[PATCH v3 2/4] drm/xe/ttm: Add PF-mem VRAM placement types for TTM

["Piórkowski, Piotr" <piotr.piorkowski@intel.com>]

From: Piotr Piórkowski <piotr.piorkowski@intel.com>

PF-mem regions are dedicated to kernel-only pinned BOs. They are not
exposed to userspace and cannot be exported through dma-buf.

Add dedicated TTM placement types for PF-mem VRAM regions and update
the BO, resource iterator, power-management, and PCI removal paths to
recognize them. Since PF-mem cannot contain user memory, its placement
types are intentionally excluded from user BO eviction.

v2:
 - Modify commit message.
 - Add comments explaining why PF-mem is excluded from user BO eviction
   (Sashiko).
 - Resolve xe_ttm_vram_mgr_alloc_sgt() physical address via
   xe_map_resource_to_region() instead of the tile's
   VRAM region (Sashiko).
 - Assert in add_vram() that the requested placement is a VRAM one before
   resolving its TTM manager.

Assisted-by: Claude:claude-5-sonnet
Signed-off-by: Piotr Piórkowski <piotr.piorkowski@intel.com>
---
 drivers/gpu/drm/xe/xe_bo.c           | 25 ++++++++++++++++++-------
 drivers/gpu/drm/xe/xe_bo.h           | 17 +++++++++++++++++
 drivers/gpu/drm/xe/xe_bo_evict.c     |  7 +++++--
 drivers/gpu/drm/xe/xe_pm.c           |  4 ++--
 drivers/gpu/drm/xe/xe_res_cursor.h   |  9 ++++++++-
 drivers/gpu/drm/xe/xe_ttm_vram_mgr.c |  9 ++++-----
 6 files changed, 54 insertions(+), 17 deletions(-)

diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c
index dde309821237..91070a0a926d 100644
--- a/drivers/gpu/drm/xe/xe_bo.c
+++ b/drivers/gpu/drm/xe/xe_bo.c
@@ -46,6 +46,8 @@ const char *const xe_mem_type_to_name[TTM_NUM_MEM_TYPES]  = {
 	[XE_PL_TT] = "gtt",
 	[XE_PL_VRAM0] = "vram0",
 	[XE_PL_VRAM1] = "vram1",
+	[XE_PL_VRAM0_PFMEM] = "vram0_pfmem",
+	[XE_PL_VRAM1_PFMEM] = "vram1_pfmem",
 	[XE_PL_STOLEN] = "stolen"
 };
 
@@ -166,11 +168,12 @@ static bool xe_bo_is_user(struct xe_bo *bo)
 static struct xe_migrate *
 mem_type_to_migrate(struct xe_device *xe, u32 mem_type)
 {
-	struct xe_tile *tile;
+	u8 tile_id;
 
 	xe_assert(xe, mem_type == XE_PL_STOLEN || mem_type_is_vram(mem_type));
-	tile = &xe->tiles[mem_type == XE_PL_STOLEN ? 0 : (mem_type - XE_PL_VRAM0)];
-	return tile->migrate;
+	tile_id = mem_type == XE_PL_STOLEN ? 0 : xe_vram_pl_to_tile_id(xe, mem_type);
+
+	return xe->tiles[tile_id].migrate;
 }
 
 static void try_add_system(struct xe_device *xe, struct xe_bo *bo,
@@ -231,13 +234,16 @@ static void add_vram(struct xe_device *xe, struct xe_bo *bo,
 		     struct ttm_place *places, u32 bo_flags, u32 mem_type, u32 *c)
 {
 	struct ttm_place place = { .mem_type = mem_type };
-	struct ttm_resource_manager *mgr = ttm_manager_type(&xe->ttm, mem_type);
-	struct xe_ttm_vram_mgr *vram_mgr = to_xe_ttm_vram_mgr(mgr);
-
+	struct ttm_resource_manager *mgr;
+	struct xe_ttm_vram_mgr *vram_mgr;
 	struct xe_vram_region *vram;
 	u64 io_size;
 
 	xe_assert(xe, *c < ARRAY_SIZE(bo->placements));
+	xe_assert(xe, mem_type_is_vram(mem_type));
+
+	mgr = ttm_manager_type(&xe->ttm, mem_type);
+	vram_mgr = to_xe_ttm_vram_mgr(mgr);
 
 	vram = container_of(vram_mgr, struct xe_vram_region, ttm);
 	xe_assert(xe, vram && vram->usable_size);
@@ -360,6 +366,8 @@ static void xe_evict_flags(struct ttm_buffer_object *tbo,
 	switch (tbo->resource->mem_type) {
 	case XE_PL_VRAM0:
 	case XE_PL_VRAM1:
+	case XE_PL_VRAM0_PFMEM:
+	case XE_PL_VRAM1_PFMEM:
 	case XE_PL_STOLEN:
 		*placement = tt_placement;
 		break;
@@ -639,7 +647,10 @@ static int xe_ttm_io_mem_reserve(struct ttm_device *bdev,
 	case XE_PL_TT:
 		return 0;
 	case XE_PL_VRAM0:
-	case XE_PL_VRAM1: {
+	case XE_PL_VRAM1:
+	case XE_PL_VRAM0_PFMEM:
+	case XE_PL_VRAM1_PFMEM:
+	{
 		struct xe_vram_region *vram = xe_map_resource_to_region(mem);
 
 		if (!xe_ttm_resource_visible(xe, mem))
diff --git a/drivers/gpu/drm/xe/xe_bo.h b/drivers/gpu/drm/xe/xe_bo.h
index e8081af5bfc1..c698f70a3b0c 100644
--- a/drivers/gpu/drm/xe/xe_bo.h
+++ b/drivers/gpu/drm/xe/xe_bo.h
@@ -83,8 +83,13 @@
 #define XE_PL_TT		TTM_PL_TT
 #define XE_PL_VRAM0		TTM_PL_VRAM
 #define XE_PL_VRAM1		(XE_PL_VRAM0 + 1)
+#define XE_PL_VRAM0_PFMEM	(XE_PL_VRAM1 + 1)
+#define XE_PL_VRAM1_PFMEM	(XE_PL_VRAM0_PFMEM + 1)
 #define XE_PL_STOLEN		(TTM_NUM_MEM_TYPES - 1)
 
+#define xe_for_each_vram_pl(i) \
+	for (i = XE_PL_VRAM0; i <= XE_PL_VRAM1_PFMEM; i++)
+
 #define XE_BO_PROPS_INVALID	(-1)
 
 #define XE_PCI_BARRIER_MMAP_OFFSET	(0x50 << XE_PTE_SHIFT)
@@ -614,4 +619,16 @@ static inline bool xe_bo_is_mem_type(struct xe_bo *bo, u32 mem_type)
 	xe_bo_assert_held(bo);
 	return bo->ttm.resource->mem_type == mem_type;
 }
+
+static inline u8 xe_vram_pl_to_tile_id(struct xe_device *xe, u32 mem_type)
+{
+	xe_assert(xe, mem_type_is_vram(mem_type));
+
+	if (mem_type >= XE_PL_VRAM0 && mem_type <= XE_PL_VRAM1)
+		return mem_type - XE_PL_VRAM0;
+
+	xe_assert(xe, mem_type >= XE_PL_VRAM0_PFMEM && mem_type <= XE_PL_VRAM1_PFMEM);
+	return mem_type - XE_PL_VRAM0_PFMEM;
+}
+
 #endif
diff --git a/drivers/gpu/drm/xe/xe_bo_evict.c b/drivers/gpu/drm/xe/xe_bo_evict.c
index 7661fca7f278..a25ec6c750da 100644
--- a/drivers/gpu/drm/xe/xe_bo_evict.c
+++ b/drivers/gpu/drm/xe/xe_bo_evict.c
@@ -118,7 +118,10 @@ int xe_bo_evict_all_user(struct xe_device *xe)
 	u32 mem_type;
 	int ret;
 
-	/* User memory */
+	/*
+	 * Iterate over placements that can contain user memory. PF-mem placements
+	 * are excluded since they are reserved for pinned kernel BOs.
+	 */
 	for (mem_type = XE_PL_TT; mem_type <= XE_PL_VRAM1; ++mem_type) {
 		struct ttm_resource_manager *man =
 			ttm_manager_type(bdev, mem_type);
@@ -303,7 +306,7 @@ void xe_bo_pci_dev_remove_all(struct xe_device *xe)
 	 * Move pagemap bos and exported dma-buf to system, and
 	 * purge everything else.
 	 */
-	for (mem_type = XE_PL_VRAM1; mem_type >= XE_PL_TT; --mem_type) {
+	for (mem_type = XE_PL_VRAM1_PFMEM; mem_type >= XE_PL_TT; --mem_type) {
 		struct ttm_resource_manager *man =
 			ttm_manager_type(&xe->ttm, mem_type);
 
diff --git a/drivers/gpu/drm/xe/xe_pm.c b/drivers/gpu/drm/xe/xe_pm.c
index f517bf453b54..a4fe6a89d9dd 100644
--- a/drivers/gpu/drm/xe/xe_pm.c
+++ b/drivers/gpu/drm/xe/xe_pm.c
@@ -975,7 +975,7 @@ int xe_pm_set_vram_threshold(struct xe_device *xe, u32 threshold)
 	u32 vram_total_mb = 0;
 	int i;
 
-	for (i = XE_PL_VRAM0; i <= XE_PL_VRAM1; ++i) {
+	xe_for_each_vram_pl(i) {
 		man = ttm_manager_type(&xe->ttm, i);
 		if (man)
 			vram_total_mb += DIV_ROUND_UP_ULL(man->size, 1024 * 1024);
@@ -1012,7 +1012,7 @@ void xe_pm_d3cold_allowed_toggle(struct xe_device *xe)
 		return;
 	}
 
-	for (i = XE_PL_VRAM0; i <= XE_PL_VRAM1; ++i) {
+	xe_for_each_vram_pl(i) {
 		man = ttm_manager_type(&xe->ttm, i);
 		if (man) {
 			vram_used = ttm_resource_manager_usage(man);
diff --git a/drivers/gpu/drm/xe/xe_res_cursor.h b/drivers/gpu/drm/xe/xe_res_cursor.h
index 0522caafd89d..4603a3356e53 100644
--- a/drivers/gpu/drm/xe/xe_res_cursor.h
+++ b/drivers/gpu/drm/xe/xe_res_cursor.h
@@ -111,7 +111,9 @@ static inline void xe_res_first(struct ttm_resource *res,
 		break;
 	}
 	case XE_PL_VRAM0:
-	case XE_PL_VRAM1: {
+	case XE_PL_VRAM1:
+	case XE_PL_VRAM0_PFMEM:
+	case XE_PL_VRAM1_PFMEM: {
 		struct gpu_buddy_block *block;
 		struct list_head *head, *next;
 		struct gpu_buddy *mm = xe_res_get_buddy(res);
@@ -303,6 +305,8 @@ static inline void xe_res_next(struct xe_res_cursor *cur, u64 size)
 		break;
 	case XE_PL_VRAM0:
 	case XE_PL_VRAM1:
+	case XE_PL_VRAM0_PFMEM:
+	case XE_PL_VRAM1_PFMEM: {
 		start = size - cur->size;
 		block = cur->node;
 
@@ -322,6 +326,7 @@ static inline void xe_res_next(struct xe_res_cursor *cur, u64 size)
 				cur->remaining);
 		cur->node = block;
 		break;
+	}
 	default:
 		return;
 	}
@@ -358,6 +363,8 @@ static inline bool xe_res_is_vram(const struct xe_res_cursor *cur)
 	case XE_PL_STOLEN:
 	case XE_PL_VRAM0:
 	case XE_PL_VRAM1:
+	case XE_PL_VRAM0_PFMEM:
+	case XE_PL_VRAM1_PFMEM:
 		return true;
 	default:
 		break;
diff --git a/drivers/gpu/drm/xe/xe_ttm_vram_mgr.c b/drivers/gpu/drm/xe/xe_ttm_vram_mgr.c
index 05911904c1f9..bb28b493fd93 100644
--- a/drivers/gpu/drm/xe/xe_ttm_vram_mgr.c
+++ b/drivers/gpu/drm/xe/xe_ttm_vram_mgr.c
@@ -18,6 +18,7 @@
 #include "xe_pm.h"
 #include "xe_res_cursor.h"
 #include "xe_ttm_vram_mgr.h"
+#include "xe_vram.h"
 #include "xe_vram_types.h"
 
 static inline struct gpu_buddy_block *
@@ -336,7 +337,6 @@ int __xe_ttm_vram_mgr_init(struct xe_device *xe, struct xe_ttm_vram_mgr *mgr,
 {
 	struct ttm_resource_manager *man = &mgr->manager;
 	struct dmem_cgroup_region *cg;
-	const char *name;
 	int err;
 
 	man->func = &xe_ttm_vram_mgr_func;
@@ -365,8 +365,7 @@ int __xe_ttm_vram_mgr_init(struct xe_device *xe, struct xe_ttm_vram_mgr *mgr,
 	if (err)
 		return err;
 
-	name = mem_type == XE_PL_VRAM0 ? "vram0" : "vram1";
-	cg = drmm_cgroup_register_region(&xe->drm, name,
+	cg = drmm_cgroup_register_region(&xe->drm, xe_mem_type_to_name[mem_type],
 					 &(struct dmem_cgroup_init){
 						.size = size,
 						.ops = &xe_ttm_vram_mgr_dmem_ops,
@@ -405,7 +404,7 @@ int xe_ttm_vram_mgr_alloc_sgt(struct xe_device *xe,
 			      enum dma_data_direction dir,
 			      struct sg_table **sgt)
 {
-	struct xe_tile *tile = &xe->tiles[res->mem_type - XE_PL_VRAM0];
+	struct xe_vram_region *vram = xe_map_resource_to_region(res);
 	struct xe_ttm_vram_mgr_resource *vres = to_xe_ttm_vram_mgr_resource(res);
 	struct xe_res_cursor cursor;
 	struct scatterlist *sg;
@@ -443,7 +442,7 @@ int xe_ttm_vram_mgr_alloc_sgt(struct xe_device *xe,
 	 */
 	xe_res_first(res, offset, length, &cursor);
 	for_each_sgtable_sg((*sgt), sg, i) {
-		phys_addr_t phys = cursor.start + xe_vram_region_io_start(tile->mem.vram);
+		phys_addr_t phys = cursor.start + xe_vram_region_io_start(vram);
 		size_t size = min_t(u64, cursor.size, SZ_2G);
 		dma_addr_t addr;
 
-- 
2.34.1