nv50: make mm available as common code

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

7 files changed, 29 insertions, 327 deletions

diff --git a/src/gallium/drivers/nv50/Makefile b/src/gallium/drivers/nv50/Makefile
index 61fb94913b..dc9ea0eeba 100644
--- a/src/gallium/drivers/nv50/Makefile
+++ b/src/gallium/drivers/nv50/Makefile
@@ -26,7 +26,6 @@ C_SOURCES = \
 	nv50_pc_optimize.c \
 	nv50_pc_regalloc.c \
 	nv50_push.c \
-	nv50_mm.c \
 	nv50_query.c
 
 LIBRARY_INCLUDES = \
diff --git a/src/gallium/drivers/nv50/nv50_buffer.c b/src/gallium/drivers/nv50/nv50_buffer.c
index 21aad9f949..f808adb0f6 100644
--- a/src/gallium/drivers/nv50/nv50_buffer.c
+++ b/src/gallium/drivers/nv50/nv50_buffer.c
@@ -6,6 +6,7 @@
 #define NOUVEAU_NVC0
 #include "nouveau/nouveau_screen.h"
 #include "nouveau/nouveau_winsys.h"
+#include "nouveau/nouveau_mm.h"
 #undef NOUVEAU_NVC0
 
 #include "nv50_context.h"
@@ -26,14 +27,14 @@ nv50_buffer_allocate(struct nv50_screen *screen, struct nv50_resource *buf,
                      unsigned domain)
 {
    if (domain == NOUVEAU_BO_VRAM) {
-      buf->mm = nv50_mm_allocate(screen->mm_VRAM, buf->base.width0, &buf->bo,
-                                 &buf->offset);
+      buf->mm = nouveau_mm_allocate(screen->mm_VRAM, buf->base.width0, &buf->bo,
+                                    &buf->offset);
       if (!buf->bo)
          return nv50_buffer_allocate(screen, buf, NOUVEAU_BO_GART);
    } else
    if (domain == NOUVEAU_BO_GART) {
-      buf->mm = nv50_mm_allocate(screen->mm_GART, buf->base.width0, &buf->bo,
-                                 &buf->offset);
+      buf->mm = nouveau_mm_allocate(screen->mm_GART, buf->base.width0, &buf->bo,
+                                    &buf->offset);
       if (!buf->bo)
          return FALSE;
    }
@@ -49,9 +50,9 @@ nv50_buffer_allocate(struct nv50_screen *screen, struct nv50_resource *buf,
 }
 
 static INLINE void
-release_allocation(struct nv50_mm_allocation **mm, struct nouveau_fence *fence)
+release_allocation(struct nouveau_mm_allocation **mm, struct nouveau_fence *fence)
 {
-   nouveau_fence_work(fence, nv50_mm_free, *mm);
+   nouveau_fence_work(fence, nouveau_mm_free, *mm);
    (*mm) = NULL;
 }
 
@@ -94,13 +95,13 @@ boolean
 nv50_buffer_download(struct nv50_context *nv50, struct nv50_resource *buf,
                      unsigned start, unsigned size)
 {
-   struct nv50_mm_allocation *mm;
+   struct nouveau_mm_allocation *mm;
    struct nouveau_bo *bounce = NULL;
    uint32_t offset;
 
    assert(buf->domain == NOUVEAU_BO_VRAM);
 
-   mm = nv50_mm_allocate(nv50->screen->mm_GART, size, &bounce, &offset);
+   mm = nouveau_mm_allocate(nv50->screen->mm_GART, size, &bounce, &offset);
    if (!bounce)
       return FALSE;
 
@@ -117,7 +118,7 @@ nv50_buffer_download(struct nv50_context *nv50, struct nv50_resource *buf,
 
    nouveau_bo_ref(NULL, &bounce);
    if (mm)
-      nv50_mm_free(mm);
+      nouveau_mm_free(mm);
    return TRUE;
 }
 
@@ -125,7 +126,7 @@ static boolean
 nv50_buffer_upload(struct nv50_context *nv50, struct nv50_resource *buf,
                    unsigned start, unsigned size)
 {
-   struct nv50_mm_allocation *mm;
+   struct nouveau_mm_allocation *mm;
    struct nouveau_bo *bounce = NULL;
    uint32_t offset;
 
@@ -135,7 +136,7 @@ nv50_buffer_upload(struct nv50_context *nv50, struct nv50_resource *buf,
       return TRUE;
    }
 
-   mm = nv50_mm_allocate(nv50->screen->mm_GART, size, &bounce, &offset);
+   mm = nouveau_mm_allocate(nv50->screen->mm_GART, size, &bounce, &offset);
    if (!bounce)
       return FALSE;
 
@@ -429,7 +430,7 @@ nv50_buffer_migrate(struct nv50_context *nv50,
       FREE(buf->data);
    } else
    if (old_domain != 0 && new_domain != 0) {
-      struct nv50_mm_allocation *mm = buf->mm;
+      struct nouveau_mm_allocation *mm = buf->mm;
 
       if (new_domain == NOUVEAU_BO_VRAM) {
          /* keep a system memory copy of our data in case we hit a fallback */
diff --git a/src/gallium/drivers/nv50/nv50_mm.c b/src/gallium/drivers/nv50/nv50_mm.c
deleted file mode 100644
index f991d6192e..0000000000
--- a/src/gallium/drivers/nv50/nv50_mm.c
+++ /dev/null
@@ -1,277 +0,0 @@
-
-#include "util/u_inlines.h"
-#include "util/u_memory.h"
-#include "util/u_double_list.h"
-
-#include "nv50_screen.h"
-
-#define MM_MIN_ORDER 7
-#define MM_MAX_ORDER 20
-
-#define MM_NUM_BUCKETS (MM_MAX_ORDER - MM_MIN_ORDER + 1)
-
-#define MM_MIN_SIZE (1 << MM_MIN_ORDER)
-#define MM_MAX_SIZE (1 << MM_MAX_ORDER)
-
-struct mm_bucket {
-   struct list_head free;
-   struct list_head used;
-   struct list_head full;
-   int num_free;
-};
-
-struct nv50_mman {
-   struct nouveau_device *dev;
-   struct mm_bucket bucket[MM_NUM_BUCKETS];
-   uint32_t storage_type;
-   uint32_t domain;
-   uint64_t allocated;
-};
-
-struct mm_slab {
-   struct list_head head;
-   struct nouveau_bo *bo;
-   struct nv50_mman *cache;
-   int order;
-   int count;
-   int free;
-   uint32_t bits[0];
-};
-
-static int
-mm_slab_alloc(struct mm_slab *slab)
-{
-   int i, n, b;
-
-   if (slab->free == 0)
-      return -1;
-
-   for (i = 0; i < (slab->count + 31) / 32; ++i) {
-      b = ffs(slab->bits[i]) - 1;
-      if (b >= 0) {
-         n = i * 32 + b;
-         assert(n < slab->count);
-         slab->free--;
-         slab->bits[i] &= ~(1 << b);
-         return n;
-      }
-   }
-   return -1;
-}
-
-static INLINE void
-mm_slab_free(struct mm_slab *slab, int i)
-{
-   assert(i < slab->count);
-   slab->bits[i / 32] |= 1 << (i % 32);
-   slab->free++;
-   assert(slab->free <= slab->count);
-}
-
-static INLINE int
-mm_get_order(uint32_t size)
-{
-   int s = __builtin_clz(size) ^ 31;
-
-   if (size > (1 << s))
-      s += 1;
-   return s;
-}
-
-static struct mm_bucket *
-mm_bucket_by_order(struct nv50_mman *cache, int order)
-{
-   if (order > MM_MAX_ORDER)
-      return NULL;
-   return &cache->bucket[MAX2(order, MM_MIN_ORDER) - MM_MIN_ORDER];
-}
-
-static struct mm_bucket *
-mm_bucket_by_size(struct nv50_mman *cache, unsigned size)
-{
-   return mm_bucket_by_order(cache, mm_get_order(size));
-}
-
-/* size of bo allocation for slab with chunks of (1 << chunk_order) bytes */
-static INLINE uint32_t
-mm_default_slab_size(unsigned chunk_order)
-{
-   static const int8_t slab_order[MM_MAX_ORDER - MM_MIN_ORDER + 1] =
-   {
-      12, 12, 13, 14, 14, 17, 17, 17, 17, 19, 19, 20, 21, 22
-   };
-
-   assert(chunk_order <= MM_MAX_ORDER && chunk_order >= MM_MIN_ORDER);
-
-   return 1 << slab_order[chunk_order - MM_MIN_ORDER];
-}
-
-static int
-mm_slab_new(struct nv50_mman *cache, int chunk_order)
-{
-   struct mm_slab *slab;
-   int words, ret;
-   const uint32_t size = mm_default_slab_size(chunk_order);
-
-   words = ((size >> chunk_order) + 31) / 32;
-   assert(words);
-
-   slab = MALLOC(sizeof(struct mm_slab) + words * 4);
-   if (!slab)
-      return PIPE_ERROR_OUT_OF_MEMORY;
-
-   memset(&slab->bits[0], ~0, words * 4);
-
-   slab->bo = NULL;
-   ret = nouveau_bo_new_tile(cache->dev, cache->domain, 0, size,
-                             0, cache->storage_type, &slab->bo);
-   if (ret) {
-      FREE(slab);
-      return PIPE_ERROR_OUT_OF_MEMORY;
-   }
-
-   LIST_INITHEAD(&slab->head);
-
-   slab->cache = cache;
-   slab->order = chunk_order;
-   slab->count = slab->free = size >> chunk_order;
-
-   LIST_ADD(&slab->head, &mm_bucket_by_order(cache, chunk_order)->free);
-
-   cache->allocated += size;
-
-   debug_printf("MM: new slab, total memory = %lu KiB\n",
-                cache->allocated / 1024);
-
-   return PIPE_OK;
-}
-
-/* @return token to identify slab or NULL if we just allocated a new bo */
-struct nv50_mm_allocation *
-nv50_mm_allocate(struct nv50_mman *cache,
-                 uint32_t size, struct nouveau_bo **bo, uint32_t *offset)
-{
-   struct mm_bucket *bucket;
-   struct mm_slab *slab;
-   struct nv50_mm_allocation *alloc;
-   int ret;
-
-   bucket = mm_bucket_by_size(cache, size);
-   if (!bucket) {
-      ret = nouveau_bo_new_tile(cache->dev, cache->domain, 0, size,
-                                0, cache->storage_type, bo);
-      if (ret)
-         debug_printf("bo_new(%x, %x): %i\n", size, cache->storage_type, ret);
-
-      *offset = 0;
-      return NULL;
-   }
-
-   if (!LIST_IS_EMPTY(&bucket->used)) {
-      slab = LIST_ENTRY(struct mm_slab, bucket->used.next, head);
-   } else {
-      if (LIST_IS_EMPTY(&bucket->free)) {
-         mm_slab_new(cache, MAX2(mm_get_order(size), MM_MIN_ORDER));
-      }
-      slab = LIST_ENTRY(struct mm_slab, bucket->free.next, head);
-
-      LIST_DEL(&slab->head);
-      LIST_ADD(&slab->head, &bucket->used);
-   }
-
-   *offset = mm_slab_alloc(slab) << slab->order;
-
-   alloc = MALLOC_STRUCT(nv50_mm_allocation);
-   if (!alloc)
-      return NULL;
-
-   nouveau_bo_ref(slab->bo, bo);
-
-   if (slab->free == 0) {
-      LIST_DEL(&slab->head);
-      LIST_ADD(&slab->head, &bucket->full);
-   }
-
-   alloc->next = NULL;
-   alloc->offset = *offset;
-   alloc->priv = (void *)slab;
-
-   return alloc;
-}
-
-void
-nv50_mm_free(struct nv50_mm_allocation *alloc)
-{
-   struct mm_slab *slab = (struct mm_slab *)alloc->priv;
-   struct mm_bucket *bucket = mm_bucket_by_order(slab->cache, slab->order);
-
-   mm_slab_free(slab, alloc->offset >> slab->order);
-
-   if (slab->free == 1) {
-      LIST_DEL(&slab->head);
-
-      if (slab->count > 1)
-         LIST_ADDTAIL(&slab->head, &bucket->used);
-      else
-         LIST_ADDTAIL(&slab->head, &bucket->free);
-   }
-
-   FREE(alloc);
-}
-
-struct nv50_mman *
-nv50_mm_create(struct nouveau_device *dev, uint32_t domain,
-               uint32_t storage_type)
-{
-   struct nv50_mman *cache = MALLOC_STRUCT(nv50_mman);
-   int i;
-
-   if (!cache)
-      return NULL;
-
-   cache->dev = dev;
-   cache->domain = domain;
-   cache->storage_type = storage_type;
-   cache->allocated = 0;
-
-   for (i = 0; i < MM_NUM_BUCKETS; ++i) {
-      LIST_INITHEAD(&cache->bucket[i].free);
-      LIST_INITHEAD(&cache->bucket[i].used);
-      LIST_INITHEAD(&cache->bucket[i].full);
-   }
-
-   return cache;
-}
-
-static INLINE void
-nv50_mm_free_slabs(struct list_head *head)
-{
-   struct mm_slab *slab, *next;
-
-   LIST_FOR_EACH_ENTRY_SAFE(slab, next, head, head) {
-      LIST_DEL(&slab->head);
-      nouveau_bo_ref(NULL, &slab->bo);
-      FREE(slab);
-   }
-}
-
-void
-nv50_mm_destroy(struct nv50_mman *cache)
-{
-   int i;
-
-   if (!cache)
-      return;
-
-   for (i = 0; i < MM_NUM_BUCKETS; ++i) {
-      if (!LIST_IS_EMPTY(&cache->bucket[i].used) ||
-          !LIST_IS_EMPTY(&cache->bucket[i].full))
-         debug_printf("WARNING: destroying GPU memory cache "
-                      "with some buffers still in use\n");
-
-      nv50_mm_free_slabs(&cache->bucket[i].free);
-      nv50_mm_free_slabs(&cache->bucket[i].used);
-      nv50_mm_free_slabs(&cache->bucket[i].full);
-   }
-}
-
diff --git a/src/gallium/drivers/nv50/nv50_query.c b/src/gallium/drivers/nv50/nv50_query.c
index 42391ec5b1..8a2bca6850 100644
--- a/src/gallium/drivers/nv50/nv50_query.c
+++ b/src/gallium/drivers/nv50/nv50_query.c
@@ -41,7 +41,7 @@ struct nv50_query {
    uint32_t offset; /* base + i * 16 */
    boolean ready;
    boolean is64bit;
-   struct nv50_mm_allocation *mm;
+   struct nouveau_mm_allocation *mm;
 };
 
 #define NV50_QUERY_ALLOC_SPACE 128
@@ -62,13 +62,13 @@ nv50_query_allocate(struct nv50_context *nv50, struct nv50_query *q, int size)
       nouveau_bo_ref(NULL, &q->bo);
       if (q->mm) {
          if (q->ready)
-            nv50_mm_free(q->mm);
+            nouveau_mm_free(q->mm);
          else
-            nouveau_fence_work(screen->base.fence.current, nv50_mm_free, q->mm);
+            nouveau_fence_work(screen->base.fence.current, nouveau_mm_free, q->mm);
       }
    }
    if (size) {
-      q->mm = nv50_mm_allocate(screen->mm_GART, size, &q->bo, &q->base);
+      q->mm = nouveau_mm_allocate(screen->mm_GART, size, &q->bo, &q->base);
       if (!q->bo)
          return FALSE;
       q->offset = q->base;
diff --git a/src/gallium/drivers/nv50/nv50_resource.h b/src/gallium/drivers/nv50/nv50_resource.h
index f42179c88f..64563421fd 100644
--- a/src/gallium/drivers/nv50/nv50_resource.h
+++ b/src/gallium/drivers/nv50/nv50_resource.h
@@ -46,7 +46,7 @@ struct nv50_resource {
    struct nouveau_fence *fence;
    struct nouveau_fence *fence_wr;
 
-   struct nv50_mm_allocation *mm;
+   struct nouveau_mm_allocation *mm;
 };
 
 void
diff --git a/src/gallium/drivers/nv50/nv50_screen.c b/src/gallium/drivers/nv50/nv50_screen.c
index e5b50103ef..bd645b8716 100644
--- a/src/gallium/drivers/nv50/nv50_screen.c
+++ b/src/gallium/drivers/nv50/nv50_screen.c
@@ -229,9 +229,9 @@ nv50_screen_destroy(struct pipe_screen *pscreen)
    if (screen->tic.entries)
       FREE(screen->tic.entries);
 
-   nv50_mm_destroy(screen->mm_GART);
-   nv50_mm_destroy(screen->mm_VRAM);
-   nv50_mm_destroy(screen->mm_VRAM_fe0);
+   nouveau_mm_destroy(screen->mm_GART);
+   nouveau_mm_destroy(screen->mm_VRAM);
+   nouveau_mm_destroy(screen->mm_VRAM_fe0);
 
    nouveau_grobj_free(&screen->tesla);
    nouveau_grobj_free(&screen->eng2d);
@@ -586,10 +586,10 @@ nv50_screen_create(struct pipe_winsys *ws, struct nouveau_device *dev)
    screen->tic.entries = CALLOC(4096, sizeof(void *));
    screen->tsc.entries = screen->tic.entries + 2048;
 
-   screen->mm_GART = nv50_mm_create(dev, NOUVEAU_BO_GART | NOUVEAU_BO_MAP,
-                                    0x000);
-   screen->mm_VRAM = nv50_mm_create(dev, NOUVEAU_BO_VRAM, 0x000);
-   screen->mm_VRAM_fe0 = nv50_mm_create(dev, NOUVEAU_BO_VRAM, 0xfe0);
+   screen->mm_GART = nouveau_mm_create(dev, NOUVEAU_BO_GART | NOUVEAU_BO_MAP,
+                                       0x000);
+   screen->mm_VRAM = nouveau_mm_create(dev, NOUVEAU_BO_VRAM, 0x000);
+   screen->mm_VRAM_fe0 = nouveau_mm_create(dev, NOUVEAU_BO_VRAM, 0xfe0);
 
    nouveau_fence_new(&screen->base, &screen->base.fence.current, FALSE);
 
diff --git a/src/gallium/drivers/nv50/nv50_screen.h b/src/gallium/drivers/nv50/nv50_screen.h
index c2ec3b58dc..672891b6b7 100644
--- a/src/gallium/drivers/nv50/nv50_screen.h
+++ b/src/gallium/drivers/nv50/nv50_screen.h
@@ -4,6 +4,7 @@
 #define NOUVEAU_NVC0
 #include "nouveau/nouveau_screen.h"
 #include "nouveau/nouveau_fence.h"
+#include "nouveau/nouveau_mm.h"
 #undef NOUVEAU_NVC0
 #include "nv50_winsys.h"
 #include "nv50_stateobj.h"
@@ -11,7 +12,6 @@
 #define NV50_TIC_MAX_ENTRIES 2048
 #define NV50_TSC_MAX_ENTRIES 2048
 
-struct nv50_mman;
 struct nv50_context;
 
 #define NV50_SCRATCH_SIZE (2 << 20)
@@ -54,9 +54,9 @@ struct nv50_screen {
 
    struct nouveau_notifier *sync;
 
-   struct nv50_mman *mm_GART;
-   struct nv50_mman *mm_VRAM;
-   struct nv50_mman *mm_VRAM_fe0;
+   struct nouveau_mman *mm_GART;
+   struct nouveau_mman *mm_VRAM;
+   struct nouveau_mman *mm_VRAM_fe0;
 
    struct nouveau_grobj *tesla;
    struct nouveau_grobj *eng2d;
@@ -69,27 +69,6 @@ nv50_screen(struct pipe_screen *screen)
    return (struct nv50_screen *)screen;
 }
 
-/* Since a resource can be migrated, we need to decouple allocations from
- * them. This struct is linked with fences for delayed freeing of allocs.
- */
-struct nv50_mm_allocation {
-   struct nv50_mm_allocation *next;
-   void *priv;
-   uint32_t offset;
-};
-
-extern struct nv50_mman *
-nv50_mm_create(struct nouveau_device *, uint32_t domain, uint32_t storage_type);
-
-extern void
-nv50_mm_destroy(struct nv50_mman *);
-
-extern struct nv50_mm_allocation *
-nv50_mm_allocate(struct nv50_mman *,
-                 uint32_t size, struct nouveau_bo **, uint32_t *offset);
-extern void
-nv50_mm_free(struct nv50_mm_allocation *);
-
 void nv50_screen_make_buffers_resident(struct nv50_screen *);
 
 int nv50_screen_tic_alloc(struct nv50_screen *, void *);