_vram(gt, vfid,
+				       src_vram, src_vram_offset,
+				       data->bo, 0, size, true);
+
+	ret = dma_fence_wait_timeout(fence, false, PF_VRAM_SAVE_RESTORE_TIMEOUT);
+	dma_fence_put(fence);
+	if (!ret) {
+		ret = -ETIME;
+		goto fail;
+	}
+
+	pf_dump_mig_data(gt, vfid, data, "VRAM data save");
+
+	ret = xe_gt_sriov_pf_migration_save_produce(gt, vfid, data);
+	if (ret)
+		goto fail;
+
+	return 0;
+
+fail:
+	xe_sriov_packet_free(data);
+	return ret;
+}
+
+#define VF_VRAM_STATE_CHUNK_MAX_SIZE SZ_512M
+static int pf_save_vf_vram_mig_data(struct xe_gt *gt, unsigned int vfid)
+{
+	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
+	loff_t *offset = &migration->save.vram_offset;
+	struct xe_bo *vram;
+	size_t vram_size, chunk_size;
+	int ret;
+
+	vram = xe_gt_sriov_pf_config_get_lmem_obj(gt, vfid);
+	if (!vram)
+		return -ENXIO;
+
+	vram_size = xe_bo_size(vram);
+
+	xe_gt_assert(gt, *offset < vram_size);
+
+	chunk_size = min(vram_size - *offset, VF_VRAM_STATE_CHUNK_MAX_SIZE);
+
+	ret = pf_save_vram_chunk(gt, vfid, vram, *offset, chunk_size);
+	if (ret)
+		goto fail;
+
+	*offset += chunk_size;
+
+	xe_bo_put(vram);
+
+	if (*offset < vram_size)
+		return -EAGAIN;
+
+	return 0;
+
+fail:
+	xe_bo_put(vram);
+	xe_gt_sriov_err(gt, "Failed to save VF%u VRAM data (%pe)\n", vfid, ERR_PTR(ret));
+	return ret;
+}
+
+static int pf_restore_vf_vram_mig_data(struct xe_gt *gt, unsigned int vfid,
+				       struct xe_sriov_packet *data)
+{
+	u64 end = data->hdr.offset + data->hdr.size;
+	struct dma_fence *fence;
+	struct xe_bo *vram;
+	size_t size;
+	int ret = 0;
+
+	vram = xe_gt_sriov_pf_config_get_lmem_obj(gt, vfid);
+	if (!vram)
+		return -ENXIO;
+
+	size = xe_bo_size(vram);
+
+	if (end > size || end < data->hdr.size) {
+		ret = -EINVAL;
+		goto err;
+	}
+
+	pf_dump_mig_data(gt, vfid, data, "VRAM data restore");
+
+	fence = __pf_save_restore_vram(gt, vfid, vram, data->hdr.offset,
+				       data->bo, 0, data->hdr.size, false);
+	ret = dma_fence_wait_timeout(fence, false, PF_VRAM_SAVE_RESTORE_TIMEOUT);
+	dma_fence_put(fence);
+	if (!ret) {
+		ret = -ETIME;
+		goto err;
+	}
+
+	return 0;
+err:
+	xe_bo_put(vram);
+	xe_gt_sriov_err(gt, "Failed to restore VF%u VRAM data (%pe)\n", vfid, ERR_PTR(ret));
+	return ret;
+}
+
+/**
+ * xe_gt_sriov_pf_migration_vram_save() - Save VF VRAM migration data.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_gt_sriov_pf_migration_vram_save(struct xe_gt *gt, unsigned int vfid)
+{
+	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
+	xe_gt_assert(gt, vfid != PFID);
+	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
+
+	return pf_save_vf_vram_mig_data(gt, vfid);
+}
+
+/**
+ * xe_gt_sriov_pf_migration_vram_restore() - Restore VF VRAM migration data.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier (can't be 0)
+ * @data: the &xe_sriov_packet containing migration data
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_gt_sriov_pf_migration_vram_restore(struct xe_gt *gt, unsigned int vfid,
+					  struct xe_sriov_packet *data)
+{
+	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
+	xe_gt_assert(gt, vfid != PFID);
+	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
+
+	return pf_restore_vf_vram_mig_data(gt, vfid, data);
+}
+
 /**
  * xe_gt_sriov_pf_migration_size() - Total size of migration data from all components within a GT.
  * @gt: the &xe_gt
@@ -544,6 +745,13 @@ ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid)
 		size += sizeof(struct xe_sriov_packet_hdr);
 	total += size;
 
+	size = pf_migration_vram_size(gt, vfid);
+	if (size < 0)
+		return size;
+	if (size > 0)
+		size += sizeof(struct xe_sriov_packet_hdr);
+	total += size;
+
 	return total;
 }
 
@@ -606,6 +814,7 @@ void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid)
 	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
 
 	migration->save.data_remaining = 0;
+	migration->save.vram_offset = 0;
 
 	xe_gt_assert(gt, pf_migration_guc_size(gt, vfid) > 0);
 	pf_migration_save_data_todo(gt, vfid, XE_SRIOV_PACKET_TYPE_GUC);
@@ -615,6 +824,9 @@ void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid)
 
 	xe_gt_assert(gt, pf_migration_mmio_size(gt, vfid) > 0);
 	pf_migration_save_data_todo(gt, vfid, XE_SRIOV_PACKET_TYPE_MMIO);
+
+	if (pf_migration_vram_size(gt, vfid) > 0)
+		pf_migration_save_data_todo(gt, vfid, XE_SRIOV_PACKET_TYPE_VRAM);
 }
 
 /**

drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h

@@ -25,6 +25,9 @@ int xe_gt_sriov_pf_migration_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
 int xe_gt_sriov_pf_migration_mmio_save(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_migration_mmio_restore(struct xe_gt *gt, unsigned int vfid,
 					  struct xe_sriov_packet *data);
+int xe_gt_sriov_pf_migration_vram_save(struct xe_gt *gt, unsigned int vfid);
+int xe_gt_sriov_pf_migration_vram_restore(struct xe_gt *gt, unsigned int vfid,
+					  struct xe_sriov_packet *data);
 
 ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid);
 

drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h

@@ -20,6 +20,8 @@ struct xe_gt_sriov_migration_data {
 	struct {
 		/** @save.data_remaining: bitmap of migration types that need to be saved */
 		unsigned long data_remaining;
+		/** @save.vram_offset: last saved offset within VRAM, used for chunked VRAM save */
+		loff_t vram_offset;
 	} save;
 };
 

drivers/gpu/drm/xe/xe_sriov_pf_control.c

@@ -5,6 +5,7 @@
 
 #include "xe_device.h"
 #include "xe_gt_sriov_pf_control.h"
+#include "xe_gt_sriov_pf_migration.h"
 #include "xe_sriov_packet.h"
 #include "xe_sriov_pf_control.h"
 #include "xe_sriov_printk.h"
@@ -171,6 +172,8 @@ int xe_sriov_pf_control_trigger_save_vf(struct xe_device *xe, unsigned int vfid)
 		return ret;
 
 	for_each_gt(gt, xe, id) {
+		xe_gt_sriov_pf_migration_save_init(gt, vfid);
+
 		ret = xe_gt_sriov_pf_control_trigger_save_vf(gt, vfid);
 		if (ret)
 			return ret;