diff --git a/src/asahi/lib/agx_bo.c b/src/asahi/lib/agx_bo.c
index a306d824f97..a028c60b125 100644
--- a/src/asahi/lib/agx_bo.c
+++ b/src/asahi/lib/agx_bo.c
@@ -97,9 +97,8 @@ agx_bo_cache_evict_stale_bos(struct agx_device *dev, unsigned tv_sec)
 }
 
 static void
-agx_bo_cache_put_locked(struct agx_bo *bo)
+agx_bo_cache_put_locked(struct agx_device *dev, struct agx_bo *bo)
 {
-   struct agx_device *dev = bo->dev;
    struct list_head *bucket = agx_bucket(dev, bo->size);
    struct timespec time;
 
@@ -132,15 +131,13 @@ agx_bo_cache_put_locked(struct agx_bo *bo)
 
 /* Tries to add a BO to the cache. Returns if it was successful */
 static bool
-agx_bo_cache_put(struct agx_bo *bo)
+agx_bo_cache_put(struct agx_device *dev, struct agx_bo *bo)
 {
-   struct agx_device *dev = bo->dev;
-
    if (bo->flags & AGX_BO_SHARED) {
       return false;
    } else {
       simple_mtx_lock(&dev->bo_cache.lock);
-      agx_bo_cache_put_locked(bo);
+      agx_bo_cache_put_locked(dev, bo);
       simple_mtx_unlock(&dev->bo_cache.lock);
 
       return true;
@@ -172,7 +169,7 @@ agx_bo_reference(struct agx_bo *bo)
 }
 
 void
-agx_bo_unreference(struct agx_bo *bo)
+agx_bo_unreference(struct agx_device *dev, struct agx_bo *bo)
 {
    if (!bo)
       return;
@@ -181,8 +178,6 @@ agx_bo_unreference(struct agx_bo *bo)
    if (p_atomic_dec_return(&bo->refcnt))
       return;
 
-   struct agx_device *dev = bo->dev;
-
    pthread_mutex_lock(&dev->bo_map_lock);
 
    /* Someone might have imported this BO while we were waiting for the
@@ -194,7 +189,7 @@ agx_bo_unreference(struct agx_bo *bo)
       if (dev->debug & AGX_DBG_TRACE)
          agxdecode_track_free(dev->agxdecode, bo);
 
-      if (!agx_bo_cache_put(bo))
+      if (!agx_bo_cache_put(dev, bo))
          agx_bo_free(dev, bo);
    }
 
diff --git a/src/asahi/lib/agx_bo.h b/src/asahi/lib/agx_bo.h
index 665e1d68549..27a7f4791a6 100644
--- a/src/asahi/lib/agx_bo.h
+++ b/src/asahi/lib/agx_bo.h
@@ -83,9 +83,6 @@ struct agx_bo {
    /* Current writer, if any (queue in upper 32 bits, syncobj in lower 32 bits) */
    uint64_t writer;
 
-   /* Owner */
-   struct agx_device *dev;
-
    /* Update atomically */
    int32_t refcnt;
 
@@ -132,9 +129,9 @@ agx_bo_create(struct agx_device *dev, unsigned size, enum agx_bo_flags flags,
 }
 
 void agx_bo_reference(struct agx_bo *bo);
-void agx_bo_unreference(struct agx_bo *bo);
+void agx_bo_unreference(struct agx_device *dev, struct agx_bo *bo);
 struct agx_bo *agx_bo_import(struct agx_device *dev, int fd);
-int agx_bo_export(struct agx_bo *bo);
+int agx_bo_export(struct agx_device *dev, struct agx_bo *bo);
 
 void agx_bo_free(struct agx_device *dev, struct agx_bo *bo);
 struct agx_bo *agx_bo_cache_fetch(struct agx_device *dev, size_t size,
diff --git a/src/asahi/lib/agx_device.c b/src/asahi/lib/agx_device.c
index 7eba78d7034..d5bf9a1056d 100644
--- a/src/asahi/lib/agx_device.c
+++ b/src/asahi/lib/agx_device.c
@@ -175,7 +175,6 @@ agx_bo_alloc(struct agx_device *dev, size_t size, size_t align,
    bo->size = gem_create.size;
    bo->align = MAX2(dev->params.vm_page_size, align);
    bo->flags = flags;
-   bo->dev = dev;
    bo->handle = handle;
    bo->prime_fd = -1;
 
@@ -207,7 +206,7 @@ agx_bo_alloc(struct agx_device *dev, size_t size, size_t align,
       return NULL;
    }
 
-   dev->ops.bo_mmap(bo);
+   dev->ops.bo_mmap(dev, bo);
 
    if (flags & AGX_BO_LOW_VA)
       bo->ptr.gpu -= dev->shader_base;
@@ -218,7 +217,7 @@ agx_bo_alloc(struct agx_device *dev, size_t size, size_t align,
 }
 
 static void
-agx_bo_mmap(struct agx_bo *bo)
+agx_bo_mmap(struct agx_device *dev, struct agx_bo *bo)
 {
    struct drm_asahi_gem_mmap_offset gem_mmap_offset = {.handle = bo->handle};
    int ret;
@@ -226,20 +225,20 @@ agx_bo_mmap(struct agx_bo *bo)
    if (bo->ptr.cpu)
       return;
 
-   ret =
-      drmIoctl(bo->dev->fd, DRM_IOCTL_ASAHI_GEM_MMAP_OFFSET, &gem_mmap_offset);
+   ret = drmIoctl(dev->fd, DRM_IOCTL_ASAHI_GEM_MMAP_OFFSET, &gem_mmap_offset);
    if (ret) {
       fprintf(stderr, "DRM_IOCTL_ASAHI_MMAP_BO failed: %m\n");
       assert(0);
    }
 
    bo->ptr.cpu = os_mmap(NULL, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED,
-                         bo->dev->fd, gem_mmap_offset.offset);
+                         dev->fd, gem_mmap_offset.offset);
    if (bo->ptr.cpu == MAP_FAILED) {
       bo->ptr.cpu = NULL;
+
       fprintf(stderr,
               "mmap failed: result=%p size=0x%llx fd=%i offset=0x%llx %m\n",
-              bo->ptr.cpu, (long long)bo->size, bo->dev->fd,
+              bo->ptr.cpu, (long long)bo->size, dev->fd,
               (long long)gem_mmap_offset.offset);
    }
 }
@@ -263,9 +262,9 @@ agx_bo_import(struct agx_device *dev, int fd)
    bo = agx_lookup_bo(dev, gem_handle);
    dev->max_handle = MAX2(dev->max_handle, gem_handle);
 
-   if (!bo->dev) {
-      bo->dev = dev;
+   if (!bo->size) {
       bo->size = lseek(fd, 0, SEEK_END);
+      bo->align = dev->params.vm_page_size;
 
       /* Sometimes this can fail and return -1. size of -1 is not
        * a nice thing for mmap to try mmap. Be more robust also
@@ -345,13 +344,13 @@ error:
 }
 
 int
-agx_bo_export(struct agx_bo *bo)
+agx_bo_export(struct agx_device *dev, struct agx_bo *bo)
 {
    int fd;
 
    assert(bo->flags & AGX_BO_SHAREABLE);
 
-   if (drmPrimeHandleToFD(bo->dev->fd, bo->handle, DRM_CLOEXEC, &fd))
+   if (drmPrimeHandleToFD(dev->fd, bo->handle, DRM_CLOEXEC, &fd))
       return -1;
 
    if (!(bo->flags & AGX_BO_SHARED)) {
@@ -366,11 +365,11 @@ agx_bo_export(struct agx_bo *bo)
       if (writer) {
          int out_sync_fd = -1;
          int ret = drmSyncobjExportSyncFile(
-            bo->dev->fd, agx_bo_writer_syncobj(writer), &out_sync_fd);
+            dev->fd, agx_bo_writer_syncobj(writer), &out_sync_fd);
          assert(ret >= 0);
          assert(out_sync_fd >= 0);
 
-         ret = agx_import_sync_file(bo->dev, bo, out_sync_fd);
+         ret = agx_import_sync_file(dev, bo, out_sync_fd);
          assert(ret >= 0);
          close(out_sync_fd);
       }
@@ -615,7 +614,7 @@ void
 agx_close_device(struct agx_device *dev)
 {
    ralloc_free((void *)dev->libagx);
-   agx_bo_unreference(dev->helper);
+   agx_bo_unreference(dev, dev->helper);
    agx_bo_cache_evict_all(dev);
    util_sparse_array_finish(&dev->bo_map);
    agxdecode_destroy_context(dev->agxdecode);
@@ -725,7 +724,7 @@ agx_debug_fault(struct agx_device *dev, uint64_t addr)
       if (bo->flags & AGX_BO_LOW_VA)
          bo_addr += dev->shader_base;
 
-      if (!bo->dev || bo_addr > addr)
+      if (!bo->size || bo_addr > addr)
          continue;
 
       if (!best || bo_addr > best->ptr.gpu)
diff --git a/src/asahi/lib/agx_device.h b/src/asahi/lib/agx_device.h
index f34b34ec9dd..2eda82700c8 100644
--- a/src/asahi/lib/agx_device.h
+++ b/src/asahi/lib/agx_device.h
@@ -64,7 +64,7 @@ typedef struct {
                               enum agx_bo_flags flags);
    int (*bo_bind)(struct agx_device *dev, struct agx_bo *bo, uint64_t addr,
                   uint32_t flags);
-   void (*bo_mmap)(struct agx_bo *bo);
+   void (*bo_mmap)(struct agx_device *dev, struct agx_bo *bo);
    ssize_t (*get_params)(struct agx_device *dev, void *buf, size_t size);
    int (*submit)(struct agx_device *dev, struct drm_asahi_submit *submit,
                  uint32_t vbo_res_id);
diff --git a/src/asahi/lib/agx_device_virtio.c b/src/asahi/lib/agx_device_virtio.c
index 8a31b8bb383..0bececefe7d 100644
--- a/src/asahi/lib/agx_device_virtio.c
+++ b/src/asahi/lib/agx_device_virtio.c
@@ -122,14 +122,13 @@ agx_virtio_bo_alloc(struct agx_device *dev, size_t size, size_t align,
    bo->size = size;
    bo->align = MAX2(dev->params.vm_page_size, align);
    bo->flags = flags;
-   bo->dev = dev;
    bo->handle = handle;
    bo->prime_fd = -1;
    bo->blob_id = blob_id;
    bo->ptr.gpu = ptr_gpu;
    bo->vbo_res_id = vdrm_handle_to_res_id(dev->vdrm, handle);
 
-   dev->ops.bo_mmap(bo);
+   dev->ops.bo_mmap(dev, bo);
 
    if (flags & AGX_BO_LOW_VA)
       bo->ptr.gpu -= dev->shader_base;
@@ -164,17 +163,17 @@ agx_virtio_bo_bind(struct agx_device *dev, struct agx_bo *bo, uint64_t addr,
 }
 
 static void
-agx_virtio_bo_mmap(struct agx_bo *bo)
+agx_virtio_bo_mmap(struct agx_device *dev, struct agx_bo *bo)
 {
    if (bo->ptr.cpu) {
       return;
    }
 
-   bo->ptr.cpu = vdrm_bo_map(bo->dev->vdrm, bo->handle, bo->size, NULL);
+   bo->ptr.cpu = vdrm_bo_map(dev->vdrm, bo->handle, bo->size, NULL);
    if (bo->ptr.cpu == MAP_FAILED) {
       bo->ptr.cpu = NULL;
       fprintf(stderr, "mmap failed: result=%p size=0x%llx fd=%i\n", bo->ptr.cpu,
-              (long long)bo->size, bo->dev->fd);
+              (long long)bo->size, dev->fd);
    }
 }
 
diff --git a/src/asahi/lib/agx_scratch.c b/src/asahi/lib/agx_scratch.c
index 6d05e2c01bf..1715178230f 100644
--- a/src/asahi/lib/agx_scratch.c
+++ b/src/asahi/lib/agx_scratch.c
@@ -92,7 +92,7 @@ static void
 agx_scratch_realloc(struct agx_scratch *scratch)
 {
    if (scratch->buf)
-      agx_bo_unreference(scratch->buf);
+      agx_bo_unreference(scratch->dev, scratch->buf);
 
    struct spill_size size = agx_scratch_get_spill_size(scratch->size_dwords);
 
@@ -301,6 +301,6 @@ void
 agx_scratch_fini(struct agx_scratch *scratch)
 {
    if (scratch->buf)
-      agx_bo_unreference(scratch->buf);
+      agx_bo_unreference(scratch->dev, scratch->buf);
    scratch->buf = NULL;
 }
diff --git a/src/asahi/lib/pool.c b/src/asahi/lib/pool.c
index 7a8a528ccd0..c931574c322 100644
--- a/src/asahi/lib/pool.c
+++ b/src/asahi/lib/pool.c
@@ -45,7 +45,7 @@ void
 agx_pool_cleanup(struct agx_pool *pool)
 {
    util_dynarray_foreach(&pool->bos, struct agx_bo *, bo) {
-      agx_bo_unreference(*bo);
+      agx_bo_unreference(pool->dev, *bo);
    }
 
    util_dynarray_fini(&pool->bos);
diff --git a/src/asahi/vulkan/hk_cmd_buffer.c b/src/asahi/vulkan/hk_cmd_buffer.c
index b3b362bf2b7..03fd005adc9 100644
--- a/src/asahi/vulkan/hk_cmd_buffer.c
+++ b/src/asahi/vulkan/hk_cmd_buffer.c
@@ -48,6 +48,7 @@ static void
 hk_free_resettable_cmd_buffer(struct hk_cmd_buffer *cmd)
 {
    struct hk_cmd_pool *pool = hk_cmd_buffer_pool(cmd);
+   struct hk_device *dev = hk_cmd_pool_device(pool);
 
    hk_descriptor_state_fini(cmd, &cmd->state.gfx.descriptors);
    hk_descriptor_state_fini(cmd, &cmd->state.cs.descriptors);
@@ -61,7 +62,7 @@ hk_free_resettable_cmd_buffer(struct hk_cmd_buffer *cmd)
    }
 
    util_dynarray_foreach(&cmd->large_bos, struct agx_bo *, bo) {
-      agx_bo_unreference(*bo);
+      agx_bo_unreference(&dev->dev, *bo);
    }
 
    util_dynarray_clear(&cmd->large_bos);
diff --git a/src/asahi/vulkan/hk_cmd_pool.c b/src/asahi/vulkan/hk_cmd_pool.c
index a3f2a85468a..354d5e0e085 100644
--- a/src/asahi/vulkan/hk_cmd_pool.c
+++ b/src/asahi/vulkan/hk_cmd_pool.c
@@ -38,7 +38,8 @@ hk_cmd_bo_create(struct hk_cmd_pool *pool, bool usc, struct hk_cmd_bo **bo_out)
 static void
 hk_cmd_bo_destroy(struct hk_cmd_pool *pool, struct hk_cmd_bo *bo)
 {
-   agx_bo_unreference(bo->bo);
+   struct hk_device *dev = hk_cmd_pool_device(pool);
+   agx_bo_unreference(&dev->dev, bo->bo);
    vk_free(&pool->vk.alloc, bo);
 }
 
diff --git a/src/asahi/vulkan/hk_descriptor_set.c b/src/asahi/vulkan/hk_descriptor_set.c
index b59a9ac4b57..4ce45fb8385 100644
--- a/src/asahi/vulkan/hk_descriptor_set.c
+++ b/src/asahi/vulkan/hk_descriptor_set.c
@@ -419,7 +419,7 @@ hk_destroy_descriptor_pool(struct hk_device *dev,
       hk_descriptor_set_destroy(dev, pool, set);
 
    util_vma_heap_finish(&pool->heap);
-   agx_bo_unreference(pool->bo);
+   agx_bo_unreference(&dev->dev, pool->bo);
 
    vk_object_free(&dev->vk, pAllocator, pool);
 }
diff --git a/src/asahi/vulkan/hk_descriptor_table.c b/src/asahi/vulkan/hk_descriptor_table.c
index 6d07ac6f384..968b2c234c8 100644
--- a/src/asahi/vulkan/hk_descriptor_table.c
+++ b/src/asahi/vulkan/hk_descriptor_table.c
@@ -88,7 +88,7 @@ void
 hk_descriptor_table_finish(struct hk_device *dev,
                            struct hk_descriptor_table *table)
 {
-   agx_bo_unreference(table->bo);
+   agx_bo_unreference(&dev->dev, table->bo);
    vk_free(&dev->vk.alloc, table->free_table);
    simple_mtx_destroy(&table->mutex);
 }
diff --git a/src/asahi/vulkan/hk_device.c b/src/asahi/vulkan/hk_device.c
index f5c4535aca2..08e7a344df8 100644
--- a/src/asahi/vulkan/hk_device.c
+++ b/src/asahi/vulkan/hk_device.c
@@ -440,7 +440,7 @@ fail_mem_cache:
 fail_queue:
    hk_queue_finish(dev, &dev->queue);
 fail_rodata:
-   agx_bo_unreference(dev->rodata.bo);
+   agx_bo_unreference(&dev->dev, dev->rodata.bo);
 fail_bg_eot:
    agx_bg_eot_cleanup(&dev->bg_eot);
 fail_internal_shaders_2:
@@ -487,8 +487,8 @@ hk_DestroyDevice(VkDevice _device, const VkAllocationCallbacks *pAllocator)
    hk_destroy_sampler_heap(dev, &dev->samplers);
    hk_descriptor_table_finish(dev, &dev->images);
    hk_descriptor_table_finish(dev, &dev->occlusion_queries);
-   agx_bo_unreference(dev->rodata.bo);
-   agx_bo_unreference(dev->heap);
+   agx_bo_unreference(&dev->dev, dev->rodata.bo);
+   agx_bo_unreference(&dev->dev, dev->heap);
    agx_bg_eot_cleanup(&dev->bg_eot);
    agx_close_device(&dev->dev);
    vk_free(&dev->vk.alloc, dev);
diff --git a/src/asahi/vulkan/hk_device_memory.c b/src/asahi/vulkan/hk_device_memory.c
index 0d10a55f5df..5b70379270c 100644
--- a/src/asahi/vulkan/hk_device_memory.c
+++ b/src/asahi/vulkan/hk_device_memory.c
@@ -80,7 +80,7 @@ hk_GetMemoryFdPropertiesKHR(VkDevice device,
 
    pMemoryFdProperties->memoryTypeBits = type_bits;
 
-   agx_bo_unreference(bo);
+   agx_bo_unreference(&dev->dev, bo);
 
    return VK_SUCCESS;
 }
@@ -190,7 +190,7 @@ hk_FreeMemory(VkDevice device, VkDeviceMemory _mem,
    struct hk_memory_heap *heap = &pdev->mem_heaps[type->heapIndex];
    p_atomic_add(&heap->used, -((int64_t)mem->bo->size));
 
-   agx_bo_unreference(mem->bo);
+   agx_bo_unreference(&dev->dev, mem->bo);
 
    vk_device_memory_destroy(&dev->vk, pAllocator, &mem->vk);
 }
@@ -312,7 +312,7 @@ hk_GetMemoryFdKHR(VkDevice device, const VkMemoryGetFdInfoKHR *pGetFdInfo,
    switch (pGetFdInfo->handleType) {
    case VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT:
    case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
-      *pFD = agx_bo_export(memory->bo);
+      *pFD = agx_bo_export(&dev->dev, memory->bo);
       return VK_SUCCESS;
    default:
       assert(!"unsupported handle type");
diff --git a/src/asahi/vulkan/hk_event.c b/src/asahi/vulkan/hk_event.c
index aadbb272e76..83f4e86af69 100644
--- a/src/asahi/vulkan/hk_event.c
+++ b/src/asahi/vulkan/hk_event.c
@@ -52,7 +52,7 @@ hk_DestroyEvent(VkDevice device, VkEvent _event,
    if (!event)
       return;
 
-   agx_bo_unreference(event->bo);
+   agx_bo_unreference(&dev->dev, event->bo);
    vk_object_free(&dev->vk, pAllocator, event);
 }
 
diff --git a/src/asahi/vulkan/hk_query_pool.c b/src/asahi/vulkan/hk_query_pool.c
index 5762c69419c..b88baac2a3a 100644
--- a/src/asahi/vulkan/hk_query_pool.c
+++ b/src/asahi/vulkan/hk_query_pool.c
@@ -140,7 +140,7 @@ hk_DestroyQueryPool(VkDevice device, VkQueryPool queryPool,
       hk_descriptor_table_remove(dev, &dev->occlusion_queries, oq_index[i]);
    }
 
-   agx_bo_unreference(pool->bo);
+   agx_bo_unreference(&dev->dev, pool->bo);
    vk_query_pool_destroy(&dev->vk, pAllocator, &pool->vk);
 }
 
diff --git a/src/asahi/vulkan/hk_shader.c b/src/asahi/vulkan/hk_shader.c
index cfe67c28382..1eea687cc87 100644
--- a/src/asahi/vulkan/hk_shader.c
+++ b/src/asahi/vulkan/hk_shader.c
@@ -869,30 +869,30 @@ hk_compile_nir(struct hk_device *dev, const VkAllocationCallbacks *pAllocator,
 static const struct vk_shader_ops hk_shader_ops;
 
 static void
-hk_destroy_linked_shader(struct hk_linked_shader *linked)
+hk_destroy_linked_shader(struct hk_device *dev, struct hk_linked_shader *linked)
 {
-   agx_bo_unreference(linked->b.bo);
+   agx_bo_unreference(&dev->dev, linked->b.bo);
    ralloc_free(linked);
 }
 
 static void
-hk_destroy_linked_shader_ht(struct hash_entry *he)
-{
-   hk_destroy_linked_shader(he->data);
-}
-
-static void
-hk_shader_destroy(struct hk_shader *s)
+hk_shader_destroy(struct hk_device *dev, struct hk_shader *s)
 {
    free((void *)s->code_ptr);
    free((void *)s->data_ptr);
-   agx_bo_unreference(s->bo);
+   agx_bo_unreference(&dev->dev, s->bo);
 
    simple_mtx_destroy(&s->linked.lock);
-   _mesa_hash_table_destroy(s->linked.ht, hk_destroy_linked_shader_ht);
 
    if (s->only_linked)
-      hk_destroy_linked_shader(s->only_linked);
+      hk_destroy_linked_shader(dev, s->only_linked);
+
+   if (s->linked.ht) {
+      hash_table_foreach(s->linked.ht, entry) {
+         hk_destroy_linked_shader(dev, entry->data);
+      }
+      _mesa_hash_table_destroy(s->linked.ht, NULL);
+   }
 }
 
 void
@@ -904,7 +904,7 @@ hk_api_shader_destroy(struct vk_device *vk_dev, struct vk_shader *vk_shader,
       container_of(vk_shader, struct hk_api_shader, vk);
 
    hk_foreach_variant(obj, shader) {
-      hk_shader_destroy(shader);
+      hk_shader_destroy(dev, shader);
    }
 
    vk_shader_free(&dev->vk, pAllocator, &obj->vk);
diff --git a/src/gallium/drivers/asahi/agx_batch.c b/src/gallium/drivers/asahi/agx_batch.c
index 3981a20ce7c..9b23cd485d7 100644
--- a/src/gallium/drivers/asahi/agx_batch.c
+++ b/src/gallium/drivers/asahi/agx_batch.c
@@ -136,7 +136,7 @@ agx_batch_init(struct agx_context *ctx,
    batch->initialized = false;
    batch->draws = 0;
    batch->incoherent_writes = false;
-   agx_bo_unreference(batch->sampler_heap.bo);
+   agx_bo_unreference(dev, batch->sampler_heap.bo);
    batch->sampler_heap.bo = NULL;
    batch->sampler_heap.count = 0;
    batch->vs_scratch = false;
@@ -347,7 +347,7 @@ agx_batch_cleanup(struct agx_context *ctx, struct agx_batch *batch, bool reset)
          /* We should write no buffers if this is an empty batch */
          assert(agx_writer_get(ctx, handle) != batch);
 
-         agx_bo_unreference(agx_lookup_bo(dev, handle));
+         agx_bo_unreference(dev, agx_lookup_bo(dev, handle));
       }
    } else {
       int handle;
@@ -363,12 +363,12 @@ agx_batch_cleanup(struct agx_context *ctx, struct agx_batch *batch, bool reset)
          p_atomic_cmpxchg(&bo->writer,
                           agx_bo_writer(ctx->queue_id, batch->syncobj), 0);
 
-         agx_bo_unreference(agx_lookup_bo(dev, handle));
+         agx_bo_unreference(dev, agx_lookup_bo(dev, handle));
       }
    }
 
-   agx_bo_unreference(batch->vdm.bo);
-   agx_bo_unreference(batch->cdm.bo);
+   agx_bo_unreference(dev, batch->vdm.bo);
+   agx_bo_unreference(dev, batch->cdm.bo);
    agx_pool_cleanup(&batch->pool);
    agx_pool_cleanup(&batch->pipeline_pool);
 
diff --git a/src/gallium/drivers/asahi/agx_pipe.c b/src/gallium/drivers/asahi/agx_pipe.c
index 259212df6cb..70d0cfea2c1 100644
--- a/src/gallium/drivers/asahi/agx_pipe.c
+++ b/src/gallium/drivers/asahi/agx_pipe.c
@@ -281,7 +281,7 @@ agx_resource_get_handle(struct pipe_screen *pscreen, struct pipe_context *ctx,
 
       handle->handle = rsrc->bo->handle;
    } else if (handle->type == WINSYS_HANDLE_TYPE_FD) {
-      int fd = agx_bo_export(rsrc->bo);
+      int fd = agx_bo_export(dev, rsrc->bo);
 
       if (fd < 0)
          return false;
@@ -631,7 +631,7 @@ agx_resource_destroy(struct pipe_screen *screen, struct pipe_resource *prsrc)
    if (rsrc->scanout)
       renderonly_scanout_destroy(rsrc->scanout, agx_screen->dev.ro);
 
-   agx_bo_unreference(rsrc->bo);
+   agx_bo_unreference(&agx_screen->dev, rsrc->bo);
    FREE(rsrc);
 }
 
@@ -717,7 +717,7 @@ agx_shadow(struct agx_context *ctx, struct agx_resource *rsrc, bool needs_copy)
    }
 
    /* Swap the pointers, dropping a reference */
-   agx_bo_unreference(rsrc->bo);
+   agx_bo_unreference(dev, rsrc->bo);
    rsrc->bo = new_;
 
    /* Reemit descriptors using this resource */
@@ -1000,11 +1000,11 @@ agx_transfer_map(struct pipe_context *pctx, struct pipe_resource *resource,
          agx_sync_writer(ctx, staging, "GPU read staging blit");
       }
 
-      dev->ops.bo_mmap(staging->bo);
+      dev->ops.bo_mmap(dev, staging->bo);
       return staging->bo->ptr.cpu;
    }
 
-   dev->ops.bo_mmap(rsrc->bo);
+   dev->ops.bo_mmap(dev, rsrc->bo);
 
    if (ail_is_level_twiddled_uncompressed(&rsrc->layout, level)) {
       /* Should never happen for buffers, and it's not safe */
@@ -1791,7 +1791,7 @@ agx_destroy_context(struct pipe_context *pctx)
    agx_bg_eot_cleanup(&ctx->bg_eot);
    agx_destroy_meta_shaders(ctx);
 
-   agx_bo_unreference(ctx->result_buf);
+   agx_bo_unreference(dev, ctx->result_buf);
 
    /* Lock around the syncobj destruction, to avoid racing
     * command submission in another context.
diff --git a/src/gallium/drivers/asahi/agx_query.c b/src/gallium/drivers/asahi/agx_query.c
index 65fc3731e17..9dd49ad6eb5 100644
--- a/src/gallium/drivers/asahi/agx_query.c
+++ b/src/gallium/drivers/asahi/agx_query.c
@@ -48,6 +48,7 @@ is_timer(struct agx_query *query)
 
 struct agx_oq_heap {
    /* The GPU allocation itself */
+   struct agx_device *dev;
    struct agx_bo *bo;
 
    /* Bitset of query indices that are in use */
@@ -58,7 +59,7 @@ static void
 agx_destroy_oq_heap(void *heap_)
 {
    struct agx_oq_heap *heap = heap_;
-   agx_bo_unreference(heap->bo);
+   agx_bo_unreference(heap->dev, heap->bo);
 }
 
 static struct agx_oq_heap *
@@ -67,9 +68,10 @@ agx_alloc_oq_heap(struct agx_context *ctx)
    struct agx_oq_heap *heap = rzalloc(ctx, struct agx_oq_heap);
    ralloc_set_destructor(heap, agx_destroy_oq_heap);
 
-   heap->bo = agx_bo_create(agx_device(ctx->base.screen),
-                            AGX_MAX_OCCLUSION_QUERIES * sizeof(uint64_t),
-                            AGX_BO_WRITEBACK, "Occlusion query heap");
+   heap->dev = agx_device(ctx->base.screen);
+   heap->bo =
+      agx_bo_create(heap->dev, AGX_MAX_OCCLUSION_QUERIES * sizeof(uint64_t),
+                    AGX_BO_WRITEBACK, "Occlusion query heap");
 
    /* At the start, everything is available */
    BITSET_ONES(heap->available);
@@ -216,6 +218,7 @@ agx_destroy_query(struct pipe_context *pctx, struct pipe_query *pquery)
 {
    struct agx_context *ctx = agx_context(pctx);
    struct agx_query *query = (struct agx_query *)pquery;
+   struct agx_device *dev = agx_device(pctx->screen);
 
    /* We don't reference count the occlusion query allocations, so we need to
     * sync writers when destroying so we can freely write from the CPU after
@@ -228,7 +231,7 @@ agx_destroy_query(struct pipe_context *pctx, struct pipe_query *pquery)
       sync_query_writers(ctx, query, "Occlusion query destroy");
       agx_free_oq(ctx, query);
    } else {
-      agx_bo_unreference(query->bo);
+      agx_bo_unreference(dev, query->bo);
    }
 
    free(pquery);
diff --git a/src/gallium/drivers/asahi/agx_state.c b/src/gallium/drivers/asahi/agx_state.c
index 5f28927356e..71d4f281059 100644
--- a/src/gallium/drivers/asahi/agx_state.c
+++ b/src/gallium/drivers/asahi/agx_state.c
@@ -2463,34 +2463,35 @@ agx_bind_cs_state(struct pipe_context *pctx, void *cso)
 }
 
 /* Forward declare because of the recursion hit with geometry shaders */
-static void agx_delete_uncompiled_shader(struct agx_uncompiled_shader *so);
+static void agx_delete_uncompiled_shader(struct agx_device *dev,
+                                         struct agx_uncompiled_shader *so);
 
 static void
-agx_delete_compiled_shader_internal(struct agx_compiled_shader *so)
+agx_delete_compiled_shader(struct agx_device *dev,
+                           struct agx_compiled_shader *so)
 {
    if (so->gs_count)
-      agx_delete_compiled_shader_internal(so->gs_count);
+      agx_delete_compiled_shader(dev, so->gs_count);
 
    if (so->pre_gs)
-      agx_delete_compiled_shader_internal(so->pre_gs);
+      agx_delete_compiled_shader(dev, so->pre_gs);
 
    if (so->gs_copy)
-      agx_delete_compiled_shader_internal(so->gs_copy);
+      agx_delete_compiled_shader(dev, so->gs_copy);
 
-   agx_bo_unreference(so->bo);
+   agx_bo_unreference(dev, so->bo);
    FREE(so);
 }
 
 static void
-agx_delete_compiled_shader(struct hash_entry *ent)
+agx_delete_uncompiled_shader(struct agx_device *dev,
+                             struct agx_uncompiled_shader *so)
 {
-   agx_delete_compiled_shader_internal(ent->data);
-}
+   hash_table_foreach(so->variants, ent) {
+      agx_delete_compiled_shader(dev, ent->data);
+   }
 
-static void
-agx_delete_uncompiled_shader(struct agx_uncompiled_shader *so)
-{
-   _mesa_hash_table_destroy(so->variants, agx_delete_compiled_shader);
+   _mesa_hash_table_destroy(so->variants, NULL);
    blob_finish(&so->serialized_nir);
    blob_finish(&so->early_serialized_nir);
 
@@ -2498,14 +2499,15 @@ agx_delete_uncompiled_shader(struct agx_uncompiled_shader *so)
       for (unsigned j = 0; j < 3; ++j) {
          for (unsigned k = 0; k < 2; ++k) {
             if (so->passthrough_progs[i][j][k])
-               agx_delete_uncompiled_shader(so->passthrough_progs[i][j][k]);
+               agx_delete_uncompiled_shader(dev,
+                                            so->passthrough_progs[i][j][k]);
          }
       }
    }
 
    for (unsigned i = 0; i < ARRAY_SIZE(so->passthrough_tcs); ++i) {
       if (so->passthrough_tcs[i])
-         agx_delete_uncompiled_shader(so->passthrough_tcs[i]);
+         agx_delete_uncompiled_shader(dev, so->passthrough_tcs[i]);
    }
 
    ralloc_free(so);
@@ -2514,7 +2516,8 @@ agx_delete_uncompiled_shader(struct agx_uncompiled_shader *so)
 static void
 agx_delete_shader_state(struct pipe_context *ctx, void *cso)
 {
-   agx_delete_uncompiled_shader(cso);
+   struct agx_device *dev = agx_device(ctx->screen);
+   agx_delete_uncompiled_shader(dev, cso);
 }
 
 struct agx_generic_meta_key {
@@ -2552,7 +2555,12 @@ agx_init_meta_shaders(struct agx_context *ctx)
 void
 agx_destroy_meta_shaders(struct agx_context *ctx)
 {
-   _mesa_hash_table_destroy(ctx->generic_meta, agx_delete_compiled_shader);
+   struct agx_device *dev = agx_device(ctx->base.screen);
+   hash_table_foreach(ctx->generic_meta, ent) {
+      agx_delete_compiled_shader(dev, ent->data);
+   }
+
+   _mesa_hash_table_destroy(ctx->generic_meta, NULL);
 }
 
 static struct agx_compiled_shader *