nv50: move HW queries to nv50_query_hw.c/h files

Signed-off-by: Samuel Pitoiset <[email protected]>
Reviewed-by: Pierre Moreau <[email protected]>

8 files changed, 476 insertions, 349 deletions

diff --git a/src/gallium/drivers/nouveau/Makefile.sources b/src/gallium/drivers/nouveau/Makefile.sources
index 06d9d979e34..83f81135590 100644
--- a/src/gallium/drivers/nouveau/Makefile.sources
+++ b/src/gallium/drivers/nouveau/Makefile.sources
@@ -74,6 +74,8 @@ NV50_C_SOURCES := \
 	nv50/nv50_push.c \
 	nv50/nv50_query.c \
 	nv50/nv50_query.h \
+	nv50/nv50_query_hw.c \
+	nv50/nv50_query_hw.h \
 	nv50/nv50_resource.c \
 	nv50/nv50_resource.h \
 	nv50/nv50_screen.c \
diff --git a/src/gallium/drivers/nouveau/nv50/nv50_query.c b/src/gallium/drivers/nouveau/nv50/nv50_query.c
index 1b4abdbb99b..dd9b85b7208 100644
--- a/src/gallium/drivers/nouveau/nv50/nv50_query.c
+++ b/src/gallium/drivers/nouveau/nv50/nv50_query.c
@@ -26,334 +26,45 @@
 
 #include "nv50/nv50_context.h"
 #include "nv50/nv50_query.h"
-#include "nv_object.xml.h"
-
-#define NV50_QUERY_STATE_READY   0
-#define NV50_QUERY_STATE_ACTIVE  1
-#define NV50_QUERY_STATE_ENDED   2
-#define NV50_QUERY_STATE_FLUSHED 3
-
-/* XXX: Nested queries, and simultaneous queries on multiple gallium contexts
- * (since we use only a single GPU channel per screen) will not work properly.
- *
- * The first is not that big of an issue because OpenGL does not allow nested
- * queries anyway.
- */
-
-#define NV50_QUERY_ALLOC_SPACE 256
-
-static bool
-nv50_query_allocate(struct nv50_context *nv50, struct nv50_query *q, int size)
-{
-   struct nv50_screen *screen = nv50->screen;
-   int ret;
-
-   if (q->bo) {
-      nouveau_bo_ref(NULL, &q->bo);
-      if (q->mm) {
-         if (q->state == NV50_QUERY_STATE_READY)
-            nouveau_mm_free(q->mm);
-         else
-            nouveau_fence_work(screen->base.fence.current, nouveau_mm_free_work,
-                               q->mm);
-      }
-   }
-   if (size) {
-      q->mm = nouveau_mm_allocate(screen->base.mm_GART, size, &q->bo, &q->base);
-      if (!q->bo)
-         return false;
-      q->offset = q->base;
-
-      ret = nouveau_bo_map(q->bo, 0, screen->base.client);
-      if (ret) {
-         nv50_query_allocate(nv50, q, 0);
-         return false;
-      }
-      q->data = (uint32_t *)((uint8_t *)q->bo->map + q->base);
-   }
-   return true;
-}
-
-static void
-nv50_query_destroy(struct pipe_context *pipe, struct pipe_query *pq)
-{
-   nv50_query_allocate(nv50_context(pipe), nv50_query(pq), 0);
-   nouveau_fence_ref(NULL, &nv50_query(pq)->fence);
-   FREE(nv50_query(pq));
-}
+#include "nv50/nv50_query_hw.h"
 
 static struct pipe_query *
-nv50_query_create(struct pipe_context *pipe, unsigned type, unsigned index)
+nv50_create_query(struct pipe_context *pipe, unsigned type, unsigned index)
 {
    struct nv50_context *nv50 = nv50_context(pipe);
    struct nv50_query *q;
 
-   q = CALLOC_STRUCT(nv50_query);
-   if (!q)
-      return NULL;
-
-   if (!nv50_query_allocate(nv50, q, NV50_QUERY_ALLOC_SPACE)) {
-      FREE(q);
-      return NULL;
-   }
-
-   q->is64bit = (type == PIPE_QUERY_PRIMITIVES_GENERATED ||
-                 type == PIPE_QUERY_PRIMITIVES_EMITTED ||
-                 type == PIPE_QUERY_SO_STATISTICS ||
-                 type == PIPE_QUERY_PIPELINE_STATISTICS);
-   q->type = type;
-
-   if (q->type == PIPE_QUERY_OCCLUSION_COUNTER) {
-      q->offset -= 32;
-      q->data -= 32 / sizeof(*q->data); /* we advance before query_begin ! */
-   }
-
+   q = nv50_hw_create_query(nv50, type, index);
    return (struct pipe_query *)q;
 }
 
 static void
-nv50_query_get(struct nouveau_pushbuf *push, struct nv50_query *q,
-               unsigned offset, uint32_t get)
+nv50_destroy_query(struct pipe_context *pipe, struct pipe_query *pq)
 {
-   offset += q->offset;
-
-   PUSH_SPACE(push, 5);
-   PUSH_REFN (push, q->bo, NOUVEAU_BO_GART | NOUVEAU_BO_WR);
-   BEGIN_NV04(push, NV50_3D(QUERY_ADDRESS_HIGH), 4);
-   PUSH_DATAh(push, q->bo->offset + offset);
-   PUSH_DATA (push, q->bo->offset + offset);
-   PUSH_DATA (push, q->sequence);
-   PUSH_DATA (push, get);
+   struct nv50_query *q = nv50_query(pq);
+   q->funcs->destroy_query(nv50_context(pipe), q);
 }
 
 static boolean
-nv50_query_begin(struct pipe_context *pipe, struct pipe_query *pq)
+nv50_begin_query(struct pipe_context *pipe, struct pipe_query *pq)
 {
-   struct nv50_context *nv50 = nv50_context(pipe);
-   struct nouveau_pushbuf *push = nv50->base.pushbuf;
    struct nv50_query *q = nv50_query(pq);
-
-   /* For occlusion queries we have to change the storage, because a previous
-    * query might set the initial render conition to false even *after* we re-
-    * initialized it to true.
-    */
-   if (q->type == PIPE_QUERY_OCCLUSION_COUNTER) {
-      q->offset += 32;
-      q->data += 32 / sizeof(*q->data);
-      if (q->offset - q->base == NV50_QUERY_ALLOC_SPACE)
-         nv50_query_allocate(nv50, q, NV50_QUERY_ALLOC_SPACE);
-
-      /* XXX: can we do this with the GPU, and sync with respect to a previous
-       *  query ?
-       */
-      q->data[0] = q->sequence; /* initialize sequence */
-      q->data[1] = 1; /* initial render condition = true */
-      q->data[4] = q->sequence + 1; /* for comparison COND_MODE */
-      q->data[5] = 0;
-   }
-   if (!q->is64bit)
-      q->data[0] = q->sequence++; /* the previously used one */
-
-   switch (q->type) {
-   case PIPE_QUERY_OCCLUSION_COUNTER:
-      q->nesting = nv50->screen->num_occlusion_queries_active++;
-      if (q->nesting) {
-         nv50_query_get(push, q, 0x10, 0x0100f002);
-      } else {
-         PUSH_SPACE(push, 4);
-         BEGIN_NV04(push, NV50_3D(COUNTER_RESET), 1);
-         PUSH_DATA (push, NV50_3D_COUNTER_RESET_SAMPLECNT);
-         BEGIN_NV04(push, NV50_3D(SAMPLECNT_ENABLE), 1);
-         PUSH_DATA (push, 1);
-      }
-      break;
-   case PIPE_QUERY_PRIMITIVES_GENERATED:
-      nv50_query_get(push, q, 0x10, 0x06805002);
-      break;
-   case PIPE_QUERY_PRIMITIVES_EMITTED:
-      nv50_query_get(push, q, 0x10, 0x05805002);
-      break;
-   case PIPE_QUERY_SO_STATISTICS:
-      nv50_query_get(push, q, 0x20, 0x05805002);
-      nv50_query_get(push, q, 0x30, 0x06805002);
-      break;
-   case PIPE_QUERY_PIPELINE_STATISTICS:
-      nv50_query_get(push, q, 0x80, 0x00801002); /* VFETCH, VERTICES */
-      nv50_query_get(push, q, 0x90, 0x01801002); /* VFETCH, PRIMS */
-      nv50_query_get(push, q, 0xa0, 0x02802002); /* VP, LAUNCHES */
-      nv50_query_get(push, q, 0xb0, 0x03806002); /* GP, LAUNCHES */
-      nv50_query_get(push, q, 0xc0, 0x04806002); /* GP, PRIMS_OUT */
-      nv50_query_get(push, q, 0xd0, 0x07804002); /* RAST, PRIMS_IN */
-      nv50_query_get(push, q, 0xe0, 0x08804002); /* RAST, PRIMS_OUT */
-      nv50_query_get(push, q, 0xf0, 0x0980a002); /* ROP, PIXELS */
-      break;
-   case PIPE_QUERY_TIME_ELAPSED:
-      nv50_query_get(push, q, 0x10, 0x00005002);
-      break;
-   default:
-      break;
-   }
-   q->state = NV50_QUERY_STATE_ACTIVE;
-   return true;
+   return q->funcs->begin_query(nv50_context(pipe), q);
 }
 
 static void
-nv50_query_end(struct pipe_context *pipe, struct pipe_query *pq)
+nv50_end_query(struct pipe_context *pipe, struct pipe_query *pq)
 {
-   struct nv50_context *nv50 = nv50_context(pipe);
-   struct nouveau_pushbuf *push = nv50->base.pushbuf;
    struct nv50_query *q = nv50_query(pq);
-
-   q->state = NV50_QUERY_STATE_ENDED;
-
-   switch (q->type) {
-   case PIPE_QUERY_OCCLUSION_COUNTER:
-      nv50_query_get(push, q, 0, 0x0100f002);
-      if (--nv50->screen->num_occlusion_queries_active == 0) {
-         PUSH_SPACE(push, 2);
-         BEGIN_NV04(push, NV50_3D(SAMPLECNT_ENABLE), 1);
-         PUSH_DATA (push, 0);
-      }
-      break;
-   case PIPE_QUERY_PRIMITIVES_GENERATED:
-      nv50_query_get(push, q, 0, 0x06805002);
-      break;
-   case PIPE_QUERY_PRIMITIVES_EMITTED:
-      nv50_query_get(push, q, 0, 0x05805002);
-      break;
-   case PIPE_QUERY_SO_STATISTICS:
-      nv50_query_get(push, q, 0x00, 0x05805002);
-      nv50_query_get(push, q, 0x10, 0x06805002);
-      break;
-   case PIPE_QUERY_PIPELINE_STATISTICS:
-      nv50_query_get(push, q, 0x00, 0x00801002); /* VFETCH, VERTICES */
-      nv50_query_get(push, q, 0x10, 0x01801002); /* VFETCH, PRIMS */
-      nv50_query_get(push, q, 0x20, 0x02802002); /* VP, LAUNCHES */
-      nv50_query_get(push, q, 0x30, 0x03806002); /* GP, LAUNCHES */
-      nv50_query_get(push, q, 0x40, 0x04806002); /* GP, PRIMS_OUT */
-      nv50_query_get(push, q, 0x50, 0x07804002); /* RAST, PRIMS_IN */
-      nv50_query_get(push, q, 0x60, 0x08804002); /* RAST, PRIMS_OUT */
-      nv50_query_get(push, q, 0x70, 0x0980a002); /* ROP, PIXELS */
-      break;
-   case PIPE_QUERY_TIMESTAMP:
-      q->sequence++;
-      /* fall through */
-   case PIPE_QUERY_TIME_ELAPSED:
-      nv50_query_get(push, q, 0, 0x00005002);
-      break;
-   case PIPE_QUERY_GPU_FINISHED:
-      q->sequence++;
-      nv50_query_get(push, q, 0, 0x1000f010);
-      break;
-   case NVA0_QUERY_STREAM_OUTPUT_BUFFER_OFFSET:
-      q->sequence++;
-      nv50_query_get(push, q, 0, 0x0d005002 | (q->index << 5));
-      break;
-   case PIPE_QUERY_TIMESTAMP_DISJOINT:
-      /* This query is not issued on GPU because disjoint is forced to false */
-      q->state = NV50_QUERY_STATE_READY;
-      break;
-   default:
-      assert(0);
-      break;
-   }
-
-   if (q->is64bit)
-      nouveau_fence_ref(nv50->screen->base.fence.current, &q->fence);
-}
-
-static inline void
-nv50_query_update(struct nv50_query *q)
-{
-   if (q->is64bit) {
-      if (nouveau_fence_signalled(q->fence))
-         q->state = NV50_QUERY_STATE_READY;
-   } else {
-      if (q->data[0] == q->sequence)
-         q->state = NV50_QUERY_STATE_READY;
-   }
+   q->funcs->end_query(nv50_context(pipe), q);
 }
 
 static boolean
-nv50_query_result(struct pipe_context *pipe, struct pipe_query *pq,
-                  boolean wait, union pipe_query_result *result)
+nv50_get_query_result(struct pipe_context *pipe, struct pipe_query *pq,
+                      boolean wait, union pipe_query_result *result)
 {
-   struct nv50_context *nv50 = nv50_context(pipe);
    struct nv50_query *q = nv50_query(pq);
-   uint64_t *res64 = (uint64_t *)result;
-   uint32_t *res32 = (uint32_t *)result;
-   uint8_t *res8 = (uint8_t *)result;
-   uint64_t *data64 = (uint64_t *)q->data;
-   int i;
-
-   if (q->state != NV50_QUERY_STATE_READY)
-      nv50_query_update(q);
-
-   if (q->state != NV50_QUERY_STATE_READY) {
-      if (!wait) {
-         /* for broken apps that spin on GL_QUERY_RESULT_AVAILABLE */
-         if (q->state != NV50_QUERY_STATE_FLUSHED) {
-            q->state = NV50_QUERY_STATE_FLUSHED;
-            PUSH_KICK(nv50->base.pushbuf);
-         }
-         return false;
-      }
-      if (nouveau_bo_wait(q->bo, NOUVEAU_BO_RD, nv50->screen->base.client))
-         return false;
-   }
-   q->state = NV50_QUERY_STATE_READY;
-
-   switch (q->type) {
-   case PIPE_QUERY_GPU_FINISHED:
-      res8[0] = true;
-      break;
-   case PIPE_QUERY_OCCLUSION_COUNTER: /* u32 sequence, u32 count, u64 time */
-      res64[0] = q->data[1] - q->data[5];
-      break;
-   case PIPE_QUERY_PRIMITIVES_GENERATED: /* u64 count, u64 time */
-   case PIPE_QUERY_PRIMITIVES_EMITTED: /* u64 count, u64 time */
-      res64[0] = data64[0] - data64[2];
-      break;
-   case PIPE_QUERY_SO_STATISTICS:
-      res64[0] = data64[0] - data64[4];
-      res64[1] = data64[2] - data64[6];
-      break;
-   case PIPE_QUERY_PIPELINE_STATISTICS:
-      for (i = 0; i < 8; ++i)
-         res64[i] = data64[i * 2] - data64[16 + i * 2];
-      break;
-   case PIPE_QUERY_TIMESTAMP:
-      res64[0] = data64[1];
-      break;
-   case PIPE_QUERY_TIMESTAMP_DISJOINT:
-      res64[0] = 1000000000;
-      res8[8] = false;
-      break;
-   case PIPE_QUERY_TIME_ELAPSED:
-      res64[0] = data64[1] - data64[3];
-      break;
-   case NVA0_QUERY_STREAM_OUTPUT_BUFFER_OFFSET:
-      res32[0] = q->data[1];
-      break;
-   default:
-      return false;
-   }
-
-   return true;
-}
-
-void
-nv84_query_fifo_wait(struct nouveau_pushbuf *push, struct nv50_query *q)
-{
-   unsigned offset = q->offset;
-
-   PUSH_SPACE(push, 5);
-   PUSH_REFN (push, q->bo, NOUVEAU_BO_GART | NOUVEAU_BO_RD);
-   BEGIN_NV04(push, SUBC_3D(NV84_SUBCHAN_SEMAPHORE_ADDRESS_HIGH), 4);
-   PUSH_DATAh(push, q->bo->offset + offset);
-   PUSH_DATA (push, q->bo->offset + offset);
-   PUSH_DATA (push, q->sequence);
-   PUSH_DATA (push, NV84_SUBCHAN_SEMAPHORE_TRIGGER_ACQUIRE_EQUAL);
+   return q->funcs->get_query_result(nv50_context(pipe), q, wait, result);
 }
 
 static void
@@ -363,7 +74,8 @@ nv50_render_condition(struct pipe_context *pipe,
 {
    struct nv50_context *nv50 = nv50_context(pipe);
    struct nouveau_pushbuf *push = nv50->base.pushbuf;
-   struct nv50_query *q;
+   struct nv50_query *q = nv50_query(pq);
+   struct nv50_hw_query *hq = nv50_hw_query(q);
    uint32_t cond;
    bool wait =
       mode != PIPE_RENDER_COND_NO_WAIT &&
@@ -373,7 +85,6 @@ nv50_render_condition(struct pipe_context *pipe,
       cond = NV50_3D_COND_MODE_ALWAYS;
    }
    else {
-      q = nv50_query(pq);
       /* NOTE: comparison of 2 queries only works if both have completed */
       switch (q->type) {
       case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
@@ -384,7 +95,7 @@ nv50_render_condition(struct pipe_context *pipe,
       case PIPE_QUERY_OCCLUSION_COUNTER:
       case PIPE_QUERY_OCCLUSION_PREDICATE:
          if (likely(!condition)) {
-            if (unlikely(q->nesting))
+            if (unlikely(hq->nesting))
                cond = wait ? NV50_3D_COND_MODE_NOT_EQUAL :
                              NV50_3D_COND_MODE_ALWAYS;
             else
@@ -419,28 +130,15 @@ nv50_render_condition(struct pipe_context *pipe,
       PUSH_DATA (push, 0);
    }
 
-   PUSH_REFN (push, q->bo, NOUVEAU_BO_GART | NOUVEAU_BO_RD);
+   PUSH_REFN (push, hq->bo, NOUVEAU_BO_GART | NOUVEAU_BO_RD);
    BEGIN_NV04(push, NV50_3D(COND_ADDRESS_HIGH), 3);
-   PUSH_DATAh(push, q->bo->offset + q->offset);
-   PUSH_DATA (push, q->bo->offset + q->offset);
+   PUSH_DATAh(push, hq->bo->offset + hq->offset);
+   PUSH_DATA (push, hq->bo->offset + hq->offset);
    PUSH_DATA (push, cond);
 
    BEGIN_NV04(push, NV50_2D(COND_ADDRESS_HIGH), 2);
-   PUSH_DATAh(push, q->bo->offset + q->offset);
-   PUSH_DATA (push, q->bo->offset + q->offset);
-}
-
-void
-nv50_query_pushbuf_submit(struct nouveau_pushbuf *push, uint16_t method,
-                          struct nv50_query *q, unsigned result_offset)
-{
-   nv50_query_update(q);
-   if (q->state != NV50_QUERY_STATE_READY)
-      nouveau_bo_wait(q->bo, NOUVEAU_BO_RD, push->client);
-   q->state = NV50_QUERY_STATE_READY;
-
-   BEGIN_NV04(push, SUBC_3D(method), 1);
-   PUSH_DATA (push, q->data[result_offset / 4]);
+   PUSH_DATAh(push, hq->bo->offset + hq->offset);
+   PUSH_DATA (push, hq->bo->offset + hq->offset);
 }
 
 void
@@ -448,10 +146,10 @@ nv50_init_query_functions(struct nv50_context *nv50)
 {
    struct pipe_context *pipe = &nv50->base.pipe;
 
-   pipe->create_query = nv50_query_create;
-   pipe->destroy_query = nv50_query_destroy;
-   pipe->begin_query = nv50_query_begin;
-   pipe->end_query = nv50_query_end;
-   pipe->get_query_result = nv50_query_result;
+   pipe->create_query = nv50_create_query;
+   pipe->destroy_query = nv50_destroy_query;
+   pipe->begin_query = nv50_begin_query;
+   pipe->end_query = nv50_end_query;
+   pipe->get_query_result = nv50_get_query_result;
    pipe->render_condition = nv50_render_condition;
 }
diff --git a/src/gallium/drivers/nouveau/nv50/nv50_query.h b/src/gallium/drivers/nouveau/nv50/nv50_query.h
index a7030133f56..d990285c857 100644
--- a/src/gallium/drivers/nouveau/nv50/nv50_query.h
+++ b/src/gallium/drivers/nouveau/nv50/nv50_query.h
@@ -4,23 +4,22 @@
 #include "pipe/p_context.h"
 
 #include "nouveau_context.h"
-#include "nouveau_mm.h"
 
-#define NVA0_QUERY_STREAM_OUTPUT_BUFFER_OFFSET (PIPE_QUERY_TYPES + 0)
+struct nv50_context;
+struct nv50_query;
+
+struct nv50_query_funcs {
+   void (*destroy_query)(struct nv50_context *, struct nv50_query *);
+   boolean (*begin_query)(struct nv50_context *, struct nv50_query *);
+   void (*end_query)(struct nv50_context *, struct nv50_query *);
+   boolean (*get_query_result)(struct nv50_context *, struct nv50_query *,
+                               boolean, union pipe_query_result *);
+};
 
 struct nv50_query {
-   uint32_t *data;
+   const struct nv50_query_funcs *funcs;
    uint16_t type;
    uint16_t index;
-   uint32_t sequence;
-   struct nouveau_bo *bo;
-   uint32_t base;
-   uint32_t offset; /* base + i * 32 */
-   uint8_t state;
-   bool is64bit;
-   int nesting; /* only used for occlusion queries */
-   struct nouveau_mm_allocation *mm;
-   struct nouveau_fence *fence;
 };
 
 static inline struct nv50_query *
@@ -30,8 +29,5 @@ nv50_query(struct pipe_query *pipe)
 }
 
 void nv50_init_query_functions(struct nv50_context *);
-void nv50_query_pushbuf_submit(struct nouveau_pushbuf *, uint16_t,
-                               struct nv50_query *, unsigned result_offset);
-void nv84_query_fifo_wait(struct nouveau_pushbuf *, struct nv50_query *);
 
 #endif
diff --git a/src/gallium/drivers/nouveau/nv50/nv50_query_hw.c b/src/gallium/drivers/nouveau/nv50/nv50_query_hw.c
new file mode 100644
index 00000000000..fcdd183e88a
--- /dev/null
+++ b/src/gallium/drivers/nouveau/nv50/nv50_query_hw.c
@@ -0,0 +1,389 @@
+/*
+ * Copyright 2011 Christoph Bumiller
+ * Copyright 2015 Samuel Pitoiset
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#define NV50_PUSH_EXPLICIT_SPACE_CHECKING
+
+#include "nv50/nv50_context.h"
+#include "nv50/nv50_query_hw.h"
+#include "nv_object.xml.h"
+
+#define NV50_HW_QUERY_STATE_READY   0
+#define NV50_HW_QUERY_STATE_ACTIVE  1
+#define NV50_HW_QUERY_STATE_ENDED   2
+#define NV50_HW_QUERY_STATE_FLUSHED 3
+
+/* XXX: Nested queries, and simultaneous queries on multiple gallium contexts
+ * (since we use only a single GPU channel per screen) will not work properly.
+ *
+ * The first is not that big of an issue because OpenGL does not allow nested
+ * queries anyway.
+ */
+
+#define NV50_HW_QUERY_ALLOC_SPACE 256
+
+static bool
+nv50_hw_query_allocate(struct nv50_context *nv50, struct nv50_query *q,
+                       int size)
+{
+   struct nv50_screen *screen = nv50->screen;
+   struct nv50_hw_query *hq = nv50_hw_query(q);
+   int ret;
+
+   if (hq->bo) {
+      nouveau_bo_ref(NULL, &hq->bo);
+      if (hq->mm) {
+         if (hq->state == NV50_HW_QUERY_STATE_READY)
+            nouveau_mm_free(hq->mm);
+         else
+            nouveau_fence_work(screen->base.fence.current,
+                               nouveau_mm_free_work, hq->mm);
+      }
+   }
+   if (size) {
+      hq->mm = nouveau_mm_allocate(screen->base.mm_GART, size,
+                                   &hq->bo, &hq->base_offset);
+      if (!hq->bo)
+         return false;
+      hq->offset = hq->base_offset;
+
+      ret = nouveau_bo_map(hq->bo, 0, screen->base.client);
+      if (ret) {
+         nv50_hw_query_allocate(nv50, q, 0);
+         return false;
+      }
+      hq->data = (uint32_t *)((uint8_t *)hq->bo->map + hq->base_offset);
+   }
+   return true;
+}
+
+static void
+nv50_hw_query_get(struct nouveau_pushbuf *push, struct nv50_query *q,
+               unsigned offset, uint32_t get)
+{
+   struct nv50_hw_query *hq = nv50_hw_query(q);
+
+   offset += hq->offset;
+
+   PUSH_SPACE(push, 5);
+   PUSH_REFN (push, hq->bo, NOUVEAU_BO_GART | NOUVEAU_BO_WR);
+   BEGIN_NV04(push, NV50_3D(QUERY_ADDRESS_HIGH), 4);
+   PUSH_DATAh(push, hq->bo->offset + offset);
+   PUSH_DATA (push, hq->bo->offset + offset);
+   PUSH_DATA (push, hq->sequence);
+   PUSH_DATA (push, get);
+}
+
+static inline void
+nv50_hw_query_update(struct nv50_query *q)
+{
+   struct nv50_hw_query *hq = nv50_hw_query(q);
+
+   if (hq->is64bit) {
+      if (nouveau_fence_signalled(hq->fence))
+         hq->state = NV50_HW_QUERY_STATE_READY;
+   } else {
+      if (hq->data[0] == hq->sequence)
+         hq->state = NV50_HW_QUERY_STATE_READY;
+   }
+}
+
+static void
+nv50_hw_destroy_query(struct nv50_context *nv50, struct nv50_query *q)
+{
+   struct nv50_hw_query *hq = nv50_hw_query(q);
+   nv50_hw_query_allocate(nv50, q, 0);
+   nouveau_fence_ref(NULL, &hq->fence);
+   FREE(hq);
+}
+
+static boolean
+nv50_hw_begin_query(struct nv50_context *nv50, struct nv50_query *q)
+{
+   struct nouveau_pushbuf *push = nv50->base.pushbuf;
+   struct nv50_hw_query *hq = nv50_hw_query(q);
+
+   /* For occlusion queries we have to change the storage, because a previous
+    * query might set the initial render condition to false even *after* we re-
+    * initialized it to true.
+    */
+   if (q->type == PIPE_QUERY_OCCLUSION_COUNTER) {
+      hq->offset += 32;
+      hq->data += 32 / sizeof(*hq->data);
+      if (hq->offset - hq->base_offset == NV50_HW_QUERY_ALLOC_SPACE)
+         nv50_hw_query_allocate(nv50, q, NV50_HW_QUERY_ALLOC_SPACE);
+
+      /* XXX: can we do this with the GPU, and sync with respect to a previous
+       *  query ?
+       */
+      hq->data[0] = hq->sequence; /* initialize sequence */
+      hq->data[1] = 1; /* initial render condition = true */
+      hq->data[4] = hq->sequence + 1; /* for comparison COND_MODE */
+      hq->data[5] = 0;
+   }
+   if (!hq->is64bit)
+      hq->data[0] = hq->sequence++; /* the previously used one */
+
+   switch (q->type) {
+   case PIPE_QUERY_OCCLUSION_COUNTER:
+      hq->nesting = nv50->screen->num_occlusion_queries_active++;
+      if (hq->nesting) {
+         nv50_hw_query_get(push, q, 0x10, 0x0100f002);
+      } else {
+         PUSH_SPACE(push, 4);
+         BEGIN_NV04(push, NV50_3D(COUNTER_RESET), 1);
+         PUSH_DATA (push, NV50_3D_COUNTER_RESET_SAMPLECNT);
+         BEGIN_NV04(push, NV50_3D(SAMPLECNT_ENABLE), 1);
+         PUSH_DATA (push, 1);
+      }
+      break;
+   case PIPE_QUERY_PRIMITIVES_GENERATED:
+      nv50_hw_query_get(push, q, 0x10, 0x06805002);
+      break;
+   case PIPE_QUERY_PRIMITIVES_EMITTED:
+      nv50_hw_query_get(push, q, 0x10, 0x05805002);
+      break;
+   case PIPE_QUERY_SO_STATISTICS:
+      nv50_hw_query_get(push, q, 0x20, 0x05805002);
+      nv50_hw_query_get(push, q, 0x30, 0x06805002);
+      break;
+   case PIPE_QUERY_PIPELINE_STATISTICS:
+      nv50_hw_query_get(push, q, 0x80, 0x00801002); /* VFETCH, VERTICES */
+      nv50_hw_query_get(push, q, 0x90, 0x01801002); /* VFETCH, PRIMS */
+      nv50_hw_query_get(push, q, 0xa0, 0x02802002); /* VP, LAUNCHES */
+      nv50_hw_query_get(push, q, 0xb0, 0x03806002); /* GP, LAUNCHES */
+      nv50_hw_query_get(push, q, 0xc0, 0x04806002); /* GP, PRIMS_OUT */
+      nv50_hw_query_get(push, q, 0xd0, 0x07804002); /* RAST, PRIMS_IN */
+      nv50_hw_query_get(push, q, 0xe0, 0x08804002); /* RAST, PRIMS_OUT */
+      nv50_hw_query_get(push, q, 0xf0, 0x0980a002); /* ROP, PIXELS */
+      break;
+   case PIPE_QUERY_TIME_ELAPSED:
+      nv50_hw_query_get(push, q, 0x10, 0x00005002);
+      break;
+   default:
+      assert(0);
+      return false;
+   }
+   hq->state = NV50_HW_QUERY_STATE_ACTIVE;
+   return true;
+}
+
+static void
+nv50_hw_end_query(struct nv50_context *nv50, struct nv50_query *q)
+{
+   struct nouveau_pushbuf *push = nv50->base.pushbuf;
+   struct nv50_hw_query *hq = nv50_hw_query(q);
+
+   hq->state = NV50_HW_QUERY_STATE_ENDED;
+
+   switch (q->type) {
+   case PIPE_QUERY_OCCLUSION_COUNTER:
+      nv50_hw_query_get(push, q, 0, 0x0100f002);
+      if (--nv50->screen->num_occlusion_queries_active == 0) {
+         PUSH_SPACE(push, 2);
+         BEGIN_NV04(push, NV50_3D(SAMPLECNT_ENABLE), 1);
+         PUSH_DATA (push, 0);
+      }
+      break;
+   case PIPE_QUERY_PRIMITIVES_GENERATED:
+      nv50_hw_query_get(push, q, 0, 0x06805002);
+      break;
+   case PIPE_QUERY_PRIMITIVES_EMITTED:
+      nv50_hw_query_get(push, q, 0, 0x05805002);
+      break;
+   case PIPE_QUERY_SO_STATISTICS:
+      nv50_hw_query_get(push, q, 0x00, 0x05805002);
+      nv50_hw_query_get(push, q, 0x10, 0x06805002);
+      break;
+   case PIPE_QUERY_PIPELINE_STATISTICS:
+      nv50_hw_query_get(push, q, 0x00, 0x00801002); /* VFETCH, VERTICES */
+      nv50_hw_query_get(push, q, 0x10, 0x01801002); /* VFETCH, PRIMS */
+      nv50_hw_query_get(push, q, 0x20, 0x02802002); /* VP, LAUNCHES */
+      nv50_hw_query_get(push, q, 0x30, 0x03806002); /* GP, LAUNCHES */
+      nv50_hw_query_get(push, q, 0x40, 0x04806002); /* GP, PRIMS_OUT */
+      nv50_hw_query_get(push, q, 0x50, 0x07804002); /* RAST, PRIMS_IN */
+      nv50_hw_query_get(push, q, 0x60, 0x08804002); /* RAST, PRIMS_OUT */
+      nv50_hw_query_get(push, q, 0x70, 0x0980a002); /* ROP, PIXELS */
+      break;
+   case PIPE_QUERY_TIMESTAMP:
+      hq->sequence++;
+      /* fall through */
+   case PIPE_QUERY_TIME_ELAPSED:
+      nv50_hw_query_get(push, q, 0, 0x00005002);
+      break;
+   case PIPE_QUERY_GPU_FINISHED:
+      hq->sequence++;
+      nv50_hw_query_get(push, q, 0, 0x1000f010);
+      break;
+   case NVA0_HW_QUERY_STREAM_OUTPUT_BUFFER_OFFSET:
+      hq->sequence++;
+      nv50_hw_query_get(push, q, 0, 0x0d005002 | (q->index << 5));
+      break;
+   case PIPE_QUERY_TIMESTAMP_DISJOINT:
+      /* This query is not issued on GPU because disjoint is forced to false */
+      hq->state = NV50_HW_QUERY_STATE_READY;
+      break;
+   default:
+      assert(0);
+      break;
+   }
+   if (hq->is64bit)
+      nouveau_fence_ref(nv50->screen->base.fence.current, &hq->fence);
+}
+
+static boolean
+nv50_hw_get_query_result(struct nv50_context *nv50, struct nv50_query *q,
+                         boolean wait, union pipe_query_result *result)
+{
+   struct nv50_hw_query *hq = nv50_hw_query(q);
+   uint64_t *res64 = (uint64_t *)result;
+   uint32_t *res32 = (uint32_t *)result;
+   uint8_t *res8 = (uint8_t *)result;
+   uint64_t *data64 = (uint64_t *)hq->data;
+   int i;
+
+   if (hq->state != NV50_HW_QUERY_STATE_READY)
+      nv50_hw_query_update(q);
+
+   if (hq->state != NV50_HW_QUERY_STATE_READY) {
+      if (!wait) {
+         /* for broken apps that spin on GL_QUERY_RESULT_AVAILABLE */
+         if (hq->state != NV50_HW_QUERY_STATE_FLUSHED) {
+            hq->state = NV50_HW_QUERY_STATE_FLUSHED;
+            PUSH_KICK(nv50->base.pushbuf);
+         }
+         return false;
+      }
+      if (nouveau_bo_wait(hq->bo, NOUVEAU_BO_RD, nv50->screen->base.client))
+         return false;
+   }
+   hq->state = NV50_HW_QUERY_STATE_READY;
+
+   switch (q->type) {
+   case PIPE_QUERY_GPU_FINISHED:
+      res8[0] = true;
+      break;
+   case PIPE_QUERY_OCCLUSION_COUNTER: /* u32 sequence, u32 count, u64 time */
+      res64[0] = hq->data[1] - hq->data[5];
+      break;
+   case PIPE_QUERY_PRIMITIVES_GENERATED: /* u64 count, u64 time */
+   case PIPE_QUERY_PRIMITIVES_EMITTED: /* u64 count, u64 time */
+      res64[0] = data64[0] - data64[2];
+      break;
+   case PIPE_QUERY_SO_STATISTICS:
+      res64[0] = data64[0] - data64[4];
+      res64[1] = data64[2] - data64[6];
+      break;
+   case PIPE_QUERY_PIPELINE_STATISTICS:
+      for (i = 0; i < 8; ++i)
+         res64[i] = data64[i * 2] - data64[16 + i * 2];
+      break;
+   case PIPE_QUERY_TIMESTAMP:
+      res64[0] = data64[1];
+      break;
+   case PIPE_QUERY_TIMESTAMP_DISJOINT:
+      res64[0] = 1000000000;
+      res8[8] = false;
+      break;
+   case PIPE_QUERY_TIME_ELAPSED:
+      res64[0] = data64[1] - data64[3];
+      break;
+   case NVA0_HW_QUERY_STREAM_OUTPUT_BUFFER_OFFSET:
+      res32[0] = hq->data[1];
+      break;
+   default:
+      assert(0);
+      return false;
+   }
+
+   return true;
+}
+
+static const struct nv50_query_funcs hw_query_funcs = {
+   .destroy_query = nv50_hw_destroy_query,
+   .begin_query = nv50_hw_begin_query,
+   .end_query = nv50_hw_end_query,
+   .get_query_result = nv50_hw_get_query_result,
+};
+
+struct nv50_query *
+nv50_hw_create_query(struct nv50_context *nv50, unsigned type, unsigned index)
+{
+   struct nv50_hw_query *hq;
+   struct nv50_query *q;
+
+   hq = CALLOC_STRUCT(nv50_hw_query);
+   if (!hq)
+      return NULL;
+
+   q = &hq->base;
+   q->funcs = &hw_query_funcs;
+   q->type = type;
+
+   if (!nv50_hw_query_allocate(nv50, q, NV50_HW_QUERY_ALLOC_SPACE)) {
+      FREE(hq);
+      return NULL;
+   }
+
+   if (q->type == PIPE_QUERY_OCCLUSION_COUNTER) {
+      /* we advance before query_begin ! */
+      hq->offset -= 32;
+      hq->data -= 32 / sizeof(*hq->data);
+   }
+
+   hq->is64bit = (type == PIPE_QUERY_PRIMITIVES_GENERATED ||
+                 type == PIPE_QUERY_PRIMITIVES_EMITTED ||
+                 type == PIPE_QUERY_SO_STATISTICS ||
+                 type == PIPE_QUERY_PIPELINE_STATISTICS);
+
+   return q;
+}
+
+void
+nv50_hw_query_pushbuf_submit(struct nouveau_pushbuf *push, uint16_t method,
+                             struct nv50_query *q, unsigned result_offset)
+{
+   struct nv50_hw_query *hq = nv50_hw_query(q);
+
+   nv50_hw_query_update(q);
+   if (hq->state != NV50_HW_QUERY_STATE_READY)
+      nouveau_bo_wait(hq->bo, NOUVEAU_BO_RD, push->client);
+   hq->state = NV50_HW_QUERY_STATE_READY;
+
+   BEGIN_NV04(push, SUBC_3D(method), 1);
+   PUSH_DATA (push, hq->data[result_offset / 4]);
+}
+
+void
+nv84_hw_query_fifo_wait(struct nouveau_pushbuf *push, struct nv50_query *q)
+{
+   struct nv50_hw_query *hq = nv50_hw_query(q);
+   unsigned offset = hq->offset;
+
+   PUSH_SPACE(push, 5);
+   PUSH_REFN (push, hq->bo, NOUVEAU_BO_GART | NOUVEAU_BO_RD);
+   BEGIN_NV04(push, SUBC_3D(NV84_SUBCHAN_SEMAPHORE_ADDRESS_HIGH), 4);
+   PUSH_DATAh(push, hq->bo->offset + offset);
+   PUSH_DATA (push, hq->bo->offset + offset);
+   PUSH_DATA (push, hq->sequence);
+   PUSH_DATA (push, NV84_SUBCHAN_SEMAPHORE_TRIGGER_ACQUIRE_EQUAL);
+}
diff --git a/src/gallium/drivers/nouveau/nv50/nv50_query_hw.h b/src/gallium/drivers/nouveau/nv50/nv50_query_hw.h
new file mode 100644
index 00000000000..fe518a55895
--- /dev/null
+++ b/src/gallium/drivers/nouveau/nv50/nv50_query_hw.h
@@ -0,0 +1,39 @@
+#ifndef __NV50_QUERY_HW_H__
+#define __NV50_QUERY_HW_H__
+
+#include "nouveau_fence.h"
+#include "nouveau_mm.h"
+
+#include "nv50_query.h"
+
+#define NVA0_HW_QUERY_STREAM_OUTPUT_BUFFER_OFFSET (PIPE_QUERY_TYPES + 0)
+
+struct nv50_hw_query {
+   struct nv50_query base;
+   uint32_t *data;
+   uint32_t sequence;
+   struct nouveau_bo *bo;
+   uint32_t base_offset;
+   uint32_t offset; /* base + i * 32 */
+   uint8_t state;
+   bool is64bit;
+   int nesting; /* only used for occlusion queries */
+   struct nouveau_mm_allocation *mm;
+   struct nouveau_fence *fence;
+};
+
+static inline struct nv50_hw_query *
+nv50_hw_query(struct nv50_query *q)
+{
+   return (struct nv50_hw_query *)q;
+}
+
+struct nv50_query *
+nv50_hw_create_query(struct nv50_context *, unsigned, unsigned);
+void
+nv50_hw_query_pushbuf_submit(struct nouveau_pushbuf *, uint16_t,
+                             struct nv50_query *, unsigned);
+void
+nv84_hw_query_fifo_wait(struct nouveau_pushbuf *, struct nv50_query *);
+
+#endif
diff --git a/src/gallium/drivers/nouveau/nv50/nv50_shader_state.c b/src/gallium/drivers/nouveau/nv50/nv50_shader_state.c
index 6b3ecbf3771..9b911043132 100644
--- a/src/gallium/drivers/nouveau/nv50/nv50_shader_state.c
+++ b/src/gallium/drivers/nouveau/nv50/nv50_shader_state.c
@@ -27,6 +27,7 @@
 #include "util/u_inlines.h"
 
 #include "nv50/nv50_context.h"
+#include "nv50/nv50_query_hw.h"
 
 void
 nv50_constbufs_validate(struct nv50_context *nv50)
@@ -641,7 +642,7 @@ nv50_stream_output_validate(struct nv50_context *nv50)
       const unsigned n = nv50->screen->base.class_3d >= NVA0_3D_CLASS ? 4 : 3;
 
       if (n == 4 && !targ->clean)
-         nv84_query_fifo_wait(push, nv50_query(targ->pq));
+         nv84_hw_query_fifo_wait(push, nv50_query(targ->pq));
       BEGIN_NV04(push, NV50_3D(STRMOUT_ADDRESS_HIGH(i)), n);
       PUSH_DATAh(push, buf->address + targ->pipe.buffer_offset);
       PUSH_DATA (push, buf->address + targ->pipe.buffer_offset);
@@ -650,8 +651,8 @@ nv50_stream_output_validate(struct nv50_context *nv50)
          PUSH_DATA(push, targ->pipe.buffer_size);
          if (!targ->clean) {
             assert(targ->pq);
-            nv50_query_pushbuf_submit(push, NVA0_3D_STRMOUT_OFFSET(i),
-                                      nv50_query(targ->pq), 0x4);
+            nv50_hw_query_pushbuf_submit(push, NVA0_3D_STRMOUT_OFFSET(i),
+                                         nv50_query(targ->pq), 0x4);
          } else {
             BEGIN_NV04(push, NVA0_3D(STRMOUT_OFFSET(i)), 1);
             PUSH_DATA(push, 0);
diff --git a/src/gallium/drivers/nouveau/nv50/nv50_state.c b/src/gallium/drivers/nouveau/nv50/nv50_state.c
index a78afc9d216..6c8c9f0b4e6 100644
--- a/src/gallium/drivers/nouveau/nv50/nv50_state.c
+++ b/src/gallium/drivers/nouveau/nv50/nv50_state.c
@@ -30,6 +30,7 @@
 
 #include "nv50/nv50_stateobj.h"
 #include "nv50/nv50_context.h"
+#include "nv50/nv50_query_hw.h"
 
 #include "nv50/nv50_3d.xml.h"
 #include "nv50/nv50_texture.xml.h"
@@ -1036,7 +1037,7 @@ nv50_so_target_create(struct pipe_context *pipe,
 
    if (nouveau_context(pipe)->screen->class_3d >= NVA0_3D_CLASS) {
       targ->pq = pipe->create_query(pipe,
-                                    NVA0_QUERY_STREAM_OUTPUT_BUFFER_OFFSET, 0);
+                                    NVA0_HW_QUERY_STREAM_OUTPUT_BUFFER_OFFSET, 0);
       if (!targ->pq) {
          FREE(targ);
          return NULL;
diff --git a/src/gallium/drivers/nouveau/nv50/nv50_vbo.c b/src/gallium/drivers/nouveau/nv50/nv50_vbo.c
index dbc6632c248..9fa6fceeefa 100644
--- a/src/gallium/drivers/nouveau/nv50/nv50_vbo.c
+++ b/src/gallium/drivers/nouveau/nv50/nv50_vbo.c
@@ -27,6 +27,7 @@
 #include "translate/translate.h"
 
 #include "nv50/nv50_context.h"
+#include "nv50/nv50_query_hw.h"
 #include "nv50/nv50_resource.h"
 
 #include "nv50/nv50_3d.xml.h"
@@ -745,8 +746,8 @@ nva0_draw_stream_output(struct nv50_context *nv50,
       PUSH_DATA (push, 0);
       BEGIN_NV04(push, NVA0_3D(DRAW_TFB_STRIDE), 1);
       PUSH_DATA (push, so->stride);
-      nv50_query_pushbuf_submit(push, NVA0_3D_DRAW_TFB_BYTES,
-                                nv50_query(so->pq), 0x4);
+      nv50_hw_query_pushbuf_submit(push, NVA0_3D_DRAW_TFB_BYTES,
+                                   nv50_query(so->pq), 0x4);
       BEGIN_NV04(push, NV50_3D(VERTEX_END_GL), 1);
       PUSH_DATA (push, 0);