virgl/drm/virgl_drm_winsys.c

/*
 * Copyright 2014, 2015 Red Hat.
 *
 * 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
 * on the rights to use, copy, modify, merge, publish, distribute, sub
 * license, 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 (including the next
 * paragraph) 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 NON-INFRINGEMENT. IN NO EVENT SHALL
 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS 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.
 */

#include <errno.h>
#include <fcntl.h>
#include <limits.h>
#include <stdio.h>
#include <sys/ioctl.h>
#include <sys/stat.h>

#include "os/os_mman.h"
#include "util/os_file.h"
#include "util/os_time.h"
#include "util/u_memory.h"
#include "util/format/u_format.h"
#include "util/u_hash_table.h"
#include "util/u_inlines.h"
#include "util/u_pointer.h"
#include "frontend/drm_driver.h"
#include "virgl/virgl_screen.h"
#include "virgl/virgl_public.h"
#include "virtio-gpu/virgl_protocol.h"

#include <xf86drm.h>
#include <libsync.h>
#include "drm-uapi/virtgpu_drm.h"

#include "virgl_drm_winsys.h"
#include "virgl_drm_public.h"

// Delete local definitions when virglrenderer_hw.h becomes public
#define VIRGL_DRM_CAPSET_VIRGL  1
#define VIRGL_DRM_CAPSET_VIRGL2 2

#define VIRGL_DRM_VERSION(major, minor) ((major) << 16 | (minor))
#define VIRGL_DRM_VERSION_FENCE_FD      VIRGL_DRM_VERSION(0, 1)

/* Gets a pointer to the virgl_hw_res containing the pointed to cache entry. */
#define cache_entry_container_res(ptr) \
    (struct virgl_hw_res*)((char*)ptr - offsetof(struct virgl_hw_res, cache_entry))

static inline boolean can_cache_resource(uint32_t bind)
{
   return bind == VIRGL_BIND_CONSTANT_BUFFER ||
          bind == VIRGL_BIND_INDEX_BUFFER ||
          bind == VIRGL_BIND_VERTEX_BUFFER ||
          bind == VIRGL_BIND_CUSTOM ||
          bind == VIRGL_BIND_STAGING ||
          bind == VIRGL_BIND_DEPTH_STENCIL ||
          bind == VIRGL_BIND_RENDER_TARGET ||
          bind == 0;
}

static void virgl_hw_res_destroy(struct virgl_drm_winsys *qdws,
                                 struct virgl_hw_res *res)
{
      struct drm_gem_close args;

      mtx_lock(&qdws->bo_handles_mutex);

      /* We intentionally avoid taking the lock in
       * virgl_drm_resource_reference. Now that the
       * lock is taken, we need to check the refcount
       * again. */
      if (pipe_is_referenced(&res->reference)) {
         mtx_unlock(&qdws->bo_handles_mutex);
         return;
      }

      _mesa_hash_table_remove_key(qdws->bo_handles,
                             (void *)(uintptr_t)res->bo_handle);
      if (res->flink_name)
         _mesa_hash_table_remove_key(qdws->bo_names,
                                (void *)(uintptr_t)res->flink_name);
      mtx_unlock(&qdws->bo_handles_mutex);
      if (res->ptr)
         os_munmap(res->ptr, res->size);

      memset(&args, 0, sizeof(args));
      args.handle = res->bo_handle;
      drmIoctl(qdws->fd, DRM_IOCTL_GEM_CLOSE, &args);
      FREE(res);
}

static boolean virgl_drm_resource_is_busy(struct virgl_winsys *vws,
                                          struct virgl_hw_res *res)
{
   struct virgl_drm_winsys *vdws = virgl_drm_winsys(vws);
   struct drm_virtgpu_3d_wait waitcmd;
   int ret;

   if (!p_atomic_read(&res->maybe_busy) && !p_atomic_read(&res->external))
      return false;

   memset(&waitcmd, 0, sizeof(waitcmd));
   waitcmd.handle = res->bo_handle;
   waitcmd.flags = VIRTGPU_WAIT_NOWAIT;

   ret = drmIoctl(vdws->fd, DRM_IOCTL_VIRTGPU_WAIT, &waitcmd);
   if (ret && errno == EBUSY)
      return TRUE;

   p_atomic_set(&res->maybe_busy, false);

   return FALSE;
}

static void
virgl_drm_winsys_destroy(struct virgl_winsys *qws)
{
   struct virgl_drm_winsys *qdws = virgl_drm_winsys(qws);

   virgl_resource_cache_flush(&qdws->cache);

   _mesa_hash_table_destroy(qdws->bo_handles, NULL);
   _mesa_hash_table_destroy(qdws->bo_names, NULL);
   mtx_destroy(&qdws->bo_handles_mutex);
   mtx_destroy(&qdws->mutex);

   FREE(qdws);
}

static void virgl_drm_resource_reference(struct virgl_winsys *qws,
                                         struct virgl_hw_res **dres,
                                         struct virgl_hw_res *sres)
{
   struct virgl_drm_winsys *qdws = virgl_drm_winsys(qws);
   struct virgl_hw_res *old = *dres;

   if (pipe_reference(&(*dres)->reference, &sres->reference)) {

      if (!can_cache_resource(old->bind) ||
          p_atomic_read(&old->external)) {
         virgl_hw_res_destroy(qdws, old);
      } else {
         mtx_lock(&qdws->mutex);
         virgl_resource_cache_add(&qdws->cache, &old->cache_entry);
         mtx_unlock(&qdws->mutex);
      }
   }
   *dres = sres;
}

static struct virgl_hw_res *
virgl_drm_winsys_resource_create_blob(struct virgl_winsys *qws,
                                      enum pipe_texture_target target,
                                      uint32_t format,
                                      uint32_t bind,
                                      uint32_t width,
                                      uint32_t height,
                                      uint32_t depth,
                                      uint32_t array_size,
                                      uint32_t last_level,
                                      uint32_t nr_samples,
                                      uint32_t flags,
                                      uint32_t size)
{
   int ret;
   int32_t blob_id;
   uint32_t cmd[VIRGL_PIPE_RES_CREATE_SIZE + 1] = { 0 };
   struct virgl_drm_winsys *qdws = virgl_drm_winsys(qws);
   struct drm_virtgpu_resource_create_blob drm_rc_blob = { 0 };
   struct virgl_hw_res *res;
   struct virgl_resource_params params = { .size = size,
                                           .bind = bind,
                                           .format = format,
                                           .flags = flags,
                                           .nr_samples = nr_samples,
                                           .width = width,
                                           .height = height,
                                           .depth = depth,
                                           .array_size = array_size,
                                           .last_level = last_level,
                                           .target = target };

   res = CALLOC_STRUCT(virgl_hw_res);
   if (!res)
      return NULL;

   /* Make sure blob is page aligned. */
   if (flags & (VIRGL_RESOURCE_FLAG_MAP_PERSISTENT |
                VIRGL_RESOURCE_FLAG_MAP_COHERENT)) {
      width = ALIGN(width, getpagesize());
      size = ALIGN(size, getpagesize());
   }

   blob_id = p_atomic_inc_return(&qdws->blob_id);
   cmd[0] = VIRGL_CMD0(VIRGL_CCMD_PIPE_RESOURCE_CREATE, 0, VIRGL_PIPE_RES_CREATE_SIZE);
   cmd[VIRGL_PIPE_RES_CREATE_FORMAT] = format;
   cmd[VIRGL_PIPE_RES_CREATE_BIND] = bind;
   cmd[VIRGL_PIPE_RES_CREATE_TARGET] = target;
   cmd[VIRGL_PIPE_RES_CREATE_WIDTH] = width;
   cmd[VIRGL_PIPE_RES_CREATE_HEIGHT] = height;
   cmd[VIRGL_PIPE_RES_CREATE_DEPTH] = depth;
   cmd[VIRGL_PIPE_RES_CREATE_ARRAY_SIZE] = array_size;
   cmd[VIRGL_PIPE_RES_CREATE_LAST_LEVEL] = last_level;
   cmd[VIRGL_PIPE_RES_CREATE_NR_SAMPLES] = nr_samples;
   cmd[VIRGL_PIPE_RES_CREATE_FLAGS] = flags;
   cmd[VIRGL_PIPE_RES_CREATE_BLOB_ID] = blob_id;

   drm_rc_blob.cmd = (unsigned long)(void *)&cmd;
   drm_rc_blob.cmd_size = 4 * (VIRGL_PIPE_RES_CREATE_SIZE + 1);
   drm_rc_blob.size = size;
   drm_rc_blob.blob_mem = VIRTGPU_BLOB_MEM_HOST3D;
   drm_rc_blob.blob_flags = VIRTGPU_BLOB_FLAG_USE_MAPPABLE;
   drm_rc_blob.blob_id = (uint64_t) blob_id;

   ret = drmIoctl(qdws->fd, DRM_IOCTL_VIRTGPU_RESOURCE_CREATE_BLOB, &drm_rc_blob);
   if (ret != 0) {
      FREE(res);
      return NULL;
   }

   res->bind = bind;
   res->res_handle = drm_rc_blob.res_handle;
   res->bo_handle = drm_rc_blob.bo_handle;
   res->size = size;
   res->flags = flags;
   res->maybe_untyped = false;
   pipe_reference_init(&res->reference, 1);
   p_atomic_set(&res->external, false);
   p_atomic_set(&res->num_cs_references, 0);
   virgl_resource_cache_entry_init(&res->cache_entry, params);
   return res;
}

static struct virgl_hw_res *
virgl_drm_winsys_resource_create(struct virgl_winsys *qws,
                                 enum pipe_texture_target target,
                                 uint32_t format,
                                 uint32_t bind,
                                 uint32_t width,
                                 uint32_t height,
                                 uint32_t depth,
                                 uint32_t array_size,
                                 uint32_t last_level,
                                 uint32_t nr_samples,
                                 uint32_t size,
                                 bool for_fencing)
{
   struct virgl_drm_winsys *qdws = virgl_drm_winsys(qws);
   struct drm_virtgpu_resource_create createcmd;
   int ret;
   struct virgl_hw_res *res;
   uint32_t stride = width * util_format_get_blocksize(format);
   struct virgl_resource_params params = { .size = size,
                                           .bind = bind,
                                           .format = format,
                                           .flags = 0,
                                           .nr_samples = nr_samples,
                                           .width = width,
                                           .height = height,
                                           .depth = depth,
                                           .array_size = array_size,
                                           .last_level = last_level,
                                           .target = target };

   res = CALLOC_STRUCT(virgl_hw_res);
   if (!res)
      return NULL;

   memset(&createcmd, 0, sizeof(createcmd));
   createcmd.target = target;
   createcmd.format = pipe_to_virgl_format(format);
   createcmd.bind = bind;
   createcmd.width = width;
   createcmd.height = height;
   createcmd.depth = depth;
   createcmd.array_size = array_size;
   createcmd.last_level = last_level;
   createcmd.nr_samples = nr_samples;
   createcmd.stride = stride;
   createcmd.size = size;

   ret = drmIoctl(qdws->fd, DRM_IOCTL_VIRTGPU_RESOURCE_CREATE, &createcmd);
   if (ret != 0) {
      FREE(res);
      return NULL;
   }

   res->bind = bind;

   res->res_handle = createcmd.res_handle;
   res->bo_handle = createcmd.bo_handle;
   res->size = size;
   res->target = target;
   res->maybe_untyped = false;
   pipe_reference_init(&res->reference, 1);
   p_atomic_set(&res->external, false);
   p_atomic_set(&res->num_cs_references, 0);

   /* A newly created resource is considered busy by the kernel until the
    * command is retired.  But for our purposes, we can consider it idle
    * unless it is used for fencing.
    */
   p_atomic_set(&res->maybe_busy, for_fencing);

   virgl_resource_cache_entry_init(&res->cache_entry, params);

   return res;
}

/*
 * Previously, with DRM_IOCTL_VIRTGPU_RESOURCE_CREATE, all host resources had
 * a guest memory shadow resource with size = stride * bpp.  Virglrenderer
 * would guess the stride implicitly when performing transfer operations, if
 * the stride wasn't specified.  Interestingly, vtest would specify the stride.
 *
 * Guessing the stride breaks down with YUV images, which may be imported into
 * Mesa as 3R8 images. It also doesn't work if an external allocator
 * (i.e, minigbm) decides to use a stride not equal to stride * bpp. With blob
 * resources, the size = stride * bpp restriction no longer holds, so use
 * explicit strides passed into Mesa.
 */
static inline bool use_explicit_stride(struct virgl_hw_res *res, uint32_t level,
				       uint32_t depth)
{
   return (params[param_resource_blob].value &&
           res->blob_mem == VIRTGPU_BLOB_MEM_HOST3D_GUEST &&
           res->target == PIPE_TEXTURE_2D &&
           level == 0 && depth == 1);
}

static int
virgl_bo_transfer_put(struct virgl_winsys *vws,
                      struct virgl_hw_res *res,
                      const struct pipe_box *box,
                      uint32_t stride, uint32_t layer_stride,
                      uint32_t buf_offset, uint32_t level)
{
   struct virgl_drm_winsys *vdws = virgl_drm_winsys(vws);
   struct drm_virtgpu_3d_transfer_to_host tohostcmd;

   p_atomic_set(&res->maybe_busy, true);

   memset(&tohostcmd, 0, sizeof(tohostcmd));
   tohostcmd.bo_handle = res->bo_handle;
   tohostcmd.box.x = box->x;
   tohostcmd.box.y = box->y;
   tohostcmd.box.z = box->z;
   tohostcmd.box.w = box->width;
   tohostcmd.box.h = box->height;
   tohostcmd.box.d = box->depth;
   tohostcmd.offset = buf_offset;
   tohostcmd.level = level;

   if (use_explicit_stride(res, level, box->depth))
      tohostcmd.stride = stride;

   return drmIoctl(vdws->fd, DRM_IOCTL_VIRTGPU_TRANSFER_TO_HOST, &tohostcmd);
}

static int
virgl_bo_transfer_get(struct virgl_winsys *vws,
                      struct virgl_hw_res *res,
                      const struct pipe_box *box,
                      uint32_t stride, uint32_t layer_stride,
                      uint32_t buf_offset, uint32_t level)
{
   struct virgl_drm_winsys *vdws = virgl_drm_winsys(vws);
   struct drm_virtgpu_3d_transfer_from_host fromhostcmd;

   p_atomic_set(&res->maybe_busy, true);

   memset(&fromhostcmd, 0, sizeof(fromhostcmd));
   fromhostcmd.bo_handle = res->bo_handle;
   fromhostcmd.level = level;
   fromhostcmd.offset = buf_offset;
   fromhostcmd.box.x = box->x;
   fromhostcmd.box.y = box->y;
   fromhostcmd.box.z = box->z;
   fromhostcmd.box.w = box->width;
   fromhostcmd.box.h = box->height;
   fromhostcmd.box.d = box->depth;

   if (use_explicit_stride(res, level, box->depth))
      fromhostcmd.stride = stride;

   return drmIoctl(vdws->fd, DRM_IOCTL_VIRTGPU_TRANSFER_FROM_HOST, &fromhostcmd);
}

static struct virgl_hw_res *
virgl_drm_winsys_resource_cache_create(struct virgl_winsys *qws,
                                       enum pipe_texture_target target,
                                       const void *map_front_private,
                                       uint32_t format,
                                       uint32_t bind,
                                       uint32_t width,
                                       uint32_t height,
                                       uint32_t depth,
                                       uint32_t array_size,
                                       uint32_t last_level,
                                       uint32_t nr_samples,
                                       uint32_t flags,
                                       uint32_t size)
{
   struct virgl_drm_winsys *qdws = virgl_drm_winsys(qws);
   struct virgl_hw_res *res;
   struct virgl_resource_cache_entry *entry;
   struct virgl_resource_params params = { .size = size,
                                     .bind = bind,
                                     .format = format,
                                     .flags = flags,
                                     .nr_samples = nr_samples,
                                     .width = width,
                                     .height = height,
                                     .depth = depth,
                                     .array_size = array_size,
                                     .last_level = last_level,
                                     .target = target };

   if (!can_cache_resource(bind))
      goto alloc;

   mtx_lock(&qdws->mutex);

   entry = virgl_resource_cache_remove_compatible(&qdws->cache, params);
   if (entry) {
      res = cache_entry_container_res(entry);
      mtx_unlock(&qdws->mutex);
      pipe_reference_init(&res->reference, 1);
      return res;
   }

   mtx_unlock(&qdws->mutex);

alloc:
   if (flags & (VIRGL_RESOURCE_FLAG_MAP_PERSISTENT |
                VIRGL_RESOURCE_FLAG_MAP_COHERENT))
      res = virgl_drm_winsys_resource_create_blob(qws, target, format, bind,
                                                  width, height, depth,
                                                  array_size, last_level,
                                                  nr_samples, flags, size);
   else
      res = virgl_drm_winsys_resource_create(qws, target, format, bind, width,
                                             height, depth, array_size,
                                             last_level, nr_samples, size,
                                             true);
   return res;
}

static struct virgl_hw_res *
virgl_drm_winsys_resource_create_handle(struct virgl_winsys *qws,
                                        struct winsys_handle *whandle,
                                        uint32_t *plane,
                                        uint32_t *stride,
                                        uint32_t *plane_offset,
                                        uint64_t *modifier,
                                        uint32_t *blob_mem)
{
   struct virgl_drm_winsys *qdws = virgl_drm_winsys(qws);
   struct drm_gem_open open_arg = {};
   struct drm_virtgpu_resource_info info_arg = {};
   struct virgl_hw_res *res = NULL;
   uint32_t handle = whandle->handle;

   if (whandle->plane >= VIRGL_MAX_PLANE_COUNT) {
      return NULL;
   }

   if (whandle->offset != 0 && whandle->type == WINSYS_HANDLE_TYPE_SHARED) {
      _debug_printf("attempt to import unsupported winsys offset %u\n",
                    whandle->offset);
      return NULL;
   } else if (whandle->type == WINSYS_HANDLE_TYPE_FD) {
      *plane = whandle->plane;
      *stride = whandle->stride;
      *plane_offset = whandle->offset;
      *modifier = whandle->modifier;
   }

   mtx_lock(&qdws->bo_handles_mutex);

   /* We must maintain a list of pairs <handle, bo>, so that we always return
    * the same BO for one particular handle. If we didn't do that and created
    * more than one BO for the same handle and then relocated them in a CS,
    * we would hit a deadlock in the kernel.
    *
    * The list of pairs is guarded by a mutex, of course. */
   if (whandle->type == WINSYS_HANDLE_TYPE_SHARED) {
      res = util_hash_table_get(qdws->bo_names, (void*)(uintptr_t)handle);
   } else if (whandle->type == WINSYS_HANDLE_TYPE_FD) {
      int r;
      r = drmPrimeFDToHandle(qdws->fd, whandle->handle, &handle);
      if (r)
         goto done;
      res = util_hash_table_get(qdws->bo_handles, (void*)(uintptr_t)handle);
   } else {
      /* Unknown handle type */
      goto done;
   }

   if (res) {
      /* qdws->bo_{names,handles} hold weak pointers to virgl_hw_res. Because
       * virgl_drm_resource_reference does not take qdws->bo_handles_mutex
       * until it enters virgl_hw_res_destroy, there is a small window that
       * the refcount can drop to zero. Call p_atomic_inc directly instead of
       * virgl_drm_resource_reference to avoid hitting assert failures.
       */
      p_atomic_inc(&res->reference.count);
      goto done;
   }

   res = CALLOC_STRUCT(virgl_hw_res);
   if (!res)
      goto done;

   if (whandle->type == WINSYS_HANDLE_TYPE_FD) {
      res->bo_handle = handle;
   } else {
      memset(&open_arg, 0, sizeof(open_arg));
      open_arg.name = whandle->handle;
      if (drmIoctl(qdws->fd, DRM_IOCTL_GEM_OPEN, &open_arg)) {
         FREE(res);
         res = NULL;
         goto done;
      }
      res->bo_handle = open_arg.handle;
      res->flink_name = whandle->handle;
   }

   memset(&info_arg, 0, sizeof(info_arg));
   info_arg.bo_handle = res->bo_handle;

   if (drmIoctl(qdws->fd, DRM_IOCTL_VIRTGPU_RESOURCE_INFO, &info_arg)) {
      /* close */
      FREE(res);
      res = NULL;
      goto done;
   }

   res->res_handle = info_arg.res_handle;
   res->blob_mem = info_arg.blob_mem;
   *blob_mem = info_arg.blob_mem;

   res->size = info_arg.size;
   res->maybe_untyped = info_arg.blob_mem ? true : false;
   pipe_reference_init(&res->reference, 1);
   p_atomic_set(&res->external, true);
   res->num_cs_references = 0;

   if (res->flink_name)
      _mesa_hash_table_insert(qdws->bo_names, (void *)(uintptr_t)res->flink_name, res);
   _mesa_hash_table_insert(qdws->bo_handles, (void *)(uintptr_t)res->bo_handle, res);

done:
   mtx_unlock(&qdws->bo_handles_mutex);
   return res;
}

static void
virgl_drm_winsys_resource_set_type(struct virgl_winsys *qws,
                                   struct virgl_hw_res *res,
                                   uint32_t format, uint32_t bind,
                                   uint32_t width, uint32_t height,
                                   uint32_t usage, uint64_t modifier,
                                   uint32_t plane_count,
                                   const uint32_t *plane_strides,
                                   const uint32_t *plane_offsets)
{
   struct virgl_drm_winsys *qdws = virgl_drm_winsys(qws);
   uint32_t cmd[VIRGL_PIPE_RES_SET_TYPE_SIZE(VIRGL_MAX_PLANE_COUNT)];
   struct drm_virtgpu_execbuffer eb;
   int ret;

   mtx_lock(&qdws->bo_handles_mutex);

   if (!res->maybe_untyped) {
      mtx_unlock(&qdws->bo_handles_mutex);
      return;
   }
   res->maybe_untyped = false;

   assert(plane_count && plane_count <= VIRGL_MAX_PLANE_COUNT);

   cmd[0] = VIRGL_CMD0(VIRGL_CCMD_PIPE_RESOURCE_SET_TYPE, 0, VIRGL_PIPE_RES_SET_TYPE_SIZE(plane_count));
   cmd[VIRGL_PIPE_RES_SET_TYPE_RES_HANDLE] = res->res_handle,
   cmd[VIRGL_PIPE_RES_SET_TYPE_FORMAT] = format;
   cmd[VIRGL_PIPE_RES_SET_TYPE_BIND] = bind;
   cmd[VIRGL_PIPE_RES_SET_TYPE_WIDTH] = width;
   cmd[VIRGL_PIPE_RES_SET_TYPE_HEIGHT] = height;
   cmd[VIRGL_PIPE_RES_SET_TYPE_USAGE] = usage;
   cmd[VIRGL_PIPE_RES_SET_TYPE_MODIFIER_LO] = (uint32_t)modifier;
   cmd[VIRGL_PIPE_RES_SET_TYPE_MODIFIER_HI] = (uint32_t)(modifier >> 32);
   for (uint32_t i = 0; i < plane_count; i++) {
      cmd[VIRGL_PIPE_RES_SET_TYPE_PLANE_STRIDE(i)] = plane_strides[i];
      cmd[VIRGL_PIPE_RES_SET_TYPE_PLANE_OFFSET(i)] = plane_offsets[i];
   }

   memset(&eb, 0, sizeof(eb));
   eb.command = (uintptr_t)cmd;
   eb.size = (1 + VIRGL_PIPE_RES_SET_TYPE_SIZE(plane_count)) * 4;
   eb.num_bo_handles = 1;
   eb.bo_handles = (uintptr_t)&res->bo_handle;

   ret = drmIoctl(qdws->fd, DRM_IOCTL_VIRTGPU_EXECBUFFER, &eb);
   if (ret == -1)
      _debug_printf("failed to set resource type: %s", strerror(errno));

   mtx_unlock(&qdws->bo_handles_mutex);
}

static boolean virgl_drm_winsys_resource_get_handle(struct virgl_winsys *qws,
                                                    struct virgl_hw_res *res,
                                                    uint32_t stride,
                                                    struct winsys_handle *whandle)
 {
   struct virgl_drm_winsys *qdws = virgl_drm_winsys(qws);
   struct drm_gem_flink flink;

   if (!res)
       return FALSE;

   if (whandle->type == WINSYS_HANDLE_TYPE_SHARED) {
      if (!res->flink_name) {
         memset(&flink, 0, sizeof(flink));
         flink.handle = res->bo_handle;

         if (drmIoctl(qdws->fd, DRM_IOCTL_GEM_FLINK, &flink)) {
            return FALSE;
         }
         res->flink_name = flink.name;

         mtx_lock(&qdws->bo_handles_mutex);
         _mesa_hash_table_insert(qdws->bo_names, (void *)(uintptr_t)res->flink_name, res);
         mtx_unlock(&qdws->bo_handles_mutex);
      }
      whandle->handle = res->flink_name;
   } else if (whandle->type == WINSYS_HANDLE_TYPE_KMS) {
      whandle->handle = res->bo_handle;
   } else if (whandle->type == WINSYS_HANDLE_TYPE_FD) {
      if (drmPrimeHandleToFD(qdws->fd, res->bo_handle, DRM_CLOEXEC, (int*)&whandle->handle))
            return FALSE;
      mtx_lock(&qdws->bo_handles_mutex);
      _mesa_hash_table_insert(qdws->bo_handles, (void *)(uintptr_t)res->bo_handle, res);
      mtx_unlock(&qdws->bo_handles_mutex);
   }

   p_atomic_set(&res->external, true);

   whandle->stride = stride;
   return TRUE;
}

static void *virgl_drm_resource_map(struct virgl_winsys *qws,
                                    struct virgl_hw_res *res)
{
   struct virgl_drm_winsys *qdws = virgl_drm_winsys(qws);
   struct drm_virtgpu_map mmap_arg;
   void *ptr;

   if (res->ptr)
      return res->ptr;

   memset(&mmap_arg, 0, sizeof(mmap_arg));
   mmap_arg.handle = res->bo_handle;
   if (drmIoctl(qdws->fd, DRM_IOCTL_VIRTGPU_MAP, &mmap_arg))
      return NULL;

   ptr = os_mmap(0, res->size, PROT_READ|PROT_WRITE, MAP_SHARED,
                 qdws->fd, mmap_arg.offset);
   if (ptr == MAP_FAILED)
      return NULL;

   res->ptr = ptr;
   return ptr;

}

static void virgl_drm_resource_wait(struct virgl_winsys *qws,
                                    struct virgl_hw_res *res)
{
   struct virgl_drm_winsys *qdws = virgl_drm_winsys(qws);
   struct drm_virtgpu_3d_wait waitcmd;
   int ret;

   if (!p_atomic_read(&res->maybe_busy) && !p_atomic_read(&res->external))
      return;

   memset(&waitcmd, 0, sizeof(waitcmd));
   waitcmd.handle = res->bo_handle;

   ret = drmIoctl(qdws->fd, DRM_IOCTL_VIRTGPU_WAIT, &waitcmd);
   if (ret)
      _debug_printf("waiting got error - %d, slow gpu or hang?\n", errno);

   p_atomic_set(&res->maybe_busy, false);
}

static bool virgl_drm_alloc_res_list(struct virgl_drm_cmd_buf *cbuf,
                                     int initial_size)
{
   cbuf->nres = initial_size;
   cbuf->cres = 0;

   cbuf->res_bo = CALLOC(cbuf->nres, sizeof(struct virgl_hw_buf*));
   if (!cbuf->res_bo)
      return false;

   cbuf->res_hlist = MALLOC(cbuf->nres * sizeof(uint32_t));
   if (!cbuf->res_hlist) {
      FREE(cbuf->res_bo);
      return false;
   }

   return true;
}

static void virgl_drm_free_res_list(struct virgl_drm_cmd_buf *cbuf)
{
   int i;

   for (i = 0; i < cbuf->cres; i++) {
      p_atomic_dec(&cbuf->res_bo[i]->num_cs_references);
      virgl_drm_resource_reference(cbuf->ws, &cbuf->res_bo[i], NULL);
   }
   FREE(cbuf->res_hlist);
   FREE(cbuf->res_bo);
}

static boolean virgl_drm_lookup_res(struct virgl_drm_cmd_buf *cbuf,
                                    struct virgl_hw_res *res)
{
   unsigned hash = res->res_handle & (sizeof(cbuf->is_handle_added)-1);
   int i;

   if (cbuf->is_handle_added[hash]) {
      i = cbuf->reloc_indices_hashlist[hash];
      if (cbuf->res_bo[i] == res)
         return true;

      for (i = 0; i < cbuf->cres; i++) {
         if (cbuf->res_bo[i] == res) {
            cbuf->reloc_indices_hashlist[hash] = i;
            return true;
         }
      }
   }
   return false;
}

static void virgl_drm_add_res(struct virgl_drm_winsys *qdws,
                              struct virgl_drm_cmd_buf *cbuf,
                              struct virgl_hw_res *res)
{
   unsigned hash = res->res_handle & (sizeof(cbuf->is_handle_added)-1);

   if (cbuf->cres >= cbuf->nres) {
      unsigned new_nres = cbuf->nres + 256;
      void *new_ptr = REALLOC(cbuf->res_bo,
                              cbuf->nres * sizeof(struct virgl_hw_buf*),
                              new_nres * sizeof(struct virgl_hw_buf*));
      if (!new_ptr) {
          _debug_printf("failure to add relocation %d, %d\n", cbuf->cres, new_nres);
          return;
      }
      cbuf->res_bo = new_ptr;

      new_ptr = REALLOC(cbuf->res_hlist,
                        cbuf->nres * sizeof(uint32_t),
                        new_nres * sizeof(uint32_t));
      if (!new_ptr) {
          _debug_printf("failure to add hlist relocation %d, %d\n", cbuf->cres, cbuf->nres);
          return;
      }
      cbuf->res_hlist = new_ptr;
      cbuf->nres = new_nres;
   }

   cbuf->res_bo[cbuf->cres] = NULL;
   virgl_drm_resource_reference(&qdws->base, &cbuf->res_bo[cbuf->cres], res);
   cbuf->res_hlist[cbuf->cres] = res->bo_handle;
   cbuf->is_handle_added[hash] = TRUE;

   cbuf->reloc_indices_hashlist[hash] = cbuf->cres;
   p_atomic_inc(&res->num_cs_references);
   cbuf->cres++;
}

/* This is called after the cbuf is submitted. */
static void virgl_drm_clear_res_list(struct virgl_drm_cmd_buf *cbuf)
{
   int i;

   for (i = 0; i < cbuf->cres; i++) {
      /* mark all BOs busy after submission */
      p_atomic_set(&cbuf->res_bo[i]->maybe_busy, true);

      p_atomic_dec(&cbuf->res_bo[i]->num_cs_references);
      virgl_drm_resource_reference(cbuf->ws, &cbuf->res_bo[i], NULL);
   }

   cbuf->cres = 0;

   memset(cbuf->is_handle_added, 0, sizeof(cbuf->is_handle_added));
}

static void virgl_drm_emit_res(struct virgl_winsys *qws,
                               struct virgl_cmd_buf *_cbuf,
                               struct virgl_hw_res *res, boolean write_buf)
{
   struct virgl_drm_winsys *qdws = virgl_drm_winsys(qws);
   struct virgl_drm_cmd_buf *cbuf = virgl_drm_cmd_buf(_cbuf);
   boolean already_in_list = virgl_drm_lookup_res(cbuf, res);

   if (write_buf)
      cbuf->base.buf[cbuf->base.cdw++] = res->res_handle;

   if (!already_in_list)
      virgl_drm_add_res(qdws, cbuf, res);
}

static boolean virgl_drm_res_is_ref(struct virgl_winsys *qws,
                                    struct virgl_cmd_buf *_cbuf,
                                    struct virgl_hw_res *res)
{
   if (!p_atomic_read(&res->num_cs_references))
      return FALSE;

   return TRUE;
}

static struct virgl_cmd_buf *virgl_drm_cmd_buf_create(struct virgl_winsys *qws,
                                                      uint32_t size)
{
   struct virgl_drm_cmd_buf *cbuf;

   cbuf = CALLOC_STRUCT(virgl_drm_cmd_buf);
   if (!cbuf)
      return NULL;

   cbuf->ws = qws;

   if (!virgl_drm_alloc_res_list(cbuf, 512)) {
      FREE(cbuf);
      return NULL;
   }

   cbuf->buf = CALLOC(size, sizeof(uint32_t));
   if (!cbuf->buf) {
      FREE(cbuf->res_hlist);
      FREE(cbuf->res_bo);
      FREE(cbuf);
      return NULL;
   }

   cbuf->in_fence_fd = -1;
   cbuf->base.buf = cbuf->buf;
   return &cbuf->base;
}

static void virgl_drm_cmd_buf_destroy(struct virgl_cmd_buf *_cbuf)
{
   struct virgl_drm_cmd_buf *cbuf = virgl_drm_cmd_buf(_cbuf);

   virgl_drm_free_res_list(cbuf);

   FREE(cbuf->buf);
   FREE(cbuf);
}

static struct pipe_fence_handle *
virgl_drm_fence_create(struct virgl_winsys *vws, int fd, bool external)
{
   struct virgl_drm_fence *fence;

   assert(vws->supports_fences);

   if (external) {
      fd = os_dupfd_cloexec(fd);
      if (fd < 0)
         return NULL;
   }

   fence = CALLOC_STRUCT(virgl_drm_fence);
   if (!fence) {
      close(fd);
      return NULL;
   }

   fence->fd = fd;
   fence->external = external;

   pipe_reference_init(&fence->reference, 1);

   return (struct pipe_fence_handle *)fence;
}

static struct pipe_fence_handle *
virgl_drm_fence_create_legacy(struct virgl_winsys *vws)
{
   struct virgl_drm_fence *fence;

   assert(!vws->supports_fences);

   fence = CALLOC_STRUCT(virgl_drm_fence);
   if (!fence)
      return NULL;
   fence->fd = -1;

   /* Resources for fences should not be from the cache, since we are basing
    * the fence status on the resource creation busy status.
    */
   fence->hw_res = virgl_drm_winsys_resource_create(vws, PIPE_BUFFER,
         PIPE_FORMAT_R8_UNORM, VIRGL_BIND_CUSTOM, 8, 1, 1, 0, 0, 0, 8, true);
   if (!fence->hw_res) {
      FREE(fence);
      return NULL;
   }

   pipe_reference_init(&fence->reference, 1);

   return (struct pipe_fence_handle *)fence;
}

static int virgl_drm_winsys_submit_cmd(struct virgl_winsys *qws,
                                       struct virgl_cmd_buf *_cbuf,
                                       struct pipe_fence_handle **fence)
{
   struct virgl_drm_winsys *qdws = virgl_drm_winsys(qws);
   struct virgl_drm_cmd_buf *cbuf = virgl_drm_cmd_buf(_cbuf);
   struct drm_virtgpu_execbuffer eb;
   int ret;

   if (cbuf->base.cdw == 0)
      return 0;

   memset(&eb, 0, sizeof(struct drm_virtgpu_execbuffer));
   eb.command = (unsigned long)(void*)cbuf->buf;
   eb.size = cbuf->base.cdw * 4;
   eb.num_bo_handles = cbuf->cres;
   eb.bo_handles = (unsigned long)(void *)cbuf->res_hlist;

   eb.fence_fd = -1;
   if (qws->supports_fences) {
      if (cbuf->in_fence_fd >= 0) {
         eb.flags |= VIRTGPU_EXECBUF_FENCE_FD_IN;
         eb.fence_fd = cbuf->in_fence_fd;
      }

      if (fence != NULL)
         eb.flags |= VIRTGPU_EXECBUF_FENCE_FD_OUT;
   } else {
      assert(cbuf->in_fence_fd < 0);
   }

   ret = drmIoctl(qdws->fd, DRM_IOCTL_VIRTGPU_EXECBUFFER, &eb);
   if (ret == -1)
      _debug_printf("got error from kernel - expect bad rendering %d\n", errno);
   cbuf->base.cdw = 0;

   if (qws->supports_fences) {
      if (cbuf->in_fence_fd >= 0) {
         close(cbuf->in_fence_fd);
         cbuf->in_fence_fd = -1;
      }

      if (fence != NULL && ret == 0)
         *fence = virgl_drm_fence_create(qws, eb.fence_fd, false);
   } else {
      if (fence != NULL && ret == 0)
         *fence = virgl_drm_fence_create_legacy(qws);
   }

   virgl_drm_clear_res_list(cbuf);

   return ret;
}

static int virgl_drm_get_caps(struct virgl_winsys *vws,
                              struct virgl_drm_caps *caps)
{
   struct virgl_drm_winsys *vdws = virgl_drm_winsys(vws);
   struct drm_virtgpu_get_caps args;
   int ret;

   virgl_ws_fill_new_caps_defaults(caps);

   memset(&args, 0, sizeof(args));
   if (params[param_capset_fix].value) {
      /* if we have the query fix - try and get cap set id 2 first */
      args.cap_set_id = 2;
      args.size = sizeof(union virgl_caps);
   } else {
      args.cap_set_id = 1;
      args.size = sizeof(struct virgl_caps_v1);
   }
   args.addr = (unsigned long)&caps->caps;

   ret = drmIoctl(vdws->fd, DRM_IOCTL_VIRTGPU_GET_CAPS, &args);
   if (ret == -1 && errno == EINVAL) {
      /* Fallback to v1 */
      args.cap_set_id = 1;
      args.size = sizeof(struct virgl_caps_v1);
      ret = drmIoctl(vdws->fd, DRM_IOCTL_VIRTGPU_GET_CAPS, &args);
      if (ret == -1)
          return ret;
   }
   return ret;
}

static struct pipe_fence_handle *
virgl_cs_create_fence(struct virgl_winsys *vws, int fd)
{
   if (!vws->supports_fences)
      return NULL;

   return virgl_drm_fence_create(vws, fd, true);
}

static bool virgl_fence_wait(struct virgl_winsys *vws,
                             struct pipe_fence_handle *_fence,
                             uint64_t timeout)
{
   struct virgl_drm_fence *fence = virgl_drm_fence(_fence);

   if (vws->supports_fences) {
      uint64_t timeout_ms;
      int timeout_poll;

      if (timeout == 0)
         return sync_wait(fence->fd, 0) == 0;

      timeout_ms = timeout / 1000000;
      /* round up */
      if (timeout_ms * 1000000 < timeout)
         timeout_ms++;

      timeout_poll = timeout_ms <= INT_MAX ? (int) timeout_ms : -1;

      return sync_wait(fence->fd, timeout_poll) == 0;
   }

   if (timeout == 0)
      return !virgl_drm_resource_is_busy(vws, fence->hw_res);

   if (timeout != PIPE_TIMEOUT_INFINITE) {
      int64_t start_time = os_time_get();
      timeout /= 1000;
      while (virgl_drm_resource_is_busy(vws, fence->hw_res)) {
         if (os_time_get() - start_time >= timeout)
            return FALSE;
         os_time_sleep(10);
      }
      return TRUE;
   }
   virgl_drm_resource_wait(vws, fence->hw_res);

   return TRUE;
}

static void virgl_fence_reference(struct virgl_winsys *vws,
                                  struct pipe_fence_handle **dst,
                                  struct pipe_fence_handle *src)
{
   struct virgl_drm_fence *dfence = virgl_drm_fence(*dst);
   struct virgl_drm_fence *sfence = virgl_drm_fence(src);

   if (pipe_reference(&dfence->reference, &sfence->reference)) {
      if (vws->supports_fences) {
         close(dfence->fd);
      } else {
         virgl_drm_resource_reference(vws, &dfence->hw_res, NULL);
      }
      FREE(dfence);
   }

   *dst = src;
}

static void virgl_fence_server_sync(struct virgl_winsys *vws,
                                    struct virgl_cmd_buf *_cbuf,
                                    struct pipe_fence_handle *_fence)
{
   struct virgl_drm_cmd_buf *cbuf = virgl_drm_cmd_buf(_cbuf);
   struct virgl_drm_fence *fence = virgl_drm_fence(_fence);

   if (!vws->supports_fences)
      return;

   /* if not an external fence, then nothing more to do without preemption: */
   if (!fence->external)
      return;

   sync_accumulate("virgl", &cbuf->in_fence_fd, fence->fd);
}

static int virgl_fence_get_fd(struct virgl_winsys *vws,
                              struct pipe_fence_handle *_fence)
{
   struct virgl_drm_fence *fence = virgl_drm_fence(_fence);

   if (!vws->supports_fences)
      return -1;

   return os_dupfd_cloexec(fence->fd);
}

static int virgl_drm_get_version(int fd)
{
	int ret;
	drmVersionPtr version;

	version = drmGetVersion(fd);

	if (!version)
		ret = -EFAULT;
	else if (version->version_major != 0)
		ret = -EINVAL;
	else
		ret = VIRGL_DRM_VERSION(0, version->version_minor);

	drmFreeVersion(version);

	return ret;
}

static bool
virgl_drm_resource_cache_entry_is_busy(struct virgl_resource_cache_entry *entry,
                                       void *user_data)
{
   struct virgl_drm_winsys *qdws = user_data;
   struct virgl_hw_res *res = cache_entry_container_res(entry);

   return virgl_drm_resource_is_busy(&qdws->base, res);
}

static void
virgl_drm_resource_cache_entry_release(struct virgl_resource_cache_entry *entry,
                                       void *user_data)
{
   struct virgl_drm_winsys *qdws = user_data;
   struct virgl_hw_res *res = cache_entry_container_res(entry);

   virgl_hw_res_destroy(qdws, res);
}

static int virgl_init_context(int drmFD)
{
   int ret;
   struct drm_virtgpu_context_init init = { 0 };
   struct drm_virtgpu_context_set_param ctx_set_param = { 0 };
   uint64_t supports_capset_virgl, supports_capset_virgl2;
   supports_capset_virgl = supports_capset_virgl2 = 0;

   supports_capset_virgl = ((1 << VIRGL_DRM_CAPSET_VIRGL) &
                             params[param_supported_capset_ids].value);

   supports_capset_virgl2 = ((1 << VIRGL_DRM_CAPSET_VIRGL2) &
                              params[param_supported_capset_ids].value);

   if (!supports_capset_virgl && !supports_capset_virgl2) {
      _debug_printf("No virgl contexts available on host");
      return -EINVAL;
   }

   ctx_set_param.param = VIRTGPU_CONTEXT_PARAM_CAPSET_ID;
   ctx_set_param.value = (supports_capset_virgl2) ?
                         VIRGL_DRM_CAPSET_VIRGL2 :
                         VIRGL_DRM_CAPSET_VIRGL;

   init.ctx_set_params = (unsigned long)(void *)&ctx_set_param;
   init.num_params = 1;

   ret = drmIoctl(drmFD, DRM_IOCTL_VIRTGPU_CONTEXT_INIT, &init);
   /*
    * EEXIST happens when a compositor does DUMB_CREATE before initializing
    * virgl.
    */
   if (ret && errno != EEXIST) {
      _debug_printf("DRM_IOCTL_VIRTGPU_CONTEXT_INIT failed with %s\n",
                     strerror(errno));
      return -1;
   }

   return 0;
}

static struct virgl_winsys *
virgl_drm_winsys_create(int drmFD)
{
   static const unsigned CACHE_TIMEOUT_USEC = 1000000;
   struct virgl_drm_winsys *qdws;
   int drm_version;
   int ret;

   for (uint32_t i = 0; i < ARRAY_SIZE(params); i++) {
      struct drm_virtgpu_getparam getparam = { 0 };
      uint64_t value = 0;
      getparam.param = params[i].param;
      getparam.value = (uint64_t)(uintptr_t)&value;
      ret = drmIoctl(drmFD, DRM_IOCTL_VIRTGPU_GETPARAM, &getparam);
      params[i].value = (ret == 0) ? value : 0;
   }

   if (!params[param_3d_features].value)
      return NULL;

   drm_version = virgl_drm_get_version(drmFD);
   if (drm_version < 0)
      return NULL;

   if (params[param_context_init].value) {
      ret = virgl_init_context(drmFD);
      if (ret)
         return NULL;
   }

   qdws = CALLOC_STRUCT(virgl_drm_winsys);
   if (!qdws)
      return NULL;

   qdws->fd = drmFD;
   virgl_resource_cache_init(&qdws->cache, CACHE_TIMEOUT_USEC,
                             virgl_drm_resource_cache_entry_is_busy,
                             virgl_drm_resource_cache_entry_release,
                             qdws);
   (void) mtx_init(&qdws->mutex, mtx_plain);
   (void) mtx_init(&qdws->bo_handles_mutex, mtx_plain);
   p_atomic_set(&qdws->blob_id, 0);

   qdws->bo_handles = util_hash_table_create_ptr_keys();
   qdws->bo_names = util_hash_table_create_ptr_keys();
   qdws->base.destroy = virgl_drm_winsys_destroy;

   qdws->base.transfer_put = virgl_bo_transfer_put;
   qdws->base.transfer_get = virgl_bo_transfer_get;
   qdws->base.resource_create = virgl_drm_winsys_resource_cache_create;
   qdws->base.resource_reference = virgl_drm_resource_reference;
   qdws->base.resource_create_from_handle = virgl_drm_winsys_resource_create_handle;
   qdws->base.resource_set_type = virgl_drm_winsys_resource_set_type;
   qdws->base.resource_get_handle = virgl_drm_winsys_resource_get_handle;
   qdws->base.resource_map = virgl_drm_resource_map;
   qdws->base.resource_wait = virgl_drm_resource_wait;
   qdws->base.resource_is_busy = virgl_drm_resource_is_busy;
   qdws->base.cmd_buf_create = virgl_drm_cmd_buf_create;
   qdws->base.cmd_buf_destroy = virgl_drm_cmd_buf_destroy;
   qdws->base.submit_cmd = virgl_drm_winsys_submit_cmd;
   qdws->base.emit_res = virgl_drm_emit_res;
   qdws->base.res_is_referenced = virgl_drm_res_is_ref;

   qdws->base.cs_create_fence = virgl_cs_create_fence;
   qdws->base.fence_wait = virgl_fence_wait;
   qdws->base.fence_reference = virgl_fence_reference;
   qdws->base.fence_server_sync = virgl_fence_server_sync;
   qdws->base.fence_get_fd = virgl_fence_get_fd;
   qdws->base.get_caps = virgl_drm_get_caps;
   qdws->base.supports_fences =  drm_version >= VIRGL_DRM_VERSION_FENCE_FD;
   qdws->base.supports_encoded_transfers = 1;

   qdws->base.supports_coherent = params[param_resource_blob].value &&
                                  params[param_host_visible].value;
   return &qdws->base;

}

static struct hash_table *fd_tab = NULL;
static mtx_t virgl_screen_mutex = _MTX_INITIALIZER_NP;

static void
virgl_drm_screen_destroy(struct pipe_screen *pscreen)
{
   struct virgl_screen *screen = virgl_screen(pscreen);
   boolean destroy;

   mtx_lock(&virgl_screen_mutex);
   destroy = --screen->refcnt == 0;
   if (destroy) {
      int fd = virgl_drm_winsys(screen->vws)->fd;
      _mesa_hash_table_remove_key(fd_tab, intptr_to_pointer(fd));
      close(fd);
   }
   mtx_unlock(&virgl_screen_mutex);

   if (destroy) {
      pscreen->destroy = screen->winsys_priv;
      pscreen->destroy(pscreen);
   }
}

static uint32_t
hash_fd(const void *key)
{
   int fd = pointer_to_intptr(key);

   return _mesa_hash_int(&fd);
}

static bool
equal_fd(const void *key1, const void *key2)
{
   int ret;
   int fd1 = pointer_to_intptr(key1);
   int fd2 = pointer_to_intptr(key2);

   /* Since the scope of prime handle is limited to drm_file,
    * virgl_screen is only shared at the drm_file level,
    * not at the device (/dev/dri/cardX) level.
    */
   ret = os_same_file_description(fd1, fd2);
   if (ret == 0) {
       return true;
   } else if (ret < 0) {
      static bool logged;

      if (!logged) {
         _debug_printf("virgl: os_same_file_description couldn't "
                       "determine if two DRM fds reference the same "
                       "file description.\n"
                       "If they do, bad things may happen!\n");
         logged = true;
      }
   }

   return false;
}

struct pipe_screen *
virgl_drm_screen_create(int fd, const struct pipe_screen_config *config)
{
   struct pipe_screen *pscreen = NULL;

   mtx_lock(&virgl_screen_mutex);
   if (!fd_tab) {
      fd_tab = _mesa_hash_table_create(NULL, hash_fd, equal_fd);
      if (!fd_tab)
         goto unlock;
   }

   pscreen = util_hash_table_get(fd_tab, intptr_to_pointer(fd));
   if (pscreen) {
      virgl_screen(pscreen)->refcnt++;
   } else {
      struct virgl_winsys *vws;
      int dup_fd = os_dupfd_cloexec(fd);

      vws = virgl_drm_winsys_create(dup_fd);
      if (!vws) {
         close(dup_fd);
         goto unlock;
      }

      pscreen = virgl_create_screen(vws, config);
      if (pscreen) {
         _mesa_hash_table_insert(fd_tab, intptr_to_pointer(dup_fd), pscreen);

         /* Bit of a hack, to avoid circular linkage dependency,
          * ie. pipe driver having to call in to winsys, we
          * override the pipe drivers screen->destroy():
          */
         virgl_screen(pscreen)->winsys_priv = pscreen->destroy;
         pscreen->destroy = virgl_drm_screen_destroy;
      }
   }

unlock:
   mtx_unlock(&virgl_screen_mutex);
   return pscreen;
}