xref: /external/mesa3d/src/mesa/drivers/dri/i965/brw_bufmgr.c

/*
 * Copyright © 2007 Red Hat Inc.
 * Copyright © 2007-2017 Intel Corporation
 * Copyright © 2006 VMware, Inc.
 * All Rights Reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (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 NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 * IN THE SOFTWARE.
 */

/*
 * Authors: Thomas Hellström <thellstrom@vmware.com>
 *          Keith Whitwell <keithw@vmware.com>
 *          Eric Anholt <eric@anholt.net>
 *          Dave Airlie <airlied@linux.ie>
 */

#include <xf86drm.h>
#include <util/u_atomic.h>
#include <fcntl.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <assert.h>
#include <sys/ioctl.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <stdbool.h>

#include "errno.h"
#include "common/gen_clflush.h"
#include "dev/gen_debug.h"
#include "common/gen_gem.h"
#include "dev/gen_device_info.h"
#include "libdrm_macros.h"
#include "main/macros.h"
#include "util/macros.h"
#include "util/hash_table.h"
#include "util/list.h"
#include "util/os_file.h"
#include "util/u_dynarray.h"
#include "util/vma.h"
#include "brw_bufmgr.h"
#include "brw_context.h"
#include "string.h"

#include "drm-uapi/i915_drm.h"

#ifdef HAVE_VALGRIND
#include <valgrind.h>
#include <memcheck.h>
#define VG(x) x
#else
#define VG(x)
#endif

/* Bufmgr is not aware of brw_context. */
#undef WARN_ONCE
#define WARN_ONCE(cond, fmt...) do {                            \
   if (unlikely(cond)) {                                        \
      static bool _warned = false;                              \
      if (!_warned) {                                           \
         fprintf(stderr, "WARNING: ");                          \
         fprintf(stderr, fmt);                                  \
         _warned = true;                                        \
      }                                                         \
   }                                                            \
} while (0)


/* VALGRIND_FREELIKE_BLOCK unfortunately does not actually undo the earlier
 * VALGRIND_MALLOCLIKE_BLOCK but instead leaves vg convinced the memory is
 * leaked. All because it does not call VG(cli_free) from its
 * VG_USERREQ__FREELIKE_BLOCK handler. Instead of treating the memory like
 * and allocation, we mark it available for use upon mmapping and remove
 * it upon unmapping.
 */
#define VG_DEFINED(ptr, size) VG(VALGRIND_MAKE_MEM_DEFINED(ptr, size))
#define VG_NOACCESS(ptr, size) VG(VALGRIND_MAKE_MEM_NOACCESS(ptr, size))

#define PAGE_SIZE 4096

#define FILE_DEBUG_FLAG DEBUG_BUFMGR

static inline int
atomic_add_unless(int *v, int add, int unless)
{
   int c, old;
   c = p_atomic_read(v);
   while (c != unless && (old = p_atomic_cmpxchg(v, c, c + add)) != c)
      c = old;
   return c == unless;
}

/**
 * i965 fixed-size bucketing VMA allocator.
 *
 * The BO cache maintains "cache buckets" for buffers of various sizes.
 * All buffers in a given bucket are identically sized - when allocating,
 * we always round up to the bucket size.  This means that virtually all
 * allocations are fixed-size; only buffers which are too large to fit in
 * a bucket can be variably-sized.
 *
 * We create an allocator for each bucket.  Each contains a free-list, where
 * each node contains a <starting address, 64-bit bitmap> pair.  Each bit
 * represents a bucket-sized block of memory.  (At the first level, each
 * bit corresponds to a page.  For the second bucket, bits correspond to
 * two pages, and so on.)  1 means a block is free, and 0 means it's in-use.
 * The lowest bit in the bitmap is for the first block.
 *
 * This makes allocations cheap - any bit of any node will do.  We can pick
 * the head of the list and use ffs() to find a free block.  If there are
 * none, we allocate 64 blocks from a larger allocator - either a bigger
 * bucketing allocator, or a fallback top-level allocator for large objects.
 */
struct vma_bucket_node {
   uint64_t start_address;
   uint64_t bitmap;
};

struct bo_cache_bucket {
   /** List of cached BOs. */
   struct list_head head;

   /** Size of this bucket, in bytes. */
   uint64_t size;

   /** List of vma_bucket_nodes. */
   struct util_dynarray vma_list[BRW_MEMZONE_COUNT];
};

struct bo_export {
   /** File descriptor associated with a handle export. */
   int drm_fd;

   /** GEM handle in drm_fd */
   uint32_t gem_handle;

   struct list_head link;
};

struct brw_bufmgr {
   uint32_t refcount;

   struct list_head link;

   int fd;

   mtx_t lock;

   /** Array of lists of cached gem objects of power-of-two sizes */
   struct bo_cache_bucket cache_bucket[14 * 4];
   int num_buckets;
   time_t time;

   struct hash_table *name_table;
   struct hash_table *handle_table;

   struct util_vma_heap vma_allocator[BRW_MEMZONE_COUNT];

   bool has_llc:1;
   bool has_mmap_wc:1;
   bool has_mmap_offset:1;
   bool bo_reuse:1;

   uint64_t initial_kflags;
};

static mtx_t global_bufmgr_list_mutex = _MTX_INITIALIZER_NP;
static struct list_head global_bufmgr_list = {
   .next = &global_bufmgr_list,
   .prev = &global_bufmgr_list,
};

static int bo_set_tiling_internal(struct brw_bo *bo, uint32_t tiling_mode,
                                  uint32_t stride);

static void bo_free(struct brw_bo *bo);

static uint64_t vma_alloc(struct brw_bufmgr *bufmgr,
                          enum brw_memory_zone memzone,
                          uint64_t size, uint64_t alignment);

static struct brw_bo *
hash_find_bo(struct hash_table *ht, unsigned int key)
{
   struct hash_entry *entry = _mesa_hash_table_search(ht, &key);
   return entry ? (struct brw_bo *) entry->data : NULL;
}

static uint64_t
bo_tile_size(struct brw_bufmgr *bufmgr, uint64_t size, uint32_t tiling)
{
   if (tiling == I915_TILING_NONE)
      return size;

   /* 965+ just need multiples of page size for tiling */
   return ALIGN(size, PAGE_SIZE);
}

/*
 * Round a given pitch up to the minimum required for X tiling on a
 * given chip.  We use 512 as the minimum to allow for a later tiling
 * change.
 */
static uint32_t
bo_tile_pitch(struct brw_bufmgr *bufmgr, uint32_t pitch, uint32_t tiling)
{
   unsigned long tile_width;

   /* If untiled, then just align it so that we can do rendering
    * to it with the 3D engine.
    */
   if (tiling == I915_TILING_NONE)
      return ALIGN(pitch, 64);

   if (tiling == I915_TILING_X)
      tile_width = 512;
   else
      tile_width = 128;

   /* 965 is flexible */
   return ALIGN(pitch, tile_width);
}

/**
 * This function finds the correct bucket fit for the input size.
 * The function works with O(1) complexity when the requested size
 * was queried instead of iterating the size through all the buckets.
 */
static struct bo_cache_bucket *
bucket_for_size(struct brw_bufmgr *bufmgr, uint64_t size)
{
   /* Calculating the pages and rounding up to the page size. */
   const unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;

   /* Row  Bucket sizes    clz((x-1) | 3)   Row    Column
    *        in pages                      stride   size
    *   0:   1  2  3  4 -> 30 30 30 30        4       1
    *   1:   5  6  7  8 -> 29 29 29 29        4       1
    *   2:  10 12 14 16 -> 28 28 28 28        8       2
    *   3:  20 24 28 32 -> 27 27 27 27       16       4
    */
   const unsigned row = 30 - __builtin_clz((pages - 1) | 3);
   const unsigned row_max_pages = 4 << row;

   /* The '& ~2' is the special case for row 1. In row 1, max pages /
    * 2 is 2, but the previous row maximum is zero (because there is
    * no previous row). All row maximum sizes are power of 2, so that
    * is the only case where that bit will be set.
    */
   const unsigned prev_row_max_pages = (row_max_pages / 2) & ~2;
   int col_size_log2 = row - 1;
   col_size_log2 += (col_size_log2 < 0);

   const unsigned col = (pages - prev_row_max_pages +
                        ((1 << col_size_log2) - 1)) >> col_size_log2;

   /* Calculating the index based on the row and column. */
   const unsigned index = (row * 4) + (col - 1);

   return (index < bufmgr->num_buckets) ?
          &bufmgr->cache_bucket[index] : NULL;
}

static enum brw_memory_zone
memzone_for_address(uint64_t address)
{
   const uint64_t _4GB = 1ull << 32;

   if (address >= _4GB)
      return BRW_MEMZONE_OTHER;

   return BRW_MEMZONE_LOW_4G;
}

static uint64_t
bucket_vma_alloc(struct brw_bufmgr *bufmgr,
                 struct bo_cache_bucket *bucket,
                 enum brw_memory_zone memzone)
{
   struct util_dynarray *vma_list = &bucket->vma_list[memzone];
   struct vma_bucket_node *node;

   if (vma_list->size == 0) {
      /* This bucket allocator is out of space - allocate a new block of
       * memory for 64 blocks from a larger allocator (either a larger
       * bucket or util_vma).
       *
       * We align the address to the node size (64 blocks) so that
       * bucket_vma_free can easily compute the starting address of this
       * block by rounding any address we return down to the node size.
       *
       * Set the first bit used, and return the start address.
       */
      uint64_t node_size = 64ull * bucket->size;
      node = util_dynarray_grow(vma_list, struct vma_bucket_node, 1);

      if (unlikely(!node))
         return 0ull;

      uint64_t addr = vma_alloc(bufmgr, memzone, node_size, node_size);
      node->start_address = gen_48b_address(addr);
      node->bitmap = ~1ull;
      return node->start_address;
   }

   /* Pick any bit from any node - they're all the right size and free. */
   node = util_dynarray_top_ptr(vma_list, struct vma_bucket_node);
   int bit = ffsll(node->bitmap) - 1;
   assert(bit >= 0 && bit <= 63);

   /* Reserve the memory by clearing the bit. */
   assert((node->bitmap & (1ull << bit)) != 0ull);
   node->bitmap &= ~(1ull << bit);

   uint64_t addr = node->start_address + bit * bucket->size;

   /* If this node is now completely full, remove it from the free list. */
   if (node->bitmap == 0ull) {
      (void) util_dynarray_pop(vma_list, struct vma_bucket_node);
   }

   return addr;
}

static void
bucket_vma_free(struct bo_cache_bucket *bucket, uint64_t address)
{
   enum brw_memory_zone memzone = memzone_for_address(address);
   struct util_dynarray *vma_list = &bucket->vma_list[memzone];
   const uint64_t node_bytes = 64ull * bucket->size;
   struct vma_bucket_node *node = NULL;

   /* bucket_vma_alloc allocates 64 blocks at a time, and aligns it to
    * that 64 block size.  So, we can round down to get the starting address.
    */
   uint64_t start = (address / node_bytes) * node_bytes;

   /* Dividing the offset from start by bucket size gives us the bit index. */
   int bit = (address - start) / bucket->size;

   assert(start + bit * bucket->size == address);

   util_dynarray_foreach(vma_list, struct vma_bucket_node, cur) {
      if (cur->start_address == start) {
         node = cur;
         break;
      }
   }

   if (!node) {
      /* No node - the whole group of 64 blocks must have been in-use. */
      node = util_dynarray_grow(vma_list, struct vma_bucket_node, 1);

      if (unlikely(!node))
         return; /* bogus, leaks some GPU VMA, but nothing we can do... */

      node->start_address = start;
      node->bitmap = 0ull;
   }

   /* Set the bit to return the memory. */
   assert((node->bitmap & (1ull << bit)) == 0ull);
   node->bitmap |= 1ull << bit;

   /* The block might be entirely free now, and if so, we could return it
    * to the larger allocator.  But we may as well hang on to it, in case
    * we get more allocations at this block size.
    */
}

static struct bo_cache_bucket *
get_bucket_allocator(struct brw_bufmgr *bufmgr, uint64_t size)
{
   /* Skip using the bucket allocator for very large sizes, as it allocates
    * 64 of them and this can balloon rather quickly.
    */
   if (size > 1024 * PAGE_SIZE)
      return NULL;

   struct bo_cache_bucket *bucket = bucket_for_size(bufmgr, size);

   if (bucket && bucket->size == size)
      return bucket;

   return NULL;
}

/**
 * Allocate a section of virtual memory for a buffer, assigning an address.
 *
 * This uses either the bucket allocator for the given size, or the large
 * object allocator (util_vma).
 */
static uint64_t
vma_alloc(struct brw_bufmgr *bufmgr,
          enum brw_memory_zone memzone,
          uint64_t size,
          uint64_t alignment)
{
   /* Without softpin support, we let the kernel assign addresses. */
   assert(brw_using_softpin(bufmgr));

   alignment = ALIGN(alignment, PAGE_SIZE);

   struct bo_cache_bucket *bucket = get_bucket_allocator(bufmgr, size);
   uint64_t addr;

   if (bucket) {
      addr = bucket_vma_alloc(bufmgr, bucket, memzone);
   } else {
      addr = util_vma_heap_alloc(&bufmgr->vma_allocator[memzone], size,
                                 alignment);
   }

   assert((addr >> 48ull) == 0);
   assert((addr % alignment) == 0);

   return gen_canonical_address(addr);
}

/**
 * Free a virtual memory area, allowing the address to be reused.
 */
static void
vma_free(struct brw_bufmgr *bufmgr,
         uint64_t address,
         uint64_t size)
{
   assert(brw_using_softpin(bufmgr));

   /* Un-canonicalize the address. */
   address = gen_48b_address(address);

   if (address == 0ull)
      return;

   struct bo_cache_bucket *bucket = get_bucket_allocator(bufmgr, size);

   if (bucket) {
      bucket_vma_free(bucket, address);
   } else {
      enum brw_memory_zone memzone = memzone_for_address(address);
      util_vma_heap_free(&bufmgr->vma_allocator[memzone], address, size);
   }
}

int
brw_bo_busy(struct brw_bo *bo)
{
   struct brw_bufmgr *bufmgr = bo->bufmgr;
   struct drm_i915_gem_busy busy = { .handle = bo->gem_handle };

   int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_BUSY, &busy);
   if (ret == 0) {
      bo->idle = !busy.busy;
      return busy.busy;
   }
   return false;
}

int
brw_bo_madvise(struct brw_bo *bo, int state)
{
   struct drm_i915_gem_madvise madv = {
      .handle = bo->gem_handle,
      .madv = state,
      .retained = 1,
   };

   drmIoctl(bo->bufmgr->fd, DRM_IOCTL_I915_GEM_MADVISE, &madv);

   return madv.retained;
}

/* drop the oldest entries that have been purged by the kernel */
static void
brw_bo_cache_purge_bucket(struct brw_bufmgr *bufmgr,
                          struct bo_cache_bucket *bucket)
{
   list_for_each_entry_safe(struct brw_bo, bo, &bucket->head, head) {
      if (brw_bo_madvise(bo, I915_MADV_DONTNEED))
         break;

      list_del(&bo->head);
      bo_free(bo);
   }
}

static struct brw_bo *
bo_calloc(void)
{
   struct brw_bo *bo = calloc(1, sizeof(*bo));
   if (!bo)
      return NULL;

   list_inithead(&bo->exports);

   return bo;
}

static struct brw_bo *
bo_alloc_internal(struct brw_bufmgr *bufmgr,
                  const char *name,
                  uint64_t size,
                  enum brw_memory_zone memzone,
                  unsigned flags,
                  uint32_t tiling_mode,
                  uint32_t stride)
{
   struct brw_bo *bo;
   int ret;
   struct bo_cache_bucket *bucket;
   bool alloc_from_cache;
   uint64_t bo_size;
   bool busy = false;
   bool zeroed = false;

   if (flags & BO_ALLOC_BUSY)
      busy = true;

   if (flags & BO_ALLOC_ZEROED)
      zeroed = true;

   /* BUSY does doesn't really jive with ZEROED as we have to wait for it to
    * be idle before we can memset.  Just disallow that combination.
    */
   assert(!(busy && zeroed));

   /* Round the allocated size up to a power of two number of pages. */
   bucket = bucket_for_size(bufmgr, size);

   /* If we don't have caching at this size, don't actually round the
    * allocation up.
    */
   if (bucket == NULL) {
      unsigned int page_size = getpagesize();
      bo_size = size == 0 ? page_size : ALIGN(size, page_size);
   } else {
      bo_size = bucket->size;
   }
   assert(bo_size);

   mtx_lock(&bufmgr->lock);
   /* Get a buffer out of the cache if available */
retry:
   alloc_from_cache = false;
   if (bucket != NULL && !list_is_empty(&bucket->head)) {
      if (busy && !zeroed) {
         /* Allocate new render-target BOs from the tail (MRU)
          * of the list, as it will likely be hot in the GPU
          * cache and in the aperture for us.  If the caller
          * asked us to zero the buffer, we don't want this
          * because we are going to mmap it.
          */
         bo = LIST_ENTRY(struct brw_bo, bucket->head.prev, head);
         list_del(&bo->head);
         alloc_from_cache = true;
      } else {
         /* For non-render-target BOs (where we're probably
          * going to map it first thing in order to fill it
          * with data), check if the last BO in the cache is
          * unbusy, and only reuse in that case. Otherwise,
          * allocating a new buffer is probably faster than
          * waiting for the GPU to finish.
          */
         bo = LIST_ENTRY(struct brw_bo, bucket->head.next, head);
         if (!brw_bo_busy(bo)) {
            alloc_from_cache = true;
            list_del(&bo->head);
         }
      }

      if (alloc_from_cache) {
         assert(list_is_empty(&bo->exports));
         if (!brw_bo_madvise(bo, I915_MADV_WILLNEED)) {
            bo_free(bo);
            brw_bo_cache_purge_bucket(bufmgr, bucket);
            goto retry;
         }

         if (bo_set_tiling_internal(bo, tiling_mode, stride)) {
            bo_free(bo);
            goto retry;
         }

         if (zeroed) {
            void *map = brw_bo_map(NULL, bo, MAP_WRITE | MAP_RAW);
            if (!map) {
               bo_free(bo);
               goto retry;
            }
            memset(map, 0, bo_size);
         }
      }
   }

   if (alloc_from_cache) {
      /* If the cache BO isn't in the right memory zone, free the old
       * memory and assign it a new address.
       */
      if ((bo->kflags & EXEC_OBJECT_PINNED) &&
          memzone != memzone_for_address(bo->gtt_offset)) {
         vma_free(bufmgr, bo->gtt_offset, bo->size);
         bo->gtt_offset = 0ull;
      }
   } else {
      bo = bo_calloc();
      if (!bo)
         goto err;

      bo->size = bo_size;
      bo->idle = true;

      struct drm_i915_gem_create create = { .size = bo_size };

      /* All new BOs we get from the kernel are zeroed, so we don't need to
       * worry about that here.
       */
      ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_CREATE, &create);
      if (ret != 0) {
         free(bo);
         goto err;
      }

      bo->gem_handle = create.handle;

      bo->bufmgr = bufmgr;

      bo->tiling_mode = I915_TILING_NONE;
      bo->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
      bo->stride = 0;

      if (bo_set_tiling_internal(bo, tiling_mode, stride))
         goto err_free;

      /* Calling set_domain() will allocate pages for the BO outside of the
       * struct mutex lock in the kernel, which is more efficient than waiting
       * to create them during the first execbuf that uses the BO.
       */
      struct drm_i915_gem_set_domain sd = {
         .handle = bo->gem_handle,
         .read_domains = I915_GEM_DOMAIN_CPU,
         .write_domain = 0,
      };

      if (drmIoctl(bo->bufmgr->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &sd) != 0)
         goto err_free;
   }

   bo->name = name;
   p_atomic_set(&bo->refcount, 1);
   bo->reusable = true;
   bo->cache_coherent = bufmgr->has_llc;
   bo->index = -1;
   bo->kflags = bufmgr->initial_kflags;

   if ((bo->kflags & EXEC_OBJECT_PINNED) && bo->gtt_offset == 0ull) {
      bo->gtt_offset = vma_alloc(bufmgr, memzone, bo->size, 1);

      if (bo->gtt_offset == 0ull)
         goto err_free;
   }

   mtx_unlock(&bufmgr->lock);

   DBG("bo_create: buf %d (%s) %llub\n", bo->gem_handle, bo->name,
       (unsigned long long) size);

   return bo;

err_free:
   bo_free(bo);
err:
   mtx_unlock(&bufmgr->lock);
   return NULL;
}

struct brw_bo *
brw_bo_alloc(struct brw_bufmgr *bufmgr,
             const char *name, uint64_t size,
             enum brw_memory_zone memzone)
{
   return bo_alloc_internal(bufmgr, name, size, memzone,
                            0, I915_TILING_NONE, 0);
}

struct brw_bo *
brw_bo_alloc_tiled(struct brw_bufmgr *bufmgr, const char *name,
                   uint64_t size, enum brw_memory_zone memzone,
                   uint32_t tiling_mode, uint32_t pitch,
                   unsigned flags)
{
   return bo_alloc_internal(bufmgr, name, size, memzone,
                            flags, tiling_mode, pitch);
}

struct brw_bo *
brw_bo_alloc_tiled_2d(struct brw_bufmgr *bufmgr, const char *name,
                      int x, int y, int cpp, enum brw_memory_zone memzone,
                      uint32_t tiling, uint32_t *pitch, unsigned flags)
{
   uint64_t size;
   uint32_t stride;
   unsigned long aligned_y, height_alignment;

   /* If we're tiled, our allocations are in 8 or 32-row blocks,
    * so failure to align our height means that we won't allocate
    * enough pages.
    *
    * If we're untiled, we still have to align to 2 rows high
    * because the data port accesses 2x2 blocks even if the
    * bottom row isn't to be rendered, so failure to align means
    * we could walk off the end of the GTT and fault.  This is
    * documented on 965, and may be the case on older chipsets
    * too so we try to be careful.
    */
   aligned_y = y;
   height_alignment = 2;

   if (tiling == I915_TILING_X)
      height_alignment = 8;
   else if (tiling == I915_TILING_Y)
      height_alignment = 32;
   aligned_y = ALIGN(y, height_alignment);

   stride = x * cpp;
   stride = bo_tile_pitch(bufmgr, stride, tiling);
   size = stride * aligned_y;
   size = bo_tile_size(bufmgr, size, tiling);
   *pitch = stride;

   if (tiling == I915_TILING_NONE)
      stride = 0;

   return bo_alloc_internal(bufmgr, name, size, memzone,
                            flags, tiling, stride);
}

/**
 * Returns a brw_bo wrapping the given buffer object handle.
 *
 * This can be used when one application needs to pass a buffer object
 * to another.
 */
struct brw_bo *
brw_bo_gem_create_from_name(struct brw_bufmgr *bufmgr,
                            const char *name, unsigned int handle)
{
   struct brw_bo *bo;

   /* At the moment most applications only have a few named bo.
    * For instance, in a DRI client only the render buffers passed
    * between X and the client are named. And since X returns the
    * alternating names for the front/back buffer a linear search
    * provides a sufficiently fast match.
    */
   mtx_lock(&bufmgr->lock);
   bo = hash_find_bo(bufmgr->name_table, handle);
   if (bo) {
      brw_bo_reference(bo);
      goto out;
   }

   struct drm_gem_open open_arg = { .name = handle };
   int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_GEM_OPEN, &open_arg);
   if (ret != 0) {
      DBG("Couldn't reference %s handle 0x%08x: %s\n",
          name, handle, strerror(errno));
      bo = NULL;
      goto out;
   }
   /* Now see if someone has used a prime handle to get this
    * object from the kernel before by looking through the list
    * again for a matching gem_handle
    */
   bo = hash_find_bo(bufmgr->handle_table, open_arg.handle);
   if (bo) {
      assert(list_is_empty(&bo->exports));
      brw_bo_reference(bo);
      goto out;
   }

   bo = bo_calloc();
   if (!bo)
      goto out;

   p_atomic_set(&bo->refcount, 1);

   bo->size = open_arg.size;
   bo->gtt_offset = 0;
   bo->bufmgr = bufmgr;
   bo->gem_handle = open_arg.handle;
   bo->name = name;
   bo->global_name = handle;
   bo->reusable = false;
   bo->external = true;
   bo->kflags = bufmgr->initial_kflags;

   if (bo->kflags & EXEC_OBJECT_PINNED)
      bo->gtt_offset = vma_alloc(bufmgr, BRW_MEMZONE_OTHER, bo->size, 1);

   _mesa_hash_table_insert(bufmgr->handle_table, &bo->gem_handle, bo);
   _mesa_hash_table_insert(bufmgr->name_table, &bo->global_name, bo);

   struct drm_i915_gem_get_tiling get_tiling = { .handle = bo->gem_handle };
   ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_GET_TILING, &get_tiling);
   if (ret != 0)
      goto err_unref;

   bo->tiling_mode = get_tiling.tiling_mode;
   bo->swizzle_mode = get_tiling.swizzle_mode;
   /* XXX stride is unknown */
   DBG("bo_create_from_handle: %d (%s)\n", handle, bo->name);

out:
   mtx_unlock(&bufmgr->lock);
   return bo;

err_unref:
   bo_free(bo);
   mtx_unlock(&bufmgr->lock);
   return NULL;
}

static void
bo_free(struct brw_bo *bo)
{
   struct brw_bufmgr *bufmgr = bo->bufmgr;

   if (bo->map_cpu) {
      VG_NOACCESS(bo->map_cpu, bo->size);
      drm_munmap(bo->map_cpu, bo->size);
   }
   if (bo->map_wc) {
      VG_NOACCESS(bo->map_wc, bo->size);
      drm_munmap(bo->map_wc, bo->size);
   }
   if (bo->map_gtt) {
      VG_NOACCESS(bo->map_gtt, bo->size);
      drm_munmap(bo->map_gtt, bo->size);
   }

   if (bo->external) {
      struct hash_entry *entry;

      if (bo->global_name) {
         entry = _mesa_hash_table_search(bufmgr->name_table, &bo->global_name);
         _mesa_hash_table_remove(bufmgr->name_table, entry);
      }

      entry = _mesa_hash_table_search(bufmgr->handle_table, &bo->gem_handle);
      _mesa_hash_table_remove(bufmgr->handle_table, entry);
   } else {
      assert(list_is_empty(&bo->exports));
   }

   /* Close this object */
   struct drm_gem_close close = { .handle = bo->gem_handle };
   int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_GEM_CLOSE, &close);
   if (ret != 0) {
      DBG("DRM_IOCTL_GEM_CLOSE %d failed (%s): %s\n",
          bo->gem_handle, bo->name, strerror(errno));
   }

   if (bo->kflags & EXEC_OBJECT_PINNED)
      vma_free(bo->bufmgr, bo->gtt_offset, bo->size);

   free(bo);
}

/** Frees all cached buffers significantly older than @time. */
static void
cleanup_bo_cache(struct brw_bufmgr *bufmgr, time_t time)
{
   int i;

   if (bufmgr->time == time)
      return;

   for (i = 0; i < bufmgr->num_buckets; i++) {
      struct bo_cache_bucket *bucket = &bufmgr->cache_bucket[i];

      list_for_each_entry_safe(struct brw_bo, bo, &bucket->head, head) {
         if (time - bo->free_time <= 1)
            break;

         list_del(&bo->head);

         bo_free(bo);
      }
   }

   bufmgr->time = time;
}

static void
bo_unreference_final(struct brw_bo *bo, time_t time)
{
   struct brw_bufmgr *bufmgr = bo->bufmgr;
   struct bo_cache_bucket *bucket;

   DBG("bo_unreference final: %d (%s)\n", bo->gem_handle, bo->name);

   list_for_each_entry_safe(struct bo_export, export, &bo->exports, link) {
      struct drm_gem_close close = { .handle = export->gem_handle };
      gen_ioctl(export->drm_fd, DRM_IOCTL_GEM_CLOSE, &close);

      list_del(&export->link);
      free(export);
   }

   bucket = bucket_for_size(bufmgr, bo->size);
   /* Put the buffer into our internal cache for reuse if we can. */
   if (bufmgr->bo_reuse && bo->reusable && bucket != NULL &&
       brw_bo_madvise(bo, I915_MADV_DONTNEED)) {
      bo->free_time = time;

      bo->name = NULL;

      list_addtail(&bo->head, &bucket->head);
   } else {
      bo_free(bo);
   }
}

void
brw_bo_unreference(struct brw_bo *bo)
{
   if (bo == NULL)
      return;

   assert(p_atomic_read(&bo->refcount) > 0);

   if (atomic_add_unless(&bo->refcount, -1, 1)) {
      struct brw_bufmgr *bufmgr = bo->bufmgr;
      struct timespec time;

      clock_gettime(CLOCK_MONOTONIC, &time);

      mtx_lock(&bufmgr->lock);

      if (p_atomic_dec_zero(&bo->refcount)) {
         bo_unreference_final(bo, time.tv_sec);
         cleanup_bo_cache(bufmgr, time.tv_sec);
      }

      mtx_unlock(&bufmgr->lock);
   }
}

static void
bo_wait_with_stall_warning(struct brw_context *brw,
                           struct brw_bo *bo,
                           const char *action)
{
   bool busy = brw && brw->perf_debug && !bo->idle;
   double elapsed = unlikely(busy) ? -get_time() : 0.0;

   brw_bo_wait_rendering(bo);

   if (unlikely(busy)) {
      elapsed += get_time();
      if (elapsed > 1e-5) /* 0.01ms */
         perf_debug("%s a busy \"%s\" BO stalled and took %.03f ms.\n",
                    action, bo->name, elapsed * 1000);
   }
}

static void
print_flags(unsigned flags)
{
   if (flags & MAP_READ)
      DBG("READ ");
   if (flags & MAP_WRITE)
      DBG("WRITE ");
   if (flags & MAP_ASYNC)
      DBG("ASYNC ");
   if (flags & MAP_PERSISTENT)
      DBG("PERSISTENT ");
   if (flags & MAP_COHERENT)
      DBG("COHERENT ");
   if (flags & MAP_RAW)
      DBG("RAW ");
   DBG("\n");
}

static void *
brw_bo_gem_mmap_legacy(struct brw_context *brw, struct brw_bo *bo, bool wc)
{
   struct brw_bufmgr *bufmgr = bo->bufmgr;

   struct drm_i915_gem_mmap mmap_arg = {
      .handle = bo->gem_handle,
      .size = bo->size,
      .flags = wc ? I915_MMAP_WC : 0,
   };

   int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_MMAP, &mmap_arg);
   if (ret != 0) {
      DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
          __FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
      return NULL;
   }
   void *map = (void *) (uintptr_t) mmap_arg.addr_ptr;

   return map;
}

static void *
brw_bo_gem_mmap_offset(struct brw_context *brw, struct brw_bo *bo, bool wc)
{
   struct brw_bufmgr *bufmgr = bo->bufmgr;

   struct drm_i915_gem_mmap_offset mmap_arg = {
      .handle = bo->gem_handle,
      .flags = wc ? I915_MMAP_OFFSET_WC : I915_MMAP_OFFSET_WB,
   };

   /* Get the fake offset back */
   int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_MMAP_OFFSET, &mmap_arg);
   if (ret != 0) {
      DBG("%s:%d: Error preparing buffer %d (%s): %s .\n",
          __FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
      return NULL;
   }

   /* And map it */
   void *map = drm_mmap(0, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED,
                        bufmgr->fd, mmap_arg.offset);
   if (map == MAP_FAILED) {
      DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
          __FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
      return NULL;
   }

   return map;
}

static void *
brw_bo_gem_mmap(struct brw_context *brw, struct brw_bo *bo, bool wc)
{
   struct brw_bufmgr *bufmgr = bo->bufmgr;

   if (bufmgr->has_mmap_offset)
      return brw_bo_gem_mmap_offset(brw, bo, wc);
   else
      return brw_bo_gem_mmap_legacy(brw, bo, wc);
}

static void *
brw_bo_map_cpu(struct brw_context *brw, struct brw_bo *bo, unsigned flags)
{
   /* We disallow CPU maps for writing to non-coherent buffers, as the
    * CPU map can become invalidated when a batch is flushed out, which
    * can happen at unpredictable times.  You should use WC maps instead.
    */
   assert(bo->cache_coherent || !(flags & MAP_WRITE));

   if (!bo->map_cpu) {
      DBG("brw_bo_map_cpu: %d (%s)\n", bo->gem_handle, bo->name);

      void *map = brw_bo_gem_mmap(brw, bo, false);
      VG_DEFINED(map, bo->size);

      if (p_atomic_cmpxchg(&bo->map_cpu, NULL, map)) {
         VG_NOACCESS(map, bo->size);
         drm_munmap(map, bo->size);
      }
   }
   assert(bo->map_cpu);

   DBG("brw_bo_map_cpu: %d (%s) -> %p, ", bo->gem_handle, bo->name,
       bo->map_cpu);
   print_flags(flags);

   if (!(flags & MAP_ASYNC)) {
      bo_wait_with_stall_warning(brw, bo, "CPU mapping");
   }

   if (!bo->cache_coherent && !bo->bufmgr->has_llc) {
      /* If we're reusing an existing CPU mapping, the CPU caches may
       * contain stale data from the last time we read from that mapping.
       * (With the BO cache, it might even be data from a previous buffer!)
       * Even if it's a brand new mapping, the kernel may have zeroed the
       * buffer via CPU writes.
       *
       * We need to invalidate those cachelines so that we see the latest
       * contents, and so long as we only read from the CPU mmap we do not
       * need to write those cachelines back afterwards.
       *
       * On LLC, the emprical evidence suggests that writes from the GPU
       * that bypass the LLC (i.e. for scanout) do *invalidate* the CPU
       * cachelines. (Other reads, such as the display engine, bypass the
       * LLC entirely requiring us to keep dirty pixels for the scanout
       * out of any cache.)
       */
      gen_invalidate_range(bo->map_cpu, bo->size);
   }

   return bo->map_cpu;
}

static void *
brw_bo_map_wc(struct brw_context *brw, struct brw_bo *bo, unsigned flags)
{
   struct brw_bufmgr *bufmgr = bo->bufmgr;

   if (!bufmgr->has_mmap_wc)
      return NULL;

   if (!bo->map_wc) {
      DBG("brw_bo_map_wc: %d (%s)\n", bo->gem_handle, bo->name);
      void *map = brw_bo_gem_mmap(brw, bo, true);
      VG_DEFINED(map, bo->size);

      if (p_atomic_cmpxchg(&bo->map_wc, NULL, map)) {
         VG_NOACCESS(map, bo->size);
         drm_munmap(map, bo->size);
      }
   }
   assert(bo->map_wc);

   DBG("brw_bo_map_wc: %d (%s) -> %p\n", bo->gem_handle, bo->name, bo->map_wc);
   print_flags(flags);

   if (!(flags & MAP_ASYNC)) {
      bo_wait_with_stall_warning(brw, bo, "WC mapping");
   }

   return bo->map_wc;
}

/**
 * Perform an uncached mapping via the GTT.
 *
 * Write access through the GTT is not quite fully coherent. On low power
 * systems especially, like modern Atoms, we can observe reads from RAM before
 * the write via GTT has landed. A write memory barrier that flushes the Write
 * Combining Buffer (i.e. sfence/mfence) is not sufficient to order the later
 * read after the write as the GTT write suffers a small delay through the GTT
 * indirection. The kernel uses an uncached mmio read to ensure the GTT write
 * is ordered with reads (either by the GPU, WB or WC) and unconditionally
 * flushes prior to execbuf submission. However, if we are not informing the
 * kernel about our GTT writes, it will not flush before earlier access, such
 * as when using the cmdparser. Similarly, we need to be careful if we should
 * ever issue a CPU read immediately following a GTT write.
 *
 * Telling the kernel about write access also has one more important
 * side-effect. Upon receiving notification about the write, it cancels any
 * scanout buffering for FBC/PSR and friends. Later FBC/PSR is then flushed by
 * either SW_FINISH or DIRTYFB. The presumption is that we never write to the
 * actual scanout via a mmaping, only to a backbuffer and so all the FBC/PSR
 * tracking is handled on the buffer exchange instead.
 */
static void *
brw_bo_map_gtt(struct brw_context *brw, struct brw_bo *bo, unsigned flags)
{
   struct brw_bufmgr *bufmgr = bo->bufmgr;

   /* Get a mapping of the buffer if we haven't before. */
   if (bo->map_gtt == NULL) {
      DBG("bo_map_gtt: mmap %d (%s)\n", bo->gem_handle, bo->name);

      struct drm_i915_gem_mmap_gtt mmap_arg = { .handle = bo->gem_handle };

      /* Get the fake offset back... */
      int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_MMAP_GTT, &mmap_arg);
      if (ret != 0) {
         DBG("%s:%d: Error preparing buffer map %d (%s): %s .\n",
             __FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
         return NULL;
      }

      /* and mmap it. */
      void *map = drm_mmap(0, bo->size, PROT_READ | PROT_WRITE,
                           MAP_SHARED, bufmgr->fd, mmap_arg.offset);
      if (map == MAP_FAILED) {
         DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
             __FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
         return NULL;
      }

      /* We don't need to use VALGRIND_MALLOCLIKE_BLOCK because Valgrind will
       * already intercept this mmap call. However, for consistency between
       * all the mmap paths, we mark the pointer as defined now and mark it
       * as inaccessible afterwards.
       */
      VG_DEFINED(map, bo->size);

      if (p_atomic_cmpxchg(&bo->map_gtt, NULL, map)) {
         VG_NOACCESS(map, bo->size);
         drm_munmap(map, bo->size);
      }
   }
   assert(bo->map_gtt);

   DBG("bo_map_gtt: %d (%s) -> %p, ", bo->gem_handle, bo->name, bo->map_gtt);
   print_flags(flags);

   if (!(flags & MAP_ASYNC)) {
      bo_wait_with_stall_warning(brw, bo, "GTT mapping");
   }

   return bo->map_gtt;
}

static bool
can_map_cpu(struct brw_bo *bo, unsigned flags)
{
   if (bo->cache_coherent)
      return true;

   /* Even if the buffer itself is not cache-coherent (such as a scanout), on
    * an LLC platform reads always are coherent (as they are performed via the
    * central system agent). It is just the writes that we need to take special
    * care to ensure that land in main memory and not stick in the CPU cache.
    */
   if (!(flags & MAP_WRITE) && bo->bufmgr->has_llc)
      return true;

   /* If PERSISTENT or COHERENT are set, the mmapping needs to remain valid
    * across batch flushes where the kernel will change cache domains of the
    * bo, invalidating continued access to the CPU mmap on non-LLC device.
    *
    * Similarly, ASYNC typically means that the buffer will be accessed via
    * both the CPU and the GPU simultaneously.  Batches may be executed that
    * use the BO even while it is mapped.  While OpenGL technically disallows
    * most drawing while non-persistent mappings are active, we may still use
    * the GPU for blits or other operations, causing batches to happen at
    * inconvenient times.
    */
   if (flags & (MAP_PERSISTENT | MAP_COHERENT | MAP_ASYNC))
      return false;

   return !(flags & MAP_WRITE);
}

void *
brw_bo_map(struct brw_context *brw, struct brw_bo *bo, unsigned flags)
{
   if (bo->tiling_mode != I915_TILING_NONE && !(flags & MAP_RAW))
      return brw_bo_map_gtt(brw, bo, flags);

   void *map;

   if (can_map_cpu(bo, flags))
      map = brw_bo_map_cpu(brw, bo, flags);
   else
      map = brw_bo_map_wc(brw, bo, flags);

   /* Allow the attempt to fail by falling back to the GTT where necessary.
    *
    * Not every buffer can be mmaped directly using the CPU (or WC), for
    * example buffers that wrap stolen memory or are imported from other
    * devices. For those, we have little choice but to use a GTT mmapping.
    * However, if we use a slow GTT mmapping for reads where we expected fast
    * access, that order of magnitude difference in throughput will be clearly
    * expressed by angry users.
    *
    * We skip MAP_RAW because we want to avoid map_gtt's fence detiling.
    */
   if (!map && !(flags & MAP_RAW)) {
      if (brw) {
         perf_debug("Fallback GTT mapping for %s with access flags %x\n",
                    bo->name, flags);
      }
      map = brw_bo_map_gtt(brw, bo, flags);
   }

   return map;
}

int
brw_bo_subdata(struct brw_bo *bo, uint64_t offset,
               uint64_t size, const void *data)
{
   struct brw_bufmgr *bufmgr = bo->bufmgr;

   struct drm_i915_gem_pwrite pwrite = {
      .handle = bo->gem_handle,
      .offset = offset,
      .size = size,
      .data_ptr = (uint64_t) (uintptr_t) data,
   };

   int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_PWRITE, &pwrite);
   if (ret != 0) {
      ret = -errno;
      DBG("%s:%d: Error writing data to buffer %d: "
          "(%"PRIu64" %"PRIu64") %s .\n",
          __FILE__, __LINE__, bo->gem_handle, offset, size, strerror(errno));
   }

   return ret;
}

/** Waits for all GPU rendering with the object to have completed. */
void
brw_bo_wait_rendering(struct brw_bo *bo)
{
   /* We require a kernel recent enough for WAIT_IOCTL support.
    * See intel_init_bufmgr()
    */
   brw_bo_wait(bo, -1);
}

/**
 * Waits on a BO for the given amount of time.
 *
 * @bo: buffer object to wait for
 * @timeout_ns: amount of time to wait in nanoseconds.
 *   If value is less than 0, an infinite wait will occur.
 *
 * Returns 0 if the wait was successful ie. the last batch referencing the
 * object has completed within the allotted time. Otherwise some negative return
 * value describes the error. Of particular interest is -ETIME when the wait has
 * failed to yield the desired result.
 *
 * Similar to brw_bo_wait_rendering except a timeout parameter allows
 * the operation to give up after a certain amount of time. Another subtle
 * difference is the internal locking semantics are different (this variant does
 * not hold the lock for the duration of the wait). This makes the wait subject
 * to a larger userspace race window.
 *
 * The implementation shall wait until the object is no longer actively
 * referenced within a batch buffer at the time of the call. The wait will
 * not guarantee that the buffer is re-issued via another thread, or an flinked
 * handle. Userspace must make sure this race does not occur if such precision
 * is important.
 *
 * Note that some kernels have broken the inifite wait for negative values
 * promise, upgrade to latest stable kernels if this is the case.
 */
int
brw_bo_wait(struct brw_bo *bo, int64_t timeout_ns)
{
   struct brw_bufmgr *bufmgr = bo->bufmgr;

   /* If we know it's idle, don't bother with the kernel round trip */
   if (bo->idle && !bo->external)
      return 0;

   struct drm_i915_gem_wait wait = {
      .bo_handle = bo->gem_handle,
      .timeout_ns = timeout_ns,
   };
   int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_WAIT, &wait);
   if (ret != 0)
      return -errno;

   bo->idle = true;

   return ret;
}

void
brw_bufmgr_unref(struct brw_bufmgr *bufmgr)
{
   mtx_lock(&global_bufmgr_list_mutex);
   if (p_atomic_dec_zero(&bufmgr->refcount)) {
      list_del(&bufmgr->link);
   } else {
      bufmgr = NULL;
   }
   mtx_unlock(&global_bufmgr_list_mutex);

   if (!bufmgr)
      return;

   mtx_destroy(&bufmgr->lock);

   /* Free any cached buffer objects we were going to reuse */
   for (int i = 0; i < bufmgr->num_buckets; i++) {
      struct bo_cache_bucket *bucket = &bufmgr->cache_bucket[i];

      list_for_each_entry_safe(struct brw_bo, bo, &bucket->head, head) {
         list_del(&bo->head);

         bo_free(bo);
      }

      if (brw_using_softpin(bufmgr)) {
         for (int z = 0; z < BRW_MEMZONE_COUNT; z++) {
            util_dynarray_fini(&bucket->vma_list[z]);
         }
      }
   }

   _mesa_hash_table_destroy(bufmgr->name_table, NULL);
   _mesa_hash_table_destroy(bufmgr->handle_table, NULL);

   if (brw_using_softpin(bufmgr)) {
      for (int z = 0; z < BRW_MEMZONE_COUNT; z++) {
         util_vma_heap_finish(&bufmgr->vma_allocator[z]);
      }
   }

   close(bufmgr->fd);
   bufmgr->fd = -1;

   free(bufmgr);
}

static int
bo_set_tiling_internal(struct brw_bo *bo, uint32_t tiling_mode,
                       uint32_t stride)
{
   struct brw_bufmgr *bufmgr = bo->bufmgr;
   struct drm_i915_gem_set_tiling set_tiling;
   int ret;

   if (bo->global_name == 0 &&
       tiling_mode == bo->tiling_mode && stride == bo->stride)
      return 0;

   memset(&set_tiling, 0, sizeof(set_tiling));
   do {
      /* set_tiling is slightly broken and overwrites the
       * input on the error path, so we have to open code
       * rmIoctl.
       */
      set_tiling.handle = bo->gem_handle;
      set_tiling.tiling_mode = tiling_mode;
      set_tiling.stride = stride;

      ret = ioctl(bufmgr->fd, DRM_IOCTL_I915_GEM_SET_TILING, &set_tiling);
   } while (ret == -1 && (errno == EINTR || errno == EAGAIN));
   if (ret == -1)
      return -errno;

   bo->tiling_mode = set_tiling.tiling_mode;
   bo->swizzle_mode = set_tiling.swizzle_mode;
   bo->stride = set_tiling.stride;
   return 0;
}

int
brw_bo_get_tiling(struct brw_bo *bo, uint32_t *tiling_mode,
                  uint32_t *swizzle_mode)
{
   *tiling_mode = bo->tiling_mode;
   *swizzle_mode = bo->swizzle_mode;
   return 0;
}

static struct brw_bo *
brw_bo_gem_create_from_prime_internal(struct brw_bufmgr *bufmgr, int prime_fd,
                                      int tiling_mode, uint32_t stride)
{
   uint32_t handle;
   struct brw_bo *bo;

   mtx_lock(&bufmgr->lock);
   int ret = drmPrimeFDToHandle(bufmgr->fd, prime_fd, &handle);
   if (ret) {
      DBG("create_from_prime: failed to obtain handle from fd: %s\n",
          strerror(errno));
      mtx_unlock(&bufmgr->lock);
      return NULL;
   }

   /*
    * See if the kernel has already returned this buffer to us. Just as
    * for named buffers, we must not create two bo's pointing at the same
    * kernel object
    */
   bo = hash_find_bo(bufmgr->handle_table, handle);
   if (bo) {
      assert(list_is_empty(&bo->exports));
      brw_bo_reference(bo);
      goto out;
   }

   bo = bo_calloc();
   if (!bo)
      goto out;

   p_atomic_set(&bo->refcount, 1);

   /* Determine size of bo.  The fd-to-handle ioctl really should
    * return the size, but it doesn't.  If we have kernel 3.12 or
    * later, we can lseek on the prime fd to get the size.  Older
    * kernels will just fail, in which case we fall back to the
    * provided (estimated or guess size). */
   ret = lseek(prime_fd, 0, SEEK_END);
   if (ret != -1)
      bo->size = ret;

   bo->bufmgr = bufmgr;

   bo->gem_handle = handle;
   _mesa_hash_table_insert(bufmgr->handle_table, &bo->gem_handle, bo);

   bo->name = "prime";
   bo->reusable = false;
   bo->external = true;
   bo->kflags = bufmgr->initial_kflags;

   if (bo->kflags & EXEC_OBJECT_PINNED) {
      assert(bo->size > 0);
      bo->gtt_offset = vma_alloc(bufmgr, BRW_MEMZONE_OTHER, bo->size, 1);
   }

   if (tiling_mode < 0) {
      struct drm_i915_gem_get_tiling get_tiling = { .handle = bo->gem_handle };
      if (drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_GET_TILING, &get_tiling))
         goto err;

      bo->tiling_mode = get_tiling.tiling_mode;
      bo->swizzle_mode = get_tiling.swizzle_mode;
      /* XXX stride is unknown */
   } else {
      bo_set_tiling_internal(bo, tiling_mode, stride);
   }

out:
   mtx_unlock(&bufmgr->lock);
   return bo;

err:
   bo_free(bo);
   mtx_unlock(&bufmgr->lock);
   return NULL;
}

struct brw_bo *
brw_bo_gem_create_from_prime(struct brw_bufmgr *bufmgr, int prime_fd)
{
   return brw_bo_gem_create_from_prime_internal(bufmgr, prime_fd, -1, 0);
}

struct brw_bo *
brw_bo_gem_create_from_prime_tiled(struct brw_bufmgr *bufmgr, int prime_fd,
                                   uint32_t tiling_mode, uint32_t stride)
{
   assert(tiling_mode == I915_TILING_NONE ||
          tiling_mode == I915_TILING_X ||
          tiling_mode == I915_TILING_Y);

   return brw_bo_gem_create_from_prime_internal(bufmgr, prime_fd,
                                                tiling_mode, stride);
}

static void
brw_bo_make_external(struct brw_bo *bo)
{
   struct brw_bufmgr *bufmgr = bo->bufmgr;

   if (!bo->external) {
      mtx_lock(&bufmgr->lock);
      if (!bo->external) {
         _mesa_hash_table_insert(bufmgr->handle_table, &bo->gem_handle, bo);
         bo->external = true;
      }
      mtx_unlock(&bufmgr->lock);
   }
}

int
brw_bo_gem_export_to_prime(struct brw_bo *bo, int *prime_fd)
{
   struct brw_bufmgr *bufmgr = bo->bufmgr;

   brw_bo_make_external(bo);

   if (drmPrimeHandleToFD(bufmgr->fd, bo->gem_handle,
                          DRM_CLOEXEC, prime_fd) != 0)
      return -errno;

   bo->reusable = false;

   return 0;
}

uint32_t
brw_bo_export_gem_handle(struct brw_bo *bo)
{
   brw_bo_make_external(bo);

   return bo->gem_handle;
}

int
brw_bo_flink(struct brw_bo *bo, uint32_t *name)
{
   struct brw_bufmgr *bufmgr = bo->bufmgr;

   if (!bo->global_name) {
      struct drm_gem_flink flink = { .handle = bo->gem_handle };

      if (drmIoctl(bufmgr->fd, DRM_IOCTL_GEM_FLINK, &flink))
         return -errno;

      brw_bo_make_external(bo);
      mtx_lock(&bufmgr->lock);
      if (!bo->global_name) {
         bo->global_name = flink.name;
         _mesa_hash_table_insert(bufmgr->name_table, &bo->global_name, bo);
      }
      mtx_unlock(&bufmgr->lock);

      bo->reusable = false;
   }

   *name = bo->global_name;
   return 0;
}

int
brw_bo_export_gem_handle_for_device(struct brw_bo *bo, int drm_fd,
                                    uint32_t *out_handle)
{
   struct brw_bufmgr *bufmgr = bo->bufmgr;

   /* Only add the new GEM handle to the list of export if it belongs to a
    * different GEM device. Otherwise we might close the same buffer multiple
    * times.
    */
   int ret = os_same_file_description(drm_fd, bufmgr->fd);
   WARN_ONCE(ret < 0,
             "Kernel has no file descriptor comparison support: %s\n",
             strerror(errno));
   if (ret == 0) {
      *out_handle = brw_bo_export_gem_handle(bo);
      return 0;
   }

   struct bo_export *export = calloc(1, sizeof(*export));
   if (!export)
      return -ENOMEM;

   export->drm_fd = drm_fd;

   int dmabuf_fd = -1;
   int err = brw_bo_gem_export_to_prime(bo, &dmabuf_fd);
   if (err) {
      free(export);
      return err;
   }

   mtx_lock(&bufmgr->lock);
   err = drmPrimeFDToHandle(drm_fd, dmabuf_fd, &export->gem_handle);
   close(dmabuf_fd);
   if (err) {
      mtx_unlock(&bufmgr->lock);
      free(export);
      return err;
   }

   bool found = false;
   list_for_each_entry(struct bo_export, iter, &bo->exports, link) {
      if (iter->drm_fd != drm_fd)
         continue;
      /* Here we assume that for a given DRM fd, we'll always get back the
       * same GEM handle for a given buffer.
       */
      assert(iter->gem_handle == export->gem_handle);
      free(export);
      export = iter;
      found = true;
      break;
   }
   if (!found)
      list_addtail(&export->link, &bo->exports);

   mtx_unlock(&bufmgr->lock);

   *out_handle = export->gem_handle;

   return 0;
}

static void
add_bucket(struct brw_bufmgr *bufmgr, int size)
{
   unsigned int i = bufmgr->num_buckets;

   assert(i < ARRAY_SIZE(bufmgr->cache_bucket));

   list_inithead(&bufmgr->cache_bucket[i].head);
   if (brw_using_softpin(bufmgr)) {
      for (int z = 0; z < BRW_MEMZONE_COUNT; z++)
         util_dynarray_init(&bufmgr->cache_bucket[i].vma_list[z], NULL);
   }
   bufmgr->cache_bucket[i].size = size;
   bufmgr->num_buckets++;

   assert(bucket_for_size(bufmgr, size) == &bufmgr->cache_bucket[i]);
   assert(bucket_for_size(bufmgr, size - 2048) == &bufmgr->cache_bucket[i]);
   assert(bucket_for_size(bufmgr, size + 1) != &bufmgr->cache_bucket[i]);
}

static void
init_cache_buckets(struct brw_bufmgr *bufmgr)
{
   uint64_t size, cache_max_size = 64 * 1024 * 1024;

   /* OK, so power of two buckets was too wasteful of memory.
    * Give 3 other sizes between each power of two, to hopefully
    * cover things accurately enough.  (The alternative is
    * probably to just go for exact matching of sizes, and assume
    * that for things like composited window resize the tiled
    * width/height alignment and rounding of sizes to pages will
    * get us useful cache hit rates anyway)
    */
   add_bucket(bufmgr, PAGE_SIZE);
   add_bucket(bufmgr, PAGE_SIZE * 2);
   add_bucket(bufmgr, PAGE_SIZE * 3);

   /* Initialize the linked lists for BO reuse cache. */
   for (size = 4 * PAGE_SIZE; size <= cache_max_size; size *= 2) {
      add_bucket(bufmgr, size);

      add_bucket(bufmgr, size + size * 1 / 4);
      add_bucket(bufmgr, size + size * 2 / 4);
      add_bucket(bufmgr, size + size * 3 / 4);
   }
}

uint32_t
brw_create_hw_context(struct brw_bufmgr *bufmgr)
{
   struct drm_i915_gem_context_create create = { };
   int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_CONTEXT_CREATE, &create);
   if (ret != 0) {
      DBG("DRM_IOCTL_I915_GEM_CONTEXT_CREATE failed: %s\n", strerror(errno));
      return 0;
   }

   return create.ctx_id;
}

int
brw_hw_context_set_priority(struct brw_bufmgr *bufmgr,
                            uint32_t ctx_id,
                            int priority)
{
   struct drm_i915_gem_context_param p = {
      .ctx_id = ctx_id,
      .param = I915_CONTEXT_PARAM_PRIORITY,
      .value = priority,
   };
   int err;

   err = 0;
   if (drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_CONTEXT_SETPARAM, &p))
      err = -errno;

   return err;
}

void
brw_destroy_hw_context(struct brw_bufmgr *bufmgr, uint32_t ctx_id)
{
   struct drm_i915_gem_context_destroy d = { .ctx_id = ctx_id };

   if (ctx_id != 0 &&
       drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_CONTEXT_DESTROY, &d) != 0) {
      fprintf(stderr, "DRM_IOCTL_I915_GEM_CONTEXT_DESTROY failed: %s\n",
              strerror(errno));
   }
}

int
brw_reg_read(struct brw_bufmgr *bufmgr, uint32_t offset, uint64_t *result)
{
   struct drm_i915_reg_read reg_read = { .offset = offset };
   int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_REG_READ, &reg_read);

   *result = reg_read.val;
   return ret;
}

static int
gem_param(int fd, int name)
{
   int v = -1; /* No param uses (yet) the sign bit, reserve it for errors */

   struct drm_i915_getparam gp = { .param = name, .value = &v };
   if (drmIoctl(fd, DRM_IOCTL_I915_GETPARAM, &gp))
      return -1;

   return v;
}

static int
gem_context_getparam(int fd, uint32_t context, uint64_t param, uint64_t *value)
{
   struct drm_i915_gem_context_param gp = {
      .ctx_id = context,
      .param = param,
   };

   if (drmIoctl(fd, DRM_IOCTL_I915_GEM_CONTEXT_GETPARAM, &gp))
      return -1;

   *value = gp.value;

   return 0;
}

bool
brw_using_softpin(struct brw_bufmgr *bufmgr)
{
   return bufmgr->initial_kflags & EXEC_OBJECT_PINNED;
}

static struct brw_bufmgr *
brw_bufmgr_ref(struct brw_bufmgr *bufmgr)
{
   p_atomic_inc(&bufmgr->refcount);
   return bufmgr;
}

/**
 * Initializes the GEM buffer manager, which uses the kernel to allocate, map,
 * and manage map buffer objections.
 *
 * \param fd File descriptor of the opened DRM device.
 */
static struct brw_bufmgr *
brw_bufmgr_create(struct gen_device_info *devinfo, int fd, bool bo_reuse)
{
   struct brw_bufmgr *bufmgr;

   bufmgr = calloc(1, sizeof(*bufmgr));
   if (bufmgr == NULL)
      return NULL;

   /* Handles to buffer objects belong to the device fd and are not
    * reference counted by the kernel.  If the same fd is used by
    * multiple parties (threads sharing the same screen bufmgr, or
    * even worse the same device fd passed to multiple libraries)
    * ownership of those handles is shared by those independent parties.
    *
    * Don't do this! Ensure that each library/bufmgr has its own device
    * fd so that its namespace does not clash with another.
    */
   bufmgr->fd = os_dupfd_cloexec(fd);
   if (bufmgr->fd < 0) {
      free(bufmgr);
      return NULL;
   }

   p_atomic_set(&bufmgr->refcount, 1);

   if (mtx_init(&bufmgr->lock, mtx_plain) != 0) {
      close(bufmgr->fd);
      free(bufmgr);
      return NULL;
   }

   uint64_t gtt_size;
   if (gem_context_getparam(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE, &gtt_size))
      gtt_size = 0;

   bufmgr->has_llc = devinfo->has_llc;
   bufmgr->has_mmap_wc = gem_param(fd, I915_PARAM_MMAP_VERSION) > 0;
   bufmgr->bo_reuse = bo_reuse;
   bufmgr->has_mmap_offset = gem_param(fd, I915_PARAM_MMAP_GTT_VERSION) >= 4;

   const uint64_t _4GB = 4ull << 30;

   /* The STATE_BASE_ADDRESS size field can only hold 1 page shy of 4GB */
   const uint64_t _4GB_minus_1 = _4GB - PAGE_SIZE;

   if (devinfo->gen >= 8 && gtt_size > _4GB) {
      bufmgr->initial_kflags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;

      /* Allocate VMA in userspace if we have softpin and full PPGTT. */
      if (gem_param(fd, I915_PARAM_HAS_EXEC_SOFTPIN) > 0 &&
          gem_param(fd, I915_PARAM_HAS_ALIASING_PPGTT) > 1) {
         bufmgr->initial_kflags |= EXEC_OBJECT_PINNED;

         util_vma_heap_init(&bufmgr->vma_allocator[BRW_MEMZONE_LOW_4G],
                            PAGE_SIZE, _4GB_minus_1);

         /* Leave the last 4GB out of the high vma range, so that no state
          * base address + size can overflow 48 bits.
          */
         util_vma_heap_init(&bufmgr->vma_allocator[BRW_MEMZONE_OTHER],
                            1 * _4GB, gtt_size - 2 * _4GB);
      } else if (devinfo->gen >= 10) {
         /* Softpin landed in 4.5, but GVT used an aliasing PPGTT until
          * kernel commit 6b3816d69628becb7ff35978aa0751798b4a940a in
          * 4.14.  Gen10+ GVT hasn't landed yet, so it's not actually a
          * problem - but extending this requirement back to earlier gens
          * might actually mean requiring 4.14.
          */
         fprintf(stderr, "i965 requires softpin (Kernel 4.5) on Gen10+.");
         close(bufmgr->fd);
         free(bufmgr);
         return NULL;
      }
   }

   init_cache_buckets(bufmgr);

   bufmgr->name_table =
      _mesa_hash_table_create(NULL, _mesa_hash_uint, _mesa_key_uint_equal);
   bufmgr->handle_table =
      _mesa_hash_table_create(NULL, _mesa_hash_uint, _mesa_key_uint_equal);

   return bufmgr;
}

struct brw_bufmgr *
brw_bufmgr_get_for_fd(struct gen_device_info *devinfo, int fd, bool bo_reuse)
{
   struct stat st;

   if (fstat(fd, &st))
      return NULL;

   struct brw_bufmgr *bufmgr = NULL;

   mtx_lock(&global_bufmgr_list_mutex);
   list_for_each_entry(struct brw_bufmgr, iter_bufmgr, &global_bufmgr_list, link) {
      struct stat iter_st;
      if (fstat(iter_bufmgr->fd, &iter_st))
         continue;

      if (st.st_rdev == iter_st.st_rdev) {
         assert(iter_bufmgr->bo_reuse == bo_reuse);
         bufmgr = brw_bufmgr_ref(iter_bufmgr);
         goto unlock;
      }
   }

   bufmgr = brw_bufmgr_create(devinfo, fd, bo_reuse);
   if (bufmgr)
      list_addtail(&bufmgr->link, &global_bufmgr_list);

 unlock:
   mtx_unlock(&global_bufmgr_list_mutex);

   return bufmgr;
}

int
brw_bufmgr_get_fd(struct brw_bufmgr *bufmgr)
{
   return bufmgr->fd;
}