drm/nouveau: directly handle comptag allocation

Another transition step to allow finer-grained patches transitioning to
new MMU backends.

Old backends will continue operate as before (accessing nvkm_mem::tag),
and new backends will get a reference to the tags allocated here.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>

drivers/gpu/drm/nouveau/include/nvkm/subdev/ltc.h

@@ -22,8 +22,6 @@ struct nvkm_ltc {
 	u32 zbc_depth[NVKM_LTC_MAX_ZBC_CNT];
 };
 
-int nvkm_ltc_tags_alloc(struct nvkm_ltc *, u32 count, struct nvkm_mm_node **);
-void nvkm_ltc_tags_free(struct nvkm_ltc *, struct nvkm_mm_node **);
 void nvkm_ltc_tags_clear(struct nvkm_device *, u32 first, u32 count);
 
 int nvkm_ltc_zbc_color_get(struct nvkm_ltc *, int index, const u32[4]);

drivers/gpu/drm/nouveau/nouveau_mem.c

@@ -23,6 +23,8 @@
 #include "nouveau_drv.h"
 #include "nouveau_bo.h"
 
+#include <subdev/ltc.h>
+
 #include <drm/ttm/ttm_bo_driver.h>
 
 int
@@ -44,6 +46,8 @@ nouveau_mem_fini(struct nouveau_mem *mem)
 		nvkm_vm_unmap(&mem->vma[0]);
 		nvkm_vm_put(&mem->vma[0]);
 	}
+	nvkm_memory_tags_put(&mem->memory, nvxx_device(&mem->cli->device),
+			     &mem->tags);
 }
 
 int
@@ -74,17 +78,47 @@ int
 nouveau_mem_vram(struct ttm_mem_reg *reg, bool contig, u8 page)
 {
 	struct nouveau_mem *mem = nouveau_mem(reg);
-	struct nvkm_ram *ram = nvxx_fb(&mem->cli->device)->ram;
+	struct nouveau_cli *cli = mem->cli;
+	struct nvkm_device *device = nvxx_device(&cli->device);
+	struct nvkm_ram *ram = nvxx_fb(&cli->device)->ram;
 	u64 size = ALIGN(reg->num_pages << PAGE_SHIFT, 1 << page);
 	int ret;
 
 	mem->mem.page = page;
 	mem->_mem->memory = &mem->memory;
 
+	if (cli->device.info.chipset < 0xc0 && mem->comp) {
+		if (page == 16) {
+			ret = nvkm_memory_tags_get(mem->_mem->memory, device,
+						   size >> page, NULL,
+						   &mem->tags);
+			WARN_ON(ret);
+		}
+		if (!mem->tags || !mem->tags->mn)
+			mem->comp = 0;
+	} else
+	if (cli->device.info.chipset >= 0xc0 &&
+	    gf100_pte_storage_type_map[mem->kind] != mem->kind) {
+		if (page == 17) {
+			ret = nvkm_memory_tags_get(mem->_mem->memory, device,
+						   size >> page,
+						   nvkm_ltc_tags_clear,
+						   &mem->tags);
+			WARN_ON(ret);
+		}
+		if (!mem->tags || !mem->tags->mn)
+			mem->kind = gf100_pte_storage_type_map[mem->kind];
+	}
+
 	ret = ram->func->get(ram, size, 1 << page, contig ? 0 : 1 << page,
 			     (mem->comp << 8) | mem->kind, &mem->_mem);
-	if (ret)
+	if (ret) {
+		nvkm_memory_tags_put(mem->_mem->memory, device, &mem->tags);
 		return ret;
+	}
+
+	if (mem->tags && mem->tags->mn)
+		mem->_mem->tag = mem->tags->mn;
 
 	reg->start = mem->_mem->offset >> PAGE_SHIFT;
 	return ret;

drivers/gpu/drm/nouveau/nouveau_mem.h

@@ -25,6 +25,7 @@ struct nouveau_mem {
 	struct nvkm_vma bar_vma;
 
 	struct nvkm_memory memory;
+	struct nvkm_tags *tags;
 };
 
 enum nvif_vmm_get {

drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgf100.c

@@ -32,7 +32,6 @@
 #include <subdev/bios/timing.h>
 #include <subdev/clk.h>
 #include <subdev/clk/pll.h>
-#include <subdev/ltc.h>
 
 struct gf100_ramfuc {
 	struct ramfuc base;
@@ -423,7 +422,6 @@ gf100_ram_tidy(struct nvkm_ram *base)
 void
 gf100_ram_put(struct nvkm_ram *ram, struct nvkm_mem **pmem)
 {
-	struct nvkm_ltc *ltc = ram->fb->subdev.device->ltc;
 	struct nvkm_mem *mem = *pmem;
 
 	*pmem = NULL;
@@ -431,8 +429,6 @@ gf100_ram_put(struct nvkm_ram *ram, struct nvkm_mem **pmem)
 		return;
 
 	mutex_lock(&ram->fb->subdev.mutex);
-	if (mem->tag)
-		nvkm_ltc_tags_free(ltc, &mem->tag);
 	__nv50_ram_put(ram, mem);
 	mutex_unlock(&ram->fb->subdev.mutex);
 
@@ -443,14 +439,11 @@ int
 gf100_ram_get(struct nvkm_ram *ram, u64 size, u32 align, u32 ncmin,
 	      u32 memtype, struct nvkm_mem **pmem)
 {
-	struct nvkm_device *device = ram->fb->subdev.device;
-	struct nvkm_ltc *ltc = ram->fb->subdev.device->ltc;
 	struct nvkm_mm *mm = &ram->vram;
 	struct nvkm_mm_node **node, *r;
 	struct nvkm_mem *mem;
 	int type = (memtype & 0x0ff);
 	int back = (memtype & 0x800);
-	const bool comp = gf100_pte_storage_type_map[type] != type;
 	int ret;
 
 	size  >>= NVKM_RAM_MM_SHIFT;
@@ -466,19 +459,6 @@ gf100_ram_get(struct nvkm_ram *ram, u64 size, u32 align, u32 ncmin,
 	mem->size = size;
 
 	mutex_lock(&ram->fb->subdev.mutex);
-	if (comp) {
-		/* compression only works with lpages */
-		if (align == (1 << (17 - NVKM_RAM_MM_SHIFT))) {
-			int n = size >> 5;
-			if (!nvkm_ltc_tags_alloc(ltc, n, &mem->tag)) {
-				nvkm_ltc_tags_clear(device, mem->tag->offset,
-							    mem->tag->length);
-			}
-		}
-
-		if (unlikely(!mem->tag))
-			type = gf100_pte_storage_type_map[type];
-	}
 	mem->memtype = type;
 
 	node = &mem->mem;

drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c

@@ -502,7 +502,6 @@ __nv50_ram_put(struct nvkm_ram *ram, struct nvkm_mem *mem)
 		next = node->next;
 		nvkm_mm_free(&ram->vram, &node);
 	}
-	nvkm_mm_free(&ram->fb->tags, &mem->tag);
 }
 
 void
@@ -526,7 +525,6 @@ nv50_ram_get(struct nvkm_ram *ram, u64 size, u32 align, u32 ncmin,
 	     u32 memtype, struct nvkm_mem **pmem)
 {
 	struct nvkm_mm *heap = &ram->vram;
-	struct nvkm_mm *tags = &ram->fb->tags;
 	struct nvkm_mm_node **node, *r;
 	struct nvkm_mem *mem;
 	int comp = (memtype & 0x300) >> 8;
@@ -543,19 +541,6 @@ nv50_ram_get(struct nvkm_ram *ram, u64 size, u32 align, u32 ncmin,
 		return -ENOMEM;
 
 	mutex_lock(&ram->fb->subdev.mutex);
-	if (comp) {
-		if (align == (1 << (16 - NVKM_RAM_MM_SHIFT))) {
-			int n = (max >> 4) * comp;
-
-			ret = nvkm_mm_head(tags, 0, 1, n, n, 1, &mem->tag);
-			if (ret)
-				mem->tag = NULL;
-		}
-
-		if (unlikely(!mem->tag))
-			comp = 0;
-	}
-
 	mem->memtype = (comp << 7) | type;
 	mem->size = max;
 

drivers/gpu/drm/nouveau/nvkm/subdev/ltc/base.c

@@ -24,24 +24,6 @@
 #include "priv.h"
 
 #include <core/memory.h>
-#include <subdev/fb.h>
-
-int
-nvkm_ltc_tags_alloc(struct nvkm_ltc *ltc, u32 n, struct nvkm_mm_node **pnode)
-{
-	struct nvkm_fb *fb = ltc->subdev.device->fb;
-	int ret = nvkm_mm_head(&fb->tags, 0, 1, n, n, 1, pnode);
-	if (ret)
-		*pnode = NULL;
-	return ret;
-}
-
-void
-nvkm_ltc_tags_free(struct nvkm_ltc *ltc, struct nvkm_mm_node **pnode)
-{
-	struct nvkm_fb *fb = ltc->subdev.device->fb;
-	nvkm_mm_free(&fb->tags, pnode);
-}
 
 void
 nvkm_ltc_tags_clear(struct nvkm_device *device, u32 first, u32 count)