intel: Protect bufmgr objects with a pthread mutex.

We want to be able to use the bufmgr from multiple threads for GL, and thus
we need to protect the internal structures.

The pthread-stubs package is used so that programs not linked against
pthreads get weak symbols to stubs and don't eat most of the cost.

5 files changed, 154 insertions, 29 deletions

diff --git a/configure.ac b/configure.ac
index 0cf09744..92507cb5 100644
--- a/configure.ac
+++ b/configure.ac
@@ -32,6 +32,10 @@ AC_PROG_CC
 AC_HEADER_STDC
 AC_SYS_LARGEFILE
 
+PKG_CHECK_MODULES(PTHREADSTUBS, pthread-stubs)
+AC_SUBST(PTHREADSTUBS_CFLAGS)
+AC_SUBST(PTHREADSTUBS_LIBS)
+
 pkgconfigdir=${libdir}/pkgconfig
 AC_SUBST(pkgconfigdir)
 AC_ARG_ENABLE(udev,	AS_HELP_STRING([--enable-udev],
diff --git a/libdrm/Makefile.am b/libdrm/Makefile.am
index f82d1917..543b2786 100644
--- a/libdrm/Makefile.am
+++ b/libdrm/Makefile.am
@@ -27,6 +27,7 @@ libdrm_la_LDFLAGS = -version-number 2:3:0 -no-undefined
 AM_CFLAGS = -I$(top_srcdir)/shared-core
 libdrm_la_SOURCES = xf86drm.c xf86drmHash.c xf86drmRandom.c xf86drmSL.c \
 	libdrm_lists.h
+libdrm_la_LIBADD = intel/libdrm_intel.la @PTHREADSTUBS_LIBS@
 
 libdrmincludedir = ${includedir}
 libdrminclude_HEADERS = xf86drm.h
diff --git a/libdrm/intel/Makefile.am b/libdrm/intel/Makefile.am
index 607c4765..92388c24 100644
--- a/libdrm/intel/Makefile.am
+++ b/libdrm/intel/Makefile.am
@@ -26,6 +26,7 @@ AM_CFLAGS = \
 	$(WARN_CFLAGS) \
 	-I$(top_srcdir)/libdrm \
 	-I$(top_srcdir)/libdrm/intel \
+	$(PTHREADSTUBS_CFLAGS) \
 	-I$(top_srcdir)/shared-core
 
 libdrm_intel_la_LTLIBRARIES = libdrm_intel.la
diff --git a/libdrm/intel/intel_bufmgr_fake.c b/libdrm/intel/intel_bufmgr_fake.c
index 22273c5b..c8f643db 100644
--- a/libdrm/intel/intel_bufmgr_fake.c
+++ b/libdrm/intel/intel_bufmgr_fake.c
@@ -43,6 +43,7 @@
 #include <assert.h>
 #include <errno.h>
 #include <xf86drm.h>
+#include <pthread.h>
 #include "intel_bufmgr.h"
 #include "intel_bufmgr_priv.h"
 #include "drm.h"
@@ -112,6 +113,8 @@ struct block {
 typedef struct _bufmgr_fake {
    dri_bufmgr bufmgr;
 
+   pthread_mutex_t lock;
+
    unsigned long low_offset;
    unsigned long size;
    void *virtual;
@@ -716,10 +719,16 @@ dri_fake_bo_wait_rendering(dri_bo *bo)
    dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr;
    dri_bo_fake *bo_fake = (dri_bo_fake *)bo;
 
-   if (bo_fake->block == NULL || !bo_fake->block->fenced)
+   pthread_mutex_lock(&bufmgr_fake->lock);
+
+   if (bo_fake->block == NULL || !bo_fake->block->fenced) {
+      pthread_mutex_unlock(&bufmgr_fake->lock);
       return;
+   }
 
    _fence_wait_internal(bufmgr_fake, bo_fake->block->fence);
+
+   pthread_mutex_unlock(&bufmgr_fake->lock);
 }
 
 /* Specifically ignore texture memory sharing.
@@ -732,6 +741,8 @@ intel_bufmgr_fake_contended_lock_take(dri_bufmgr *bufmgr)
    dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bufmgr;
    struct block *block, *tmp;
 
+   pthread_mutex_lock(&bufmgr_fake->lock);
+
    bufmgr_fake->need_fence = 1;
    bufmgr_fake->fail = 0;
 
@@ -751,6 +762,8 @@ intel_bufmgr_fake_contended_lock_take(dri_bufmgr *bufmgr)
       assert(_fence_test(bufmgr_fake, block->fence));
       set_dirty(block->bo);
    }
+
+   pthread_mutex_unlock(&bufmgr_fake->lock);
 }
 
 static dri_bo *
@@ -825,21 +838,29 @@ intel_bo_fake_alloc_static(dri_bufmgr *bufmgr, const char *name,
 static void
 dri_fake_bo_reference(dri_bo *bo)
 {
+   dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr;
    dri_bo_fake *bo_fake = (dri_bo_fake *)bo;
 
+   pthread_mutex_lock(&bufmgr_fake->lock);
    bo_fake->refcount++;
+   pthread_mutex_unlock(&bufmgr_fake->lock);
 }
 
 static void
-dri_fake_bo_unreference(dri_bo *bo)
+dri_fake_bo_reference_locked(dri_bo *bo)
+{
+   dri_bo_fake *bo_fake = (dri_bo_fake *)bo;
+
+   bo_fake->refcount++;
+}
+
+static void
+dri_fake_bo_unreference_locked(dri_bo *bo)
 {
    dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr;
    dri_bo_fake *bo_fake = (dri_bo_fake *)bo;
    int i;
 
-   if (!bo)
-      return;
-
    if (--bo_fake->refcount == 0) {
       assert(bo_fake->map_count == 0);
       /* No remaining references, so free it */
@@ -848,17 +869,25 @@ dri_fake_bo_unreference(dri_bo *bo)
       free_backing_store(bo);
 
       for (i = 0; i < bo_fake->nr_relocs; i++)
-	 dri_bo_unreference(bo_fake->relocs[i].target_buf);
+	 dri_fake_bo_unreference_locked(bo_fake->relocs[i].target_buf);
 
       DBG("drm_bo_unreference: free buf %d %s\n", bo_fake->id, bo_fake->name);
 
       free(bo_fake->relocs);
       free(bo);
-
-      return;
    }
 }
 
+static void
+dri_fake_bo_unreference(dri_bo *bo)
+{
+   dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr;
+
+   pthread_mutex_lock(&bufmgr_fake->lock);
+   dri_fake_bo_unreference_locked(bo);
+   pthread_mutex_unlock(&bufmgr_fake->lock);
+}
+
 /**
  * Set the buffer as not requiring backing store, and instead get the callback
  * invoked whenever it would be set dirty.
@@ -871,6 +900,8 @@ void intel_bo_fake_disable_backing_store(dri_bo *bo,
    dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr;
    dri_bo_fake *bo_fake = (dri_bo_fake *)bo;
 
+   pthread_mutex_lock(&bufmgr_fake->lock);
+
    if (bo_fake->backing_store)
       free_backing_store(bo);
 
@@ -887,6 +918,8 @@ void intel_bo_fake_disable_backing_store(dri_bo *bo,
     */
    if (invalidate_cb != NULL)
       invalidate_cb(bo, ptr);
+
+   pthread_mutex_unlock(&bufmgr_fake->lock);
 }
 
 /**
@@ -894,7 +927,7 @@ void intel_bo_fake_disable_backing_store(dri_bo *bo,
  * BM_NO_BACKING_STORE or BM_PINNED) or backing store, as necessary.
  */
 static int
-dri_fake_bo_map(dri_bo *bo, int write_enable)
+dri_fake_bo_map_locked(dri_bo *bo, int write_enable)
 {
    dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr;
    dri_bo_fake *bo_fake = (dri_bo_fake *)bo;
@@ -965,7 +998,20 @@ dri_fake_bo_map(dri_bo *bo, int write_enable)
 }
 
 static int
-dri_fake_bo_unmap(dri_bo *bo)
+dri_fake_bo_map(dri_bo *bo, int write_enable)
+{
+   dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr;
+   int ret;
+
+   pthread_mutex_lock(&bufmgr_fake->lock);
+   ret = dri_fake_bo_map_locked(bo, write_enable);
+   pthread_mutex_unlock(&bufmgr_fake->lock);
+
+   return ret;
+}
+
+static int
+dri_fake_bo_unmap_locked(dri_bo *bo)
 {
    dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr;
    dri_bo_fake *bo_fake = (dri_bo_fake *)bo;
@@ -986,11 +1032,26 @@ dri_fake_bo_unmap(dri_bo *bo)
    return 0;
 }
 
+static int
+dri_fake_bo_unmap(dri_bo *bo)
+{
+   dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr;
+   int ret;
+
+   pthread_mutex_lock(&bufmgr_fake->lock);
+   ret = dri_fake_bo_unmap_locked(bo);
+   pthread_mutex_unlock(&bufmgr_fake->lock);
+
+   return ret;
+}
+
 static void
 dri_fake_kick_all(dri_bufmgr_fake *bufmgr_fake)
 {
    struct block *block, *tmp;
 
+   pthread_mutex_lock(&bufmgr_fake->lock);
+
    bufmgr_fake->performed_rendering = 0;
    /* okay for ever BO that is on the HW kick it off.
       seriously not afraid of the POLICE right now */
@@ -1004,6 +1065,8 @@ dri_fake_kick_all(dri_bufmgr_fake *bufmgr_fake)
       if (!(bo_fake->flags & BM_NO_BACKING_STORE))
          bo_fake->dirty = 1;
    }
+
+   pthread_mutex_unlock(&bufmgr_fake->lock);
 }
 
 static int
@@ -1012,9 +1075,6 @@ dri_fake_bo_validate(dri_bo *bo)
    dri_bufmgr_fake *bufmgr_fake;
    dri_bo_fake *bo_fake = (dri_bo_fake *)bo;
 
-   /* XXX: Sanity-check whether we've already validated this one under
-    * different flags.  See drmAddValidateItem().
-    */
    bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr;
 
    DBG("drm_bo_validate: (buf %d: %s, %d kb)\n", bo_fake->id, bo_fake->name,
@@ -1097,6 +1157,7 @@ dri_fake_destroy(dri_bufmgr *bufmgr)
 {
    dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bufmgr;
 
+   pthread_mutex_destroy(&bufmgr_fake->lock);
    mmDestroy(bufmgr_fake->heap);
    free(bufmgr);
 }
@@ -1112,6 +1173,8 @@ dri_fake_emit_reloc(dri_bo *reloc_buf,
    dri_bo_fake *target_fake = (dri_bo_fake *)target_buf;
    int i;
 
+   pthread_mutex_lock(&bufmgr_fake->lock);
+
    assert(reloc_buf);
    assert(target_buf);
 
@@ -1124,7 +1187,7 @@ dri_fake_emit_reloc(dri_bo *reloc_buf,
 
    assert(reloc_fake->nr_relocs <= MAX_RELOCS);
 
-   dri_bo_reference(target_buf);
+   dri_fake_bo_reference_locked(target_buf);
 
    if (!target_fake->is_static)
       reloc_fake->child_size += ALIGN(target_buf->size, target_fake->alignment);
@@ -1145,6 +1208,8 @@ dri_fake_emit_reloc(dri_bo *reloc_buf,
       }
    }
 
+   pthread_mutex_unlock(&bufmgr_fake->lock);
+
    return 0;
 }
 
@@ -1190,7 +1255,7 @@ dri_fake_reloc_and_validate_buffer(dri_bo *bo)
          ret = dri_fake_reloc_and_validate_buffer(r->target_buf);
          if (ret != 0) {
             if (bo->virtual != NULL)
-                dri_bo_unmap(bo);
+                dri_fake_bo_unmap_locked(bo);
             return ret;
          }
       }
@@ -1200,7 +1265,7 @@ dri_fake_reloc_and_validate_buffer(dri_bo *bo)
 	 reloc_data = r->target_buf->offset + r->delta;
 
 	 if (bo->virtual == NULL)
-	    dri_bo_map(bo, 1);
+	    dri_fake_bo_map_locked(bo, 1);
 
 	 *(uint32_t *)((uint8_t *)bo->virtual + r->offset) = reloc_data;
 
@@ -1209,7 +1274,7 @@ dri_fake_reloc_and_validate_buffer(dri_bo *bo)
    }
 
    if (bo->virtual != NULL)
-      dri_bo_unmap(bo);
+      dri_fake_bo_unmap_locked(bo);
 
    if (bo_fake->write_domain != 0) {
       if (!(bo_fake->flags & (BM_NO_BACKING_STORE|BM_PINNED))) {
@@ -1272,6 +1337,8 @@ dri_fake_bo_exec(dri_bo *bo, int used,
    int ret;
    int retry_count = 0;
 
+   pthread_mutex_lock(&bufmgr_fake->lock);
+
    bufmgr_fake->performed_rendering = 0;
 
    dri_fake_calculate_domains(bo);
@@ -1316,6 +1383,8 @@ dri_fake_bo_exec(dri_bo *bo, int used,
 
    dri_bo_fake_post_submit(bo);
 
+   pthread_mutex_unlock(&bufmgr_fake->lock);
+
    return 0;
 }
 
@@ -1368,6 +1437,8 @@ intel_bufmgr_fake_evict_all(dri_bufmgr *bufmgr)
    dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bufmgr;
    struct block *block, *tmp;
 
+   pthread_mutex_lock(&bufmgr_fake->lock);
+
    bufmgr_fake->need_fence = 1;
    bufmgr_fake->fail = 0;
 
@@ -1387,6 +1458,8 @@ intel_bufmgr_fake_evict_all(dri_bufmgr *bufmgr)
       /* Releases the memory, and memcpys dirty contents out if necessary. */
       free_block(bufmgr_fake, block);
    }
+
+   pthread_mutex_unlock(&bufmgr_fake->lock);
 }
 void intel_bufmgr_fake_set_last_dispatch(dri_bufmgr *bufmgr,
 					 volatile unsigned int *last_dispatch)
@@ -1406,6 +1479,11 @@ intel_bufmgr_fake_init(int fd,
 
    bufmgr_fake = calloc(1, sizeof(*bufmgr_fake));
 
+   if (pthread_mutex_init(&bufmgr_fake->lock, NULL) != 0) {
+      free(bufmgr_fake);
+      return NULL;
+   }
+
    /* Initialize allocator */
    DRMINITLISTHEAD(&bufmgr_fake->fenced);
    DRMINITLISTHEAD(&bufmgr_fake->on_hardware);
diff --git a/libdrm/intel/intel_bufmgr_gem.c b/libdrm/intel/intel_bufmgr_gem.c
index cd36cdc7..9bd44417 100644
--- a/libdrm/intel/intel_bufmgr_gem.c
+++ b/libdrm/intel/intel_bufmgr_gem.c
@@ -44,6 +44,7 @@
 #include <string.h>
 #include <unistd.h>
 #include <assert.h>
+#include <pthread.h>
 #include <sys/ioctl.h>
 #include <sys/mman.h>
 
@@ -84,6 +85,8 @@ typedef struct _dri_bufmgr_gem {
 
     int max_relocs;
 
+    pthread_mutex_t lock;
+
     struct drm_i915_gem_exec_object *exec_objects;
     dri_bo **exec_bos;
     int exec_size;
@@ -133,6 +136,8 @@ struct _dri_bo_gem {
     dri_bo_gem *next;
 };
 
+static void dri_gem_bo_reference_locked(dri_bo *bo);
+
 static int
 logbase2(int n)
 {
@@ -237,7 +242,7 @@ intel_add_validate_buffer(dri_bo *bo)
     bufmgr_gem->exec_objects[index].alignment = 0;
     bufmgr_gem->exec_objects[index].offset = 0;
     bufmgr_gem->exec_bos[index] = bo;
-    dri_bo_reference(bo);
+    dri_gem_bo_reference_locked(bo);
     bufmgr_gem->exec_count++;
 }
 
@@ -285,6 +290,7 @@ dri_gem_bo_alloc(dri_bufmgr *bufmgr, const char *name,
 	    bo_size = page_size;
     }
 
+    pthread_mutex_lock(&bufmgr_gem->lock);
     /* Get a buffer out of the cache if available */
     if (bucket != NULL && bucket->num_entries > 0) {
 	struct drm_i915_gem_busy busy;
@@ -302,6 +308,7 @@ dri_gem_bo_alloc(dri_bufmgr *bufmgr, const char *name,
 	    bucket->num_entries--;
 	}
     }
+    pthread_mutex_unlock(&bufmgr_gem->lock);
 
     if (!alloc_from_cache) {
 	struct drm_i915_gem_create create;
@@ -379,6 +386,17 @@ intel_bo_gem_create_from_name(dri_bufmgr *bufmgr, const char *name,
 static void
 dri_gem_bo_reference(dri_bo *bo)
 {
+    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)bo->bufmgr;
+    dri_bo_gem *bo_gem = (dri_bo_gem *)bo;
+
+    pthread_mutex_lock(&bufmgr_gem->lock);
+    bo_gem->refcount++;
+    pthread_mutex_unlock(&bufmgr_gem->lock);
+}
+
+static void
+dri_gem_bo_reference_locked(dri_bo *bo)
+{
     dri_bo_gem *bo_gem = (dri_bo_gem *)bo;
 
     bo_gem->refcount++;
@@ -407,14 +425,11 @@ dri_gem_bo_free(dri_bo *bo)
 }
 
 static void
-dri_gem_bo_unreference(dri_bo *bo)
+dri_gem_bo_unreference_locked(dri_bo *bo)
 {
     dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)bo->bufmgr;
     dri_bo_gem *bo_gem = (dri_bo_gem *)bo;
 
-    if (!bo)
-	return;
-
     if (--bo_gem->refcount == 0) {
 	struct dri_gem_bo_bucket *bucket;
 
@@ -423,7 +438,7 @@ dri_gem_bo_unreference(dri_bo *bo)
 
 	    /* Unreference all the target buffers */
 	    for (i = 0; i < bo_gem->reloc_count; i++)
-		 dri_bo_unreference(bo_gem->reloc_target_bo[i]);
+		 dri_gem_bo_unreference_locked(bo_gem->reloc_target_bo[i]);
 	    free(bo_gem->reloc_target_bo);
 	    free(bo_gem->relocs);
 	}
@@ -451,20 +466,28 @@ dri_gem_bo_unreference(dri_bo *bo)
 	} else {
 	    dri_gem_bo_free(bo);
 	}
-
-	return;
     }
 }
 
+static void
+dri_gem_bo_unreference(dri_bo *bo)
+{
+    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)bo->bufmgr;
+
+    pthread_mutex_lock(&bufmgr_gem->lock);
+    dri_gem_bo_unreference_locked(bo);
+    pthread_mutex_unlock(&bufmgr_gem->lock);
+}
+
 static int
 dri_gem_bo_map(dri_bo *bo, int write_enable)
 {
-    dri_bufmgr_gem *bufmgr_gem;
+    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)bo->bufmgr;
     dri_bo_gem *bo_gem = (dri_bo_gem *)bo;
     struct drm_i915_gem_set_domain set_domain;
     int ret;
 
-    bufmgr_gem = (dri_bufmgr_gem *)bo->bufmgr;
+    pthread_mutex_lock(&bufmgr_gem->lock);
 
     /* Allow recursive mapping. Mesa may recursively map buffers with
      * nested display loops.
@@ -514,6 +537,8 @@ dri_gem_bo_map(dri_bo *bo, int write_enable)
 	bo_gem->swrast = 1;
     }
 
+    pthread_mutex_unlock(&bufmgr_gem->lock);
+
     return 0;
 }
 
@@ -530,6 +555,7 @@ dri_gem_bo_unmap(dri_bo *bo)
 
     assert(bo_gem->mapped);
 
+    pthread_mutex_lock(&bufmgr_gem->lock);
     if (bo_gem->swrast) {
 	sw_finish.handle = bo_gem->gem_handle;
 	do {
@@ -538,6 +564,7 @@ dri_gem_bo_unmap(dri_bo *bo)
 	} while (ret == -1 && errno == EINTR);
 	bo_gem->swrast = 0;
     }
+    pthread_mutex_unlock(&bufmgr_gem->lock);
     return 0;
 }
 
@@ -622,6 +649,8 @@ dri_bufmgr_gem_destroy(dri_bufmgr *bufmgr)
     free(bufmgr_gem->exec_objects);
     free(bufmgr_gem->exec_bos);
 
+    pthread_mutex_destroy(&bufmgr_gem->lock);
+
     /* Free any cached buffer objects we were going to reuse */
     for (i = 0; i < INTEL_GEM_BO_BUCKETS; i++) {
 	struct dri_gem_bo_bucket *bucket = &bufmgr_gem->cache_bucket[i];
@@ -657,6 +686,8 @@ dri_gem_bo_emit_reloc(dri_bo *bo, uint32_t read_domains, uint32_t write_domain,
     dri_bo_gem *bo_gem = (dri_bo_gem *)bo;
     dri_bo_gem *target_bo_gem = (dri_bo_gem *)target_bo;
 
+    pthread_mutex_lock(&bufmgr_gem->lock);
+
     /* Create a new relocation list if needed */
     if (bo_gem->relocs == NULL)
 	intel_setup_reloc_list(bo);
@@ -677,9 +708,12 @@ dri_gem_bo_emit_reloc(dri_bo *bo, uint32_t read_domains, uint32_t write_domain,
     bo_gem->relocs[bo_gem->reloc_count].presumed_offset = target_bo->offset;
 
     bo_gem->reloc_target_bo[bo_gem->reloc_count] = target_bo;
-    dri_bo_reference(target_bo);
+    dri_gem_bo_reference_locked(target_bo);
 
     bo_gem->reloc_count++;
+
+    pthread_mutex_unlock(&bufmgr_gem->lock);
+
     return 0;
 }
 
@@ -736,6 +770,7 @@ dri_gem_bo_exec(dri_bo *bo, int used,
     struct drm_i915_gem_execbuffer execbuf;
     int ret, i;
 
+    pthread_mutex_lock(&bufmgr_gem->lock);
     /* Update indices and set up the validate list. */
     dri_gem_bo_process_reloc(bo);
 
@@ -771,10 +806,11 @@ dri_gem_bo_exec(dri_bo *bo, int used,
 
 	/* Disconnect the buffer from the validate list */
 	bo_gem->validate_index = -1;
-	dri_bo_unreference(bo);
+	dri_gem_bo_unreference_locked(bo);
 	bufmgr_gem->exec_bos[i] = NULL;
     }
     bufmgr_gem->exec_count = 0;
+    pthread_mutex_unlock(&bufmgr_gem->lock);
 
     return 0;
 }
@@ -899,6 +935,11 @@ intel_bufmgr_gem_init(int fd, int batch_size)
     bufmgr_gem = calloc(1, sizeof(*bufmgr_gem));
     bufmgr_gem->fd = fd;
 
+    if (pthread_mutex_init(&bufmgr_gem->lock, NULL) != 0) {
+      free(bufmgr_gem);
+      return NULL;
+   }
+
     /* Let's go with one relocation per every 2 dwords (but round down a bit
      * since a power of two will mean an extra page allocation for the reloc
      * buffer).