Merge tag 'drm-xe-fixes-2025-04-18' of https://gitlab.freedesktop.org/drm/xe/kernel into drm-fixes

Driver Changes:
- Fix LRC address being written too late for GuC
- Fix notifier vs folio deadlock
- Fix race betwen dma_buf unmap and vram eviction
- Fix debugfs handling PXP terminations unconditionally

Signed-off-by: Dave Airlie <airlied@redhat.com>

From: Lucas De Marchi <lucas.demarchi@intel.com>
Link: https://lore.kernel.org/r/ndinq644zenywaaycxyfqqivsb2xer4z7err3dlpalbz33jfkm@ttabzsg6wnet

4 files changed, 57 insertions, 60 deletions

diff --git a/drivers/gpu/drm/xe/xe_dma_buf.c b/drivers/gpu/drm/xe/xe_dma_buf.c
index f67803e15a0e..f7a20264ea33 100644
--- a/drivers/gpu/drm/xe/xe_dma_buf.c
+++ b/drivers/gpu/drm/xe/xe_dma_buf.c
@@ -145,10 +145,7 @@ static void xe_dma_buf_unmap(struct dma_buf_attachment *attach,
 			     struct sg_table *sgt,
 			     enum dma_data_direction dir)
 {
-	struct dma_buf *dma_buf = attach->dmabuf;
-	struct xe_bo *bo = gem_to_xe_bo(dma_buf->priv);
-
-	if (!xe_bo_is_vram(bo)) {
+	if (sg_page(sgt->sgl)) {
 		dma_unmap_sgtable(attach->dev, sgt, dir, 0);
 		sg_free_table(sgt);
 		kfree(sgt);
diff --git a/drivers/gpu/drm/xe/xe_guc_ads.c b/drivers/gpu/drm/xe/xe_guc_ads.c
index e7c9e095a19f..7031542a70ce 100644
--- a/drivers/gpu/drm/xe/xe_guc_ads.c
+++ b/drivers/gpu/drm/xe/xe_guc_ads.c
@@ -490,24 +490,52 @@ static void fill_engine_enable_masks(struct xe_gt *gt,
 		       engine_enable_mask(gt, XE_ENGINE_CLASS_OTHER));
 }
 
-static void guc_prep_golden_lrc_null(struct xe_guc_ads *ads)
+/*
+ * Write the offsets corresponding to the golden LRCs. The actual data is
+ * populated later by guc_golden_lrc_populate()
+ */
+static void guc_golden_lrc_init(struct xe_guc_ads *ads)
 {
 	struct xe_device *xe = ads_to_xe(ads);
+	struct xe_gt *gt = ads_to_gt(ads);
 	struct iosys_map info_map = IOSYS_MAP_INIT_OFFSET(ads_to_map(ads),
 			offsetof(struct __guc_ads_blob, system_info));
-	u8 guc_class;
+	size_t alloc_size, real_size;
+	u32 addr_ggtt, offset;
+	int class;
+
+	offset = guc_ads_golden_lrc_offset(ads);
+	addr_ggtt = xe_bo_ggtt_addr(ads->bo) + offset;
+
+	for (class = 0; class < XE_ENGINE_CLASS_MAX; ++class) {
+		u8 guc_class;
+
+		guc_class = xe_engine_class_to_guc_class(class);
 
-	for (guc_class = 0; guc_class <= GUC_MAX_ENGINE_CLASSES; ++guc_class) {
 		if (!info_map_read(xe, &info_map,
 				   engine_enabled_masks[guc_class]))
 			continue;
 
+		real_size = xe_gt_lrc_size(gt, class);
+		alloc_size = PAGE_ALIGN(real_size);
+
+		/*
+		 * This interface is slightly confusing. We need to pass the
+		 * base address of the full golden context and the size of just
+		 * the engine state, which is the section of the context image
+		 * that starts after the execlists LRC registers. This is
+		 * required to allow the GuC to restore just the engine state
+		 * when a watchdog reset occurs.
+		 * We calculate the engine state size by removing the size of
+		 * what comes before it in the context image (which is identical
+		 * on all engines).
+		 */
 		ads_blob_write(ads, ads.eng_state_size[guc_class],
-			       guc_ads_golden_lrc_size(ads) -
-			       xe_lrc_skip_size(xe));
+			       real_size - xe_lrc_skip_size(xe));
 		ads_blob_write(ads, ads.golden_context_lrca[guc_class],
-			       xe_bo_ggtt_addr(ads->bo) +
-			       guc_ads_golden_lrc_offset(ads));
+			       addr_ggtt);
+
+		addr_ggtt += alloc_size;
 	}
 }
 
@@ -857,7 +885,7 @@ void xe_guc_ads_populate_minimal(struct xe_guc_ads *ads)
 
 	xe_map_memset(ads_to_xe(ads), ads_to_map(ads), 0, 0, ads->bo->size);
 	guc_policies_init(ads);
-	guc_prep_golden_lrc_null(ads);
+	guc_golden_lrc_init(ads);
 	guc_mapping_table_init_invalid(gt, &info_map);
 	guc_doorbell_init(ads);
 
@@ -883,7 +911,7 @@ void xe_guc_ads_populate(struct xe_guc_ads *ads)
 	guc_policies_init(ads);
 	fill_engine_enable_masks(gt, &info_map);
 	guc_mmio_reg_state_init(ads);
-	guc_prep_golden_lrc_null(ads);
+	guc_golden_lrc_init(ads);
 	guc_mapping_table_init(gt, &info_map);
 	guc_capture_prep_lists(ads);
 	guc_doorbell_init(ads);
@@ -903,18 +931,22 @@ void xe_guc_ads_populate(struct xe_guc_ads *ads)
 		       guc_ads_private_data_offset(ads));
 }
 
-static void guc_populate_golden_lrc(struct xe_guc_ads *ads)
+/*
+ * After the golden LRC's are recorded for each engine class by the first
+ * submission, copy them to the ADS, as initialized earlier by
+ * guc_golden_lrc_init().
+ */
+static void guc_golden_lrc_populate(struct xe_guc_ads *ads)
 {
 	struct xe_device *xe = ads_to_xe(ads);
 	struct xe_gt *gt = ads_to_gt(ads);
 	struct iosys_map info_map = IOSYS_MAP_INIT_OFFSET(ads_to_map(ads),
 			offsetof(struct __guc_ads_blob, system_info));
 	size_t total_size = 0, alloc_size, real_size;
-	u32 addr_ggtt, offset;
+	u32 offset;
 	int class;
 
 	offset = guc_ads_golden_lrc_offset(ads);
-	addr_ggtt = xe_bo_ggtt_addr(ads->bo) + offset;
 
 	for (class = 0; class < XE_ENGINE_CLASS_MAX; ++class) {
 		u8 guc_class;
@@ -931,26 +963,9 @@ static void guc_populate_golden_lrc(struct xe_guc_ads *ads)
 		alloc_size = PAGE_ALIGN(real_size);
 		total_size += alloc_size;
 
-		/*
-		 * This interface is slightly confusing. We need to pass the
-		 * base address of the full golden context and the size of just
-		 * the engine state, which is the section of the context image
-		 * that starts after the execlists LRC registers. This is
-		 * required to allow the GuC to restore just the engine state
-		 * when a watchdog reset occurs.
-		 * We calculate the engine state size by removing the size of
-		 * what comes before it in the context image (which is identical
-		 * on all engines).
-		 */
-		ads_blob_write(ads, ads.eng_state_size[guc_class],
-			       real_size - xe_lrc_skip_size(xe));
-		ads_blob_write(ads, ads.golden_context_lrca[guc_class],
-			       addr_ggtt);
-
 		xe_map_memcpy_to(xe, ads_to_map(ads), offset,
 				 gt->default_lrc[class], real_size);
 
-		addr_ggtt += alloc_size;
 		offset += alloc_size;
 	}
 
@@ -959,7 +974,7 @@ static void guc_populate_golden_lrc(struct xe_guc_ads *ads)
 
 void xe_guc_ads_populate_post_load(struct xe_guc_ads *ads)
 {
-	guc_populate_golden_lrc(ads);
+	guc_golden_lrc_populate(ads);
 }
 
 static int guc_ads_action_update_policies(struct xe_guc_ads *ads, u32 policy_offset)
diff --git a/drivers/gpu/drm/xe/xe_hmm.c b/drivers/gpu/drm/xe/xe_hmm.c
index c3cc0fa105e8..57b71956ddf4 100644
--- a/drivers/gpu/drm/xe/xe_hmm.c
+++ b/drivers/gpu/drm/xe/xe_hmm.c
@@ -19,29 +19,6 @@ static u64 xe_npages_in_range(unsigned long start, unsigned long end)
 	return (end - start) >> PAGE_SHIFT;
 }
 
-/**
- * xe_mark_range_accessed() - mark a range is accessed, so core mm
- * have such information for memory eviction or write back to
- * hard disk
- * @range: the range to mark
- * @write: if write to this range, we mark pages in this range
- * as dirty
- */
-static void xe_mark_range_accessed(struct hmm_range *range, bool write)
-{
-	struct page *page;
-	u64 i, npages;
-
-	npages = xe_npages_in_range(range->start, range->end);
-	for (i = 0; i < npages; i++) {
-		page = hmm_pfn_to_page(range->hmm_pfns[i]);
-		if (write)
-			set_page_dirty_lock(page);
-
-		mark_page_accessed(page);
-	}
-}
-
 static int xe_alloc_sg(struct xe_device *xe, struct sg_table *st,
 		       struct hmm_range *range, struct rw_semaphore *notifier_sem)
 {
@@ -331,7 +308,6 @@ int xe_hmm_userptr_populate_range(struct xe_userptr_vma *uvma,
 	if (ret)
 		goto out_unlock;
 
-	xe_mark_range_accessed(&hmm_range, write);
 	userptr->sg = &userptr->sgt;
 	xe_hmm_userptr_set_mapped(uvma);
 	userptr->notifier_seq = hmm_range.notifier_seq;
diff --git a/drivers/gpu/drm/xe/xe_pxp_debugfs.c b/drivers/gpu/drm/xe/xe_pxp_debugfs.c
index ccfbacf08efc..525a2f6bb076 100644
--- a/drivers/gpu/drm/xe/xe_pxp_debugfs.c
+++ b/drivers/gpu/drm/xe/xe_pxp_debugfs.c
@@ -66,9 +66,18 @@ static int pxp_terminate(struct seq_file *m, void *data)
 {
 	struct xe_pxp *pxp = node_to_pxp(m->private);
 	struct drm_printer p = drm_seq_file_printer(m);
+	int ready = xe_pxp_get_readiness_status(pxp);
 
-	if (!xe_pxp_is_enabled(pxp))
-		return -ENODEV;
+	if (ready < 0)
+		return ready; /* disabled or error occurred */
+	else if (!ready)
+		return -EBUSY; /* init still in progress */
+
+	/* no need for a termination if PXP is not active */
+	if (pxp->status != XE_PXP_ACTIVE) {
+		drm_printf(&p, "PXP not active\n");
+		return 0;
+	}
 
 	/* simulate a termination interrupt */
 	spin_lock_irq(&pxp->xe->irq.lock);