WIP nouveau: add locking

src/gallium/drivers/nouveau/nouveau_buffer.c

@@ -380,6 +380,7 @@ nouveau_buffer_transfer_map(struct pipe_context *pipe,
                             struct pipe_transfer **ptransfer)
 {
    struct nouveau_context *nv = nouveau_context(pipe);
+   struct nouveau_screen *screen = nv->screen;
    struct nv04_resource *buf = nv04_resource(resource);
    struct nouveau_transfer *tx = MALLOC_STRUCT(nouveau_transfer);
    uint8_t *map;
@@ -427,14 +428,19 @@ nouveau_buffer_transfer_map(struct pipe_context *pipe,
                buf->data = NULL;
             }
             nouveau_transfer_staging(nv, tx, false);
+            pipe_mutex_lock(screen->push_mutex);
             nouveau_transfer_read(nv, tx);
+            pipe_mutex_unlock(screen->push_mutex);
          } 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)
+            if (!buf->data) {
+               pipe_mutex_lock(screen->push_mutex);
                nouveau_buffer_cache(nv, buf);
+               pipe_mutex_unlock(screen->push_mutex);
+            }
          }
       }
       return buf->data ? (buf->data + box->x) : tx->map;
@@ -479,7 +485,9 @@ nouveau_buffer_transfer_map(struct pipe_context *pipe,
       if (unlikely(usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE)) {
          /* Discarding was not possible, must sync because
           * subsequent transfers might use UNSYNCHRONIZED. */
+         pipe_mutex_lock(screen->push_mutex);
          nouveau_buffer_sync(nv, buf, usage & PIPE_TRANSFER_READ_WRITE);
+         pipe_mutex_unlock(screen->push_mutex);
       } else
       if (usage & PIPE_TRANSFER_DISCARD_RANGE) {
          /* The whole range is being discarded, so it doesn't matter what was
@@ -488,10 +496,13 @@ nouveau_buffer_transfer_map(struct pipe_context *pipe,
          map = tx->map;
       } else
       if (nouveau_buffer_busy(buf, PIPE_TRANSFER_READ)) {
-         if (usage & PIPE_TRANSFER_DONTBLOCK)
+         if (usage & PIPE_TRANSFER_DONTBLOCK) {
             map = NULL;
-         else
+         } else {
+            pipe_mutex_lock(screen->push_mutex);
             nouveau_buffer_sync(nv, buf, usage & PIPE_TRANSFER_READ_WRITE);
+            pipe_mutex_unlock(screen->push_mutex);
+         }
       } 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. */
@@ -515,9 +526,13 @@ nouveau_buffer_transfer_flush_region(struct pipe_context *pipe,
 {
    struct nouveau_transfer *tx = nouveau_transfer(transfer);
    struct nv04_resource *buf = nv04_resource(transfer->resource);
+   struct nouveau_screen *screen = nouveau_context(pipe)->screen;
 
-   if (tx->map)
+   if (tx->map) {
+      pipe_mutex_lock(screen->push_mutex);
       nouveau_transfer_write(nouveau_context(pipe), tx, box->x, box->width);
+      pipe_mutex_unlock(screen->push_mutex);
+   }
 
    util_range_add(&buf->valid_buffer_range,
                   tx->base.box.x + box->x,
@@ -537,11 +552,15 @@ nouveau_buffer_transfer_unmap(struct pipe_context *pipe,
    struct nouveau_context *nv = nouveau_context(pipe);
    struct nouveau_transfer *tx = nouveau_transfer(transfer);
    struct nv04_resource *buf = nv04_resource(transfer->resource);
+   struct nouveau_screen *screen = nouveau_context(pipe)->screen;
 
    if (tx->base.usage & PIPE_TRANSFER_WRITE) {
       if (!(tx->base.usage & PIPE_TRANSFER_FLUSH_EXPLICIT)) {
-         if (tx->map)
+         if (tx->map) {
+            pipe_mutex_lock(screen->push_mutex);
             nouveau_transfer_write(nv, tx, 0, tx->base.box.width);
+            pipe_mutex_unlock(screen->push_mutex);
+         }
 
          util_range_add(&buf->valid_buffer_range,
                         tx->base.box.x, tx->base.box.x + tx->base.box.width);

src/gallium/drivers/nouveau/nouveau_fence.c

@@ -71,12 +71,14 @@ nouveau_fence_emit(struct nouveau_fence *fence)
 
    ++fence->ref;
 
+   pipe_mutex_lock(screen->fence.list_mutex);
    if (screen->fence.tail)
       screen->fence.tail->next = fence;
    else
       screen->fence.head = fence;
 
    screen->fence.tail = fence;
+   pipe_mutex_unlock(screen->fence.list_mutex);
 
    screen->fence.emit(&screen->base, &fence->sequence);
 
@@ -90,6 +92,9 @@ nouveau_fence_del(struct nouveau_fence *fence)
    struct nouveau_fence *it;
    struct nouveau_screen *screen = fence->screen;
 
+   /* XXX This can race against fence_update. But fence_update can also call
+    * into this, so ... be have to be careful.
+    */
    if (fence->state == NOUVEAU_FENCE_STATE_EMITTED ||
        fence->state == NOUVEAU_FENCE_STATE_FLUSHED) {
       if (fence == screen->fence.head) {
@@ -123,6 +128,7 @@ nouveau_fence_update(struct nouveau_screen *screen, bool flushed)
       return;
    screen->fence.sequence_ack = sequence;
 
+   pipe_mutex_lock(screen->fence.list_mutex);
    for (fence = screen->fence.head; fence; fence = next) {
       next = fence->next;
       sequence = fence->sequence;
@@ -144,6 +150,7 @@ nouveau_fence_update(struct nouveau_screen *screen, bool flushed)
          if (fence->state == NOUVEAU_FENCE_STATE_EMITTED)
             fence->state = NOUVEAU_FENCE_STATE_FLUSHED;
    }
+   pipe_mutex_unlock(screen->fence.list_mutex);
 }
 
 #define NOUVEAU_FENCE_MAX_SPINS (1 << 31)
@@ -198,18 +205,27 @@ nouveau_fence_wait(struct nouveau_fence *fence, struct pipe_debug_callback *debu
    uint32_t spins = 0;
    int64_t start = 0;
 
+   /* Fast-path for the case where the fence is already signaled to avoid
+    * messing around with mutexes and timing.
+    */
+   if (fence->state == NOUVEAU_FENCE_STATE_SIGNALLED)
+      return true;
+
    if (debug && debug->debug_message)
       start = os_time_get_nano();
 
    if (!nouveau_fence_kick(fence))
       return false;
 
+   pipe_mutex_unlock(screen->push_mutex);
+
    do {
       if (fence->state == NOUVEAU_FENCE_STATE_SIGNALLED) {
          if (debug && debug->debug_message)
             pipe_debug_message(debug, PERF_INFO,
                                "stalled %.3f ms waiting for fence",
                                (os_time_get_nano() - start) / 1000000.f);
+         pipe_mutex_lock(screen->push_mutex);
          return true;
       }
       if (!spins)
@@ -227,6 +243,8 @@ nouveau_fence_wait(struct nouveau_fence *fence, struct pipe_debug_callback *debu
                 fence->sequence,
                 screen->fence.sequence_ack, screen->fence.sequence);
 
+   pipe_mutex_lock(screen->push_mutex);
+
    return false;
 }
 

src/gallium/drivers/nouveau/nouveau_fence.h

@@ -2,6 +2,7 @@
 #ifndef __NOUVEAU_FENCE_H__
 #define __NOUVEAU_FENCE_H__
 
+#include "util/u_atomic.h"
 #include "util/u_inlines.h"
 #include "util/list.h"
 
@@ -47,10 +48,10 @@ static inline void
 nouveau_fence_ref(struct nouveau_fence *fence, struct nouveau_fence **ref)
 {
    if (fence)
-      ++fence->ref;
+      p_atomic_inc(&fence->ref);
 
    if (*ref) {
-      if (--(*ref)->ref == 0)
+      if (p_atomic_dec_zero(&(*ref)->ref))
          nouveau_fence_del(*ref);
    }
 

src/gallium/drivers/nouveau/nouveau_screen.c

@@ -73,10 +73,14 @@ nouveau_screen_fence_finish(struct pipe_screen *screen,
                             struct pipe_fence_handle *pfence,
                             uint64_t timeout)
 {
+   bool ret;
    if (!timeout)
       return nouveau_fence_signalled(nouveau_fence(pfence));
 
-   return nouveau_fence_wait(nouveau_fence(pfence), NULL);
+   pipe_mutex_lock(nouveau_screen(screen)->push_mutex);
+   ret = nouveau_fence_wait(nouveau_fence(pfence), NULL);
+   pipe_mutex_unlock(nouveau_screen(screen)->push_mutex);
+   return ret;
 }
 
 
@@ -153,6 +157,9 @@ nouveau_screen_init(struct nouveau_screen *screen, struct nouveau_device *dev)
    if (nv_dbg)
       nouveau_mesa_debug = atoi(nv_dbg);
 
+   pipe_mutex_init(screen->push_mutex);
+   pipe_mutex_init(screen->fence.list_mutex);
+
    /* These must be set before any failure is possible, as the cleanup
     * paths assume they're responsible for deleting them.
     */
@@ -253,6 +260,9 @@ nouveau_screen_fini(struct nouveau_screen *screen)
    nouveau_device_del(&screen->device);
    nouveau_drm_del(&screen->drm);
    close(fd);
+
+   pipe_mutex_destroy(screen->push_mutex);
+   pipe_mutex_destroy(screen->fence.list_mutex);
 }
 
 static void

src/gallium/drivers/nouveau/nouveau_screen.h

@@ -3,6 +3,7 @@
 
 #include "pipe/p_screen.h"
 #include "util/u_memory.h"
+#include "os/os_thread.h"
 
 #ifdef DEBUG
 # define NOUVEAU_ENABLE_DRIVER_STATISTICS
@@ -22,6 +23,7 @@ struct nouveau_screen {
    struct nouveau_object *channel;
    struct nouveau_client *client;
    struct nouveau_pushbuf *pushbuf;
+   pipe_mutex push_mutex;
 
    int refcount;
 
@@ -39,6 +41,7 @@ struct nouveau_screen {
       struct nouveau_fence *head;
       struct nouveau_fence *tail;
       struct nouveau_fence *current;
+      pipe_mutex list_mutex;
       u32 sequence;
       u32 sequence_ack;
       void (*emit)(struct pipe_screen *, u32 *sequence);

src/gallium/drivers/nouveau/nv30/nv30_clear.c

@@ -58,8 +58,11 @@ nv30_clear(struct pipe_context *pipe, unsigned buffers,
    struct pipe_framebuffer_state *fb = &nv30->framebuffer;
    uint32_t colr = 0, zeta = 0, mode = 0;
 
-   if (!nv30_state_validate(nv30, NV30_NEW_FRAMEBUFFER | NV30_NEW_SCISSOR, true))
+   pipe_mutex_lock(nv30->screen->base.push_mutex);
+   if (!nv30_state_validate(nv30, NV30_NEW_FRAMEBUFFER | NV30_NEW_SCISSOR, true)) {
+      pipe_mutex_unlock(nv30->screen->base.push_mutex);
       return;
+   }
 
    if (buffers & PIPE_CLEAR_COLOR && fb->nr_cbufs) {
       colr  = pack_rgba(fb->cbufs[0]->format, color->f);
@@ -96,6 +99,7 @@ nv30_clear(struct pipe_context *pipe, unsigned buffers,
    PUSH_DATA (push, mode);
 
    nv30_state_release(nv30);
+   pipe_mutex_unlock(nv30->screen->base.push_mutex);
 }
 
 static void
@@ -125,11 +129,15 @@ nv30_clear_render_target(struct pipe_context *pipe, struct pipe_surface *ps,
       rt_format |= NV30_3D_RT_FORMAT_TYPE_LINEAR;
    }
 
+   pipe_mutex_lock(nv30->screen->base.push_mutex);
+
    refn.bo = mt->base.bo;
    refn.flags = NOUVEAU_BO_VRAM | NOUVEAU_BO_WR;
    if (nouveau_pushbuf_space(push, 16, 1, 0) ||
-       nouveau_pushbuf_refn (push, &refn, 1))
+       nouveau_pushbuf_refn (push, &refn, 1)) {
+      pipe_mutex_unlock(nv30->screen->base.push_mutex);
       return;
+   }
 
    BEGIN_NV04(push, NV30_3D(RT_ENABLE), 1);
    PUSH_DATA (push, NV30_3D_RT_ENABLE_COLOR0);
@@ -154,6 +162,8 @@ nv30_clear_render_target(struct pipe_context *pipe, struct pipe_surface *ps,
                     NV30_3D_CLEAR_BUFFERS_COLOR_B |
                     NV30_3D_CLEAR_BUFFERS_COLOR_A);
 
+   pipe_mutex_unlock(nv30->screen->base.push_mutex);
+
    nv30->dirty |= NV30_NEW_FRAMEBUFFER | NV30_NEW_SCISSOR;
 }
 
@@ -189,11 +199,15 @@ nv30_clear_depth_stencil(struct pipe_context *pipe, struct pipe_surface *ps,
    if (buffers & PIPE_CLEAR_STENCIL)
       mode |= NV30_3D_CLEAR_BUFFERS_STENCIL;
 
+   pipe_mutex_lock(nv30->screen->base.push_mutex);
+
    refn.bo = mt->base.bo;
    refn.flags = NOUVEAU_BO_VRAM | NOUVEAU_BO_WR;
    if (nouveau_pushbuf_space(push, 32, 1, 0) ||
-       nouveau_pushbuf_refn (push, &refn, 1))
+       nouveau_pushbuf_refn (push, &refn, 1)) {
+      pipe_mutex_unlock(nv30->screen->base.push_mutex);
       return;
+   }
 
    BEGIN_NV04(push, NV30_3D(RT_ENABLE), 1);
    PUSH_DATA (push, 0);
@@ -219,6 +233,8 @@ nv30_clear_depth_stencil(struct pipe_context *pipe, struct pipe_surface *ps,
    BEGIN_NV04(push, NV30_3D(CLEAR_BUFFERS), 1);
    PUSH_DATA (push, mode);
 
+   pipe_mutex_unlock(nv30->screen->base.push_mutex);
+
    nv30->dirty |= NV30_NEW_FRAMEBUFFER | NV30_NEW_SCISSOR;
 }
 

src/gallium/drivers/nouveau/nv30/nv30_context.c

@@ -201,6 +201,8 @@ nv30_context_create(struct pipe_screen *pscreen, void *priv, unsigned ctxflags)
    if (!nv30)
       return NULL;
 
+   pipe_mutex_lock(screen->base.push_mutex);
+
    nv30->screen = screen;
    nv30->base.screen = &screen->base;
    nv30->base.copy_data = nv30_transfer_copy_data;
@@ -226,6 +228,7 @@ nv30_context_create(struct pipe_screen *pscreen, void *priv, unsigned ctxflags)
    ret = nouveau_bufctx_new(nv30->base.client, 64, &nv30->bufctx);
    if (ret) {
       nv30_context_destroy(pipe);
+      pipe_mutex_unlock(screen->base.push_mutex);
       return NULL;
    }
 
@@ -259,10 +262,13 @@ nv30_context_create(struct pipe_screen *pscreen, void *priv, unsigned ctxflags)
    nv30->blitter = util_blitter_create(pipe);
    if (!nv30->blitter) {
       nv30_context_destroy(pipe);
+      pipe_mutex_unlock(screen->base.push_mutex);
       return NULL;
    }
 
    nouveau_context_init_vdec(&nv30->base);
 
+   pipe_mutex_unlock(screen->base.push_mutex);
+
    return pipe;
 }