X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fnouveau%2Fnouveau_buffer.c;h=97305d993ff99941b154cb6c6b2896e45e4d6222;hb=09bb8c85577e761062ff8c85847bf49bad4ea86f;hp=936e2bf246ace722d4faf2117e5117785de02fb9;hpb=7435c043988dd83b430f3d3a7ca5a5a1b2f30d61;p=mesa.git

diff --git a/src/gallium/drivers/nouveau/nouveau_buffer.c b/src/gallium/drivers/nouveau/nouveau_buffer.c
index 936e2bf246a..97305d993ff 100644
--- a/src/gallium/drivers/nouveau/nouveau_buffer.c
+++ b/src/gallium/drivers/nouveau/nouveau_buffer.c
@@ -2,6 +2,7 @@
 #include "util/u_inlines.h"
 #include "util/u_memory.h"
 #include "util/u_math.h"
+#include "util/u_surface.h"
 
 #include "nouveau_screen.h"
 #include "nouveau_context.h"
@@ -12,50 +13,61 @@
 
 struct nouveau_transfer {
    struct pipe_transfer base;
+
+   uint8_t *map;
+   struct nouveau_bo *bo;
+   struct nouveau_mm_allocation *mm;
+   uint32_t offset;
 };
 
-static INLINE struct nouveau_transfer *
+static inline struct nouveau_transfer *
 nouveau_transfer(struct pipe_transfer *transfer)
 {
    return (struct nouveau_transfer *)transfer;
 }
 
-static INLINE boolean
+static inline bool
+nouveau_buffer_malloc(struct nv04_resource *buf)
+{
+   if (!buf->data)
+      buf->data = align_malloc(buf->base.width0, NOUVEAU_MIN_BUFFER_MAP_ALIGN);
+   return !!buf->data;
+}
+
+static inline bool
 nouveau_buffer_allocate(struct nouveau_screen *screen,
                         struct nv04_resource *buf, unsigned domain)
 {
-   uint32_t size = buf->base.width0;
-
-   if (buf->base.bind & PIPE_BIND_CONSTANT_BUFFER)
-      size = align(size, 0x100);
+   uint32_t size = align(buf->base.width0, 0x100);
 
    if (domain == NOUVEAU_BO_VRAM) {
       buf->mm = nouveau_mm_allocate(screen->mm_VRAM, size,
                                     &buf->bo, &buf->offset);
       if (!buf->bo)
          return nouveau_buffer_allocate(screen, buf, NOUVEAU_BO_GART);
+      NOUVEAU_DRV_STAT(screen, buf_obj_current_bytes_vid, buf->base.width0);
    } else
    if (domain == NOUVEAU_BO_GART) {
       buf->mm = nouveau_mm_allocate(screen->mm_GART, size,
                                     &buf->bo, &buf->offset);
       if (!buf->bo)
-         return FALSE;
-   }
-   if (domain != NOUVEAU_BO_GART) {
-      if (!buf->data) {
-         buf->data = MALLOC(buf->base.width0);
-         if (!buf->data)
-            return FALSE;
-      }
+         return false;
+      NOUVEAU_DRV_STAT(screen, buf_obj_current_bytes_sys, buf->base.width0);
+   } else {
+      assert(domain == 0);
+      if (!nouveau_buffer_malloc(buf))
+         return false;
    }
    buf->domain = domain;
    if (buf->bo)
       buf->address = buf->bo->offset + buf->offset;
 
-   return TRUE;
+   util_range_set_empty(&buf->valid_buffer_range);
+
+   return true;
 }
 
-static INLINE void
+static inline void
 release_allocation(struct nouveau_mm_allocation **mm,
                    struct nouveau_fence *fence)
 {
@@ -63,23 +75,38 @@ release_allocation(struct nouveau_mm_allocation **mm,
    (*mm) = NULL;
 }
 
-INLINE void
+inline void
 nouveau_buffer_release_gpu_storage(struct nv04_resource *buf)
 {
-   nouveau_bo_ref(NULL, &buf->bo);
+   if (buf->fence && buf->fence->state < NOUVEAU_FENCE_STATE_FLUSHED) {
+      nouveau_fence_work(buf->fence, nouveau_fence_unref_bo, buf->bo);
+      buf->bo = NULL;
+   } else {
+      nouveau_bo_ref(NULL, &buf->bo);
+   }
 
    if (buf->mm)
       release_allocation(&buf->mm, buf->fence);
 
+   if (buf->domain == NOUVEAU_BO_VRAM)
+      NOUVEAU_DRV_STAT_RES(buf, buf_obj_current_bytes_vid, -(uint64_t)buf->base.width0);
+   if (buf->domain == NOUVEAU_BO_GART)
+      NOUVEAU_DRV_STAT_RES(buf, buf_obj_current_bytes_sys, -(uint64_t)buf->base.width0);
+
    buf->domain = 0;
 }
 
-static INLINE boolean
+static inline bool
 nouveau_buffer_reallocate(struct nouveau_screen *screen,
                           struct nv04_resource *buf, unsigned domain)
 {
    nouveau_buffer_release_gpu_storage(buf);
 
+   nouveau_fence_ref(NULL, &buf->fence);
+   nouveau_fence_ref(NULL, &buf->fence_wr);
+
+   buf->status &= NOUVEAU_BUFFER_STATUS_REALLOC_MASK;
+
    return nouveau_buffer_allocate(screen, buf, domain);
 }
 
@@ -92,148 +119,136 @@ nouveau_buffer_destroy(struct pipe_screen *pscreen,
    nouveau_buffer_release_gpu_storage(res);
 
    if (res->data && !(res->status & NOUVEAU_BUFFER_STATUS_USER_MEMORY))
-      FREE(res->data);
+      align_free(res->data);
 
-   FREE(res);
-}
-
-/* Maybe just migrate to GART right away if we actually need to do this. */
-boolean
-nouveau_buffer_download(struct nouveau_context *nv, struct nv04_resource *buf,
-                        unsigned start, unsigned size)
-{
-   struct nouveau_mm_allocation *mm;
-   struct nouveau_bo *bounce = NULL;
-   uint32_t offset;
+   nouveau_fence_ref(NULL, &res->fence);
+   nouveau_fence_ref(NULL, &res->fence_wr);
 
-   assert(buf->domain == NOUVEAU_BO_VRAM);
+   util_range_destroy(&res->valid_buffer_range);
 
-   mm = nouveau_mm_allocate(nv->screen->mm_GART, size, &bounce, &offset);
-   if (!bounce)
-      return FALSE;
-
-   nv->copy_data(nv, bounce, offset, NOUVEAU_BO_GART,
-                 buf->bo, buf->offset + start, NOUVEAU_BO_VRAM, size);
-
-   if (nouveau_bo_map(bounce, NOUVEAU_BO_RD, nv->screen->client))
-      return FALSE;
-   memcpy(buf->data + start, (uint8_t *)bounce->map + offset, size);
-
-   buf->status &= ~NOUVEAU_BUFFER_STATUS_GPU_WRITING;
+   FREE(res);
 
-   nouveau_bo_ref(NULL, &bounce);
-   if (mm)
-      nouveau_mm_free(mm);
-   return TRUE;
+   NOUVEAU_DRV_STAT(nouveau_screen(pscreen), buf_obj_current_count, -1);
 }
 
-static boolean
-nouveau_buffer_upload(struct nouveau_context *nv, struct nv04_resource *buf,
-                      unsigned start, unsigned size)
+/* Set up a staging area for the transfer. This is either done in "regular"
+ * system memory if the driver supports push_data (nv50+) and the data is
+ * small enough (and permit_pb == true), or in GART memory.
+ */
+static uint8_t *
+nouveau_transfer_staging(struct nouveau_context *nv,
+                         struct nouveau_transfer *tx, bool permit_pb)
 {
-   struct nouveau_mm_allocation *mm;
-   struct nouveau_bo *bounce = NULL;
-   uint32_t offset;
-
-   if (size <= 192 && (nv->push_data || nv->push_cb)) {
-      if (buf->base.bind & PIPE_BIND_CONSTANT_BUFFER)
-         nv->push_cb(nv, buf->bo, buf->domain, buf->offset, buf->base.width0,
-                     start, size / 4, (const uint32_t *)(buf->data + start));
-      else
-         nv->push_data(nv, buf->bo, buf->offset + start, buf->domain,
-                       size, buf->data + start);
-      return TRUE;
-   }
+   const unsigned adj = tx->base.box.x & NOUVEAU_MIN_BUFFER_MAP_ALIGN_MASK;
+   const unsigned size = align(tx->base.box.width, 4) + adj;
 
-   mm = nouveau_mm_allocate(nv->screen->mm_GART, size, &bounce, &offset);
-   if (!bounce)
-      return FALSE;
+   if (!nv->push_data)
+      permit_pb = false;
 
-   nouveau_bo_map(bounce, 0, nv->screen->client);
-   memcpy((uint8_t *)bounce->map + offset, buf->data + start, size);
-
-   nv->copy_data(nv, buf->bo, buf->offset + start, NOUVEAU_BO_VRAM,
-                 bounce, offset, NOUVEAU_BO_GART, size);
-
-   nouveau_bo_ref(NULL, &bounce);
-   if (mm)
-      release_allocation(&mm, nv->screen->fence.current);
-
-   if (start == 0 && size == buf->base.width0)
-      buf->status &= ~NOUVEAU_BUFFER_STATUS_GPU_WRITING;
-   return TRUE;
+   if ((size <= nv->screen->transfer_pushbuf_threshold) && permit_pb) {
+      tx->map = align_malloc(size, NOUVEAU_MIN_BUFFER_MAP_ALIGN);
+      if (tx->map)
+         tx->map += adj;
+   } else {
+      tx->mm =
+         nouveau_mm_allocate(nv->screen->mm_GART, size, &tx->bo, &tx->offset);
+      if (tx->bo) {
+         tx->offset += adj;
+         if (!nouveau_bo_map(tx->bo, 0, NULL))
+            tx->map = (uint8_t *)tx->bo->map + tx->offset;
+      }
+   }
+   return tx->map;
 }
 
-static struct pipe_transfer *
-nouveau_buffer_transfer_get(struct pipe_context *pipe,
-                            struct pipe_resource *resource,
-                            unsigned level, unsigned usage,
-                            const struct pipe_box *box)
+/* Copies data from the resource into the transfer's temporary GART
+ * buffer. Also updates buf->data if present.
+ *
+ * Maybe just migrate to GART right away if we actually need to do this. */
+static bool
+nouveau_transfer_read(struct nouveau_context *nv, struct nouveau_transfer *tx)
 {
-   struct nv04_resource *buf = nv04_resource(resource);
-   struct nouveau_context *nv = nouveau_context(pipe);
-   struct nouveau_transfer *xfr = CALLOC_STRUCT(nouveau_transfer);
-   if (!xfr)
-      return NULL;
+   struct nv04_resource *buf = nv04_resource(tx->base.resource);
+   const unsigned base = tx->base.box.x;
+   const unsigned size = tx->base.box.width;
 
-   xfr->base.resource = resource;
-   xfr->base.box.x = box->x;
-   xfr->base.box.width = box->width;
-   xfr->base.usage = usage;
+   NOUVEAU_DRV_STAT(nv->screen, buf_read_bytes_staging_vid, size);
 
-   if (buf->domain == NOUVEAU_BO_VRAM) {
-      if (usage & PIPE_TRANSFER_READ) {
-         if (buf->status & NOUVEAU_BUFFER_STATUS_GPU_WRITING)
-            nouveau_buffer_download(nv, buf, 0, buf->base.width0);
-      }
-   }
+   nv->copy_data(nv, tx->bo, tx->offset, NOUVEAU_BO_GART,
+                 buf->bo, buf->offset + base, buf->domain, size);
+
+   if (nouveau_bo_wait(tx->bo, NOUVEAU_BO_RD, nv->client))
+      return false;
 
-   return &xfr->base;
+   if (buf->data)
+      memcpy(buf->data + base, tx->map, size);
+
+   return true;
 }
 
 static void
-nouveau_buffer_transfer_destroy(struct pipe_context *pipe,
-                                struct pipe_transfer *transfer)
+nouveau_transfer_write(struct nouveau_context *nv, struct nouveau_transfer *tx,
+                       unsigned offset, unsigned size)
 {
-   struct nv04_resource *buf = nv04_resource(transfer->resource);
-   struct nouveau_transfer *xfr = nouveau_transfer(transfer);
-   struct nouveau_context *nv = nouveau_context(pipe);
+   struct nv04_resource *buf = nv04_resource(tx->base.resource);
+   uint8_t *data = tx->map + offset;
+   const unsigned base = tx->base.box.x + offset;
+   const bool can_cb = !((base | size) & 3);
 
-   if (xfr->base.usage & PIPE_TRANSFER_WRITE) {
-      if (buf->domain == NOUVEAU_BO_VRAM) {
-         nouveau_buffer_upload(nv, buf, transfer->box.x, transfer->box.width);
-      }
+   if (buf->data)
+      memcpy(data, buf->data + base, size);
+   else
+      buf->status |= NOUVEAU_BUFFER_STATUS_DIRTY;
 
-      if (buf->domain != 0 && (buf->base.bind & (PIPE_BIND_VERTEX_BUFFER |
-                                                 PIPE_BIND_INDEX_BUFFER)))
-         nouveau_context(pipe)->vbo_dirty = TRUE;
-   }
+   if (buf->domain == NOUVEAU_BO_VRAM)
+      NOUVEAU_DRV_STAT(nv->screen, buf_write_bytes_staging_vid, size);
+   if (buf->domain == NOUVEAU_BO_GART)
+      NOUVEAU_DRV_STAT(nv->screen, buf_write_bytes_staging_sys, size);
 
-   FREE(xfr);
+   if (tx->bo)
+      nv->copy_data(nv, buf->bo, buf->offset + base, buf->domain,
+                    tx->bo, tx->offset + offset, NOUVEAU_BO_GART, size);
+   else
+   if (nv->push_cb && can_cb)
+      nv->push_cb(nv, buf,
+                  base, size / 4, (const uint32_t *)data);
+   else
+      nv->push_data(nv, buf->bo, buf->offset + base, buf->domain, size, data);
+
+   nouveau_fence_ref(nv->screen->fence.current, &buf->fence);
+   nouveau_fence_ref(nv->screen->fence.current, &buf->fence_wr);
 }
 
-static INLINE boolean
-nouveau_buffer_sync(struct nv04_resource *buf, unsigned rw)
+/* Does a CPU wait for the buffer's backing data to become reliably accessible
+ * for write/read by waiting on the buffer's relevant fences.
+ */
+static inline bool
+nouveau_buffer_sync(struct nouveau_context *nv,
+                    struct nv04_resource *buf, unsigned rw)
 {
    if (rw == PIPE_TRANSFER_READ) {
       if (!buf->fence_wr)
-         return TRUE;
-      if (!nouveau_fence_wait(buf->fence_wr))
-         return FALSE;
+         return true;
+      NOUVEAU_DRV_STAT_RES(buf, buf_non_kernel_fence_sync_count,
+                           !nouveau_fence_signalled(buf->fence_wr));
+      if (!nouveau_fence_wait(buf->fence_wr, &nv->debug))
+         return false;
    } else {
       if (!buf->fence)
-         return TRUE;
-      if (!nouveau_fence_wait(buf->fence))
-         return FALSE;
+         return true;
+      NOUVEAU_DRV_STAT_RES(buf, buf_non_kernel_fence_sync_count,
+                           !nouveau_fence_signalled(buf->fence));
+      if (!nouveau_fence_wait(buf->fence, &nv->debug))
+         return false;
 
       nouveau_fence_ref(NULL, &buf->fence);
    }
    nouveau_fence_ref(NULL, &buf->fence_wr);
 
-   return TRUE;
+   return true;
 }
 
-static INLINE boolean
+static inline bool
 nouveau_buffer_busy(struct nv04_resource *buf, unsigned rw)
 {
    if (rw == PIPE_TRANSFER_READ)
@@ -242,41 +257,248 @@ nouveau_buffer_busy(struct nv04_resource *buf, unsigned rw)
       return (buf->fence && !nouveau_fence_signalled(buf->fence));
 }
 
+static inline void
+nouveau_buffer_transfer_init(struct nouveau_transfer *tx,
+                             struct pipe_resource *resource,
+                             const struct pipe_box *box,
+                             unsigned usage)
+{
+   tx->base.resource = resource;
+   tx->base.level = 0;
+   tx->base.usage = usage;
+   tx->base.box.x = box->x;
+   tx->base.box.y = 0;
+   tx->base.box.z = 0;
+   tx->base.box.width = box->width;
+   tx->base.box.height = 1;
+   tx->base.box.depth = 1;
+   tx->base.stride = 0;
+   tx->base.layer_stride = 0;
+
+   tx->bo = NULL;
+   tx->map = NULL;
+}
+
+static inline void
+nouveau_buffer_transfer_del(struct nouveau_context *nv,
+                            struct nouveau_transfer *tx)
+{
+   if (tx->map) {
+      if (likely(tx->bo)) {
+         nouveau_fence_work(nv->screen->fence.current,
+                            nouveau_fence_unref_bo, tx->bo);
+         if (tx->mm)
+            release_allocation(&tx->mm, nv->screen->fence.current);
+      } else {
+         align_free(tx->map -
+                    (tx->base.box.x & NOUVEAU_MIN_BUFFER_MAP_ALIGN_MASK));
+      }
+   }
+}
+
+/* Creates a cache in system memory of the buffer data. */
+static bool
+nouveau_buffer_cache(struct nouveau_context *nv, struct nv04_resource *buf)
+{
+   struct nouveau_transfer tx;
+   bool ret;
+   tx.base.resource = &buf->base;
+   tx.base.box.x = 0;
+   tx.base.box.width = buf->base.width0;
+   tx.bo = NULL;
+   tx.map = NULL;
+
+   if (!buf->data)
+      if (!nouveau_buffer_malloc(buf))
+         return false;
+   if (!(buf->status & NOUVEAU_BUFFER_STATUS_DIRTY))
+      return true;
+   nv->stats.buf_cache_count++;
+
+   if (!nouveau_transfer_staging(nv, &tx, false))
+      return false;
+
+   ret = nouveau_transfer_read(nv, &tx);
+   if (ret) {
+      buf->status &= ~NOUVEAU_BUFFER_STATUS_DIRTY;
+      memcpy(buf->data, tx.map, buf->base.width0);
+   }
+   nouveau_buffer_transfer_del(nv, &tx);
+   return ret;
+}
+
+
+#define NOUVEAU_TRANSFER_DISCARD \
+   (PIPE_TRANSFER_DISCARD_RANGE | PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE)
+
+/* Checks whether it is possible to completely discard the memory backing this
+ * resource. This can be useful if we would otherwise have to wait for a read
+ * operation to complete on this data.
+ */
+static inline bool
+nouveau_buffer_should_discard(struct nv04_resource *buf, unsigned usage)
+{
+   if (!(usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE))
+      return false;
+   if (unlikely(buf->base.bind & PIPE_BIND_SHARED))
+      return false;
+   if (unlikely(usage & PIPE_TRANSFER_PERSISTENT))
+      return false;
+   return buf->mm && nouveau_buffer_busy(buf, PIPE_TRANSFER_WRITE);
+}
+
+/* Returns a pointer to a memory area representing a window into the
+ * resource's data.
+ *
+ * This may or may not be the _actual_ memory area of the resource. However
+ * when calling nouveau_buffer_transfer_unmap, if it wasn't the actual memory
+ * area, the contents of the returned map are copied over to the resource.
+ *
+ * The usage indicates what the caller plans to do with the map:
+ *
+ *   WRITE means that the user plans to write to it
+ *
+ *   READ means that the user plans on reading from it
+ *
+ *   DISCARD_WHOLE_RESOURCE means that the whole resource is going to be
+ *   potentially overwritten, and even if it isn't, the bits that aren't don't
+ *   need to be maintained.
+ *
+ *   DISCARD_RANGE means that all the data in the specified range is going to
+ *   be overwritten.
+ *
+ * The strategy for determining what kind of memory area to return is complex,
+ * see comments inside of the function.
+ */
 static void *
 nouveau_buffer_transfer_map(struct pipe_context *pipe,
-                            struct pipe_transfer *transfer)
+                            struct pipe_resource *resource,
+                            unsigned level, unsigned usage,
+                            const struct pipe_box *box,
+                            struct pipe_transfer **ptransfer)
 {
    struct nouveau_context *nv = nouveau_context(pipe);
-   struct nouveau_transfer *xfr = nouveau_transfer(transfer);
-   struct nv04_resource *buf = nv04_resource(transfer->resource);
-   struct nouveau_bo *bo = buf->bo;
+   struct nv04_resource *buf = nv04_resource(resource);
+   struct nouveau_transfer *tx = MALLOC_STRUCT(nouveau_transfer);
    uint8_t *map;
    int ret;
-   uint32_t offset = xfr->base.box.x;
-   uint32_t flags = 0;
 
-   if (buf->domain != NOUVEAU_BO_GART)
-      return buf->data + offset;
+   if (!tx)
+      return NULL;
+   nouveau_buffer_transfer_init(tx, resource, box, usage);
+   *ptransfer = &tx->base;
+
+   if (usage & PIPE_TRANSFER_READ)
+      NOUVEAU_DRV_STAT(nv->screen, buf_transfers_rd, 1);
+   if (usage & PIPE_TRANSFER_WRITE)
+      NOUVEAU_DRV_STAT(nv->screen, buf_transfers_wr, 1);
+
+   /* If we are trying to write to an uninitialized range, the user shouldn't
+    * care what was there before. So we can treat the write as if the target
+    * range were being discarded. Furthermore, since we know that even if this
+    * buffer is busy due to GPU activity, because the contents were
+    * uninitialized, the GPU can't care what was there, and so we can treat
+    * the write as being unsynchronized.
+    */
+   if ((usage & PIPE_TRANSFER_WRITE) &&
+       !util_ranges_intersect(&buf->valid_buffer_range, box->x, box->x + box->width))
+      usage |= PIPE_TRANSFER_DISCARD_RANGE | PIPE_TRANSFER_UNSYNCHRONIZED;
 
-   if (!buf->mm)
-      flags = nouveau_screen_transfer_flags(xfr->base.usage);
+   if (buf->domain == NOUVEAU_BO_VRAM) {
+      if (usage & NOUVEAU_TRANSFER_DISCARD) {
+         /* Set up a staging area for the user to write to. It will be copied
+          * back into VRAM on unmap. */
+         if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE)
+            buf->status &= NOUVEAU_BUFFER_STATUS_REALLOC_MASK;
+         nouveau_transfer_staging(nv, tx, true);
+      } else {
+         if (buf->status & NOUVEAU_BUFFER_STATUS_GPU_WRITING) {
+            /* The GPU is currently writing to this buffer. Copy its current
+             * contents to a staging area in the GART. This is necessary since
+             * not the whole area being mapped is being discarded.
+             */
+            if (buf->data) {
+               align_free(buf->data);
+               buf->data = NULL;
+            }
+            nouveau_transfer_staging(nv, tx, false);
+            nouveau_transfer_read(nv, tx);
+         } else {
+            /* The buffer is currently idle. Create a staging area for writes,
+             * and make sure that the cached data is up-to-date. */
+            if (usage & PIPE_TRANSFER_WRITE)
+               nouveau_transfer_staging(nv, tx, true);
+            if (!buf->data)
+               nouveau_buffer_cache(nv, buf);
+         }
+      }
+      return buf->data ? (buf->data + box->x) : tx->map;
+   } else
+   if (unlikely(buf->domain == 0)) {
+      return buf->data + box->x;
+   }
 
-   offset += buf->offset;
+   /* At this point, buf->domain == GART */
 
-   ret = nouveau_bo_map(buf->bo, flags, nv->screen->client);
-   if (ret)
-      return NULL;
-   map = (uint8_t *)bo->map + offset;
+   if (nouveau_buffer_should_discard(buf, usage)) {
+      int ref = buf->base.reference.count - 1;
+      nouveau_buffer_reallocate(nv->screen, buf, buf->domain);
+      if (ref > 0) /* any references inside context possible ? */
+         nv->invalidate_resource_storage(nv, &buf->base, ref);
+   }
 
-   if (buf->mm) {
-      if (xfr->base.usage & PIPE_TRANSFER_DONTBLOCK) {
-         if (nouveau_buffer_busy(buf, xfr->base.usage & PIPE_TRANSFER_READ_WRITE))
-            return NULL;
+   /* Note that nouveau_bo_map ends up doing a nouveau_bo_wait with the
+    * relevant flags. If buf->mm is set, that means this resource is part of a
+    * larger slab bo that holds multiple resources. So in that case, don't
+    * wait on the whole slab and instead use the logic below to return a
+    * reasonable buffer for that case.
+    */
+   ret = nouveau_bo_map(buf->bo,
+                        buf->mm ? 0 : nouveau_screen_transfer_flags(usage),
+                        nv->client);
+   if (ret) {
+      FREE(tx);
+      return NULL;
+   }
+   map = (uint8_t *)buf->bo->map + buf->offset + box->x;
+
+   /* using kernel fences only if !buf->mm */
+   if ((usage & PIPE_TRANSFER_UNSYNCHRONIZED) || !buf->mm)
+      return map;
+
+   /* If the GPU is currently reading/writing this buffer, we shouldn't
+    * interfere with its progress. So instead we either wait for the GPU to
+    * complete its operation, or set up a staging area to perform our work in.
+    */
+   if (nouveau_buffer_busy(buf, usage & PIPE_TRANSFER_READ_WRITE)) {
+      if (unlikely(usage & (PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE |
+                            PIPE_TRANSFER_PERSISTENT))) {
+         /* Discarding was not possible, must sync because
+          * subsequent transfers might use UNSYNCHRONIZED. */
+         nouveau_buffer_sync(nv, buf, usage & PIPE_TRANSFER_READ_WRITE);
       } else
-      if (!(xfr->base.usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
-         nouveau_buffer_sync(buf, xfr->base.usage & PIPE_TRANSFER_READ_WRITE);
+      if (usage & PIPE_TRANSFER_DISCARD_RANGE) {
+         /* The whole range is being discarded, so it doesn't matter what was
+          * there before. No need to copy anything over. */
+         nouveau_transfer_staging(nv, tx, true);
+         map = tx->map;
+      } else
+      if (nouveau_buffer_busy(buf, PIPE_TRANSFER_READ)) {
+         if (usage & PIPE_TRANSFER_DONTBLOCK)
+            map = NULL;
+         else
+            nouveau_buffer_sync(nv, buf, usage & PIPE_TRANSFER_READ_WRITE);
+      } else {
+         /* It is expected that the returned buffer be a representation of the
+          * data in question, so we must copy it over from the buffer. */
+         nouveau_transfer_staging(nv, tx, true);
+         if (tx->map)
+            memcpy(tx->map, map, box->width);
+         map = tx->map;
       }
    }
+   if (!map)
+      FREE(tx);
    return map;
 }
 
@@ -287,23 +509,88 @@ nouveau_buffer_transfer_flush_region(struct pipe_context *pipe,
                                      struct pipe_transfer *transfer,
                                      const struct pipe_box *box)
 {
-#if 0
-   struct nv04_resource *res = nv04_resource(transfer->resource);
-   struct nouveau_bo *bo = res->bo;
-   unsigned offset = res->offset + transfer->box.x + box->x;
+   struct nouveau_transfer *tx = nouveau_transfer(transfer);
+   struct nv04_resource *buf = nv04_resource(transfer->resource);
 
-   /* not using non-snoop system memory yet, no need for cflush */
-   if (1)
-      return;
+   if (tx->map)
+      nouveau_transfer_write(nouveau_context(pipe), tx, box->x, box->width);
 
-   /* XXX: maybe need to upload for VRAM buffers here */
-#endif
+   util_range_add(&buf->valid_buffer_range,
+                  tx->base.box.x + box->x,
+                  tx->base.box.x + box->x + box->width);
 }
 
+/* Unmap stage of the transfer. If it was a WRITE transfer and the map that
+ * was returned was not the real resource's data, this needs to transfer the
+ * data back to the resource.
+ *
+ * Also marks vbo dirty based on the buffer's binding
+ */
 static void
 nouveau_buffer_transfer_unmap(struct pipe_context *pipe,
                               struct pipe_transfer *transfer)
 {
+   struct nouveau_context *nv = nouveau_context(pipe);
+   struct nouveau_transfer *tx = nouveau_transfer(transfer);
+   struct nv04_resource *buf = nv04_resource(transfer->resource);
+
+   if (tx->base.usage & PIPE_TRANSFER_WRITE) {
+      if (!(tx->base.usage & PIPE_TRANSFER_FLUSH_EXPLICIT)) {
+         if (tx->map)
+            nouveau_transfer_write(nv, tx, 0, tx->base.box.width);
+
+         util_range_add(&buf->valid_buffer_range,
+                        tx->base.box.x, tx->base.box.x + tx->base.box.width);
+      }
+
+      if (likely(buf->domain)) {
+         const uint8_t bind = buf->base.bind;
+         /* make sure we invalidate dedicated caches */
+         if (bind & (PIPE_BIND_VERTEX_BUFFER | PIPE_BIND_INDEX_BUFFER))
+            nv->vbo_dirty = true;
+      }
+   }
+
+   if (!tx->bo && (tx->base.usage & PIPE_TRANSFER_WRITE))
+      NOUVEAU_DRV_STAT(nv->screen, buf_write_bytes_direct, tx->base.box.width);
+
+   nouveau_buffer_transfer_del(nv, tx);
+   FREE(tx);
+}
+
+
+void
+nouveau_copy_buffer(struct nouveau_context *nv,
+                    struct nv04_resource *dst, unsigned dstx,
+                    struct nv04_resource *src, unsigned srcx, unsigned size)
+{
+   assert(dst->base.target == PIPE_BUFFER && src->base.target == PIPE_BUFFER);
+
+   if (likely(dst->domain) && likely(src->domain)) {
+      nv->copy_data(nv,
+                    dst->bo, dst->offset + dstx, dst->domain,
+                    src->bo, src->offset + srcx, src->domain, size);
+
+      dst->status |= NOUVEAU_BUFFER_STATUS_GPU_WRITING;
+      nouveau_fence_ref(nv->screen->fence.current, &dst->fence);
+      nouveau_fence_ref(nv->screen->fence.current, &dst->fence_wr);
+
+      src->status |= NOUVEAU_BUFFER_STATUS_GPU_READING;
+      nouveau_fence_ref(nv->screen->fence.current, &src->fence);
+   } else {
+      struct pipe_box src_box;
+      src_box.x = srcx;
+      src_box.y = 0;
+      src_box.z = 0;
+      src_box.width = size;
+      src_box.height = 1;
+      src_box.depth = 1;
+      util_resource_copy_region(&nv->pipe,
+                                &dst->base, 0, dstx, 0, 0,
+                                &src->base, 0, &src_box);
+   }
+
+   util_range_add(&dst->valid_buffer_range, dstx, dstx + size);
 }
 
 
@@ -312,22 +599,24 @@ nouveau_resource_map_offset(struct nouveau_context *nv,
                             struct nv04_resource *res, uint32_t offset,
                             uint32_t flags)
 {
-   if ((res->domain == NOUVEAU_BO_VRAM) &&
-       (res->status & NOUVEAU_BUFFER_STATUS_GPU_WRITING))
-      nouveau_buffer_download(nv, res, 0, res->base.width0);
+   if (unlikely(res->status & NOUVEAU_BUFFER_STATUS_USER_MEMORY))
+      return res->data + offset;
 
-   if ((res->domain != NOUVEAU_BO_GART) ||
-       (res->status & NOUVEAU_BUFFER_STATUS_USER_MEMORY))
+   if (res->domain == NOUVEAU_BO_VRAM) {
+      if (!res->data || (res->status & NOUVEAU_BUFFER_STATUS_GPU_WRITING))
+         nouveau_buffer_cache(nv, res);
+   }
+   if (res->domain != NOUVEAU_BO_GART)
       return res->data + offset;
 
    if (res->mm) {
       unsigned rw;
       rw = (flags & NOUVEAU_BO_WR) ? PIPE_TRANSFER_WRITE : PIPE_TRANSFER_READ;
-      nouveau_buffer_sync(res, rw);
+      nouveau_buffer_sync(nv, res, rw);
       if (nouveau_bo_map(res->bo, 0, NULL))
          return NULL;
    } else {
-      if (nouveau_bo_map(res->bo, flags, nv->screen->client))
+      if (nouveau_bo_map(res->bo, flags, nv->client))
          return NULL;
    }
    return (uint8_t *)res->bo->map + res->offset + offset;
@@ -338,12 +627,9 @@ const struct u_resource_vtbl nouveau_buffer_vtbl =
 {
    u_default_resource_get_handle,     /* get_handle */
    nouveau_buffer_destroy,               /* resource_destroy */
-   nouveau_buffer_transfer_get,          /* get_transfer */
-   nouveau_buffer_transfer_destroy,      /* transfer_destroy */
    nouveau_buffer_transfer_map,          /* transfer_map */
    nouveau_buffer_transfer_flush_region, /* transfer_flush_region */
    nouveau_buffer_transfer_unmap,        /* transfer_unmap */
-   u_default_transfer_inline_write    /* transfer_inline_write */
 };
 
 struct pipe_resource *
@@ -352,7 +638,7 @@ nouveau_buffer_create(struct pipe_screen *pscreen,
 {
    struct nouveau_screen *screen = nouveau_screen(pscreen);
    struct nv04_resource *buffer;
-   boolean ret;
+   bool ret;
 
    buffer = CALLOC_STRUCT(nv04_resource);
    if (!buffer)
@@ -363,14 +649,50 @@ nouveau_buffer_create(struct pipe_screen *pscreen,
    pipe_reference_init(&buffer->base.reference, 1);
    buffer->base.screen = pscreen;
 
-   if ((buffer->base.bind & screen->sysmem_bindings) == screen->sysmem_bindings)
-      ret = nouveau_buffer_allocate(screen, buffer, 0);
-   else
-      ret = nouveau_buffer_allocate(screen, buffer, NOUVEAU_BO_GART);
+   if (buffer->base.flags & (PIPE_RESOURCE_FLAG_MAP_PERSISTENT |
+                             PIPE_RESOURCE_FLAG_MAP_COHERENT)) {
+      buffer->domain = NOUVEAU_BO_GART;
+   } else if (buffer->base.bind == 0 || (buffer->base.bind &
+              (screen->vidmem_bindings & screen->sysmem_bindings))) {
+      switch (buffer->base.usage) {
+      case PIPE_USAGE_DEFAULT:
+      case PIPE_USAGE_IMMUTABLE:
+         buffer->domain = NV_VRAM_DOMAIN(screen);
+         break;
+      case PIPE_USAGE_DYNAMIC:
+         /* For most apps, we'd have to do staging transfers to avoid sync
+          * with this usage, and GART -> GART copies would be suboptimal.
+          */
+         buffer->domain = NV_VRAM_DOMAIN(screen);
+         break;
+      case PIPE_USAGE_STAGING:
+      case PIPE_USAGE_STREAM:
+         buffer->domain = NOUVEAU_BO_GART;
+         break;
+      default:
+         assert(0);
+         break;
+      }
+   } else {
+      if (buffer->base.bind & screen->vidmem_bindings)
+         buffer->domain = NV_VRAM_DOMAIN(screen);
+      else
+      if (buffer->base.bind & screen->sysmem_bindings)
+         buffer->domain = NOUVEAU_BO_GART;
+   }
 
-   if (ret == FALSE)
+   ret = nouveau_buffer_allocate(screen, buffer, buffer->domain);
+
+   if (ret == false)
       goto fail;
 
+   if (buffer->domain == NOUVEAU_BO_VRAM && screen->hint_buf_keep_sysmem_copy)
+      nouveau_buffer_cache(NULL, buffer);
+
+   NOUVEAU_DRV_STAT(screen, buf_obj_current_count, 1);
+
+   util_range_init(&buffer->valid_buffer_range);
+
    return &buffer->base;
 
 fail:
@@ -402,28 +724,26 @@ nouveau_user_buffer_create(struct pipe_screen *pscreen, void *ptr,
    buffer->data = ptr;
    buffer->status = NOUVEAU_BUFFER_STATUS_USER_MEMORY;
 
+   util_range_init(&buffer->valid_buffer_range);
+   util_range_add(&buffer->valid_buffer_range, 0, bytes);
+
    return &buffer->base;
 }
 
-/* Like download, but for GART buffers. Merge ? */
-static INLINE boolean
+static inline bool
 nouveau_buffer_data_fetch(struct nouveau_context *nv, struct nv04_resource *buf,
                           struct nouveau_bo *bo, unsigned offset, unsigned size)
 {
-   if (!buf->data) {
-      buf->data = MALLOC(size);
-      if (!buf->data)
-         return FALSE;
-   }
-   if (nouveau_bo_map(bo, NOUVEAU_BO_RD, nv->screen->client))
-      return FALSE;
+   if (!nouveau_buffer_malloc(buf))
+      return false;
+   if (nouveau_bo_map(bo, NOUVEAU_BO_RD, nv->client))
+      return false;
    memcpy(buf->data, (uint8_t *)bo->map + offset, size);
-
-   return TRUE;
+   return true;
 }
 
 /* Migrate a linear buffer (vertex, index, constants) USER -> GART -> VRAM. */
-boolean
+bool
 nouveau_buffer_migrate(struct nouveau_context *nv,
                        struct nv04_resource *buf, const unsigned new_domain)
 {
@@ -438,12 +758,12 @@ nouveau_buffer_migrate(struct nouveau_context *nv,
 
    if (new_domain == NOUVEAU_BO_GART && old_domain == 0) {
       if (!nouveau_buffer_allocate(screen, buf, new_domain))
-         return FALSE;
-      ret = nouveau_bo_map(buf->bo, 0, nv->screen->client);
+         return false;
+      ret = nouveau_bo_map(buf->bo, 0, nv->client);
       if (ret)
          return ret;
       memcpy((uint8_t *)buf->bo->map + buf->offset, buf->data, size);
-      FREE(buf->data);
+      align_free(buf->data);
    } else
    if (old_domain != 0 && new_domain != 0) {
       struct nouveau_mm_allocation *mm = buf->mm;
@@ -451,7 +771,7 @@ nouveau_buffer_migrate(struct nouveau_context *nv,
       if (new_domain == NOUVEAU_BO_VRAM) {
          /* keep a system memory copy of our data in case we hit a fallback */
          if (!nouveau_buffer_data_fetch(nv, buf, buf->bo, buf->offset, size))
-            return FALSE;
+            return false;
          if (nouveau_mesa_debug)
             debug_printf("migrating %u KiB to VRAM\n", size / 1024);
       }
@@ -465,27 +785,35 @@ nouveau_buffer_migrate(struct nouveau_context *nv,
       nv->copy_data(nv, buf->bo, buf->offset, new_domain,
                     bo, offset, old_domain, buf->base.width0);
 
-      nouveau_bo_ref(NULL, &bo);
+      nouveau_fence_work(screen->fence.current, nouveau_fence_unref_bo, bo);
       if (mm)
          release_allocation(&mm, screen->fence.current);
    } else
    if (new_domain == NOUVEAU_BO_VRAM && old_domain == 0) {
+      struct nouveau_transfer tx;
       if (!nouveau_buffer_allocate(screen, buf, NOUVEAU_BO_VRAM))
-         return FALSE;
-      if (!nouveau_buffer_upload(nv, buf, 0, buf->base.width0))
-         return FALSE;
+         return false;
+      tx.base.resource = &buf->base;
+      tx.base.box.x = 0;
+      tx.base.box.width = buf->base.width0;
+      tx.bo = NULL;
+      tx.map = NULL;
+      if (!nouveau_transfer_staging(nv, &tx, false))
+         return false;
+      nouveau_transfer_write(nv, &tx, 0, tx.base.box.width);
+      nouveau_buffer_transfer_del(nv, &tx);
    } else
-      return FALSE;
+      return false;
 
    assert(buf->domain == new_domain);
-   return TRUE;
+   return true;
 }
 
 /* Migrate data from glVertexAttribPointer(non-VBO) user buffers to GART.
  * We'd like to only allocate @size bytes here, but then we'd have to rebase
  * the vertex indices ...
  */
-boolean
+bool
 nouveau_user_buffer_upload(struct nouveau_context *nv,
                            struct nv04_resource *buf,
                            unsigned base, unsigned size)
@@ -497,20 +825,53 @@ nouveau_user_buffer_upload(struct nouveau_context *nv,
 
    buf->base.width0 = base + size;
    if (!nouveau_buffer_reallocate(screen, buf, NOUVEAU_BO_GART))
-      return FALSE;
+      return false;
 
-   ret = nouveau_bo_map(buf->bo, 0, nv->screen->client);
+   ret = nouveau_bo_map(buf->bo, 0, nv->client);
    if (ret)
-      return FALSE;
+      return false;
    memcpy((uint8_t *)buf->bo->map + buf->offset + base, buf->data + base, size);
 
-   return TRUE;
+   return true;
+}
+
+/* Invalidate underlying buffer storage, reset fences, reallocate to non-busy
+ * buffer.
+ */
+void
+nouveau_buffer_invalidate(struct pipe_context *pipe,
+                          struct pipe_resource *resource)
+{
+   struct nouveau_context *nv = nouveau_context(pipe);
+   struct nv04_resource *buf = nv04_resource(resource);
+   int ref = buf->base.reference.count - 1;
+
+   /* Shared buffers shouldn't get reallocated */
+   if (unlikely(buf->base.bind & PIPE_BIND_SHARED))
+      return;
+
+   /* We can't touch persistent/coherent buffers */
+   if (buf->base.flags & (PIPE_RESOURCE_FLAG_MAP_PERSISTENT |
+                          PIPE_RESOURCE_FLAG_MAP_COHERENT))
+      return;
+
+   /* If the buffer is sub-allocated and not currently being written, just
+    * wipe the valid buffer range. Otherwise we have to create fresh
+    * storage. (We don't keep track of fences for non-sub-allocated BO's.)
+    */
+   if (buf->mm && !nouveau_buffer_busy(buf, PIPE_TRANSFER_WRITE)) {
+      util_range_set_empty(&buf->valid_buffer_range);
+   } else {
+      nouveau_buffer_reallocate(nv->screen, buf, buf->domain);
+      if (ref > 0) /* any references inside context possible ? */
+         nv->invalidate_resource_storage(nv, &buf->base, ref);
+   }
 }
 
 
 /* Scratch data allocation. */
 
-static INLINE int
+static inline int
 nouveau_scratch_bo_alloc(struct nouveau_context *nv, struct nouveau_bo **pbo,
                          unsigned size)
 {
@@ -518,17 +879,28 @@ nouveau_scratch_bo_alloc(struct nouveau_context *nv, struct nouveau_bo **pbo,
                          4096, size, NULL, pbo);
 }
 
+static void
+nouveau_scratch_unref_bos(void *d)
+{
+   struct runout *b = d;
+   int i;
+
+   for (i = 0; i < b->nr; ++i)
+      nouveau_bo_ref(NULL, &b->bo[i]);
+
+   FREE(b);
+}
+
 void
 nouveau_scratch_runout_release(struct nouveau_context *nv)
 {
-   if (!nv->scratch.nr_runout)
+   if (!nv->scratch.runout)
+      return;
+
+   if (!nouveau_fence_work(nv->screen->fence.current, nouveau_scratch_unref_bos,
+         nv->scratch.runout))
       return;
-   do {
-      --nv->scratch.nr_runout;
-      nouveau_bo_ref(NULL, &nv->scratch.runout[nv->scratch.nr_runout]);
-   } while (nv->scratch.nr_runout);
 
-   FREE(nv->scratch.runout);
    nv->scratch.end = 0;
    nv->scratch.runout = NULL;
 }
@@ -536,25 +908,30 @@ nouveau_scratch_runout_release(struct nouveau_context *nv)
 /* Allocate an extra bo if we can't fit everything we need simultaneously.
  * (Could happen for very large user arrays.)
  */
-static INLINE boolean
+static inline bool
 nouveau_scratch_runout(struct nouveau_context *nv, unsigned size)
 {
    int ret;
-   const unsigned n = nv->scratch.nr_runout++;
-
-   nv->scratch.runout = REALLOC(nv->scratch.runout,
-                                (n + 0) * sizeof(*nv->scratch.runout),
-                                (n + 1) * sizeof(*nv->scratch.runout));
-   nv->scratch.runout[n] = NULL;
+   unsigned n;
 
-   ret = nouveau_scratch_bo_alloc(nv, &nv->scratch.runout[n], size);
+   if (nv->scratch.runout)
+      n = nv->scratch.runout->nr;
+   else
+      n = 0;
+   nv->scratch.runout = REALLOC(nv->scratch.runout, n == 0 ? 0 :
+                                (sizeof(*nv->scratch.runout) + (n + 0) * sizeof(void *)),
+                                 sizeof(*nv->scratch.runout) + (n + 1) * sizeof(void *));
+   nv->scratch.runout->nr = n + 1;
+   nv->scratch.runout->bo[n] = NULL;
+
+   ret = nouveau_scratch_bo_alloc(nv, &nv->scratch.runout->bo[n], size);
    if (!ret) {
-      ret = nouveau_bo_map(nv->scratch.runout[n], 0, NULL);
+      ret = nouveau_bo_map(nv->scratch.runout->bo[n], 0, NULL);
       if (ret)
-         nouveau_bo_ref(NULL, &nv->scratch.runout[--nv->scratch.nr_runout]);
+         nouveau_bo_ref(NULL, &nv->scratch.runout->bo[--nv->scratch.runout->nr]);
    }
    if (!ret) {
-      nv->scratch.current = nv->scratch.runout[n];
+      nv->scratch.current = nv->scratch.runout->bo[n];
       nv->scratch.offset = 0;
       nv->scratch.end = size;
       nv->scratch.map = nv->scratch.current->map;
@@ -565,7 +942,7 @@ nouveau_scratch_runout(struct nouveau_context *nv, unsigned size)
 /* Continue to next scratch buffer, if available (no wrapping, large enough).
  * Allocate it if it has not yet been created.
  */
-static INLINE boolean
+static inline bool
 nouveau_scratch_next(struct nouveau_context *nv, unsigned size)
 {
    struct nouveau_bo *bo;
@@ -573,30 +950,30 @@ nouveau_scratch_next(struct nouveau_context *nv, unsigned size)
    const unsigned i = (nv->scratch.id + 1) % NOUVEAU_MAX_SCRATCH_BUFS;
 
    if ((size > nv->scratch.bo_size) || (i == nv->scratch.wrap))
-      return FALSE;
+      return false;
    nv->scratch.id = i;
 
    bo = nv->scratch.bo[i];
    if (!bo) {
       ret = nouveau_scratch_bo_alloc(nv, &bo, nv->scratch.bo_size);
       if (ret)
-         return FALSE;
+         return false;
       nv->scratch.bo[i] = bo;
    }
    nv->scratch.current = bo;
    nv->scratch.offset = 0;
    nv->scratch.end = nv->scratch.bo_size;
 
-   ret = nouveau_bo_map(bo, NOUVEAU_BO_WR, nv->screen->client);
+   ret = nouveau_bo_map(bo, NOUVEAU_BO_WR, nv->client);
    if (!ret)
       nv->scratch.map = bo->map;
    return !ret;
 }
 
-static boolean
+static bool
 nouveau_scratch_more(struct nouveau_context *nv, unsigned min_size)
 {
-   boolean ret;
+   bool ret;
 
    ret = nouveau_scratch_next(nv, min_size);
    if (!ret)
@@ -604,30 +981,28 @@ nouveau_scratch_more(struct nouveau_context *nv, unsigned min_size)
    return ret;
 }
 
-/* Upload data to scratch memory and update buffer address.
- * Returns the bo the data resides in, if successful.
- */
-struct nouveau_bo *
+
+/* Copy data to a scratch buffer and return address & bo the data resides in. */
+uint64_t
 nouveau_scratch_data(struct nouveau_context *nv,
-                     struct nv04_resource *buf, unsigned base, unsigned size)
+                     const void *data, unsigned base, unsigned size,
+                     struct nouveau_bo **bo)
 {
-   struct nouveau_bo *bo;
    unsigned bgn = MAX2(base, nv->scratch.offset);
    unsigned end = bgn + size;
 
    if (end >= nv->scratch.end) {
       end = base + size;
       if (!nouveau_scratch_more(nv, end))
-         return NULL;
+         return 0;
       bgn = base;
    }
    nv->scratch.offset = align(end, 4);
 
-   memcpy(nv->scratch.map + bgn, buf->data + base, size);
+   memcpy(nv->scratch.map + bgn, (const uint8_t *)data + base, size);
 
-   bo = nv->scratch.current;
-   buf->address = bo->offset + (bgn - base);
-   return bo;
+   *bo = nv->scratch.current;
+   return (*bo)->offset + (bgn - base);
 }
 
 void *