xref: /xsrc/external/mit/xf86-video-intel/dist/src/sna/kgem.c

/*
 * Copyright (c) 2011 Intel Corporation
 *
 * 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:
 *    Chris Wilson <chris@chris-wilson.co.uk>
 *
 */

#ifdef HAVE_CONFIG_H
#include "config.h"
#endif

#include "sna.h"
#include "sna_reg.h"

#include <unistd.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <time.h>
#include <sched.h>
#include <errno.h>
#include <fcntl.h>

#include <xf86drm.h>

#ifdef HAVE_VALGRIND
#include <valgrind.h>
#include <memcheck.h>
#endif

#ifdef HAVE_STRUCT_SYSINFO_TOTALRAM
#include <sys/sysinfo.h>
#endif

#include "sna_cpuid.h"

static struct kgem_bo *
search_linear_cache(struct kgem *kgem, unsigned int num_pages, unsigned flags);

static struct kgem_bo *
search_snoop_cache(struct kgem *kgem, unsigned int num_pages, unsigned flags);

#define DBG_NO_HW 0
#define DBG_NO_EXEC 0
#define DBG_NO_TILING 0
#define DBG_NO_CACHE 0
#define DBG_NO_SNOOP_CACHE 0
#define DBG_NO_CACHE_LEVEL 0
#define DBG_NO_CPU 0
#define DBG_NO_CREATE2 0
#define DBG_NO_USERPTR 0
#define DBG_NO_UNSYNCHRONIZED_USERPTR 0
#define DBG_NO_COHERENT_MMAP_GTT 0
#define DBG_NO_LLC 0
#define DBG_NO_SEMAPHORES 0
#define DBG_NO_MADV 0
#define DBG_NO_UPLOAD_CACHE 0
#define DBG_NO_UPLOAD_ACTIVE 0
#define DBG_NO_MAP_UPLOAD 0
#define DBG_NO_RELAXED_FENCING 0
#define DBG_NO_SECURE_BATCHES 0
#define DBG_NO_PINNED_BATCHES 0
#define DBG_NO_SHRINK_BATCHES 0
#define DBG_NO_FAST_RELOC 0
#define DBG_NO_HANDLE_LUT 0
#define DBG_NO_WT 0
#define DBG_NO_WC_MMAP 0
#define DBG_NO_BLT_Y 0
#define DBG_NO_SCANOUT_Y 0
#define DBG_NO_DIRTYFB 0
#define DBG_NO_DETILING 0
#define DBG_DUMP 0
#define DBG_NO_MALLOC_CACHE 0

#define FORCE_MMAP_SYNC 0 /* ((1 << DOMAIN_CPU) | (1 << DOMAIN_GTT)) */

#ifndef DEBUG_SYNC
#define DEBUG_SYNC 0
#endif

#define SHOW_BATCH_BEFORE 0
#define SHOW_BATCH_AFTER 0

#if 0
#define ASSERT_IDLE(kgem__, handle__) assert(!__kgem_busy(kgem__, handle__))
#define ASSERT_MAYBE_IDLE(kgem__, handle__, expect__) assert(!(expect__) || !__kgem_busy(kgem__, handle__))
#else
#define ASSERT_IDLE(kgem__, handle__)
#define ASSERT_MAYBE_IDLE(kgem__, handle__, expect__)
#endif

/* Worst case seems to be 965gm where we cannot write within a cacheline that
 * is being simultaneously being read by the GPU, or within the sampler
 * prefetch. In general, the chipsets seem to have a requirement that sampler
 * offsets be aligned to a cacheline (64 bytes).
 *
 * Actually, it turns out the BLT color pattern (BR15) has the most severe
 * alignment restrictions, 64 bytes for 8-bpp, 128 bytes for 16-bpp and 256
 * bytes for 32-bpp.
 */
#define UPLOAD_ALIGNMENT 256

#define PAGE_ALIGN(x) ALIGN(x, PAGE_SIZE)
#define NUM_PAGES(x) (((x) + PAGE_SIZE-1) / PAGE_SIZE)

#define MAX_GTT_VMA_CACHE 512
#define MAX_CPU_VMA_CACHE INT16_MAX
#define MAP_PRESERVE_TIME 10

#define MAKE_USER_MAP(ptr) ((void*)((uintptr_t)(ptr) | 1))
#define IS_USER_MAP(ptr) ((uintptr_t)(ptr) & 1)

#define LOCAL_I915_PARAM_HAS_BLT		11
#define LOCAL_I915_PARAM_HAS_RELAXED_FENCING	12
#define LOCAL_I915_PARAM_HAS_RELAXED_DELTA	15
#define LOCAL_I915_PARAM_HAS_LLC		17
#define LOCAL_I915_PARAM_HAS_SEMAPHORES		20
#define LOCAL_I915_PARAM_HAS_SECURE_BATCHES	23
#define LOCAL_I915_PARAM_HAS_PINNED_BATCHES	24
#define LOCAL_I915_PARAM_HAS_NO_RELOC		25
#define LOCAL_I915_PARAM_HAS_HANDLE_LUT		26
#define LOCAL_I915_PARAM_HAS_WT			27
#define LOCAL_I915_PARAM_MMAP_VERSION		30
#define LOCAL_I915_PARAM_MMAP_GTT_COHERENT	52

#define LOCAL_I915_EXEC_IS_PINNED		(1<<10)
#define LOCAL_I915_EXEC_NO_RELOC		(1<<11)
#define LOCAL_I915_EXEC_HANDLE_LUT		(1<<12)

#define LOCAL_I915_GEM_CREATE2       0x34
#define LOCAL_IOCTL_I915_GEM_CREATE2 DRM_IOWR (DRM_COMMAND_BASE + LOCAL_I915_GEM_CREATE2, struct local_i915_gem_create2)
struct local_i915_gem_create2 {
	uint64_t size;
	uint32_t placement;
#define LOCAL_I915_CREATE_PLACEMENT_SYSTEM 0
#define LOCAL_I915_CREATE_PLACEMENT_STOLEN 1 /* Cannot use CPU mmaps or pread/pwrite */
	uint32_t domain;
	uint32_t caching;
	uint32_t tiling_mode;
	uint32_t stride;
	uint32_t flags;
	uint32_t pad;
	uint32_t handle;
};

#define LOCAL_I915_GEM_USERPTR       0x33
#define LOCAL_IOCTL_I915_GEM_USERPTR DRM_IOWR (DRM_COMMAND_BASE + LOCAL_I915_GEM_USERPTR, struct local_i915_gem_userptr)
struct local_i915_gem_userptr {
	uint64_t user_ptr;
	uint64_t user_size;
	uint32_t flags;
#define I915_USERPTR_READ_ONLY		0x1
#define I915_USERPTR_UNSYNCHRONIZED	0x80000000
	uint32_t handle;
};

#define UNCACHED	0
#define SNOOPED		1
#define DISPLAY		2

struct local_i915_gem_caching {
	uint32_t handle;
	uint32_t caching;
};

#define LOCAL_I915_GEM_SET_CACHING	0x2f
#define LOCAL_I915_GEM_GET_CACHING	0x30
#define LOCAL_IOCTL_I915_GEM_SET_CACHING DRM_IOW(DRM_COMMAND_BASE + LOCAL_I915_GEM_SET_CACHING, struct local_i915_gem_caching)
#define LOCAL_IOCTL_I915_GEM_GET_CACHING DRM_IOW(DRM_COMMAND_BASE + LOCAL_I915_GEM_GET_CACHING, struct local_i915_gem_caching)

struct local_i915_gem_mmap {
	uint32_t handle;
	uint32_t pad;
	uint64_t offset;
	uint64_t size;
	uint64_t addr_ptr;
};
#define LOCAL_IOCTL_I915_GEM_MMAP DRM_IOWR(DRM_COMMAND_BASE + DRM_I915_GEM_MMAP, struct local_i915_gem_mmap)

struct local_i915_gem_mmap2 {
	uint32_t handle;
	uint32_t pad;
	uint64_t offset;
	uint64_t size;
	uint64_t addr_ptr;
	uint64_t flags;
#define I915_MMAP_WC 0x1
};
#define LOCAL_IOCTL_I915_GEM_MMAP_v2 DRM_IOWR(DRM_COMMAND_BASE + DRM_I915_GEM_MMAP, struct local_i915_gem_mmap2)

struct kgem_buffer {
	struct kgem_bo base;
	void *mem;
	uint32_t used;
	uint32_t need_io : 1;
	uint32_t write : 2;
	uint32_t mmapped : 2;
};
enum {
	MMAPPED_NONE,
	MMAPPED_GTT,
	MMAPPED_CPU
};

static struct kgem_bo *__kgem_freed_bo;
static struct kgem_request *__kgem_freed_request;
static struct drm_i915_gem_exec_object2 _kgem_dummy_exec;

static inline struct sna *__to_sna(struct kgem *kgem)
{
	/* minor layering violations */
	return container_of(kgem, struct sna, kgem);
}

static inline int bytes(struct kgem_bo *bo)
{
	return __kgem_bo_size(bo);
}

#define bucket(B) (B)->size.pages.bucket
#define num_pages(B) (B)->size.pages.count

static int __do_ioctl(int fd, unsigned long req, void *arg)
{
	do {
		int err;

		switch ((err = errno)) {
		case EAGAIN:
			sched_yield();
		case EINTR:
			break;
		default:
			return -err;
		}

		if (likely(ioctl(fd, req, arg) == 0))
			return 0;
	} while (1);
}

inline static int do_ioctl(int fd, unsigned long req, void *arg)
{
	if (likely(ioctl(fd, req, arg) == 0))
		return 0;

	return __do_ioctl(fd, req, arg);
}

#ifdef DEBUG_MEMORY
static void debug_alloc(struct kgem *kgem, size_t size)
{
	kgem->debug_memory.bo_allocs++;
	kgem->debug_memory.bo_bytes += size;
}
static void debug_alloc__bo(struct kgem *kgem, struct kgem_bo *bo)
{
	debug_alloc(kgem, bytes(bo));
}
#else
#define debug_alloc__bo(k, b)
#endif

#ifndef NDEBUG
static void assert_tiling(struct kgem *kgem, struct kgem_bo *bo)
{
	struct drm_i915_gem_get_tiling tiling;

	assert(bo);

	if (!kgem->can_fence && kgem->gen >= 040 && bo->tiling)
		return; /* lies */

	VG_CLEAR(tiling);
	tiling.handle = bo->handle;
	tiling.tiling_mode = bo->tiling;
	(void)do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_GET_TILING, &tiling);
	assert(tiling.tiling_mode == bo->tiling);
}

static void assert_caching(struct kgem *kgem, struct kgem_bo *bo)
{
	struct local_i915_gem_caching arg;
	int expect = kgem->has_llc ? SNOOPED : UNCACHED;

	VG_CLEAR(arg);
	arg.handle = bo->handle;
	arg.caching = expect;

	(void)do_ioctl(kgem->fd, LOCAL_IOCTL_I915_GEM_GET_CACHING, &arg);

	assert(arg.caching == expect);
}

static void assert_bo_retired(struct kgem_bo *bo)
{
	DBG(("%s: handle=%d, domain: %d exec? %d, rq? %d\n", __FUNCTION__,
	     bo->handle, bo->domain, bo->exec != NULL, bo->rq != NULL));
	assert(bo->refcnt);
	assert(bo->rq == NULL);
	assert(bo->exec == NULL);
	assert(!bo->needs_flush);
	assert(list_is_empty(&bo->request));
}
#else
#define assert_tiling(kgem, bo)
#define assert_caching(kgem, bo)
#define assert_bo_retired(bo)
#endif

static int __find_debugfs(struct kgem *kgem)
{
	int i;

	for (i = 0; i < DRM_MAX_MINOR; i++) {
		char path[80];

		sprintf(path, "/sys/kernel/debug/dri/%d/i915_wedged", i);
		if (access(path, R_OK) == 0)
			return i;

		sprintf(path, "/debug/dri/%d/i915_wedged", i);
		if (access(path, R_OK) == 0)
			return i;
	}

	return -1;
}

static int kgem_get_minor(struct kgem *kgem)
{
	struct stat st;

	if (fstat(kgem->fd, &st))
		return __find_debugfs(kgem);

	if (!S_ISCHR(st.st_mode))
		return __find_debugfs(kgem);

	return st.st_rdev & 0x63;
}

static bool find_hang_state(struct kgem *kgem, char *path, int maxlen)
{
	int minor = kgem_get_minor(kgem);

	/* Search for our hang state in a few canonical locations.
	 * In the unlikely event of having multiple devices, we
	 * will need to check which minor actually corresponds to ours.
	 */

	snprintf(path, maxlen, "/sys/class/drm/card%d/error", minor);
	if (access(path, R_OK) == 0)
		return true;

	snprintf(path, maxlen, "/sys/kernel/debug/dri/%d/i915_error_state", minor);
	if (access(path, R_OK) == 0)
		return true;

	snprintf(path, maxlen, "/debug/dri/%d/i915_error_state", minor);
	if (access(path, R_OK) == 0)
		return true;

	path[0] = '\0';
	return false;
}

static bool has_error_state(struct kgem *kgem, char *path)
{
   bool ret = false;
   char no;
   int fd;

   fd = open(path, O_RDONLY);
   if (fd >= 0) {
      ret = read(fd, &no, 1) == 1 && no != 'N';
      close(fd);
   }

   return ret;
}

static int kgem_get_screen_index(struct kgem *kgem)
{
	return __to_sna(kgem)->scrn->scrnIndex;
}

static void
__kgem_set_wedged(struct kgem *kgem)
{
	static int once;
	char path[256];

	if (kgem->wedged)
		return;

	if (!once &&
	    find_hang_state(kgem, path, sizeof(path)) &&
            has_error_state(kgem, path)) {
		xf86DrvMsg(kgem_get_screen_index(kgem), X_ERROR,
			   "When reporting this, please include %s and the full dmesg.\n",
			   path);
		once = 1;
	}

	kgem->wedged = true;
	sna_render_mark_wedged(__to_sna(kgem));
}

static void kgem_sna_reset(struct kgem *kgem)
{
	struct sna *sna = __to_sna(kgem);

	sna->render.reset(sna);
	sna->blt_state.fill_bo = 0;
}

static void kgem_sna_flush(struct kgem *kgem)
{
	struct sna *sna = __to_sna(kgem);

	sna->render.flush(sna);

	if (sna->render.solid_cache.dirty)
		sna_render_flush_solid(sna);
}

static bool kgem_bo_rmfb(struct kgem *kgem, struct kgem_bo *bo)
{
	if (bo->scanout && bo->delta) {
		DBG(("%s: releasing fb=%d for handle=%d\n",
		     __FUNCTION__, bo->delta, bo->handle));
		/* XXX will leak if we are not DRM_MASTER. *shrug* */
		do_ioctl(kgem->fd, DRM_IOCTL_MODE_RMFB, &bo->delta);
		bo->delta = 0;
		return true;
	} else
		return false;
}

static bool kgem_set_tiling(struct kgem *kgem, struct kgem_bo *bo,
			    int tiling, int stride)
{
	struct drm_i915_gem_set_tiling set_tiling;
	int err;

	if (tiling == bo->tiling) {
		if (tiling == I915_TILING_NONE) {
			bo->pitch = stride;
			return true;
		}
		if (stride == bo->pitch)
			return true;
	}

	if (DBG_NO_TILING)
		return false;

	VG_CLEAR(set_tiling);
restart:
	set_tiling.handle = bo->handle;
	set_tiling.tiling_mode = tiling;
	set_tiling.stride = tiling ? stride : 0;

	if (ioctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_TILING, &set_tiling) == 0) {
		bo->tiling = set_tiling.tiling_mode;
		bo->pitch = set_tiling.tiling_mode ? set_tiling.stride : stride;
		DBG(("%s: handle=%d, tiling=%d [%d], pitch=%d [%d]: %d\n",
		     __FUNCTION__, bo->handle,
		     bo->tiling, tiling,
		     bo->pitch, stride,
		     set_tiling.tiling_mode == tiling));
		return set_tiling.tiling_mode == tiling && bo->pitch >= stride;
	}

	err = errno;
	if (err == EINTR)
		goto restart;

	if (err == EAGAIN) {
		sched_yield();
		goto restart;
	}

	if (err == EBUSY && kgem_bo_rmfb(kgem, bo))
		goto restart;

	ERR(("%s: failed to set-tiling(tiling=%d, pitch=%d) for handle=%d: %d\n",
	     __FUNCTION__, tiling, stride, bo->handle, err));
	return false;
}

static bool gem_set_caching(int fd, uint32_t handle, int caching)
{
	struct local_i915_gem_caching arg;

	VG_CLEAR(arg);
	arg.handle = handle;
	arg.caching = caching;
	return do_ioctl(fd, LOCAL_IOCTL_I915_GEM_SET_CACHING, &arg) == 0;
}

static uint32_t gem_userptr(int fd, void *ptr, int size, int read_only)
{
	struct local_i915_gem_userptr arg;

	VG_CLEAR(arg);
	arg.user_ptr = (uintptr_t)ptr;
	arg.user_size = size;
	arg.flags = I915_USERPTR_UNSYNCHRONIZED;
	if (read_only)
		arg.flags |= I915_USERPTR_READ_ONLY;

	if (DBG_NO_UNSYNCHRONIZED_USERPTR ||
	    do_ioctl(fd, LOCAL_IOCTL_I915_GEM_USERPTR, &arg)) {
		arg.flags &= ~I915_USERPTR_UNSYNCHRONIZED;
		if (do_ioctl(fd, LOCAL_IOCTL_I915_GEM_USERPTR, &arg)) {
			DBG(("%s: failed to map %p + %d bytes: %d\n",
			     __FUNCTION__, ptr, size, errno));
			return 0;
		}
	}

	return arg.handle;
}

static bool __kgem_throttle(struct kgem *kgem, bool harder)
{
	/* Let this be woken up by sigtimer so that we don't block here
	 * too much and completely starve X. We will sleep again shortly,
	 * and so catch up or detect the hang.
	 */
	do {
		if (ioctl(kgem->fd, DRM_IOCTL_I915_GEM_THROTTLE) == 0) {
			kgem->need_throttle = 0;
			return false;
		}

		if (errno == EIO)
			return true;
	} while (harder);

	return false;
}

static bool __kgem_throttle_retire(struct kgem *kgem, unsigned flags)
{
	if (flags & CREATE_NO_RETIRE || !kgem->need_retire) {
		DBG(("%s: not retiring\n", __FUNCTION__));
		return false;
	}

	if (kgem_retire(kgem))
		return true;

	if (flags & CREATE_NO_THROTTLE || !kgem->need_throttle) {
		DBG(("%s: not throttling\n", __FUNCTION__));
		return false;
	}

	__kgem_throttle(kgem, false);
	return kgem_retire(kgem);
}

static void *__kgem_bo_map__gtt(struct kgem *kgem, struct kgem_bo *bo)
{
	struct drm_i915_gem_mmap_gtt gtt;
	void *ptr;
	int err;

	DBG(("%s(handle=%d, size=%d)\n", __FUNCTION__,
	     bo->handle, bytes(bo)));

	if (bo->tiling && !kgem->can_fence)
		return NULL;

	VG_CLEAR(gtt);
retry_gtt:
	gtt.handle = bo->handle;
	if ((err = do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_MMAP_GTT, &gtt))) {
		DBG(("%s: failed %d, throttling/cleaning caches\n",
		     __FUNCTION__, err));
		assert(err != EINVAL);

		(void)__kgem_throttle_retire(kgem, 0);
		if (kgem_expire_cache(kgem))
			goto retry_gtt;

		if (kgem_cleanup_cache(kgem))
			goto retry_gtt;

		ERR(("%s: failed to retrieve GTT offset for handle=%d: %d\n",
		     __FUNCTION__, bo->handle, -err));
		return NULL;
	}

retry_mmap:
	ptr = mmap(0, bytes(bo), PROT_READ | PROT_WRITE, MAP_SHARED,
		   kgem->fd, gtt.offset);
	if (ptr == MAP_FAILED) {
		err = errno;
		DBG(("%s: failed %d, throttling/cleaning caches\n",
		     __FUNCTION__, err));
		assert(err != EINVAL);

		if (__kgem_throttle_retire(kgem, 0))
			goto retry_mmap;

		if (kgem_cleanup_cache(kgem))
			goto retry_mmap;

		ERR(("%s: failed to mmap handle=%d, %d bytes, into GTT domain: %d\n",
		     __FUNCTION__, bo->handle, bytes(bo), err));
		ptr = NULL;
	}

	/* Cache this mapping to avoid the overhead of an
	 * excruciatingly slow GTT pagefault. This is more an
	 * issue with compositing managers which need to
	 * frequently flush CPU damage to their GPU bo.
	 */
	return bo->map__gtt = ptr;
}

static void *__kgem_bo_map__wc(struct kgem *kgem, struct kgem_bo *bo)
{
	struct local_i915_gem_mmap2 wc;
	int err;

	DBG(("%s(handle=%d, size=%d)\n", __FUNCTION__,
	     bo->handle, bytes(bo)));
	assert(kgem->has_wc_mmap);

	VG_CLEAR(wc);

retry_wc:
	wc.handle = bo->handle;
	wc.offset = 0;
	wc.size = bytes(bo);
	wc.flags = I915_MMAP_WC;
	if ((err = do_ioctl(kgem->fd, LOCAL_IOCTL_I915_GEM_MMAP_v2, &wc))) {
		DBG(("%s: failed %d, throttling/cleaning caches\n",
		     __FUNCTION__, err));
		assert(err != EINVAL);

		if (__kgem_throttle_retire(kgem, 0))
			goto retry_wc;

		if (kgem_cleanup_cache(kgem))
			goto retry_wc;

		ERR(("%s: failed to mmap handle=%d, %d bytes, into CPU(wc) domain: %d\n",
		     __FUNCTION__, bo->handle, bytes(bo), -err));
		return NULL;
	}

	VG(VALGRIND_MAKE_MEM_DEFINED(wc.addr_ptr, bytes(bo)));

	DBG(("%s: caching CPU(wc) vma for %d\n", __FUNCTION__, bo->handle));
	return bo->map__wc = (void *)(uintptr_t)wc.addr_ptr;
}

static void *__kgem_bo_map__cpu(struct kgem *kgem, struct kgem_bo *bo)
{
	struct local_i915_gem_mmap arg;
	int err;

	VG_CLEAR(arg);
	arg.offset = 0;

retry:
	arg.handle = bo->handle;
	arg.size = bytes(bo);
	if ((err = do_ioctl(kgem->fd, LOCAL_IOCTL_I915_GEM_MMAP, &arg))) {
		DBG(("%s: failed %d, throttling/cleaning caches\n",
		     __FUNCTION__, err));
		assert(err != -EINVAL || bo->prime);

		if (__kgem_throttle_retire(kgem, 0))
			goto retry;

		if (kgem_cleanup_cache(kgem))
			goto retry;

		ERR(("%s: failed to mmap handle=%d (prime? %d), %d bytes, into CPU domain: %d\n",
		     __FUNCTION__, bo->handle, bo->prime, bytes(bo), -err));
		bo->purged = 1;
		return NULL;
	}

	VG(VALGRIND_MAKE_MEM_DEFINED(arg.addr_ptr, bytes(bo)));

	DBG(("%s: caching CPU vma for %d\n", __FUNCTION__, bo->handle));
	return bo->map__cpu = (void *)(uintptr_t)arg.addr_ptr;
}

static int gem_write(int fd, uint32_t handle,
		     int offset, int length,
		     const void *src)
{
	struct drm_i915_gem_pwrite pwrite;

	DBG(("%s(handle=%d, offset=%d, len=%d)\n", __FUNCTION__,
	     handle, offset, length));

	VG_CLEAR(pwrite);
	pwrite.handle = handle;
	pwrite.offset = offset;
	pwrite.size = length;
	pwrite.data_ptr = (uintptr_t)src;
	return do_ioctl(fd, DRM_IOCTL_I915_GEM_PWRITE, &pwrite);
}

static int gem_write__cachealigned(int fd, uint32_t handle,
				   int offset, int length,
				   const void *src)
{
	struct drm_i915_gem_pwrite pwrite;

	DBG(("%s(handle=%d, offset=%d, len=%d)\n", __FUNCTION__,
	     handle, offset, length));

	VG_CLEAR(pwrite);
	pwrite.handle = handle;
	/* align the transfer to cachelines; fortuitously this is safe! */
	if ((offset | length) & 63) {
		pwrite.offset = offset & ~63;
		pwrite.size = ALIGN(offset+length, 64) - pwrite.offset;
		pwrite.data_ptr = (uintptr_t)src + pwrite.offset - offset;
	} else {
		pwrite.offset = offset;
		pwrite.size = length;
		pwrite.data_ptr = (uintptr_t)src;
	}
	return do_ioctl(fd, DRM_IOCTL_I915_GEM_PWRITE, &pwrite);
}

static int gem_read(int fd, uint32_t handle, const void *dst,
		    int offset, int length)
{
	struct drm_i915_gem_pread pread;
	int ret;

	DBG(("%s(handle=%d, len=%d)\n", __FUNCTION__,
	     handle, length));

	VG_CLEAR(pread);
	pread.handle = handle;
	pread.offset = offset;
	pread.size = length;
	pread.data_ptr = (uintptr_t)dst;
	ret = do_ioctl(fd, DRM_IOCTL_I915_GEM_PREAD, &pread);
	if (ret) {
		DBG(("%s: failed, errno=%d\n", __FUNCTION__, -ret));
		return ret;
	}

	VG(VALGRIND_MAKE_MEM_DEFINED(dst, length));
	return 0;
}

bool __kgem_busy(struct kgem *kgem, int handle)
{
	struct drm_i915_gem_busy busy;

	VG_CLEAR(busy);
	busy.handle = handle;
	busy.busy = !kgem->wedged;
	(void)do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_BUSY, &busy);
	DBG(("%s: handle=%d, busy=%d, wedged=%d\n",
	     __FUNCTION__, handle, busy.busy, kgem->wedged));

	return busy.busy;
}

static void kgem_bo_retire(struct kgem *kgem, struct kgem_bo *bo)
{
	DBG(("%s: retiring bo handle=%d (needed flush? %d), rq? %d [busy?=%d]\n",
	     __FUNCTION__, bo->handle, bo->needs_flush, bo->rq != NULL,
	     __kgem_busy(kgem, bo->handle)));
	assert(bo->exec == NULL);
	assert(list_is_empty(&bo->vma));

	if (bo->rq)
		__kgem_retire_requests_upto(kgem, bo);
	ASSERT_IDLE(kgem, bo->handle);
	assert_bo_retired(bo);
}

static void kgem_bo_maybe_retire(struct kgem *kgem, struct kgem_bo *bo)
{
	DBG(("%s: retiring bo handle=%d (needed flush? %d), rq? %d [busy?=%d]\n",
	     __FUNCTION__, bo->handle, bo->needs_flush, bo->rq != NULL,
	     __kgem_busy(kgem, bo->handle)));
	assert(bo->exec == NULL);
	assert(list_is_empty(&bo->vma));

	if (bo->rq) {
		if (!__kgem_busy(kgem, bo->handle))
			__kgem_retire_requests_upto(kgem, bo);
	} else {
		assert(!bo->needs_flush);
		ASSERT_IDLE(kgem, bo->handle);
	}
}

bool kgem_bo_write(struct kgem *kgem, struct kgem_bo *bo,
		   const void *data, int length)
{
	void *ptr;
	int err;

	assert(bo->refcnt);
	assert(bo->proxy == NULL);
	ASSERT_IDLE(kgem, bo->handle);

	assert(length <= bytes(bo));
retry:
	ptr = NULL;
	if (bo->domain == DOMAIN_CPU || (kgem->has_llc && !bo->scanout)) {
		ptr = bo->map__cpu;
		if (ptr == NULL)
			ptr = __kgem_bo_map__cpu(kgem, bo);
	} else if (kgem->has_wc_mmap) {
		ptr = bo->map__wc;
		if (ptr == NULL)
			ptr = __kgem_bo_map__wc(kgem, bo);
	}
	if (ptr) {
		/* XXX unsynchronized? */
		memcpy(ptr, data, length);
		return true;
	}

	if ((err = gem_write(kgem->fd, bo->handle, 0, length, data))) {
		DBG(("%s: failed %d, throttling/cleaning caches\n",
		     __FUNCTION__, err));
		assert(err != EINVAL);

		(void)__kgem_throttle_retire(kgem, 0);
		if (kgem_expire_cache(kgem))
			goto retry;

		if (kgem_cleanup_cache(kgem))
			goto retry;

		ERR(("%s: failed to write %d bytes into BO handle=%d: %d\n",
		     __FUNCTION__, length, bo->handle, -err));
		return false;
	}

	DBG(("%s: flush=%d, domain=%d\n", __FUNCTION__, bo->flush, bo->domain));
	if (bo->exec == NULL)
		kgem_bo_maybe_retire(kgem, bo);
	bo->domain = DOMAIN_NONE;
	bo->gtt_dirty = true;
	return true;
}

static uint32_t gem_create(int fd, int num_pages)
{
	struct drm_i915_gem_create create;

	VG_CLEAR(create);
	create.handle = 0;
	create.size = PAGE_SIZE * num_pages;
	(void)do_ioctl(fd, DRM_IOCTL_I915_GEM_CREATE, &create);

	return create.handle;
}

static void
kgem_bo_set_purgeable(struct kgem *kgem, struct kgem_bo *bo)
{
#if !DBG_NO_MADV
	struct drm_i915_gem_madvise madv;

	assert(bo->exec == NULL);

	VG_CLEAR(madv);
	madv.handle = bo->handle;
	madv.madv = I915_MADV_DONTNEED;
	if (do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_MADVISE, &madv) == 0) {
		bo->purged = true;
		kgem->need_purge |= !madv.retained && bo->domain != DOMAIN_CPU;
	}
#endif
}

static bool
kgem_bo_is_retained(struct kgem *kgem, struct kgem_bo *bo)
{
#if DBG_NO_MADV
	return true;
#else
	struct drm_i915_gem_madvise madv;

	if (!bo->purged)
		return true;

	VG_CLEAR(madv);
	madv.handle = bo->handle;
	madv.madv = I915_MADV_DONTNEED;
	if (do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_MADVISE, &madv) == 0)
		return madv.retained;

	return false;
#endif
}

static bool
kgem_bo_clear_purgeable(struct kgem *kgem, struct kgem_bo *bo)
{
#if DBG_NO_MADV
	return true;
#else
	struct drm_i915_gem_madvise madv;

	assert(bo->purged);

	VG_CLEAR(madv);
	madv.handle = bo->handle;
	madv.madv = I915_MADV_WILLNEED;
	if (do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_MADVISE, &madv) == 0) {
		bo->purged = !madv.retained;
		kgem->need_purge |= !madv.retained && bo->domain != DOMAIN_CPU;
		return madv.retained;
	}

	return false;
#endif
}

static void gem_close(int fd, uint32_t handle)
{
	struct drm_gem_close close;

	VG_CLEAR(close);
	close.handle = handle;
	(void)do_ioctl(fd, DRM_IOCTL_GEM_CLOSE, &close);
}

constant inline static unsigned long __fls(unsigned long word)
{
#if defined(__GNUC__) && (defined(__i386__) || defined(__x86__) || defined(__x86_64__))
	asm("bsr %1,%0"
	    : "=r" (word)
	    : "rm" (word));
	return word;
#else
	unsigned int v = 0;

	while (word >>= 1)
		v++;

	return v;
#endif
}

constant inline static int cache_bucket(int num_pages)
{
	return __fls(num_pages);
}

static struct kgem_bo *__kgem_bo_init(struct kgem_bo *bo,
				      int handle, int num_pages)
{
	DBG(("%s(handle=%d, num_pages=%d)\n", __FUNCTION__, handle, num_pages));

	assert(num_pages);
	memset(bo, 0, sizeof(*bo));

	bo->refcnt = 1;
	bo->handle = handle;
	bo->target_handle = -1;
	num_pages(bo) = num_pages;
	bucket(bo) = cache_bucket(num_pages);
	bo->reusable = true;
	bo->domain = DOMAIN_CPU;
	list_init(&bo->request);
	list_init(&bo->list);
	list_init(&bo->vma);

	return bo;
}

static struct kgem_bo *__kgem_bo_alloc(int handle, int num_pages)
{
	struct kgem_bo *bo;

	if (__kgem_freed_bo) {
		bo = __kgem_freed_bo;
		__kgem_freed_bo = *(struct kgem_bo **)bo;
	} else {
		bo = malloc(sizeof(*bo));
		if (bo == NULL)
			return NULL;
	}

	return __kgem_bo_init(bo, handle, num_pages);
}

static struct kgem_request *__kgem_request_alloc(struct kgem *kgem)
{
	struct kgem_request *rq;

	if (unlikely(kgem->wedged)) {
		rq = &kgem->static_request;
	} else {
		rq = __kgem_freed_request;
		if (rq) {
			__kgem_freed_request = *(struct kgem_request **)rq;
		} else {
			rq = malloc(sizeof(*rq));
			if (rq == NULL)
				rq = &kgem->static_request;
		}
	}

	list_init(&rq->buffers);
	rq->bo = NULL;
	rq->ring = 0;

	return rq;
}

static void __kgem_request_free(struct kgem_request *rq)
{
	_list_del(&rq->list);
	if (DBG_NO_MALLOC_CACHE) {
		free(rq);
	} else {
		*(struct kgem_request **)rq = __kgem_freed_request;
		__kgem_freed_request = rq;
	}
}

static struct list *inactive(struct kgem *kgem, int num_pages)
{
	assert(num_pages < MAX_CACHE_SIZE / PAGE_SIZE);
	assert(cache_bucket(num_pages) < NUM_CACHE_BUCKETS);
	return &kgem->inactive[cache_bucket(num_pages)];
}

static struct list *active(struct kgem *kgem, int num_pages, int tiling)
{
	assert(num_pages < MAX_CACHE_SIZE / PAGE_SIZE);
	assert(cache_bucket(num_pages) < NUM_CACHE_BUCKETS);
	return &kgem->active[cache_bucket(num_pages)][tiling];
}

static size_t
agp_aperture_size(struct pci_device *dev, unsigned gen)
{
	/* XXX assume that only future chipsets are unknown and follow
	 * the post gen2 PCI layout.
	 */
	return dev->regions[gen < 030 ? 0 : 2].size;
}

static size_t
total_ram_size(void)
{
#ifdef HAVE_STRUCT_SYSINFO_TOTALRAM
	struct sysinfo info;
	if (sysinfo(&info) == 0)
		return (size_t)info.totalram * info.mem_unit;
#endif

#ifdef _SC_PHYS_PAGES
	 return (size_t)sysconf(_SC_PHYS_PAGES) * sysconf(_SC_PAGE_SIZE);
#endif

	return 0;
}

static unsigned
cpu_cache_size__cpuid4(void)
{
	/* Deterministic Cache Parameters (Function 04h)":
	 *    When EAX is initialized to a value of 4, the CPUID instruction
	 *    returns deterministic cache information in the EAX, EBX, ECX
	 *    and EDX registers.  This function requires ECX be initialized
	 *    with an index which indicates which cache to return information
	 *    about. The OS is expected to call this function (CPUID.4) with
	 *    ECX = 0, 1, 2, until EAX[4:0] == 0, indicating no more caches.
	 *    The order in which the caches are returned is not specified
	 *    and may change at Intel's discretion.
	 *
	 * Calculating the Cache Size in bytes:
	 *          = (Ways +1) * (Partitions +1) * (Line Size +1) * (Sets +1)
	 */

	 unsigned int eax, ebx, ecx, edx;
	 unsigned int llc_size = 0;
	 int cnt;

	 if (__get_cpuid_max(BASIC_CPUID, NULL) < 4)
		 return 0;

	 cnt = 0;
	 do {
		 unsigned associativity, line_partitions, line_size, sets;

		 __cpuid_count(4, cnt++, eax, ebx, ecx, edx);

		 if ((eax & 0x1f) == 0)
			 break;

		 associativity = ((ebx >> 22) & 0x3ff) + 1;
		 line_partitions = ((ebx >> 12) & 0x3ff) + 1;
		 line_size = (ebx & 0xfff) + 1;
		 sets = ecx + 1;

		 llc_size = associativity * line_partitions * line_size * sets;
	 } while (1);

	 return llc_size;
}

static unsigned
cpu_cache_size(void)
{
	unsigned size;
	FILE *file;

	size = cpu_cache_size__cpuid4();
	if (size)
		return size;

	file = fopen("/proc/cpuinfo", "r");
	if (file) {
		size_t len = 0;
		char *line = NULL;
		while (getline(&line, &len, file) != -1) {
			int kb;
			if (sscanf(line, "cache size : %d KB", &kb) == 1) {
				/* Paranoid check against gargantuan caches */
				if (kb <= 1<<20)
					size = kb * 1024;
				break;
			}
		}
		free(line);
		fclose(file);
	}

	if (size == 0)
		size = 64 * 1024;

	return size;
}

static int gem_param(struct kgem *kgem, int name)
{
	drm_i915_getparam_t gp;
	int v = -1; /* No param uses the sign bit, reserve it for errors */

	VG_CLEAR(gp);
	gp.param = name;
	gp.value = &v;
	if (do_ioctl(kgem->fd, DRM_IOCTL_I915_GETPARAM, &gp))
		return -1;

	VG(VALGRIND_MAKE_MEM_DEFINED(&v, sizeof(v)));
	return v;
}

static bool test_has_execbuffer2(struct kgem *kgem)
{
	struct drm_i915_gem_execbuffer2 execbuf;

	memset(&execbuf, 0, sizeof(execbuf));
	execbuf.buffer_count = 1;

	return do_ioctl(kgem->fd,
			 DRM_IOCTL_I915_GEM_EXECBUFFER2,
			 &execbuf) == -EFAULT;
}

static bool test_has_no_reloc(struct kgem *kgem)
{
	if (DBG_NO_FAST_RELOC)
		return false;

	return gem_param(kgem, LOCAL_I915_PARAM_HAS_NO_RELOC) > 0;
}

static bool test_has_handle_lut(struct kgem *kgem)
{
	if (DBG_NO_HANDLE_LUT)
		return false;

	return gem_param(kgem, LOCAL_I915_PARAM_HAS_HANDLE_LUT) > 0;
}

static bool test_has_wt(struct kgem *kgem)
{
	if (DBG_NO_WT)
		return false;

	return gem_param(kgem, LOCAL_I915_PARAM_HAS_WT) > 0;
}

static bool test_has_semaphores_enabled(struct kgem *kgem)
{
	FILE *file;
	bool detected = false;
	int ret;

	if (DBG_NO_SEMAPHORES)
		return false;

	ret = gem_param(kgem, LOCAL_I915_PARAM_HAS_SEMAPHORES);
	if (ret != -1)
		return ret > 0;

	file = fopen("/sys/module/i915/parameters/semaphores", "r");
	if (file) {
		int value;
		if (fscanf(file, "%d", &value) == 1)
			detected = value != 0;
		fclose(file);
	}

	return detected;
}

static bool is_hw_supported(struct kgem *kgem,
			    struct pci_device *dev)
{
	if (DBG_NO_HW)
		return false;

	if (!test_has_execbuffer2(kgem))
		return false;

	if (kgem->gen == (unsigned)-1) /* unknown chipset, assume future gen */
		return kgem->has_blt;

	/* Although pre-855gm the GMCH is fubar, it works mostly. So
	 * let the user decide through "NoAccel" whether or not to risk
	 * hw acceleration.
	 */

	if (kgem->gen == 060 && dev && dev->revision < 8) {
		/* pre-production SNB with dysfunctional BLT */
		return false;
	}

	if (kgem->gen >= 060) /* Only if the kernel supports the BLT ring */
		return kgem->has_blt;

	return true;
}

static bool test_has_relaxed_fencing(struct kgem *kgem)
{
	if (kgem->gen < 040) {
		if (DBG_NO_RELAXED_FENCING)
			return false;

		return gem_param(kgem, LOCAL_I915_PARAM_HAS_RELAXED_FENCING) > 0;
	} else
		return true;
}

static bool test_has_coherent_mmap_gtt(struct kgem *kgem)
{
	if (DBG_NO_COHERENT_MMAP_GTT)
		return false;

	return gem_param(kgem, LOCAL_I915_PARAM_MMAP_GTT_COHERENT) > 0;
}

static bool test_has_llc(struct kgem *kgem)
{
	int has_llc = -1;

	if (DBG_NO_LLC)
		return false;

	has_llc = gem_param(kgem, LOCAL_I915_PARAM_HAS_LLC);
	if (has_llc == -1) {
		DBG(("%s: no kernel/drm support for HAS_LLC, assuming support for LLC based on GPU generation\n", __FUNCTION__));
		has_llc = kgem->gen >= 060;
	}

	return has_llc;
}

static bool test_has_wc_mmap(struct kgem *kgem)
{
	struct local_i915_gem_mmap2 wc;
	bool ret;

	if (DBG_NO_WC_MMAP)
		return false;

	/* XXX See https://bugs.freedesktop.org/show_bug.cgi?id=90841 */
	if (kgem->gen < 033)
		return false;

	if (gem_param(kgem, LOCAL_I915_PARAM_MMAP_VERSION) < 1)
		return false;

	VG_CLEAR(wc);
	wc.handle = gem_create(kgem->fd, 1);
	wc.offset = 0;
	wc.size = 4096;
	wc.flags = I915_MMAP_WC;
	ret = do_ioctl(kgem->fd, LOCAL_IOCTL_I915_GEM_MMAP_v2, &wc) == 0;
	gem_close(kgem->fd, wc.handle);

	return ret;
}

static bool test_has_caching(struct kgem *kgem)
{
	uint32_t handle;
	bool ret;

	if (DBG_NO_CACHE_LEVEL)
		return false;

	/* Incoherent blt and sampler hangs the GPU */
	if (kgem->gen == 040)
		return false;

	handle = gem_create(kgem->fd, 1);
	if (handle == 0)
		return false;

	ret = gem_set_caching(kgem->fd, handle, UNCACHED);
	gem_close(kgem->fd, handle);
	return ret;
}

static bool test_has_userptr(struct kgem *kgem)
{
	struct local_i915_gem_userptr arg;
	void *ptr;

	if (DBG_NO_USERPTR)
		return false;

	/* Incoherent blt and sampler hangs the GPU */
	if (kgem->gen == 040)
		return false;

	if (posix_memalign(&ptr, PAGE_SIZE, PAGE_SIZE))
		return false;

	VG_CLEAR(arg);
	arg.user_ptr = (uintptr_t)ptr;
	arg.user_size = PAGE_SIZE;
	arg.flags = I915_USERPTR_UNSYNCHRONIZED;

	if (DBG_NO_UNSYNCHRONIZED_USERPTR ||
	    do_ioctl(kgem->fd, LOCAL_IOCTL_I915_GEM_USERPTR, &arg)) {
		arg.flags &= ~I915_USERPTR_UNSYNCHRONIZED;
		if (do_ioctl(kgem->fd, LOCAL_IOCTL_I915_GEM_USERPTR, &arg))
			arg.handle = 0;
		/* Leak the userptr bo to keep the mmu_notifier alive */
	} else {
		gem_close(kgem->fd, arg.handle);
		free(ptr);
	}

	return arg.handle != 0;
}

static bool test_has_create2(struct kgem *kgem)
{
#if defined(USE_CREATE2)
	struct local_i915_gem_create2 args;

	if (DBG_NO_CREATE2)
		return false;

	memset(&args, 0, sizeof(args));
	args.size = PAGE_SIZE;
	args.caching = DISPLAY;
	if (do_ioctl(kgem->fd, LOCAL_IOCTL_I915_GEM_CREATE2, &args) == 0)
		gem_close(kgem->fd, args.handle);

	return args.handle != 0;
#else
	return false;
#endif
}

static bool test_can_blt_y(struct kgem *kgem)
{
	struct drm_i915_gem_exec_object2 object;
	uint32_t batch[] = {
#define MI_LOAD_REGISTER_IMM (0x22<<23 | (3-2))
#define BCS_SWCTRL 0x22200
#define BCS_SRC_Y (1 << 0)
#define BCS_DST_Y (1 << 1)
		MI_LOAD_REGISTER_IMM,
		BCS_SWCTRL,
		(BCS_SRC_Y | BCS_DST_Y) << 16 | (BCS_SRC_Y | BCS_DST_Y),

		MI_LOAD_REGISTER_IMM,
		BCS_SWCTRL,
		(BCS_SRC_Y | BCS_DST_Y) << 16,

		MI_BATCH_BUFFER_END,
		0,
	};
	int ret;

	if (DBG_NO_BLT_Y)
		return false;

	if (kgem->gen < 060)
		return false;

	memset(&object, 0, sizeof(object));
	object.handle = gem_create(kgem->fd, 1);

	ret = gem_write(kgem->fd, object.handle, 0, sizeof(batch), batch);
	if (ret == 0) {
		struct drm_i915_gem_execbuffer2 execbuf;

		memset(&execbuf, 0, sizeof(execbuf));
		execbuf.buffers_ptr = (uintptr_t)&object;
		execbuf.buffer_count = 1;
		execbuf.flags = KGEM_BLT;

		ret = do_ioctl(kgem->fd,
			       DRM_IOCTL_I915_GEM_EXECBUFFER2,
			       &execbuf);
	}
	gem_close(kgem->fd, object.handle);

	return ret == 0;
}

static bool gem_set_tiling(int fd, uint32_t handle, int tiling, int stride)
{
	struct drm_i915_gem_set_tiling set_tiling;

	if (DBG_NO_TILING)
		return false;

	VG_CLEAR(set_tiling);
	set_tiling.handle = handle;
	set_tiling.tiling_mode = tiling;
	set_tiling.stride = stride;

	if (ioctl(fd, DRM_IOCTL_I915_GEM_SET_TILING, &set_tiling) == 0)
		return set_tiling.tiling_mode == tiling;

	return false;
}

static bool test_can_scanout_y(struct kgem *kgem)
{
	struct drm_mode_fb_cmd arg;
	bool ret = false;

	if (DBG_NO_SCANOUT_Y)
		return false;

	VG_CLEAR(arg);
	arg.width = 32;
	arg.height = 32;
	arg.pitch = 4*32;
	arg.bpp = 32;
	arg.depth = 24;
	arg.handle = gem_create(kgem->fd, 1);

	if (gem_set_tiling(kgem->fd, arg.handle, I915_TILING_Y, arg.pitch))
		ret = do_ioctl(kgem->fd, DRM_IOCTL_MODE_ADDFB, &arg) == 0;
	if (!ret) {
		struct local_mode_fb_cmd2 {
			uint32_t fb_id;
			uint32_t width, height;
			uint32_t pixel_format;
			uint32_t flags;

			uint32_t handles[4];
			uint32_t pitches[4];
			uint32_t offsets[4];
			uint64_t modifiers[4];
		} f;
#define LOCAL_IOCTL_MODE_ADDFB2 DRM_IOWR(0xb8, struct local_mode_fb_cmd2)
		memset(&f, 0, sizeof(f));
		f.width = arg.width;
		f.height = arg.height;
		f.handles[0] = arg.handle;
		f.pitches[0] = arg.pitch;
		f.modifiers[0] = (uint64_t)1 << 56 | 2; /* MOD_Y_TILED */
		f.pixel_format = 'X' | 'R' << 8 | '2' << 16 | '4' << 24; /* XRGB8888 */
		f.flags = 1 << 1; /* + modifier */
		if (drmIoctl(kgem->fd, LOCAL_IOCTL_MODE_ADDFB2, &f) == 0) {
			ret = true;
			arg.fb_id = f.fb_id;
		}
	}
	do_ioctl(kgem->fd, DRM_IOCTL_MODE_RMFB, &arg.fb_id);
	gem_close(kgem->fd, arg.handle);

	return ret;
}

static bool test_has_dirtyfb(struct kgem *kgem)
{
	struct drm_mode_fb_cmd create;
	bool ret = false;

	if (DBG_NO_DIRTYFB)
		return false;

	VG_CLEAR(create);
	create.width = 32;
	create.height = 32;
	create.pitch = 4*32;
	create.bpp = 24;
	create.depth = 32; /* {bpp:24, depth:32} -> x8r8g8b8 */
	create.handle = gem_create(kgem->fd, 1);
	if (create.handle == 0)
		return false;

	if (drmIoctl(kgem->fd, DRM_IOCTL_MODE_ADDFB, &create) == 0) {
		struct drm_mode_fb_dirty_cmd dirty;

		memset(&dirty, 0, sizeof(dirty));
		dirty.fb_id = create.fb_id;
		ret = drmIoctl(kgem->fd,
			       DRM_IOCTL_MODE_DIRTYFB,
			       &dirty) == 0;

		/* XXX There may be multiple levels of DIRTYFB, depending on
		 * whether the kernel thinks tracking dirty regions is
		 * beneficial vs flagging the whole fb as dirty.
		 */

		drmIoctl(kgem->fd,
			 DRM_IOCTL_MODE_RMFB,
			 &create.fb_id);
	}
	gem_close(kgem->fd, create.handle);

	return ret;
}

static bool test_has_secure_batches(struct kgem *kgem)
{
	if (DBG_NO_SECURE_BATCHES)
		return false;

	return gem_param(kgem, LOCAL_I915_PARAM_HAS_SECURE_BATCHES) > 0;
}

static bool test_has_pinned_batches(struct kgem *kgem)
{
	if (DBG_NO_PINNED_BATCHES)
		return false;

	return gem_param(kgem, LOCAL_I915_PARAM_HAS_PINNED_BATCHES) > 0;
}

static bool kgem_init_pinned_batches(struct kgem *kgem)
{
	int count[2] = { 16, 4 };
	int size[2] = { 1, 4 };
	int ret = 0;
	int n, i;

	if (unlikely(kgem->wedged))
		return true;

	for (n = 0; n < ARRAY_SIZE(count); n++) {
		for (i = 0; i < count[n]; i++) {
			struct drm_i915_gem_pin pin;
			struct kgem_bo *bo;

			VG_CLEAR(pin);

			pin.handle = gem_create(kgem->fd, size[n]);
			if (pin.handle == 0)
				goto err;

			DBG(("%s: new handle=%d, num_pages=%d\n",
			     __FUNCTION__, pin.handle, size[n]));

			bo = __kgem_bo_alloc(pin.handle, size[n]);
			if (bo == NULL) {
				gem_close(kgem->fd, pin.handle);
				goto err;
			}

			pin.alignment = 0;
			ret = do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_PIN, &pin);
			if (ret) {
				gem_close(kgem->fd, pin.handle);
				free(bo);
				goto err;
			}
			bo->presumed_offset = pin.offset;
			debug_alloc__bo(kgem, bo);
			list_add(&bo->list, &kgem->pinned_batches[n]);
		}
	}

	return true;

err:
	for (n = 0; n < ARRAY_SIZE(kgem->pinned_batches); n++) {
		while (!list_is_empty(&kgem->pinned_batches[n])) {
			kgem_bo_destroy(kgem,
					list_first_entry(&kgem->pinned_batches[n],
							 struct kgem_bo, list));
		}
	}

	/* If we fail to pin some memory for 830gm/845g, we need to disable
	 * acceleration as otherwise the machine will eventually fail. However,
	 * the kernel started arbitrarily rejecting PIN, so hope for the best
	 * if the ioctl no longer works.
	 */
	if (ret != -ENODEV && kgem->gen == 020)
		return false;

	kgem->has_pinned_batches = false;

	/* For simplicity populate the lists with a single unpinned bo */
	for (n = 0; n < ARRAY_SIZE(count); n++) {
		struct kgem_bo *bo;
		uint32_t handle;

		handle = gem_create(kgem->fd, size[n]);
		if (handle == 0)
			return false;

		bo = __kgem_bo_alloc(handle, size[n]);
		if (bo == NULL) {
			gem_close(kgem->fd, handle);
			return false;
		}

		debug_alloc__bo(kgem, bo);
		list_add(&bo->list, &kgem->pinned_batches[n]);
	}
	return true;
}

static void kgem_init_swizzling(struct kgem *kgem)
{
	struct local_i915_gem_get_tiling_v2 {
		uint32_t handle;
		uint32_t tiling_mode;
		uint32_t swizzle_mode;
		uint32_t phys_swizzle_mode;
	} tiling;
#define LOCAL_IOCTL_I915_GEM_GET_TILING DRM_IOWR (DRM_COMMAND_BASE + DRM_I915_GEM_GET_TILING, struct local_i915_gem_get_tiling_v2)

	memset(&tiling, 0, sizeof(tiling));
	tiling.handle = gem_create(kgem->fd, 1);
	if (!tiling.handle)
		return;

	if (!gem_set_tiling(kgem->fd, tiling.handle, I915_TILING_X, 512))
		goto out;

	if (do_ioctl(kgem->fd, LOCAL_IOCTL_I915_GEM_GET_TILING, &tiling))
		goto out;

	DBG(("%s: swizzle_mode=%d, phys_swizzle_mode=%d\n",
	     __FUNCTION__, tiling.swizzle_mode, tiling.phys_swizzle_mode));

	kgem->can_fence =
		!DBG_NO_TILING &&
		tiling.swizzle_mode != I915_BIT_6_SWIZZLE_UNKNOWN;

	if (kgem->gen < 050 && tiling.phys_swizzle_mode != tiling.swizzle_mode)
		goto out;

	if (!DBG_NO_DETILING)
		choose_memcpy_tiled_x(kgem,
				      tiling.swizzle_mode,
				      __to_sna(kgem)->cpu_features);
out:
	gem_close(kgem->fd, tiling.handle);
	DBG(("%s: can fence?=%d\n", __FUNCTION__, kgem->can_fence));
}

static void kgem_fixup_relocs(struct kgem *kgem, struct kgem_bo *bo, int shrink)
{
	int n;

	bo->target_handle = kgem->has_handle_lut ? kgem->nexec : bo->handle;

	assert(kgem->nreloc__self <= 256);
	if (kgem->nreloc__self == 0)
		return;

	DBG(("%s: fixing up %d%s self-relocations to handle=%p, presumed-offset=%llx\n",
	     __FUNCTION__, kgem->nreloc__self,
	     kgem->nreloc__self == 256 ? "+" : "",
	     bo->handle, (long long)bo->presumed_offset));
	for (n = 0; n < kgem->nreloc__self; n++) {
		int i = kgem->reloc__self[n];
		uint64_t addr;

		assert(kgem->reloc[i].target_handle == ~0U);
		kgem->reloc[i].target_handle = bo->target_handle;
		kgem->reloc[i].presumed_offset = bo->presumed_offset;

		if (kgem->reloc[i].read_domains == I915_GEM_DOMAIN_INSTRUCTION) {
			DBG(("%s: moving base of self-reloc[%d:%d] %d -> %d\n",
			     __FUNCTION__, n, i,
			     kgem->reloc[i].delta,
			     kgem->reloc[i].delta - shrink));

			kgem->reloc[i].delta -= shrink;
		}
		addr = (int)kgem->reloc[i].delta + bo->presumed_offset;
		kgem->batch[kgem->reloc[i].offset/sizeof(uint32_t)] = addr;
		if (kgem->gen >= 0100)
			kgem->batch[kgem->reloc[i].offset/sizeof(uint32_t) + 1] = addr >> 32;
	}

	if (n == 256) {
		for (n = kgem->reloc__self[255]; n < kgem->nreloc; n++) {
			if (kgem->reloc[n].target_handle == ~0U) {
				uint64_t addr;

				kgem->reloc[n].target_handle = bo->target_handle;
				kgem->reloc[n].presumed_offset = bo->presumed_offset;

				if (kgem->reloc[n].read_domains == I915_GEM_DOMAIN_INSTRUCTION) {
					DBG(("%s: moving base of reloc[%d] %d -> %d\n",
					     __FUNCTION__, n,
					     kgem->reloc[n].delta,
					     kgem->reloc[n].delta - shrink));
					kgem->reloc[n].delta -= shrink;
				}

				addr = (int)kgem->reloc[n].delta + bo->presumed_offset;
				kgem->batch[kgem->reloc[n].offset/sizeof(uint32_t)] = addr;
				if (kgem->gen >= 0100)
					kgem->batch[kgem->reloc[n].offset/sizeof(uint32_t) + 1] = addr >> 32;
			}
		}
	}

	if (shrink) {
		DBG(("%s: shrinking by %d\n", __FUNCTION__, shrink));
		for (n = 0; n < kgem->nreloc; n++) {
			if (kgem->reloc[n].offset >= sizeof(uint32_t)*kgem->nbatch)
				kgem->reloc[n].offset -= shrink;
		}
	}
}

static int kgem_bo_wait(struct kgem *kgem, struct kgem_bo *bo)
{
	struct local_i915_gem_wait {
		uint32_t handle;
		uint32_t flags;
		int64_t timeout;
	} wait;
#define LOCAL_I915_GEM_WAIT       0x2c
#define LOCAL_IOCTL_I915_GEM_WAIT         DRM_IOWR(DRM_COMMAND_BASE + LOCAL_I915_GEM_WAIT, struct local_i915_gem_wait)
	int ret;

	DBG(("%s: waiting for handle=%d\n", __FUNCTION__, bo->handle));
	if (bo->rq == NULL)
		return 0;

	VG_CLEAR(wait);
	wait.handle = bo->handle;
	wait.flags = 0;
	wait.timeout = -1;
	ret = do_ioctl(kgem->fd, LOCAL_IOCTL_I915_GEM_WAIT, &wait);
	if (ret) {
		struct drm_i915_gem_set_domain set_domain;

		VG_CLEAR(set_domain);
		set_domain.handle = bo->handle;
		set_domain.read_domains = I915_GEM_DOMAIN_GTT;
		set_domain.write_domain = I915_GEM_DOMAIN_GTT;
		ret = do_ioctl(kgem->fd,
			       DRM_IOCTL_I915_GEM_SET_DOMAIN,
			       &set_domain);
	}

	if (ret == 0)
		__kgem_retire_requests_upto(kgem, bo);

	return ret;
}

static struct kgem_bo *kgem_new_batch(struct kgem *kgem)
{
	struct kgem_bo *last;
	unsigned flags;

	last = kgem->batch_bo;
	if (last) {
		kgem_fixup_relocs(kgem, last, 0);
		kgem->batch = NULL;
	}

	if (kgem->batch) {
		assert(last == NULL);
		return NULL;
	}

	flags = CREATE_CPU_MAP | CREATE_NO_THROTTLE;
	if (!kgem->has_llc)
		flags |= CREATE_UNCACHED;

restart:
	kgem->batch_bo = kgem_create_linear(kgem,
					    sizeof(uint32_t)*kgem->batch_size,
					    flags);
	if (kgem->batch_bo)
		kgem->batch = kgem_bo_map__cpu(kgem, kgem->batch_bo);
	if (kgem->batch == NULL) {
		int ring = kgem->ring == KGEM_BLT;
		assert(ring < ARRAY_SIZE(kgem->requests));

		if (kgem->batch_bo) {
			kgem_bo_destroy(kgem, kgem->batch_bo);
			kgem->batch_bo = NULL;
		}

		if (!list_is_empty(&kgem->requests[ring])) {
			struct kgem_request *rq;

			rq = list_first_entry(&kgem->requests[ring],
					      struct kgem_request, list);
			assert(rq->ring == ring);
			assert(rq->bo);
			assert(RQ(rq->bo->rq) == rq);
			if (kgem_bo_wait(kgem, rq->bo) == 0)
				goto restart;
		}

		if (flags & CREATE_NO_THROTTLE) {
			flags &= ~CREATE_NO_THROTTLE;
			if (kgem_cleanup_cache(kgem))
				goto restart;
		}

		DBG(("%s: unable to map batch bo, mallocing(size=%d)\n",
		     __FUNCTION__, sizeof(uint32_t)*kgem->batch_size));
		if (posix_memalign((void **)&kgem->batch, PAGE_SIZE,
				   ALIGN(sizeof(uint32_t) * kgem->batch_size, PAGE_SIZE))) {
			ERR(("%s: batch allocation failed, disabling acceleration\n", __FUNCTION__));
			__kgem_set_wedged(kgem);
		}
	} else {
		DBG(("%s: allocated and mapped batch handle=%d [size=%d]\n",
		     __FUNCTION__, kgem->batch_bo->handle,
		     sizeof(uint32_t)*kgem->batch_size));
		kgem_bo_sync__cpu(kgem, kgem->batch_bo);
	}

	DBG(("%s: using last batch handle=%d\n",
	     __FUNCTION__, last ? last->handle : 0));
	return last;
}

static void
no_retire(struct kgem *kgem)
{
	(void)kgem;
}

static void
no_expire(struct kgem *kgem)
{
	(void)kgem;
}

static void
no_context_switch(struct kgem *kgem, int new_mode)
{
	(void)kgem;
	(void)new_mode;
}

static uint64_t get_gtt_size(int fd)
{
	struct drm_i915_gem_get_aperture aperture;
	struct local_i915_gem_context_param {
		uint32_t context;
		uint32_t size;
		uint64_t param;
#define LOCAL_CONTEXT_PARAM_BAN_PERIOD	0x1
#define LOCAL_CONTEXT_PARAM_NO_ZEROMAP	0x2
#define LOCAL_CONTEXT_PARAM_GTT_SIZE	0x3
		uint64_t value;
	} p;
#define LOCAL_I915_GEM_CONTEXT_GETPARAM       0x34
#define LOCAL_IOCTL_I915_GEM_CONTEXT_GETPARAM DRM_IOWR (DRM_COMMAND_BASE + LOCAL_I915_GEM_CONTEXT_GETPARAM, struct local_i915_gem_context_param)

	memset(&aperture, 0, sizeof(aperture));

	memset(&p, 0, sizeof(p));
	p.param = LOCAL_CONTEXT_PARAM_GTT_SIZE;
	if (drmIoctl(fd, LOCAL_IOCTL_I915_GEM_CONTEXT_GETPARAM, &p) == 0)
		aperture.aper_size = p.value;
	if (aperture.aper_size == 0)
		(void)drmIoctl(fd, DRM_IOCTL_I915_GEM_GET_APERTURE, &aperture);
	if (aperture.aper_size == 0)
		aperture.aper_size = 64*1024*1024;

	DBG(("%s: aperture size %lld, available now %lld\n",
	     __FUNCTION__,
	     (long long)aperture.aper_size,
	     (long long)aperture.aper_available_size));

	/* clamp aperture to uint32_t for simplicity */
	if (aperture.aper_size > 0xc0000000)
		aperture.aper_size = 0xc0000000;

	return aperture.aper_size;
}

void kgem_init(struct kgem *kgem, int fd, struct pci_device *dev, unsigned gen)
{
	size_t totalram;
	unsigned half_gpu_max;
	unsigned int i, j;
	uint64_t gtt_size;

	DBG(("%s: fd=%d, gen=%d\n", __FUNCTION__, fd, gen));

	kgem->fd = fd;
	kgem->gen = gen;

	kgem->retire = no_retire;
	kgem->expire = no_expire;
	kgem->context_switch = no_context_switch;

	list_init(&kgem->requests[0]);
	list_init(&kgem->requests[1]);
	list_init(&kgem->batch_buffers);
	list_init(&kgem->active_buffers);
	list_init(&kgem->flushing);
	list_init(&kgem->large);
	list_init(&kgem->large_inactive);
	list_init(&kgem->snoop);
	list_init(&kgem->scanout);
	for (i = 0; i < ARRAY_SIZE(kgem->pinned_batches); i++)
		list_init(&kgem->pinned_batches[i]);
	for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++)
		list_init(&kgem->inactive[i]);
	for (i = 0; i < ARRAY_SIZE(kgem->active); i++) {
		for (j = 0; j < ARRAY_SIZE(kgem->active[i]); j++)
			list_init(&kgem->active[i][j]);
	}
	for (i = 0; i < ARRAY_SIZE(kgem->vma); i++) {
		for (j = 0; j < ARRAY_SIZE(kgem->vma[i].inactive); j++)
			list_init(&kgem->vma[i].inactive[j]);
	}
	kgem->vma[MAP_GTT].count = -MAX_GTT_VMA_CACHE;
	kgem->vma[MAP_CPU].count = -MAX_CPU_VMA_CACHE;

	kgem->has_blt = gem_param(kgem, LOCAL_I915_PARAM_HAS_BLT) > 0;
	DBG(("%s: has BLT ring? %d\n", __FUNCTION__,
	     kgem->has_blt));

	kgem->has_relaxed_delta =
		gem_param(kgem, LOCAL_I915_PARAM_HAS_RELAXED_DELTA) > 0;
	DBG(("%s: has relaxed delta? %d\n", __FUNCTION__,
	     kgem->has_relaxed_delta));

	kgem->has_relaxed_fencing = test_has_relaxed_fencing(kgem);
	DBG(("%s: has relaxed fencing? %d\n", __FUNCTION__,
	     kgem->has_relaxed_fencing));

	kgem->has_coherent_mmap_gtt = test_has_coherent_mmap_gtt(kgem);
	DBG(("%s: has coherent writes into GTT maps? %d\n", __FUNCTION__,
	     kgem->has_coherent_mmap_gtt));

	kgem->has_llc = test_has_llc(kgem);
	DBG(("%s: has shared last-level-cache? %d\n", __FUNCTION__,
	     kgem->has_llc));

	kgem->has_wt = test_has_wt(kgem);
	DBG(("%s: has write-through caching for scanouts? %d\n", __FUNCTION__,
	     kgem->has_wt));

	kgem->has_wc_mmap = test_has_wc_mmap(kgem);
	DBG(("%s: has wc-mmapping? %d\n", __FUNCTION__,
	     kgem->has_wc_mmap));

	kgem->has_caching = test_has_caching(kgem);
	DBG(("%s: has set-cache-level? %d\n", __FUNCTION__,
	     kgem->has_caching));

	kgem->has_userptr = test_has_userptr(kgem);
	DBG(("%s: has userptr? %d\n", __FUNCTION__,
	     kgem->has_userptr));

	kgem->has_create2 = test_has_create2(kgem);
	DBG(("%s: has create2? %d\n", __FUNCTION__,
	     kgem->has_create2));

	kgem->has_no_reloc = test_has_no_reloc(kgem);
	DBG(("%s: has no-reloc? %d\n", __FUNCTION__,
	     kgem->has_no_reloc));

	kgem->has_handle_lut = test_has_handle_lut(kgem);
	DBG(("%s: has handle-lut? %d\n", __FUNCTION__,
	     kgem->has_handle_lut));

	kgem->has_semaphores = false;
	if (kgem->has_blt && test_has_semaphores_enabled(kgem))
		kgem->has_semaphores = true;
	DBG(("%s: semaphores enabled? %d\n", __FUNCTION__,
	     kgem->has_semaphores));

	kgem->can_blt_cpu = gen >= 030;
	DBG(("%s: can blt to cpu? %d\n", __FUNCTION__,
	     kgem->can_blt_cpu));

	kgem->can_blt_y = test_can_blt_y(kgem);
	DBG(("%s: can blit to Y-tiled surfaces? %d\n", __FUNCTION__,
	     kgem->can_blt_y));

	kgem->can_render_y = gen != 021 && (gen >> 3) != 4;
	DBG(("%s: can render to Y-tiled surfaces? %d\n", __FUNCTION__,
	     kgem->can_render_y));

	kgem->can_scanout_y = test_can_scanout_y(kgem);
	DBG(("%s: can scanout Y-tiled surfaces? %d\n", __FUNCTION__,
	     kgem->can_scanout_y));

	kgem->has_dirtyfb = test_has_dirtyfb(kgem);
	DBG(("%s: has dirty fb? %d\n", __FUNCTION__, kgem->has_dirtyfb));

	kgem->has_secure_batches = test_has_secure_batches(kgem);
	DBG(("%s: can use privileged batchbuffers? %d\n", __FUNCTION__,
	     kgem->has_secure_batches));

	kgem->has_pinned_batches = test_has_pinned_batches(kgem);
	DBG(("%s: can use pinned batchbuffers (to avoid CS w/a)? %d\n", __FUNCTION__,
	     kgem->has_pinned_batches));

	if (!is_hw_supported(kgem, dev)) {
		xf86DrvMsg(kgem_get_screen_index(kgem), X_WARNING,
			   "Detected unsupported/dysfunctional hardware, disabling acceleration.\n");
		__kgem_set_wedged(kgem);
	} else if (__kgem_throttle(kgem, false)) {
		xf86DrvMsg(kgem_get_screen_index(kgem), X_WARNING,
			   "Detected a hung GPU, disabling acceleration.\n");
		__kgem_set_wedged(kgem);
	}

	kgem->batch_size = UINT16_MAX & ~7;
	if (gen == 020 && !kgem->has_pinned_batches)
		/* Limited to what we can pin */
		kgem->batch_size = 4*1024;
	if (gen == 022)
		/* 865g cannot handle a batch spanning multiple pages */
		kgem->batch_size = PAGE_SIZE / sizeof(uint32_t);
	if (gen >= 070)
		kgem->batch_size = 16*1024;
	if (!kgem->has_relaxed_delta && kgem->batch_size > 4*1024)
		kgem->batch_size = 4*1024;

	if (!kgem_init_pinned_batches(kgem)) {
		xf86DrvMsg(kgem_get_screen_index(kgem), X_WARNING,
			   "Unable to reserve memory for GPU, disabling acceleration.\n");
		__kgem_set_wedged(kgem);
	}

	DBG(("%s: maximum batch size? %d\n", __FUNCTION__,
	     kgem->batch_size));
	kgem_new_batch(kgem);

	kgem->half_cpu_cache_pages = cpu_cache_size() >> 13;
	DBG(("%s: last-level cache size: %d bytes, threshold in pages: %d\n",
	     __FUNCTION__, cpu_cache_size(), kgem->half_cpu_cache_pages));

	kgem->next_request = __kgem_request_alloc(kgem);

	DBG(("%s: cpu bo enabled %d: llc? %d, set-cache-level? %d, userptr? %d\n", __FUNCTION__,
	     !DBG_NO_CPU && (kgem->has_llc | kgem->has_userptr | kgem->has_caching),
	     kgem->has_llc, kgem->has_caching, kgem->has_userptr));

	gtt_size = get_gtt_size(fd);
	kgem->aperture_total = gtt_size;
	kgem->aperture_high = gtt_size * 3/4;
	kgem->aperture_low = gtt_size * 1/3;
	if (gen < 033) {
		/* Severe alignment penalties */
		kgem->aperture_high /= 2;
		kgem->aperture_low /= 2;
	}
	DBG(("%s: aperture low=%u [%u], high=%u [%u]\n", __FUNCTION__,
	     kgem->aperture_low, kgem->aperture_low / (1024*1024),
	     kgem->aperture_high, kgem->aperture_high / (1024*1024)));

	kgem->aperture_mappable = 256 * 1024 * 1024;
	if (dev != NULL)
		kgem->aperture_mappable = agp_aperture_size(dev, gen);
	if (kgem->aperture_mappable == 0 || kgem->aperture_mappable > gtt_size)
		kgem->aperture_mappable = gtt_size;
	DBG(("%s: aperture mappable=%d [%d MiB]\n", __FUNCTION__,
	     kgem->aperture_mappable, kgem->aperture_mappable / (1024*1024)));

	kgem->aperture_fenceable = MIN(256*1024*1024, kgem->aperture_mappable);
	DBG(("%s: aperture fenceable=%d [%d MiB]\n", __FUNCTION__,
	     kgem->aperture_fenceable, kgem->aperture_fenceable / (1024*1024)));

	kgem->buffer_size = 64 * 1024;
	while (kgem->buffer_size < kgem->aperture_mappable >> 10)
		kgem->buffer_size *= 2;
	if (kgem->buffer_size >> 12 > kgem->half_cpu_cache_pages)
		kgem->buffer_size = kgem->half_cpu_cache_pages << 12;
	kgem->buffer_size = 1 << __fls(kgem->buffer_size);
	DBG(("%s: buffer size=%d [%d KiB]\n", __FUNCTION__,
	     kgem->buffer_size, kgem->buffer_size / 1024));
	assert(kgem->buffer_size);

	kgem->max_object_size = 3 * (kgem->aperture_high >> 12) << 10;
	kgem->max_gpu_size = kgem->max_object_size;
	if (!kgem->has_llc && kgem->max_gpu_size > MAX_CACHE_SIZE)
		kgem->max_gpu_size = MAX_CACHE_SIZE;

	totalram = total_ram_size();
	if (totalram == 0) {
		DBG(("%s: total ram size unknown, assuming maximum of total aperture\n",
		     __FUNCTION__));
		totalram = kgem->aperture_total;
	}
	DBG(("%s: total ram=%lld\n", __FUNCTION__, (long long)totalram));
	if (kgem->max_object_size > totalram / 2)
		kgem->max_object_size = totalram / 2;
	if (kgem->max_gpu_size > totalram / 4)
		kgem->max_gpu_size = totalram / 4;

	if (kgem->aperture_high > totalram / 2) {
		kgem->aperture_high = totalram / 2;
		kgem->aperture_low = kgem->aperture_high / 4;
		DBG(("%s: reduced aperture watermaks to fit into ram; low=%d [%d], high=%d [%d]\n", __FUNCTION__,
		     kgem->aperture_low, kgem->aperture_low / (1024*1024),
		     kgem->aperture_high, kgem->aperture_high / (1024*1024)));
	}

	kgem->max_cpu_size = kgem->max_object_size;

	half_gpu_max = kgem->max_gpu_size / 2;
	kgem->max_copy_tile_size = (MAX_CACHE_SIZE + 1)/2;
	if (kgem->max_copy_tile_size > half_gpu_max)
		kgem->max_copy_tile_size = half_gpu_max;

	if (kgem->has_llc)
		kgem->max_upload_tile_size = kgem->max_copy_tile_size;
	else
		kgem->max_upload_tile_size = kgem->aperture_fenceable / 4;
	if (kgem->max_upload_tile_size > half_gpu_max)
		kgem->max_upload_tile_size = half_gpu_max;
	if (kgem->max_upload_tile_size > kgem->aperture_high/2)
		kgem->max_upload_tile_size = kgem->aperture_high/2;
	if (kgem->max_upload_tile_size > kgem->aperture_low)
		kgem->max_upload_tile_size = kgem->aperture_low;
	if (kgem->max_upload_tile_size < 16*PAGE_SIZE)
		kgem->max_upload_tile_size = 16*PAGE_SIZE;

	kgem->large_object_size = MAX_CACHE_SIZE;
	if (kgem->large_object_size > half_gpu_max)
		kgem->large_object_size = half_gpu_max;
	if (kgem->max_copy_tile_size > kgem->aperture_high/2)
		kgem->max_copy_tile_size = kgem->aperture_high/2;
	if (kgem->max_copy_tile_size > kgem->aperture_low)
		kgem->max_copy_tile_size = kgem->aperture_low;
	if (kgem->max_copy_tile_size < 16*PAGE_SIZE)
		kgem->max_copy_tile_size = 16*PAGE_SIZE;

	if (kgem->has_llc | kgem->has_caching | kgem->has_userptr) {
		if (kgem->large_object_size > kgem->max_cpu_size)
			kgem->large_object_size = kgem->max_cpu_size;
	} else
		kgem->max_cpu_size = 0;
	if (DBG_NO_CPU)
		kgem->max_cpu_size = 0;

	DBG(("%s: maximum object size=%u\n",
	     __FUNCTION__, kgem->max_object_size));
	DBG(("%s: large object thresold=%u\n",
	     __FUNCTION__, kgem->large_object_size));
	DBG(("%s: max object sizes (gpu=%u, cpu=%u, tile upload=%u, copy=%u)\n",
	     __FUNCTION__,
	     kgem->max_gpu_size, kgem->max_cpu_size,
	     kgem->max_upload_tile_size, kgem->max_copy_tile_size));

	/* Convert the aperture thresholds to pages */
	kgem->aperture_mappable /= PAGE_SIZE;
	kgem->aperture_fenceable /= PAGE_SIZE;
	kgem->aperture_low /= PAGE_SIZE;
	kgem->aperture_high /= PAGE_SIZE;
	kgem->aperture_total /= PAGE_SIZE;

	kgem->fence_max = gem_param(kgem, I915_PARAM_NUM_FENCES_AVAIL) - 2;
	if ((int)kgem->fence_max < 0)
		kgem->fence_max = 5; /* minimum safe value for all hw */
	DBG(("%s: max fences=%d\n", __FUNCTION__, kgem->fence_max));

	kgem->batch_flags_base = 0;
	if (kgem->has_no_reloc)
		kgem->batch_flags_base |= LOCAL_I915_EXEC_NO_RELOC;
	if (kgem->has_handle_lut)
		kgem->batch_flags_base |= LOCAL_I915_EXEC_HANDLE_LUT;
	if (kgem->has_pinned_batches)
		kgem->batch_flags_base |= LOCAL_I915_EXEC_IS_PINNED;

	kgem_init_swizzling(kgem);
}

/* XXX hopefully a good approximation */
static uint32_t kgem_get_unique_id(struct kgem *kgem)
{
	uint32_t id;
	id = ++kgem->unique_id;
	if (id == 0)
		id = ++kgem->unique_id;
	return id;
}

inline static uint32_t kgem_pitch_alignment(struct kgem *kgem, unsigned flags)
{
	if (flags & CREATE_PRIME)
		return 256;
	if (flags & CREATE_SCANOUT)
		return 64;
	if (kgem->gen >= 0100)
		return 32;
	return 8;
}

void kgem_get_tile_size(struct kgem *kgem, int tiling, int pitch,
			int *tile_width, int *tile_height, int *tile_size)
{
	if (kgem->gen <= 030) {
		if (tiling) {
			if (kgem->gen < 030) {
				*tile_width = 128;
				*tile_height = 16;
				*tile_size = 2048;
			} else {
				*tile_width = 512;
				*tile_height = 8;
				*tile_size = 4096;
			}
		} else {
			*tile_width = 1;
			*tile_height = 1;
			*tile_size = 1;
		}
	} else switch (tiling) {
	default:
	case I915_TILING_NONE:
		*tile_width = 1;
		*tile_height = 1;
		*tile_size = 1;
		break;
	case I915_TILING_X:
		*tile_width = 512;
		*tile_height = 8;
		*tile_size = 4096;
		break;
	case I915_TILING_Y:
		*tile_width = 128;
		*tile_height = 32;
		*tile_size = 4096;
		break;
	}

	/* Force offset alignment to tile-row */
	if (tiling && kgem->gen < 033)
		*tile_width = pitch;
}

static uint32_t kgem_surface_size(struct kgem *kgem,
				  bool relaxed_fencing,
				  unsigned flags,
				  uint32_t width,
				  uint32_t height,
				  uint32_t bpp,
				  uint32_t tiling,
				  uint32_t *pitch)
{
	uint32_t tile_width, tile_height;
	uint32_t size;

	assert(width <= MAXSHORT);
	assert(height <= MAXSHORT);
	assert(bpp >= 8);

	if (kgem->gen <= 030) {
		if (tiling) {
			if (kgem->gen < 030) {
				tile_width = 128;
				tile_height = 16;
			} else {
				tile_width = 512;
				tile_height = 8;
			}
		} else {
			tile_width = 2 * bpp >> 3;
			tile_width = ALIGN(tile_width,
					   kgem_pitch_alignment(kgem, flags));
			tile_height = 1;
		}
	} else switch (tiling) {
	default:
	case I915_TILING_NONE:
		tile_width = 2 * bpp >> 3;
		tile_width = ALIGN(tile_width,
				   kgem_pitch_alignment(kgem, flags));
		tile_height = 1;
		break;

	case I915_TILING_X:
		tile_width = 512;
		tile_height = 8;
		break;
	case I915_TILING_Y:
		tile_width = 128;
		tile_height = 32;
		break;
	}
	/* XXX align to an even tile row */
	if (!kgem->has_relaxed_fencing)
		tile_height *= 2;

	*pitch = ALIGN(width * bpp / 8, tile_width);
	height = ALIGN(height, tile_height);
	DBG(("%s: tile_width=%d, tile_height=%d => aligned pitch=%d, height=%d\n",
	     __FUNCTION__, tile_width, tile_height, *pitch, height));

	if (kgem->gen >= 040)
		return PAGE_ALIGN(*pitch * height);

	/* If it is too wide for the blitter, don't even bother.  */
	if (tiling != I915_TILING_NONE) {
		if (*pitch > 8192) {
			DBG(("%s: too wide for tiled surface (pitch=%d, limit=%d)\n",
			     __FUNCTION__, *pitch, 8192));
			return 0;
		}

		for (size = tile_width; size < *pitch; size <<= 1)
			;
		*pitch = size;
	} else {
		if (*pitch >= 32768) {
			DBG(("%s: too wide for linear surface (pitch=%d, limit=%d)\n",
			     __FUNCTION__, *pitch, 32767));
			return 0;
		}
	}

	size = *pitch * height;
	if (relaxed_fencing || tiling == I915_TILING_NONE)
		return PAGE_ALIGN(size);

	/* We need to allocate a pot fence region for a tiled buffer. */
	if (kgem->gen < 030)
		tile_width = 512 * 1024;
	else
		tile_width = 1024 * 1024;
	while (tile_width < size)
		tile_width *= 2;
	return tile_width;
}

bool kgem_check_surface_size(struct kgem *kgem,
			     uint32_t width,
			     uint32_t height,
			     uint32_t bpp,
			     uint32_t tiling,
			     uint32_t pitch,
			     uint32_t size)
{
	uint32_t min_size, min_pitch;
	int tile_width, tile_height, tile_size;

	DBG(("%s(width=%d, height=%d, bpp=%d, tiling=%d, pitch=%d, size=%d)\n",
	     __FUNCTION__, width, height, bpp, tiling, pitch, size));

	if (pitch & 3)
		return false;

	min_size = kgem_surface_size(kgem, kgem->has_relaxed_fencing, 0,
				     width, height, bpp, tiling,
				     &min_pitch);

	DBG(("%s: min_pitch=%d, min_size=%d\n", __FUNCTION__, min_pitch, min_size));

	if (size < min_size)
		return false;

	if (pitch < min_pitch)
		return false;

	kgem_get_tile_size(kgem, tiling, min_pitch,
			   &tile_width, &tile_height, &tile_size);

	DBG(("%s: tile_width=%d, tile_size=%d\n", __FUNCTION__, tile_width, tile_size));
	if (pitch & (tile_width - 1))
		return false;
	if (size & (tile_size - 1))
		return false;

	return true;
}

static uint32_t kgem_aligned_height(struct kgem *kgem,
				    uint32_t height, uint32_t tiling)
{
	uint32_t tile_height;

	if (kgem->gen <= 030) {
		tile_height = tiling ? kgem->gen < 030 ? 16 : 8 : 1;
	} else switch (tiling) {
		/* XXX align to an even tile row */
	default:
	case I915_TILING_NONE:
		tile_height = 1;
		break;
	case I915_TILING_X:
		tile_height = 8;
		break;
	case I915_TILING_Y:
		tile_height = 32;
		break;
	}

	/* XXX align to an even tile row */
	if (!kgem->has_relaxed_fencing)
		tile_height *= 2;

	return ALIGN(height, tile_height);
}

static struct drm_i915_gem_exec_object2 *
kgem_add_handle(struct kgem *kgem, struct kgem_bo *bo)
{
	struct drm_i915_gem_exec_object2 *exec;

	DBG(("%s: handle=%d, index=%d\n",
	     __FUNCTION__, bo->handle, kgem->nexec));

	assert(kgem->nexec < ARRAY_SIZE(kgem->exec));
	bo->target_handle = kgem->has_handle_lut ? kgem->nexec : bo->handle;
	exec = memset(&kgem->exec[kgem->nexec++], 0, sizeof(*exec));
	exec->handle = bo->handle;
	exec->offset = bo->presumed_offset;

	kgem->aperture += num_pages(bo);

	return exec;
}

static void kgem_add_bo(struct kgem *kgem, struct kgem_bo *bo)
{
	assert(bo->refcnt);
	assert(bo->proxy == NULL);

	bo->exec = kgem_add_handle(kgem, bo);
	bo->rq = MAKE_REQUEST(kgem->next_request, kgem->ring);

	list_move_tail(&bo->request, &kgem->next_request->buffers);
	if (bo->io && !list_is_empty(&bo->list))
		list_move(&bo->list, &kgem->batch_buffers);

	/* XXX is it worth working around gcc here? */
	kgem->flush |= bo->flush;
}

static void kgem_clear_swctrl(struct kgem *kgem)
{
	uint32_t *b;

	if (kgem->bcs_state == 0)
		return;

	DBG(("%s: clearin SWCTRL LRI from %x\n",
	     __FUNCTION__, kgem->bcs_state));

	b = kgem->batch + kgem->nbatch;
	kgem->nbatch += 7;

	*b++ = MI_FLUSH_DW;
	*b++ = 0;
	*b++ = 0;
	*b++ = 0;

	*b++ = MI_LOAD_REGISTER_IMM;
	*b++ = BCS_SWCTRL;
	*b++ = (BCS_SRC_Y | BCS_DST_Y) << 16;

	kgem->bcs_state = 0;
}

static uint32_t kgem_end_batch(struct kgem *kgem)
{
	kgem_clear_swctrl(kgem);
	kgem->batch[kgem->nbatch++] = MI_BATCH_BUFFER_END;
	if (kgem->nbatch & 1)
		kgem->batch[kgem->nbatch++] = MI_NOOP;

	return kgem->nbatch;
}

static void kgem_bo_binding_free(struct kgem *kgem, struct kgem_bo *bo)
{
	struct kgem_bo_binding *b;

	b = bo->binding.next;
	while (b) {
		struct kgem_bo_binding *next = b->next;
		free(b);
		b = next;
	}
}

static void kgem_bo_free(struct kgem *kgem, struct kgem_bo *bo)
{
	DBG(("%s: handle=%d, size=%d\n", __FUNCTION__, bo->handle, bytes(bo)));
	assert(bo->refcnt == 0);
	assert(bo->proxy == NULL);
	assert(bo->exec == NULL);
	assert(!bo->snoop || bo->rq == NULL);

#ifdef DEBUG_MEMORY
	kgem->debug_memory.bo_allocs--;
	kgem->debug_memory.bo_bytes -= bytes(bo);
#endif

	kgem_bo_binding_free(kgem, bo);
	kgem_bo_rmfb(kgem, bo);

	if (IS_USER_MAP(bo->map__cpu)) {
		assert(bo->rq == NULL);
		assert(!__kgem_busy(kgem, bo->handle));
		assert(MAP(bo->map__cpu) != bo || bo->io || bo->flush);
		if (!(bo->io || bo->flush)) {
			DBG(("%s: freeing snooped base\n", __FUNCTION__));
			assert(bo != MAP(bo->map__cpu));
			free(MAP(bo->map__cpu));
		}
		bo->map__cpu = NULL;
	}

	DBG(("%s: releasing %p:%p vma for handle=%d, count=%d\n",
	     __FUNCTION__, bo->map__gtt, bo->map__cpu,
	     bo->handle, list_is_empty(&bo->vma) ? 0 : kgem->vma[bo->map__gtt == NULL && bo->map__wc == NULL].count));

	if (!list_is_empty(&bo->vma)) {
		_list_del(&bo->vma);
		kgem->vma[bo->map__gtt == NULL && bo->map__wc == NULL].count--;
	}

	if (bo->map__gtt)
		munmap(bo->map__gtt, bytes(bo));
	if (bo->map__wc) {
		VG(VALGRIND_MAKE_MEM_NOACCESS(bo->map__wc, bytes(bo)));
		munmap(bo->map__wc, bytes(bo));
	}
	if (bo->map__cpu) {
		VG(VALGRIND_MAKE_MEM_NOACCESS(MAP(bo->map__cpu), bytes(bo)));
		munmap(MAP(bo->map__cpu), bytes(bo));
	}

	_list_del(&bo->list);
	_list_del(&bo->request);
	gem_close(kgem->fd, bo->handle);

	if (!bo->io && !DBG_NO_MALLOC_CACHE) {
		*(struct kgem_bo **)bo = __kgem_freed_bo;
		__kgem_freed_bo = bo;
	} else
		free(bo);
}

inline static void kgem_bo_move_to_inactive(struct kgem *kgem,
					    struct kgem_bo *bo)
{
	DBG(("%s: moving handle=%d to inactive\n", __FUNCTION__, bo->handle));

	assert(bo->refcnt == 0);
	assert(bo->reusable);
	assert(bo->rq == NULL);
	assert(bo->exec == NULL);
	assert(bo->domain != DOMAIN_GPU);
	assert(!bo->proxy);
	assert(!bo->io);
	assert(!bo->scanout);
	assert(!bo->snoop);
	assert(!bo->flush);
	assert(!bo->needs_flush);
	assert(!bo->delta);
	assert(list_is_empty(&bo->vma));
	assert_tiling(kgem, bo);
	assert_caching(kgem, bo);
	ASSERT_IDLE(kgem, bo->handle);

	if (bucket(bo) >= NUM_CACHE_BUCKETS) {
		if (bo->map__gtt) {
			DBG(("%s: relinquishing large GTT mapping for handle=%d\n",
			     __FUNCTION__, bo->handle));
			munmap(bo->map__gtt, bytes(bo));
			bo->map__gtt = NULL;
		}

		list_move(&bo->list, &kgem->large_inactive);
	} else {
		assert(bo->flush == false);
		assert(list_is_empty(&bo->vma));
		list_move(&bo->list, &kgem->inactive[bucket(bo)]);
		if (bo->map__gtt && !kgem_bo_can_map(kgem, bo)) {
			DBG(("%s: relinquishing old GTT mapping for handle=%d\n",
			     __FUNCTION__, bo->handle));
			munmap(bo->map__gtt, bytes(bo));
			bo->map__gtt = NULL;
		}
		if (bo->map__gtt || (bo->map__wc && !bo->tiling)) {
			list_add(&bo->vma, &kgem->vma[0].inactive[bucket(bo)]);
			kgem->vma[0].count++;
		}
		if (bo->map__cpu && list_is_empty(&bo->vma)) {
			list_add(&bo->vma, &kgem->vma[1].inactive[bucket(bo)]);
			kgem->vma[1].count++;
		}
	}

	kgem->need_expire = true;
}

static struct kgem_bo *kgem_bo_replace_io(struct kgem_bo *bo)
{
	struct kgem_bo *base;

	if (!bo->io)
		return bo;

	assert(!bo->snoop);
	assert(!bo->purged);
	assert(!bo->scanout);
	assert(!bo->delta);

	if (__kgem_freed_bo) {
		base = __kgem_freed_bo;
		__kgem_freed_bo = *(struct kgem_bo **)base;
	} else
		base = malloc(sizeof(*base));
	if (base) {
		DBG(("%s: transferring io handle=%d to bo\n",
		     __FUNCTION__, bo->handle));
		/* transfer the handle to a minimum bo */
		memcpy(base, bo, sizeof(*base));
		base->io = false;
		list_init(&base->list);
		list_replace(&bo->request, &base->request);
		list_replace(&bo->vma, &base->vma);
		free(bo);
		bo = base;
	} else
		bo->reusable = false;

	return bo;
}

inline static void kgem_bo_remove_from_inactive(struct kgem *kgem,
						struct kgem_bo *bo)
{
	DBG(("%s: removing handle=%d from inactive\n", __FUNCTION__, bo->handle));

	list_del(&bo->list);
	assert(bo->rq == NULL);
	assert(bo->exec == NULL);
	assert(!bo->purged);
	if (!list_is_empty(&bo->vma)) {
		assert(bo->map__gtt || bo->map__wc || bo->map__cpu);
		list_del(&bo->vma);
		kgem->vma[bo->map__gtt == NULL && bo->map__wc == NULL].count--;
	}
}

inline static void kgem_bo_remove_from_active(struct kgem *kgem,
					      struct kgem_bo *bo)
{
	DBG(("%s: removing handle=%d from active\n", __FUNCTION__, bo->handle));

	list_del(&bo->list);
	assert(bo->rq != NULL);
	if (RQ(bo->rq) == (void *)kgem) {
		assert(bo->exec == NULL);
		list_del(&bo->request);
	}
	assert(list_is_empty(&bo->vma));
}

static void _kgem_bo_delete_buffer(struct kgem *kgem, struct kgem_bo *bo)
{
	struct kgem_buffer *io = (struct kgem_buffer *)bo->proxy;

	DBG(("%s: size=%d, offset=%d, parent used=%d\n",
	     __FUNCTION__, bo->size.bytes, bo->delta, io->used));

	if (ALIGN(bo->delta + bo->size.bytes, UPLOAD_ALIGNMENT) == io->used)
		io->used = bo->delta;
}

static bool check_scanout_size(struct kgem *kgem,
			       struct kgem_bo *bo,
			       int width, int height)
{
	struct drm_mode_fb_cmd info;

	assert(bo->scanout);

	VG_CLEAR(info);
	info.fb_id = bo->delta;

	if (do_ioctl(kgem->fd, DRM_IOCTL_MODE_GETFB, &info))
		return false;

	gem_close(kgem->fd, info.handle);

	if (width > info.width || height > info.height) {
		DBG(("%s: not using scanout %d (%dx%d), want (%dx%d)\n",
		     __FUNCTION__,
		     info.fb_id, info.width, info.height,
		     width, height));
		return false;
	}

	return true;
}

static void kgem_bo_move_to_scanout(struct kgem *kgem, struct kgem_bo *bo)
{
	assert(bo->refcnt == 0);
	assert(bo->scanout);
	assert(!bo->flush);
	assert(!bo->snoop);
	assert(!bo->io);

	if (bo->purged) { /* for stolen fb */
		if (!bo->exec) {
			DBG(("%s: discarding purged scanout - stolen?\n",
			     __FUNCTION__));
			kgem_bo_free(kgem, bo);
		}
		return;
	}

	DBG(("%s: moving %d [fb %d] to scanout cache, active? %d\n",
	     __FUNCTION__, bo->handle, bo->delta, bo->rq != NULL));
	if (bo->rq)
		list_move_tail(&bo->list, &kgem->scanout);
	else
		list_move(&bo->list, &kgem->scanout);

	kgem->need_expire = true;
}

static void kgem_bo_move_to_snoop(struct kgem *kgem, struct kgem_bo *bo)
{
	assert(bo->reusable);
	assert(!bo->scanout);
	assert(!bo->flush);
	assert(!bo->needs_flush);
	assert(bo->refcnt == 0);
	assert(bo->exec == NULL);
	assert(!bo->purged);
	assert(!bo->delta);

	if (DBG_NO_SNOOP_CACHE) {
		kgem_bo_free(kgem, bo);
		return;
	}

	if (num_pages(bo) > kgem->max_cpu_size >> 13) {
		DBG(("%s handle=%d discarding large CPU buffer (%d >%d pages)\n",
		     __FUNCTION__, bo->handle, num_pages(bo), kgem->max_cpu_size >> 13));
		kgem_bo_free(kgem, bo);
		return;
	}

	assert(bo->tiling == I915_TILING_NONE);
	assert(bo->rq == NULL);

	DBG(("%s: moving %d to snoop cachee\n", __FUNCTION__, bo->handle));
	list_add(&bo->list, &kgem->snoop);
	kgem->need_expire = true;
}

static bool kgem_bo_move_to_cache(struct kgem *kgem, struct kgem_bo *bo)
{
	bool retired = false;

	DBG(("%s: release handle=%d\n", __FUNCTION__, bo->handle));

	if (bo->prime) {
		DBG(("%s: discarding imported prime handle=%d\n",
		     __FUNCTION__, bo->handle));
		kgem_bo_free(kgem, bo);
	} else if (bo->snoop) {
		kgem_bo_move_to_snoop(kgem, bo);
	} else if (bo->scanout) {
		kgem_bo_move_to_scanout(kgem, bo);
	} else if ((bo = kgem_bo_replace_io(bo))->reusable) {
		kgem_bo_move_to_inactive(kgem, bo);
		retired = true;
	} else
		kgem_bo_free(kgem, bo);

	return retired;
}

static struct kgem_bo *
search_snoop_cache(struct kgem *kgem, unsigned int num_pages, unsigned flags)
{
	struct kgem_bo *bo, *first = NULL;

	DBG(("%s: num_pages=%d, flags=%x\n", __FUNCTION__, num_pages, flags));

	if ((kgem->has_caching | kgem->has_userptr) == 0)
		return NULL;

	if (list_is_empty(&kgem->snoop)) {
		DBG(("%s: inactive and cache empty\n", __FUNCTION__));
		if (!__kgem_throttle_retire(kgem, flags)) {
			DBG(("%s: nothing retired\n", __FUNCTION__));
			return NULL;
		}
	}

	list_for_each_entry(bo, &kgem->snoop, list) {
		assert(bo->refcnt == 0);
		assert(bo->snoop);
		assert(!bo->scanout);
		assert(!bo->purged);
		assert(bo->proxy == NULL);
		assert(bo->tiling == I915_TILING_NONE);
		assert(bo->rq == NULL);
		assert(bo->exec == NULL);

		if (num_pages > num_pages(bo))
			continue;

		if (num_pages(bo) > 2*num_pages) {
			if (first == NULL)
				first = bo;
			continue;
		}

		list_del(&bo->list);
		bo->pitch = 0;
		bo->delta = 0;

		DBG(("  %s: found handle=%d (num_pages=%d) in snoop cache\n",
		     __FUNCTION__, bo->handle, num_pages(bo)));
		return bo;
	}

	if (first) {
		list_del(&first->list);
		first->pitch = 0;
		first->delta = 0;

		DBG(("  %s: found handle=%d (num_pages=%d) in snoop cache\n",
		     __FUNCTION__, first->handle, num_pages(first)));
		return first;
	}

	return NULL;
}

void kgem_bo_undo(struct kgem *kgem, struct kgem_bo *bo)
{
	if (kgem->nexec != 1 || bo->exec == NULL)
		return;

	assert(bo);
	DBG(("%s: only handle in batch, discarding last operations for handle=%d\n",
	     __FUNCTION__, bo->handle));

	assert(bo->exec == &_kgem_dummy_exec || bo->exec == &kgem->exec[0]);
	assert(kgem->exec[0].handle == bo->handle);
	assert(RQ(bo->rq) == kgem->next_request);

	bo->refcnt++;
	kgem_reset(kgem);
	bo->refcnt--;

	assert(kgem->nreloc == 0);
	assert(kgem->nexec == 0);
	assert(bo->exec == NULL);
}

void kgem_bo_pair_undo(struct kgem *kgem, struct kgem_bo *a, struct kgem_bo *b)
{
	if (kgem->nexec > 2)
		return;

	if (kgem->nexec == 1) {
		if (a)
			kgem_bo_undo(kgem, a);
		if (b)
			kgem_bo_undo(kgem, b);
		return;
	}

	if (a == NULL || b == NULL)
		return;
	assert(a != b);
	if (a->exec == NULL || b->exec == NULL)
		return;

	DBG(("%s: only handles in batch, discarding last operations for handle=%d (index=%d) and handle=%d (index=%d)\n",
	     __FUNCTION__,
	     a->handle, a->proxy ? -1 : a->exec - kgem->exec,
	     b->handle, b->proxy ? -1 : b->exec - kgem->exec));

	assert(a->exec == &_kgem_dummy_exec ||
	       a->exec == &kgem->exec[0] ||
	       a->exec == &kgem->exec[1]);
	assert(a->handle == kgem->exec[0].handle || a->handle == kgem->exec[1].handle);
	assert(RQ(a->rq) == kgem->next_request);
	assert(b->exec == &_kgem_dummy_exec ||
	       b->exec == &kgem->exec[0] ||
	       b->exec == &kgem->exec[1]);
	assert(b->handle == kgem->exec[0].handle || b->handle == kgem->exec[1].handle);
	assert(RQ(b->rq) == kgem->next_request);

	a->refcnt++;
	b->refcnt++;
	kgem_reset(kgem);
	b->refcnt--;
	a->refcnt--;

	assert(kgem->nreloc == 0);
	assert(kgem->nexec == 0);
	assert(a->exec == NULL);
	assert(b->exec == NULL);
}

static void __kgem_bo_destroy(struct kgem *kgem, struct kgem_bo *bo)
{
	DBG(("%s: handle=%d, size=%d\n", __FUNCTION__, bo->handle, bytes(bo)));

	assert(list_is_empty(&bo->list));
	assert(list_is_empty(&bo->vma));
	assert(bo->refcnt == 0);
	assert(bo->proxy == NULL);
	assert(bo->active_scanout == 0);
	assert_tiling(kgem, bo);

	bo->binding.offset = 0;

	if (DBG_NO_CACHE)
		goto destroy;

	if (bo->prime)
		goto destroy;

	if (bo->snoop && !bo->flush) {
		DBG(("%s: handle=%d is snooped\n", __FUNCTION__, bo->handle));
		assert(bo->reusable);
		assert(list_is_empty(&bo->list));
		if (bo->exec == NULL && bo->rq && !__kgem_busy(kgem, bo->handle))
			__kgem_bo_clear_busy(bo);
		if (bo->rq == NULL)
			kgem_bo_move_to_snoop(kgem, bo);
		return;
	}
	if (!IS_USER_MAP(bo->map__cpu))
		bo->flush = false;

	if (bo->scanout) {
		kgem_bo_move_to_scanout(kgem, bo);
		return;
	}

	if (bo->io)
		bo = kgem_bo_replace_io(bo);
	if (!bo->reusable) {
		DBG(("%s: handle=%d, not reusable\n",
		     __FUNCTION__, bo->handle));
		goto destroy;
	}

	assert(list_is_empty(&bo->vma));
	assert(list_is_empty(&bo->list));
	assert(bo->flush == false);
	assert(bo->snoop == false);
	assert(bo->io == false);
	assert(bo->scanout == false);
	assert_caching(kgem, bo);

	kgem_bo_undo(kgem, bo);
	assert(bo->refcnt == 0);

	if (bo->rq && bo->exec == NULL && !__kgem_busy(kgem, bo->handle))
		__kgem_bo_clear_busy(bo);

	if (bo->rq) {
		struct list *cache;

		DBG(("%s: handle=%d -> active\n", __FUNCTION__, bo->handle));
		if (bucket(bo) < NUM_CACHE_BUCKETS)
			cache = &kgem->active[bucket(bo)][bo->tiling];
		else
			cache = &kgem->large;
		list_add(&bo->list, cache);
		return;
	}

	assert(bo->exec == NULL);
	assert(list_is_empty(&bo->request));

	if (bo->map__cpu == NULL || bucket(bo) >= NUM_CACHE_BUCKETS) {
		if (!kgem->has_llc && bo->domain == DOMAIN_CPU)
			goto destroy;

		DBG(("%s: handle=%d, purged\n",
		     __FUNCTION__, bo->handle));
	}

	kgem_bo_move_to_inactive(kgem, bo);
	return;

destroy:
	if (!bo->exec)
		kgem_bo_free(kgem, bo);
}

static void kgem_bo_unref(struct kgem *kgem, struct kgem_bo *bo)
{
	assert(bo->refcnt);
	if (--bo->refcnt == 0)
		__kgem_bo_destroy(kgem, bo);
}

static void kgem_buffer_release(struct kgem *kgem, struct kgem_buffer *bo)
{
	assert(bo->base.io);
	while (!list_is_empty(&bo->base.vma)) {
		struct kgem_bo *cached;

		cached = list_first_entry(&bo->base.vma, struct kgem_bo, vma);
		assert(cached->proxy == &bo->base);
		assert(cached != &bo->base);
		list_del(&cached->vma);

		assert(*(struct kgem_bo **)cached->map__gtt == cached);
		*(struct kgem_bo **)cached->map__gtt = NULL;
		cached->map__gtt = NULL;

		kgem_bo_destroy(kgem, cached);
	}
}

void kgem_retire__buffers(struct kgem *kgem)
{
	while (!list_is_empty(&kgem->active_buffers)) {
		struct kgem_buffer *bo =
			list_last_entry(&kgem->active_buffers,
					struct kgem_buffer,
					base.list);

		DBG(("%s: handle=%d, busy? %d [%d]\n",
		     __FUNCTION__, bo->base.handle, bo->base.rq != NULL, bo->base.exec != NULL));

		assert(bo->base.exec == NULL || RQ(bo->base.rq) == kgem->next_request);
		if (bo->base.rq)
			break;

		DBG(("%s: releasing upload cache for handle=%d? %d\n",
		     __FUNCTION__, bo->base.handle, !list_is_empty(&bo->base.vma)));
		list_del(&bo->base.list);
		kgem_buffer_release(kgem, bo);
		kgem_bo_unref(kgem, &bo->base);
	}
}

static bool kgem_retire__flushing(struct kgem *kgem)
{
	struct kgem_bo *bo, *next;
	bool retired = false;

	list_for_each_entry_safe(bo, next, &kgem->flushing, request) {
		assert(RQ(bo->rq) == (void *)kgem);
		assert(bo->exec == NULL);

		if (__kgem_busy(kgem, bo->handle))
			break;

		__kgem_bo_clear_busy(bo);

		if (bo->refcnt)
			continue;

		retired |= kgem_bo_move_to_cache(kgem, bo);
	}
#if HAS_DEBUG_FULL
	{
		int count = 0;
		list_for_each_entry(bo, &kgem->flushing, request)
			count++;
		DBG(("%s: %d bo on flushing list, retired? %d\n", __FUNCTION__, count, retired));
	}
#endif

	kgem->need_retire |= !list_is_empty(&kgem->flushing);

	return retired;
}

static bool __kgem_bo_flush(struct kgem *kgem, struct kgem_bo *bo)
{
	struct drm_i915_gem_busy busy;

	if (!bo->needs_flush)
		return false;

	bo->needs_flush = false;

	VG_CLEAR(busy);
	busy.handle = bo->handle;
	busy.busy = !kgem->wedged;
	(void)do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_BUSY, &busy);
	DBG(("%s: handle=%d, busy=%d, wedged=%d\n",
	     __FUNCTION__, bo->handle, busy.busy, kgem->wedged));

	if (busy.busy == 0)
		return false;

	DBG(("%s: moving %d to flushing\n",
	     __FUNCTION__, bo->handle));
	list_add(&bo->request, &kgem->flushing);
	bo->rq = MAKE_REQUEST(kgem, !!(busy.busy & ~0x1ffff));
	bo->needs_flush = busy.busy & 0xffff;
	kgem->need_retire = true;
	return true;
}

static bool __kgem_retire_rq(struct kgem *kgem, struct kgem_request *rq)
{
	bool retired = false;

	DBG(("%s: request %d complete\n",
	     __FUNCTION__, rq->bo->handle));
	assert(RQ(rq->bo->rq) == rq);
	assert(rq != (struct kgem_request *)kgem);
	assert(rq != &kgem->static_request);

	if (rq == kgem->fence[rq->ring])
		kgem->fence[rq->ring] = NULL;

	while (!list_is_empty(&rq->buffers)) {
		struct kgem_bo *bo;

		bo = list_first_entry(&rq->buffers,
				      struct kgem_bo,
				      request);

		assert(RQ(bo->rq) == rq);
		assert(bo->exec == NULL);
		assert(bo->domain == DOMAIN_GPU || bo->domain == DOMAIN_NONE);

		list_del(&bo->request);

		if (unlikely(__kgem_bo_flush(kgem, bo))) {
			assert(bo != rq->bo);
			DBG(("%s: movied %d to flushing\n",
			     __FUNCTION__, bo->handle));
			continue;
		}

		bo->domain = DOMAIN_NONE;
		bo->rq = NULL;
		if (bo->refcnt)
			continue;

		retired |= kgem_bo_move_to_cache(kgem, bo);
	}

	assert(rq->bo->rq == NULL);
	assert(rq->bo->exec == NULL);
	assert(list_is_empty(&rq->bo->request));
	assert(rq->bo->refcnt > 0);

	if (--rq->bo->refcnt == 0) {
		kgem_bo_move_to_inactive(kgem, rq->bo);
		retired = true;
	}

	__kgem_request_free(rq);
	return retired;
}

static bool kgem_retire__requests_ring(struct kgem *kgem, int ring)
{
	bool retired = false;

	assert(ring < ARRAY_SIZE(kgem->requests));
	while (!list_is_empty(&kgem->requests[ring])) {
		struct kgem_request *rq;

		DBG(("%s: retiring ring %d\n", __FUNCTION__, ring));

		rq = list_first_entry(&kgem->requests[ring],
				      struct kgem_request,
				      list);
		assert(rq->ring == ring);
		assert(rq->bo);
		assert(RQ(rq->bo->rq) == rq);
		if (__kgem_busy(kgem, rq->bo->handle))
			break;

		retired |= __kgem_retire_rq(kgem, rq);
	}

#if HAS_DEBUG_FULL
	{
		struct kgem_bo *bo;
		int count = 0;

		list_for_each_entry(bo, &kgem->requests[ring], request)
			count++;

		bo = NULL;
		if (!list_is_empty(&kgem->requests[ring]))
			bo = list_first_entry(&kgem->requests[ring],
					      struct kgem_request,
					      list)->bo;

		DBG(("%s: ring=%d, %d outstanding requests, oldest=%d, retired? %d\n",
		     __FUNCTION__, ring, count, bo ? bo->handle : 0, retired));
	}
#endif

	return retired;
}

static bool kgem_retire__requests(struct kgem *kgem)
{
	bool retired = false;
	int n;

	for (n = 0; n < ARRAY_SIZE(kgem->requests); n++) {
		retired |= kgem_retire__requests_ring(kgem, n);
		kgem->need_retire |= !list_is_empty(&kgem->requests[n]);
	}

	return retired;
}

bool kgem_retire(struct kgem *kgem)
{
	bool retired = false;

	DBG(("%s, need_retire?=%d\n", __FUNCTION__, kgem->need_retire));

	kgem->need_retire = false;

	retired |= kgem_retire__flushing(kgem);
	retired |= kgem_retire__requests(kgem);

	DBG(("%s -- retired=%d, need_retire=%d\n",
	     __FUNCTION__, retired, kgem->need_retire));

	kgem->retire(kgem);

	return retired;
}

bool __kgem_ring_is_idle(struct kgem *kgem, int ring)
{
	struct kgem_request *rq;

	assert(ring < ARRAY_SIZE(kgem->requests));
	assert(!list_is_empty(&kgem->requests[ring]));

	rq = kgem->fence[ring];
	if (rq) {
		struct kgem_request *tmp;

		if (__kgem_busy(kgem, rq->bo->handle)) {
			DBG(("%s: last fence handle=%d still busy\n",
			     __FUNCTION__, rq->bo->handle));
			return false;
		}

		do {
			tmp = list_first_entry(&kgem->requests[ring],
					       struct kgem_request,
					       list);
			assert(tmp->ring == ring);
			__kgem_retire_rq(kgem, tmp);
		} while (tmp != rq);

		assert(kgem->fence[ring] == NULL);
		if (list_is_empty(&kgem->requests[ring]))
			return true;
	}

	rq = list_last_entry(&kgem->requests[ring],
			     struct kgem_request, list);
	assert(rq->ring == ring);
	assert(rq->bo);
	assert(RQ(rq->bo->rq) == rq);
	if (__kgem_busy(kgem, rq->bo->handle)) {
		DBG(("%s: last requests handle=%d still busy\n",
		     __FUNCTION__, rq->bo->handle));
		kgem->fence[ring] = rq;
		return false;
	}

	DBG(("%s: ring=%d idle (handle=%d)\n",
	     __FUNCTION__, ring, rq->bo->handle));

	while (!list_is_empty(&kgem->requests[ring])) {
		rq = list_first_entry(&kgem->requests[ring],
				      struct kgem_request,
				      list);
		assert(rq->ring == ring);
		__kgem_retire_rq(kgem, rq);
	}

	return true;
}

bool __kgem_retire_requests_upto(struct kgem *kgem, struct kgem_bo *bo)
{
	struct kgem_request * const rq = RQ(bo->rq), *tmp;
	struct list *requests = &kgem->requests[rq->ring];

	DBG(("%s(handle=%d, ring=%d)\n", __FUNCTION__, bo->handle, rq->ring));

	assert(rq != &kgem->static_request);
	if (rq == (struct kgem_request *)kgem) {
		__kgem_bo_clear_busy(bo);
		return false;
	}

	assert(rq->ring < ARRAY_SIZE(kgem->requests));
	do {
		tmp = list_first_entry(requests, struct kgem_request, list);
		assert(tmp->ring == rq->ring);
		__kgem_retire_rq(kgem, tmp);
	} while (tmp != rq);

	assert(bo->needs_flush || bo->rq == NULL);
	assert(bo->needs_flush || list_is_empty(&bo->request));
	assert(bo->needs_flush || bo->domain == DOMAIN_NONE);
	return bo->rq;
}

#if 0
static void kgem_commit__check_reloc(struct kgem *kgem)
{
	struct kgem_request *rq = kgem->next_request;
	struct kgem_bo *bo;
	bool has_64bit = kgem->gen >= 0100;
	int i;

	for (i = 0; i < kgem->nreloc; i++) {
		list_for_each_entry(bo, &rq->buffers, request) {
			if (bo->target_handle == kgem->reloc[i].target_handle) {
				uint64_t value = 0;
				gem_read(kgem->fd, rq->bo->handle, &value, kgem->reloc[i].offset, has_64bit ? 8 : 4);
				assert(bo->exec->offset == -1 || value == bo->exec->offset + (int)kgem->reloc[i].delta);
				break;
			}
		}
	}
}
#else
#define kgem_commit__check_reloc(kgem)
#endif

#ifndef NDEBUG
static void kgem_commit__check_buffers(struct kgem *kgem)
{
	struct kgem_buffer *bo;

	list_for_each_entry(bo, &kgem->active_buffers, base.list)
		assert(bo->base.exec == NULL);
}
#else
#define kgem_commit__check_buffers(kgem)
#endif

static void kgem_commit(struct kgem *kgem)
{
	struct kgem_request *rq = kgem->next_request;
	struct kgem_bo *bo, *next;

	kgem_commit__check_reloc(kgem);

	list_for_each_entry_safe(bo, next, &rq->buffers, request) {
		assert(next->request.prev == &bo->request);

		DBG(("%s: release handle=%d (proxy? %d), dirty? %d flush? %d, snoop? %d -> offset=%x\n",
		     __FUNCTION__, bo->handle, bo->proxy != NULL,
		     bo->gpu_dirty, bo->needs_flush, bo->snoop,
		     (unsigned)bo->exec->offset));

		assert(bo->exec);
		assert(bo->proxy == NULL || bo->exec == &_kgem_dummy_exec);
		assert(RQ(bo->rq) == rq || (RQ(bo->proxy->rq) == rq));

		bo->presumed_offset = bo->exec->offset;
		bo->exec = NULL;
		bo->target_handle = -1;

		if (!bo->refcnt && !bo->reusable) {
			assert(!bo->snoop);
			assert(!bo->proxy);
			kgem_bo_free(kgem, bo);
			continue;
		}

		bo->binding.offset = 0;
		bo->domain = DOMAIN_GPU;
		bo->gpu_dirty = false;
		bo->gtt_dirty = false;

		if (bo->proxy) {
			/* proxies are not used for domain tracking */
			__kgem_bo_clear_busy(bo);
		}

		kgem->scanout_busy |= bo->scanout && bo->needs_flush;
	}

	if (rq == &kgem->static_request) {
		struct drm_i915_gem_set_domain set_domain;

		DBG(("%s: syncing due to allocation failure\n", __FUNCTION__));

		VG_CLEAR(set_domain);
		set_domain.handle = rq->bo->handle;
		set_domain.read_domains = I915_GEM_DOMAIN_GTT;
		set_domain.write_domain = I915_GEM_DOMAIN_GTT;
		if (do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain)) {
			DBG(("%s: sync: GPU hang detected\n", __FUNCTION__));
			kgem_throttle(kgem);
		}

		while (!list_is_empty(&rq->buffers)) {
			bo = list_first_entry(&rq->buffers,
					      struct kgem_bo,
					      request);

			assert(RQ(bo->rq) == rq);
			assert(bo->exec == NULL);
			assert(bo->domain == DOMAIN_GPU);

			list_del(&bo->request);
			bo->domain = DOMAIN_NONE;
			bo->rq = NULL;

			if (bo->refcnt == 0)
				_kgem_bo_destroy(kgem, bo);
		}

		kgem_retire(kgem);
		assert(list_is_empty(&rq->buffers));

		assert(rq->bo->map__gtt == NULL);
		assert(rq->bo->map__wc == NULL);
		assert(rq->bo->map__cpu == NULL);
		gem_close(kgem->fd, rq->bo->handle);
		kgem_cleanup_cache(kgem);
	} else {
		assert(rq != (struct kgem_request *)kgem);
		assert(rq->ring < ARRAY_SIZE(kgem->requests));
		assert(rq->bo);
		list_add_tail(&rq->list, &kgem->requests[rq->ring]);
		kgem->need_throttle = kgem->need_retire = 1;

		if (kgem->fence[rq->ring] == NULL &&
		    __kgem_busy(kgem, rq->bo->handle))
			kgem->fence[rq->ring] = rq;
	}

	kgem->next_request = NULL;

	kgem_commit__check_buffers(kgem);
}

static void kgem_close_list(struct kgem *kgem, struct list *head)
{
	while (!list_is_empty(head))
		kgem_bo_free(kgem, list_first_entry(head, struct kgem_bo, list));
}

static void kgem_close_inactive(struct kgem *kgem)
{
	unsigned int i;

	for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) {
		kgem_close_list(kgem, &kgem->inactive[i]);
		assert(list_is_empty(&kgem->inactive[i]));
	}
}

static void kgem_finish_buffers(struct kgem *kgem)
{
	struct kgem_buffer *bo, *next;

	list_for_each_entry_safe(bo, next, &kgem->batch_buffers, base.list) {
		DBG(("%s: buffer handle=%d, used=%d, exec?=%d, write=%d, mmapped=%s, refcnt=%d\n",
		     __FUNCTION__, bo->base.handle, bo->used, bo->base.exec!=NULL,
		     bo->write, bo->mmapped == MMAPPED_CPU ? "cpu" : bo->mmapped == MMAPPED_GTT ? "gtt" : "no",
		     bo->base.refcnt));

		assert(next->base.list.prev == &bo->base.list);
		assert(bo->base.io);
		assert(bo->base.refcnt >= 1);

		if (bo->base.refcnt > 1 && !bo->base.exec) {
			DBG(("%s: skipping unattached handle=%d, used=%d, refcnt=%d\n",
			     __FUNCTION__, bo->base.handle, bo->used, bo->base.refcnt));
			continue;
		}

		if (!bo->write) {
			assert(bo->base.exec || bo->base.refcnt > 1);
			goto decouple;
		}

		if (bo->mmapped) {
			uint32_t used;

			assert(!bo->need_io);

			used = ALIGN(bo->used, PAGE_SIZE);
			if (!DBG_NO_UPLOAD_ACTIVE &&
			    used + PAGE_SIZE <= bytes(&bo->base) &&
			    (kgem->has_llc || bo->mmapped == MMAPPED_GTT || bo->base.snoop)) {
				DBG(("%s: retaining upload buffer (%d/%d): used=%d, refcnt=%d\n",
				     __FUNCTION__, bo->used, bytes(&bo->base), used, bo->base.refcnt));
				bo->used = used;
				list_move(&bo->base.list,
					  &kgem->active_buffers);
				kgem->need_retire = true;
				continue;
			}
			DBG(("%s: discarding mmapped buffer, used=%d, map type=%d\n",
			     __FUNCTION__, bo->used, bo->mmapped));
			goto decouple;
		}

		if (!bo->used || !bo->base.exec) {
			/* Unless we replace the handle in the execbuffer,
			 * then this bo will become active. So decouple it
			 * from the buffer list and track it in the normal
			 * manner.
			 */
			goto decouple;
		}

		assert(bo->need_io);
		assert(bo->base.rq == MAKE_REQUEST(kgem->next_request, kgem->ring));
		assert(bo->base.domain != DOMAIN_GPU);

		if (bo->base.refcnt == 1 &&
		    bo->base.size.pages.count > 1 &&
		    bo->used < bytes(&bo->base) / 2) {
			struct kgem_bo *shrink;
			unsigned alloc = NUM_PAGES(bo->used);

			shrink = search_snoop_cache(kgem, alloc,
						    CREATE_INACTIVE | CREATE_NO_RETIRE);
			if (shrink) {
				void *map;
				int n;

				DBG(("%s: used=%d, shrinking %d to %d, handle %d to %d\n",
				     __FUNCTION__,
				     bo->used, bytes(&bo->base), bytes(shrink),
				     bo->base.handle, shrink->handle));

				assert(bo->used <= bytes(shrink));
				map = kgem_bo_map__cpu(kgem, shrink);
				if (map) {
					memcpy(map, bo->mem, bo->used);

					shrink->target_handle =
						kgem->has_handle_lut ? bo->base.target_handle : shrink->handle;
					for (n = 0; n < kgem->nreloc; n++) {
						if (kgem->reloc[n].target_handle == bo->base.target_handle) {
							uint64_t addr = (int)kgem->reloc[n].delta + shrink->presumed_offset;
							kgem->batch[kgem->reloc[n].offset/sizeof(kgem->batch[0])] = addr;
							if (kgem->gen >= 0100)
								kgem->batch[kgem->reloc[n].offset/sizeof(kgem->batch[0]) + 1] = addr >> 32;

							kgem->reloc[n].target_handle = shrink->target_handle;
							kgem->reloc[n].presumed_offset = shrink->presumed_offset;
						}
					}

					bo->base.exec->handle = shrink->handle;
					bo->base.exec->offset = shrink->presumed_offset;
					shrink->exec = bo->base.exec;
					shrink->rq = bo->base.rq;
					list_replace(&bo->base.request,
						     &shrink->request);
					list_init(&bo->base.request);
					shrink->needs_flush = bo->base.gpu_dirty;

					bo->base.exec = NULL;
					bo->base.rq = NULL;
					bo->base.gpu_dirty = false;
					bo->base.needs_flush = false;
					bo->used = 0;

					goto decouple;
				}

				__kgem_bo_destroy(kgem, shrink);
			}

			shrink = search_linear_cache(kgem, alloc,
						     CREATE_INACTIVE | CREATE_NO_RETIRE);
			if (shrink) {
				int n;

				DBG(("%s: used=%d, shrinking %d to %d, handle %d to %d\n",
				     __FUNCTION__,
				     bo->used, bytes(&bo->base), bytes(shrink),
				     bo->base.handle, shrink->handle));

				assert(bo->used <= bytes(shrink));
				if (gem_write__cachealigned(kgem->fd, shrink->handle,
							    0, bo->used, bo->mem) == 0) {
					shrink->target_handle =
						kgem->has_handle_lut ? bo->base.target_handle : shrink->handle;
					for (n = 0; n < kgem->nreloc; n++) {
						if (kgem->reloc[n].target_handle == bo->base.target_handle) {
							uint64_t addr = (int)kgem->reloc[n].delta + shrink->presumed_offset;
							kgem->batch[kgem->reloc[n].offset/sizeof(kgem->batch[0])] = addr;
							if (kgem->gen >= 0100)
								kgem->batch[kgem->reloc[n].offset/sizeof(kgem->batch[0]) + 1] = addr >> 32;

							kgem->reloc[n].target_handle = shrink->target_handle;
							kgem->reloc[n].presumed_offset = shrink->presumed_offset;
						}
					}

					bo->base.exec->handle = shrink->handle;
					bo->base.exec->offset = shrink->presumed_offset;
					shrink->exec = bo->base.exec;
					shrink->rq = bo->base.rq;
					list_replace(&bo->base.request,
						     &shrink->request);
					list_init(&bo->base.request);
					shrink->needs_flush = bo->base.gpu_dirty;

					bo->base.exec = NULL;
					bo->base.rq = NULL;
					bo->base.gpu_dirty = false;
					bo->base.needs_flush = false;
					bo->used = 0;

					goto decouple;
				}

				__kgem_bo_destroy(kgem, shrink);
			}
		}

		DBG(("%s: handle=%d, uploading %d/%d\n",
		     __FUNCTION__, bo->base.handle, bo->used, bytes(&bo->base)));
		ASSERT_IDLE(kgem, bo->base.handle);
		assert(bo->used <= bytes(&bo->base));
		gem_write__cachealigned(kgem->fd, bo->base.handle,
					0, bo->used, bo->mem);
		bo->need_io = 0;

decouple:
		DBG(("%s: releasing handle=%d\n",
		     __FUNCTION__, bo->base.handle));
		list_del(&bo->base.list);
		kgem_bo_unref(kgem, &bo->base);
	}
}

static void kgem_cleanup(struct kgem *kgem)
{
	int n;

	for (n = 0; n < ARRAY_SIZE(kgem->requests); n++) {
		while (!list_is_empty(&kgem->requests[n])) {
			struct kgem_request *rq;

			rq = list_first_entry(&kgem->requests[n],
					      struct kgem_request,
					      list);
			assert(rq->ring == n);
			while (!list_is_empty(&rq->buffers)) {
				struct kgem_bo *bo;

				bo = list_first_entry(&rq->buffers,
						      struct kgem_bo,
						      request);

				bo->exec = NULL;
				bo->gpu_dirty = false;
				__kgem_bo_clear_busy(bo);
				if (bo->refcnt == 0)
					kgem_bo_free(kgem, bo);
			}

			if (--rq->bo->refcnt == 0)
				kgem_bo_free(kgem, rq->bo);

			__kgem_request_free(rq);
		}
	}

	kgem_close_inactive(kgem);
}

static int
kgem_batch_write(struct kgem *kgem,
		 struct kgem_bo *bo,
		 uint32_t size)
{
	char *ptr;
	int ret;

	assert(bo->exec == NULL);
	assert(bo->rq == NULL);
	assert(!__kgem_busy(kgem, bo->handle));

#if DBG_NO_EXEC
	{
		uint32_t batch[] = { MI_BATCH_BUFFER_END, 0};
		return gem_write(kgem->fd, bo->handle, 0, sizeof(batch), batch);
	}
#endif

	assert(!bo->scanout);
retry:
	ptr = NULL;
	if (bo->domain == DOMAIN_CPU || kgem->has_llc) {
		ptr = bo->map__cpu;
		if (ptr == NULL)
			ptr = __kgem_bo_map__cpu(kgem, bo);
	} else if (kgem->has_wc_mmap) {
		ptr = bo->map__wc;
		if (ptr == NULL)
			ptr = __kgem_bo_map__wc(kgem, bo);
	}
	if (ptr) {
		memcpy(ptr, kgem->batch, sizeof(uint32_t)*kgem->nbatch);
		if (kgem->surface != kgem->batch_size) {
			ret = PAGE_ALIGN(sizeof(uint32_t) * kgem->batch_size);
			ret -= sizeof(uint32_t) * kgem->surface;
			ptr += size - ret;
			memcpy(ptr, kgem->batch + kgem->surface,
			       (kgem->batch_size - kgem->surface)*sizeof(uint32_t));
		}
		return 0;
	}

	/* If there is no surface data, just upload the batch */
	if (kgem->surface == kgem->batch_size) {
		if ((ret = gem_write__cachealigned(kgem->fd, bo->handle,
						   0, sizeof(uint32_t)*kgem->nbatch,
						   kgem->batch)) == 0)
			return 0;

		goto expire;
	}

	/* Are the batch pages conjoint with the surface pages? */
	if (kgem->surface < kgem->nbatch + PAGE_SIZE/sizeof(uint32_t)) {
		assert(size == PAGE_ALIGN(kgem->batch_size*sizeof(uint32_t)));
		if ((ret = gem_write__cachealigned(kgem->fd, bo->handle,
						   0, kgem->batch_size*sizeof(uint32_t),
						   kgem->batch)) == 0)
			return 0;

		goto expire;
	}

	/* Disjoint surface/batch, upload separately */
	if ((ret = gem_write__cachealigned(kgem->fd, bo->handle,
					   0, sizeof(uint32_t)*kgem->nbatch,
					   kgem->batch)))
		goto expire;

	ret = PAGE_ALIGN(sizeof(uint32_t) * kgem->batch_size);
	ret -= sizeof(uint32_t) * kgem->surface;
	assert(size-ret >= kgem->nbatch*sizeof(uint32_t));
	if (gem_write(kgem->fd, bo->handle,
		      size - ret, (kgem->batch_size - kgem->surface)*sizeof(uint32_t),
		      kgem->batch + kgem->surface))
		goto expire;

	return 0;

expire:
	assert(ret != EINVAL);

	(void)__kgem_throttle_retire(kgem, 0);
	if (kgem_expire_cache(kgem))
		goto retry;

	if (kgem_cleanup_cache(kgem))
		goto retry;

	ERR(("%s: failed to write batch (handle=%d): %d\n",
	     __FUNCTION__, bo->handle, -ret));
	return ret;
}

void kgem_reset(struct kgem *kgem)
{
	if (kgem->next_request) {
		struct kgem_request *rq = kgem->next_request;

		while (!list_is_empty(&rq->buffers)) {
			struct kgem_bo *bo =
				list_first_entry(&rq->buffers,
						 struct kgem_bo,
						 request);
			list_del(&bo->request);

			assert(RQ(bo->rq) == rq);

			bo->binding.offset = 0;
			bo->exec = NULL;
			bo->target_handle = -1;
			bo->gpu_dirty = false;

			if (bo->needs_flush && __kgem_busy(kgem, bo->handle)) {
				assert(bo->domain == DOMAIN_GPU || bo->domain == DOMAIN_NONE);
				list_add(&bo->request, &kgem->flushing);
				bo->rq = (void *)kgem;
				kgem->need_retire = true;
			} else
				__kgem_bo_clear_busy(bo);

			if (bo->refcnt || bo->rq)
				continue;

			kgem_bo_move_to_cache(kgem, bo);
		}

		if (rq != &kgem->static_request) {
			list_init(&rq->list);
			__kgem_request_free(rq);
		}
	}

	kgem->nfence = 0;
	kgem->nexec = 0;
	kgem->nreloc = 0;
	kgem->nreloc__self = 0;
	kgem->aperture = 0;
	kgem->aperture_fenced = 0;
	kgem->aperture_max_fence = 0;
	kgem->nbatch = 0;
	kgem->surface = kgem->batch_size;
	kgem->mode = KGEM_NONE;
	kgem->needs_semaphore = false;
	kgem->needs_reservation = false;
	kgem->flush = 0;
	kgem->batch_flags = kgem->batch_flags_base;
	assert(kgem->batch);

	kgem->next_request = __kgem_request_alloc(kgem);

	kgem_sna_reset(kgem);
}

static int compact_batch_surface(struct kgem *kgem, int *shrink)
{
	int size, n;

	if (!kgem->has_relaxed_delta)
		return kgem->batch_size * sizeof(uint32_t);

	/* See if we can pack the contents into one or two pages */
	n = ALIGN(kgem->batch_size, 1024);
	size = n - kgem->surface + kgem->nbatch;
	size = ALIGN(size, 1024);

	*shrink = (n - size) * sizeof(uint32_t);
	return size * sizeof(uint32_t);
}

static struct kgem_bo *first_available(struct kgem *kgem, struct list *list)
{
	struct kgem_bo *bo;

	list_for_each_entry(bo, list, list) {
		assert(bo->refcnt > 0);

		if (bo->rq) {
			assert(RQ(bo->rq)->bo == bo);
			if (__kgem_busy(kgem, bo->handle))
				break;

			__kgem_retire_rq(kgem, RQ(bo->rq));
			assert(bo->rq == NULL);
		}

		if (bo->refcnt > 1)
			continue;

		list_move_tail(&bo->list, list);
		return kgem_bo_reference(bo);
	}

	return NULL;
}

static struct kgem_bo *
kgem_create_batch(struct kgem *kgem)
{
	struct kgem_bo *bo;
	int size, shrink = 0;

#if !DBG_NO_SHRINK_BATCHES
	if (kgem->surface != kgem->batch_size)
		size = compact_batch_surface(kgem, &shrink);
	else
		size = kgem->nbatch * sizeof(uint32_t);

	if (size <= 4096) {
		bo = first_available(kgem, &kgem->pinned_batches[0]);
		if (bo)
			goto write;
	}

	if (size <= 16384) {
		bo = first_available(kgem, &kgem->pinned_batches[1]);
		if (bo)
			goto write;
	}

	if (kgem->gen == 020) {
		bo = kgem_create_linear(kgem, size, CREATE_CACHED | CREATE_TEMPORARY);
		if (bo)
			goto write;

		/* Nothing available for reuse, rely on the kernel wa */
		if (kgem->has_pinned_batches) {
			bo = kgem_create_linear(kgem, size, CREATE_CACHED | CREATE_TEMPORARY);
			if (bo) {
				kgem->batch_flags &= ~LOCAL_I915_EXEC_IS_PINNED;
				goto write;
			}
		}

		if (size < 16384) {
			bo = list_first_entry(&kgem->pinned_batches[size > 4096],
					      struct kgem_bo,
					      list);
			list_move_tail(&bo->list, &kgem->pinned_batches[size > 4096]);

			DBG(("%s: syncing due to busy batches\n", __FUNCTION__));
			if (kgem_bo_wait(kgem, bo))
				return NULL;

			kgem_retire(kgem);
			assert(bo->rq == NULL);
			bo = kgem_bo_reference(bo);
			goto write;
		}
	}
#else
	if (kgem->surface != kgem->batch_size)
		size = kgem->batch_size * sizeof(uint32_t);
	else
		size = kgem->nbatch * sizeof(uint32_t);
#endif

	if (!kgem->batch_bo || !kgem->has_llc) {
		bo = kgem_create_linear(kgem, size, CREATE_NO_THROTTLE);
		if (bo) {
write:
			kgem_fixup_relocs(kgem, bo, shrink);
			if (kgem_batch_write(kgem, bo, size)) {
				kgem_bo_destroy(kgem, bo);
				return NULL;
			}
			return bo;
		}
	}

	return kgem_new_batch(kgem);
}

#if !NDEBUG
static bool dump_file(const char *path)
{
	FILE *file;
	size_t len = 0;
	char *line = NULL;

	file = fopen(path, "r");
	if (file == NULL)
		return false;

	while (getline(&line, &len, file) != -1)
		ErrorF("%s", line);

	free(line);
	fclose(file);
	return true;
}

static void dump_debugfs(struct kgem *kgem, const char *name)
{
	char path[80];
	int minor = kgem_get_minor(kgem);

	if (minor < 0)
		return;

	sprintf(path, "/sys/kernel/debug/dri/%d/%s", minor, name);
	if (dump_file(path))
		return;

	sprintf(path, "/debug/dri/%d/%s", minor, name);
	if (dump_file(path))
		return;
}

static void dump_gtt_info(struct kgem *kgem)
{
	dump_debugfs(kgem, "i915_gem_gtt");
}

static void dump_fence_regs(struct kgem *kgem)
{
	dump_debugfs(kgem, "i915_gem_fence_regs");
}
#endif

static int do_execbuf(struct kgem *kgem, struct drm_i915_gem_execbuffer2 *execbuf)
{
	int ret;

retry:
	ret = do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, execbuf);
	if (ret == 0)
		return 0;

	DBG(("%s: failed ret=%d, throttling and discarding cache\n", __FUNCTION__, ret));
	(void)__kgem_throttle_retire(kgem, 0);
	if (kgem_expire_cache(kgem))
		goto retry;

	if (kgem_cleanup_cache(kgem))
		goto retry;

	/* last gasp */
	ret = do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, execbuf);
	if (ret != -ENOSPC)
		return ret;

	/* One final trick up our sleeve for when we run out of space.
	 * We turn everything off to free up our pinned framebuffers,
	 * sprites and cursors, and try just one more time.
	 */

	xf86DrvMsg(kgem_get_screen_index(kgem), X_WARNING,
		   "Failed to submit rendering commands, trying again with outputs disabled.\n");

	if (sna_mode_disable(__to_sna(kgem))) {
		kgem_cleanup_cache(kgem);
		ret = do_ioctl(kgem->fd,
			       DRM_IOCTL_I915_GEM_EXECBUFFER2,
			       execbuf);
		DBG(("%s: last_gasp ret=%d\n", __FUNCTION__, ret));
		sna_mode_enable(__to_sna(kgem));
	}

	return ret;
}

void _kgem_submit(struct kgem *kgem)
{
	struct kgem_request *rq;
	uint32_t batch_end;
	int i, ret;

	assert(!DBG_NO_HW);
	assert(!kgem->wedged);

	assert(kgem->nbatch);
	assert(kgem->nbatch <= KGEM_BATCH_SIZE(kgem));
	assert(kgem->nbatch <= kgem->surface);

	batch_end = kgem_end_batch(kgem);
	kgem_sna_flush(kgem);

	DBG(("batch[%d/%d, flags=%x]: %d %d %d %d, nreloc=%d, nexec=%d, nfence=%d, aperture=%d [fenced=%d]\n",
	     kgem->mode, kgem->ring, kgem->batch_flags,
	     batch_end, kgem->nbatch, kgem->surface, kgem->batch_size,
	     kgem->nreloc, kgem->nexec, kgem->nfence, kgem->aperture, kgem->aperture_fenced));

	assert(kgem->nbatch <= kgem->batch_size);
	assert(kgem->nbatch <= kgem->surface);
	assert(kgem->nreloc <= ARRAY_SIZE(kgem->reloc));
	assert(kgem->nexec < ARRAY_SIZE(kgem->exec));
	assert(kgem->nfence <= kgem->fence_max);

	kgem_finish_buffers(kgem);

#if SHOW_BATCH_BEFORE
	__kgem_batch_debug(kgem, batch_end);
#endif

	rq = kgem->next_request;
	assert(rq->bo == NULL);

	rq->bo = kgem_create_batch(kgem);
	if (rq->bo) {
		struct drm_i915_gem_execbuffer2 execbuf;

		assert(!rq->bo->needs_flush);

		i = kgem->nexec++;
		kgem->exec[i].handle = rq->bo->handle;
		kgem->exec[i].relocation_count = kgem->nreloc;
		kgem->exec[i].relocs_ptr = (uintptr_t)kgem->reloc;
		kgem->exec[i].alignment = 0;
		kgem->exec[i].offset = rq->bo->presumed_offset;
		/* Make sure the kernel releases any fence, ignored if gen4+ */
		kgem->exec[i].flags = EXEC_OBJECT_NEEDS_FENCE;
		kgem->exec[i].rsvd1 = 0;
		kgem->exec[i].rsvd2 = 0;

		rq->bo->exec = &kgem->exec[i];
		rq->bo->rq = MAKE_REQUEST(rq, kgem->ring); /* useful sanity check */
		list_add(&rq->bo->request, &rq->buffers);
		rq->ring = kgem->ring == KGEM_BLT;

		memset(&execbuf, 0, sizeof(execbuf));
		execbuf.buffers_ptr = (uintptr_t)kgem->exec;
		execbuf.buffer_count = kgem->nexec;
		if (kgem->gen < 030)
			execbuf.batch_len = batch_end*sizeof(uint32_t);
		execbuf.flags = kgem->ring | kgem->batch_flags;

		if (DBG_DUMP) {
			int fd = open("/tmp/i915-batchbuffers.dump",
				      O_WRONLY | O_CREAT | O_APPEND,
				      0666);
			if (fd != -1) {
				ret = write(fd, kgem->batch, batch_end*sizeof(uint32_t));
				fd = close(fd);
			}
		}

		ret = do_execbuf(kgem, &execbuf);
	} else
		ret = -ENOMEM;

	if (ret < 0) {
		kgem_throttle(kgem);
		if (!kgem->wedged) {
			xf86DrvMsg(kgem_get_screen_index(kgem), X_ERROR,
				   "Failed to submit rendering commands (%s), disabling acceleration.\n",
				   strerror(-ret));
			__kgem_set_wedged(kgem);
		}

#if !NDEBUG
		ErrorF("batch[%d/%d]: %d %d %d, nreloc=%d, nexec=%d, nfence=%d, aperture=%d, fenced=%d, high=%d,%d: errno=%d\n",
		       kgem->mode, kgem->ring, batch_end, kgem->nbatch, kgem->surface,
		       kgem->nreloc, kgem->nexec, kgem->nfence, kgem->aperture, kgem->aperture_fenced, kgem->aperture_high, kgem->aperture_total, -ret);

		for (i = 0; i < kgem->nexec; i++) {
			struct kgem_bo *bo, *found = NULL;

			list_for_each_entry(bo, &kgem->next_request->buffers, request) {
				if (bo->handle == kgem->exec[i].handle) {
					found = bo;
					break;
				}
			}
			ErrorF("exec[%d] = handle:%d, presumed offset: %x, size: %d, tiling %d, fenced %d, snooped %d, deleted %d\n",
			       i,
			       kgem->exec[i].handle,
			       (int)kgem->exec[i].offset,
			       found ? kgem_bo_size(found) : -1,
			       found ? found->tiling : -1,
			       (int)(kgem->exec[i].flags & EXEC_OBJECT_NEEDS_FENCE),
			       found ? found->snoop : -1,
			       found ? found->purged : -1);
		}
		for (i = 0; i < kgem->nreloc; i++) {
			ErrorF("reloc[%d] = pos:%d, target:%d, delta:%d, read:%x, write:%x, offset:%x\n",
			       i,
			       (int)kgem->reloc[i].offset,
			       kgem->reloc[i].target_handle,
			       kgem->reloc[i].delta,
			       kgem->reloc[i].read_domains,
			       kgem->reloc[i].write_domain,
			       (int)kgem->reloc[i].presumed_offset);
		}

		{
			struct drm_i915_gem_get_aperture aperture;
			if (do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_GET_APERTURE, &aperture) == 0)
				ErrorF("Aperture size %lld, available %lld\n",
				       (long long)aperture.aper_size,
				       (long long)aperture.aper_available_size);
		}

		if (ret == -ENOSPC)
			dump_gtt_info(kgem);
		if (ret == -EDEADLK)
			dump_fence_regs(kgem);

		if (DEBUG_SYNC) {
			int fd = open("/tmp/batchbuffer", O_WRONLY | O_CREAT | O_APPEND, 0666);
			if (fd != -1) {
				int ignored = write(fd, kgem->batch, batch_end*sizeof(uint32_t));
				assert(ignored == batch_end*sizeof(uint32_t));
				close(fd);
			}

			FatalError("SNA: failed to submit batchbuffer, errno=%d\n", -ret);
		}
#endif
	} else {
		if (DEBUG_SYNC) {
			struct drm_i915_gem_set_domain set_domain;

			VG_CLEAR(set_domain);
			set_domain.handle = rq->bo->handle;
			set_domain.read_domains = I915_GEM_DOMAIN_GTT;
			set_domain.write_domain = I915_GEM_DOMAIN_GTT;

			ret = do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain);
		}

#if SHOW_BATCH_AFTER
		if (gem_read(kgem->fd, rq->bo->handle, kgem->batch, 0, batch_end*sizeof(uint32_t)) == 0)
			__kgem_batch_debug(kgem, batch_end);
#endif

		kgem_commit(kgem);
	}

	if (unlikely(kgem->wedged))
		kgem_cleanup(kgem);

	kgem_reset(kgem);

	assert(kgem->next_request != NULL);
}

void kgem_throttle(struct kgem *kgem)
{
	if (unlikely(kgem->wedged))
		return;

	if (__kgem_throttle(kgem, true)) {
		xf86DrvMsg(kgem_get_screen_index(kgem), X_ERROR,
			   "Detected a hung GPU, disabling acceleration.\n");
		__kgem_set_wedged(kgem);
		kgem->need_throttle = false;
	}
}

int kgem_is_wedged(struct kgem *kgem)
{
	return __kgem_throttle(kgem, true);
}

static void kgem_purge_cache(struct kgem *kgem)
{
	struct kgem_bo *bo, *next;
	int i;

	for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) {
		list_for_each_entry_safe(bo, next, &kgem->inactive[i], list) {
			if (!kgem_bo_is_retained(kgem, bo)) {
				DBG(("%s: purging %d\n",
				     __FUNCTION__, bo->handle));
				kgem_bo_free(kgem, bo);
			}
		}
	}

	kgem->need_purge = false;
}

void kgem_clean_scanout_cache(struct kgem *kgem)
{
	while (!list_is_empty(&kgem->scanout)) {
		struct kgem_bo *bo;

		bo = list_first_entry(&kgem->scanout, struct kgem_bo, list);

		assert(bo->scanout);
		assert(!bo->refcnt);
		assert(!bo->prime);
		assert(bo->proxy == NULL);

		if (bo->exec || __kgem_busy(kgem, bo->handle))
			break;

		DBG(("%s: handle=%d, fb=%d (reusable=%d)\n",
		     __FUNCTION__, bo->handle, bo->delta, bo->reusable));
		list_del(&bo->list);

		kgem_bo_rmfb(kgem, bo);
		bo->scanout = false;

		if (!bo->purged) {
			bo->reusable = true;
			if (kgem->has_llc &&
			    !gem_set_caching(kgem->fd, bo->handle, SNOOPED))
				bo->reusable = false;

		}

		__kgem_bo_destroy(kgem, bo);
	}
}

void kgem_clean_large_cache(struct kgem *kgem)
{
	while (!list_is_empty(&kgem->large_inactive)) {
		kgem_bo_free(kgem,
			     list_first_entry(&kgem->large_inactive,
					      struct kgem_bo, list));

	}
}

bool kgem_expire_cache(struct kgem *kgem)
{
	time_t now, expire;
	struct kgem_bo *bo;
	unsigned int size = 0, count = 0;
	bool idle;
	unsigned int i;

	if (!time(&now))
		return false;

	while (__kgem_freed_bo) {
		bo = __kgem_freed_bo;
		__kgem_freed_bo = *(struct kgem_bo **)bo;
		free(bo);
	}

	while (__kgem_freed_request) {
		struct kgem_request *rq = __kgem_freed_request;
		__kgem_freed_request = *(struct kgem_request **)rq;
		free(rq);
	}

	kgem_clean_large_cache(kgem);
	if (__to_sna(kgem)->scrn->vtSema)
		kgem_clean_scanout_cache(kgem);

	expire = 0;
	list_for_each_entry(bo, &kgem->snoop, list) {
		if (bo->delta) {
			expire = now - MAX_INACTIVE_TIME/2;
			break;
		}

		assert(now);
		bo->delta = now;
	}
	if (expire) {
		while (!list_is_empty(&kgem->snoop)) {
			bo = list_last_entry(&kgem->snoop, struct kgem_bo, list);

			if (bo->delta > expire)
				break;

			kgem_bo_free(kgem, bo);
		}
	}
#ifdef DEBUG_MEMORY
	{
		long snoop_size = 0;
		int snoop_count = 0;
		list_for_each_entry(bo, &kgem->snoop, list)
			snoop_count++, snoop_size += bytes(bo);
		DBG(("%s: still allocated %d bo, %ld bytes, in snoop cache\n",
		     __FUNCTION__, snoop_count, snoop_size));
	}
#endif

	kgem_retire(kgem);
	if (unlikely(kgem->wedged))
		kgem_cleanup(kgem);

	kgem->expire(kgem);

	if (kgem->need_purge)
		kgem_purge_cache(kgem);

	if (kgem->need_retire)
		kgem_retire(kgem);

	expire = 0;
	idle = true;
	for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) {
		idle &= list_is_empty(&kgem->inactive[i]);
		list_for_each_entry(bo, &kgem->inactive[i], list) {
			if (bo->delta) {
				expire = now - MAX_INACTIVE_TIME;
				break;
			}

			assert(now);
			kgem_bo_set_purgeable(kgem, bo);
			bo->delta = now;
		}
	}
	if (expire == 0) {
		DBG(("%s: idle? %d\n", __FUNCTION__, idle));
		kgem->need_expire = !idle;
		return false;
	}

	idle = true;
	for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) {
		struct list preserve;

		list_init(&preserve);
		while (!list_is_empty(&kgem->inactive[i])) {
			bo = list_last_entry(&kgem->inactive[i],
					     struct kgem_bo, list);

			if (bo->delta > expire) {
				idle = false;
				break;
			}

			if (bo->map__cpu && bo->delta + MAP_PRESERVE_TIME > expire) {
				idle = false;
				list_move_tail(&bo->list, &preserve);
			} else {
				count++;
				size += bytes(bo);
				kgem_bo_free(kgem, bo);
				DBG(("%s: expiring handle=%d\n",
				     __FUNCTION__, bo->handle));
			}
		}
		list_splice_tail(&preserve, &kgem->inactive[i]);
	}

#ifdef DEBUG_MEMORY
	{
		long inactive_size = 0;
		int inactive_count = 0;
		for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++)
			list_for_each_entry(bo, &kgem->inactive[i], list)
				inactive_count++, inactive_size += bytes(bo);
		DBG(("%s: still allocated %d bo, %ld bytes, in inactive cache\n",
		     __FUNCTION__, inactive_count, inactive_size));
	}
#endif

	DBG(("%s: expired %d objects, %d bytes, idle? %d\n",
	     __FUNCTION__, count, size, idle));

	kgem->need_expire = !idle;
	return count;
	(void)count;
	(void)size;
}

bool kgem_cleanup_cache(struct kgem *kgem)
{
	unsigned int i;
	int n;

	DBG(("%s\n", __FUNCTION__));

	/* sync to the most recent request */
	for (n = 0; n < ARRAY_SIZE(kgem->requests); n++) {
		if (!list_is_empty(&kgem->requests[n])) {
			struct kgem_request *rq;

			rq = list_last_entry(&kgem->requests[n],
					     struct kgem_request,
					     list);

			DBG(("%s: sync on cleanup\n", __FUNCTION__));
			assert(rq->ring == n);
			assert(rq->bo);
			assert(RQ(rq->bo->rq) == rq);
			kgem_bo_wait(kgem, rq->bo);
		}
		assert(list_is_empty(&kgem->requests[n]));
	}

	kgem_retire(kgem);
	kgem_cleanup(kgem);

	DBG(("%s: need_expire?=%d\n", __FUNCTION__, kgem->need_expire));
	if (!kgem->need_expire)
		return false;

	for (i = 0; i < ARRAY_SIZE(kgem->inactive); i++) {
		while (!list_is_empty(&kgem->inactive[i]))
			kgem_bo_free(kgem,
				     list_last_entry(&kgem->inactive[i],
						     struct kgem_bo, list));
	}

	kgem_clean_large_cache(kgem);
	kgem_clean_scanout_cache(kgem);

	while (!list_is_empty(&kgem->snoop))
		kgem_bo_free(kgem,
			     list_last_entry(&kgem->snoop,
					     struct kgem_bo, list));

	while (__kgem_freed_bo) {
		struct kgem_bo *bo = __kgem_freed_bo;
		__kgem_freed_bo = *(struct kgem_bo **)bo;
		free(bo);
	}

	kgem->need_purge = false;
	kgem->need_expire = false;

	DBG(("%s: complete\n", __FUNCTION__));
	return true;
}

static struct kgem_bo *
search_linear_cache(struct kgem *kgem, unsigned int num_pages, unsigned flags)
{
	struct kgem_bo *bo, *first = NULL;
	bool use_active = (flags & CREATE_INACTIVE) == 0;
	struct list *cache;

	DBG(("%s: num_pages=%d, flags=%x, use_active? %d, use_large=%d [max=%d]\n",
	     __FUNCTION__, num_pages, flags, use_active,
	     num_pages >= MAX_CACHE_SIZE / PAGE_SIZE,
	     MAX_CACHE_SIZE / PAGE_SIZE));

	assert(num_pages);

	if (num_pages >= MAX_CACHE_SIZE / PAGE_SIZE) {
		DBG(("%s: searching large buffers\n", __FUNCTION__));
retry_large:
		cache = use_active ? &kgem->large : &kgem->large_inactive;
		list_for_each_entry_safe(bo, first, cache, list) {
			assert(bo->refcnt == 0);
			assert(bo->reusable);
			assert(!bo->scanout);

			if (num_pages > num_pages(bo))
				goto discard;

			if (bo->tiling != I915_TILING_NONE) {
				if (use_active && kgem->gen < 040)
					goto discard;

				if (!kgem_set_tiling(kgem, bo,
						    I915_TILING_NONE, 0))
					goto discard;
			}
			assert(bo->tiling == I915_TILING_NONE);
			bo->pitch = 0;

			if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo))
				goto discard;

			list_del(&bo->list);
			if (RQ(bo->rq) == (void *)kgem) {
				assert(bo->exec == NULL);
				list_del(&bo->request);
			}

			bo->delta = 0;
			assert_tiling(kgem, bo);
			return bo;

discard:
			if (!use_active)
				kgem_bo_free(kgem, bo);
		}

		if (use_active) {
			use_active = false;
			goto retry_large;
		}

		if (__kgem_throttle_retire(kgem, flags))
			goto retry_large;

		return NULL;
	}

	if (!use_active && list_is_empty(inactive(kgem, num_pages))) {
		DBG(("%s: inactive and cache bucket empty\n",
		     __FUNCTION__));

		if (flags & CREATE_NO_RETIRE) {
			DBG(("%s: can not retire\n", __FUNCTION__));
			return NULL;
		}

		if (list_is_empty(active(kgem, num_pages, I915_TILING_NONE))) {
			DBG(("%s: active cache bucket empty\n", __FUNCTION__));
			return NULL;
		}

		if (!__kgem_throttle_retire(kgem, flags)) {
			DBG(("%s: nothing retired\n", __FUNCTION__));
			return NULL;
		}

		if (list_is_empty(inactive(kgem, num_pages))) {
			DBG(("%s: active cache bucket still empty after retire\n",
			     __FUNCTION__));
			return NULL;
		}
	}

	if (!use_active && flags & (CREATE_CPU_MAP | CREATE_GTT_MAP)) {
		int for_cpu = !!(flags & CREATE_CPU_MAP);
		DBG(("%s: searching for inactive %s map\n",
		     __FUNCTION__, for_cpu ? "cpu" : "gtt"));
		cache = &kgem->vma[for_cpu].inactive[cache_bucket(num_pages)];
		list_for_each_entry(bo, cache, vma) {
			assert(for_cpu ? !!bo->map__cpu : (bo->map__gtt || bo->map__wc));
			assert(bucket(bo) == cache_bucket(num_pages));
			assert(bo->proxy == NULL);
			assert(bo->rq == NULL);
			assert(bo->exec == NULL);
			assert(!bo->scanout);

			if (num_pages > num_pages(bo)) {
				DBG(("inactive too small: %d < %d\n",
				     num_pages(bo), num_pages));
				continue;
			}

			if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo)) {
				kgem_bo_free(kgem, bo);
				break;
			}

			if (!kgem_set_tiling(kgem, bo, I915_TILING_NONE, 0)) {
				kgem_bo_free(kgem, bo);
				break;
			}

			kgem_bo_remove_from_inactive(kgem, bo);
			assert(list_is_empty(&bo->vma));
			assert(list_is_empty(&bo->list));

			assert(bo->tiling == I915_TILING_NONE);
			assert(bo->pitch == 0);
			bo->delta = 0;
			DBG(("  %s: found handle=%d (num_pages=%d) in linear vma cache\n",
			     __FUNCTION__, bo->handle, num_pages(bo)));
			assert(use_active || bo->domain != DOMAIN_GPU);
			assert(!bo->needs_flush);
			assert_tiling(kgem, bo);
			ASSERT_MAYBE_IDLE(kgem, bo->handle, !use_active);
			return bo;
		}

		if (flags & CREATE_EXACT)
			return NULL;

		if (flags & CREATE_CPU_MAP && !kgem->has_llc)
			return NULL;
	}

	cache = use_active ? active(kgem, num_pages, I915_TILING_NONE) : inactive(kgem, num_pages);
	list_for_each_entry(bo, cache, list) {
		assert(bo->refcnt == 0);
		assert(bo->reusable);
		assert(!!bo->rq == !!use_active);
		assert(bo->proxy == NULL);
		assert(!bo->scanout);

		if (num_pages > num_pages(bo))
			continue;

		if (use_active &&
		    kgem->gen <= 040 &&
		    bo->tiling != I915_TILING_NONE)
			continue;

		if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo)) {
			kgem_bo_free(kgem, bo);
			break;
		}

		if (I915_TILING_NONE != bo->tiling) {
			if (flags & (CREATE_CPU_MAP | CREATE_GTT_MAP))
				continue;

			if (first)
				continue;

			if (!kgem_set_tiling(kgem, bo, I915_TILING_NONE, 0))
				continue;
		}
		assert(bo->tiling == I915_TILING_NONE);
		bo->pitch = 0;

		if (bo->map__gtt || bo->map__wc || bo->map__cpu) {
			if (flags & (CREATE_CPU_MAP | CREATE_GTT_MAP)) {
				int for_cpu = !!(flags & CREATE_CPU_MAP);
				if (for_cpu ? !!bo->map__cpu : (bo->map__gtt || bo->map__wc)){
					if (first != NULL)
						break;

					first = bo;
					continue;
				}
			} else {
				if (first != NULL)
					break;

				first = bo;
				continue;
			}
		} else {
			if (flags & CREATE_GTT_MAP && !kgem_bo_can_map(kgem, bo))
				continue;

			if (flags & (CREATE_CPU_MAP | CREATE_GTT_MAP)) {
				if (first != NULL)
					break;

				first = bo;
				continue;
			}
		}

		if (use_active)
			kgem_bo_remove_from_active(kgem, bo);
		else
			kgem_bo_remove_from_inactive(kgem, bo);

		assert(bo->tiling == I915_TILING_NONE);
		assert(bo->pitch == 0);
		bo->delta = 0;
		DBG(("  %s: found handle=%d (num_pages=%d) in linear %s cache\n",
		     __FUNCTION__, bo->handle, num_pages(bo),
		     use_active ? "active" : "inactive"));
		assert(list_is_empty(&bo->list));
		assert(list_is_empty(&bo->vma));
		assert(use_active || bo->domain != DOMAIN_GPU);
		assert(!bo->needs_flush || use_active);
		assert_tiling(kgem, bo);
		ASSERT_MAYBE_IDLE(kgem, bo->handle, !use_active);
		return bo;
	}

	if (first) {
		assert(first->tiling == I915_TILING_NONE);

		if (use_active)
			kgem_bo_remove_from_active(kgem, first);
		else
			kgem_bo_remove_from_inactive(kgem, first);

		first->pitch = 0;
		first->delta = 0;
		DBG(("  %s: found handle=%d (near-miss) (num_pages=%d) in linear %s cache\n",
		     __FUNCTION__, first->handle, num_pages(first),
		     use_active ? "active" : "inactive"));
		assert(list_is_empty(&first->list));
		assert(list_is_empty(&first->vma));
		assert(use_active || first->domain != DOMAIN_GPU);
		assert(!first->needs_flush || use_active);
		ASSERT_MAYBE_IDLE(kgem, first->handle, !use_active);
		return first;
	}

	return NULL;
}

struct kgem_bo *kgem_create_for_name(struct kgem *kgem, uint32_t name)
{
	struct drm_gem_open open_arg;
	struct drm_i915_gem_get_tiling tiling;
	struct kgem_bo *bo;

	DBG(("%s(name=%d)\n", __FUNCTION__, name));

	VG_CLEAR(open_arg);
	open_arg.name = name;
	if (do_ioctl(kgem->fd, DRM_IOCTL_GEM_OPEN, &open_arg))
		return NULL;

	DBG(("%s: new handle=%d\n", __FUNCTION__, open_arg.handle));

	VG_CLEAR(tiling);
	tiling.handle = open_arg.handle;
	if (do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_GET_TILING, &tiling)) {
		DBG(("%s(name=%d) get-tiling failed, ret=%d\n", __FUNCTION__, name, errno));
		gem_close(kgem->fd, open_arg.handle);
		return NULL;
	}

	DBG(("%s: handle=%d, tiling=%d\n", __FUNCTION__, tiling.handle, tiling.tiling_mode));

	bo = __kgem_bo_alloc(open_arg.handle, open_arg.size / PAGE_SIZE);
	if (bo == NULL) {
		gem_close(kgem->fd, open_arg.handle);
		return NULL;
	}

	bo->unique_id = kgem_get_unique_id(kgem);
	bo->tiling = tiling.tiling_mode;
	bo->prime = true;
	bo->reusable = false;
	kgem_bo_unclean(kgem, bo);

	debug_alloc__bo(kgem, bo);
	return bo;
}

struct kgem_bo *kgem_create_for_prime(struct kgem *kgem, int name, uint32_t size)
{
#ifdef DRM_IOCTL_PRIME_FD_TO_HANDLE
	struct drm_prime_handle args;
	struct drm_i915_gem_get_tiling tiling;
	struct local_i915_gem_caching caching;
	struct kgem_bo *bo;
	off_t seek;

	DBG(("%s(name=%d)\n", __FUNCTION__, name));

	VG_CLEAR(args);
	args.fd = name;
	args.flags = 0;
	if (do_ioctl(kgem->fd, DRM_IOCTL_PRIME_FD_TO_HANDLE, &args)) {
		DBG(("%s(name=%d) fd-to-handle failed, ret=%d\n", __FUNCTION__, name, errno));
		return NULL;
	}

	VG_CLEAR(tiling);
	tiling.handle = args.handle;
	if (do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_GET_TILING, &tiling)) {
		DBG(("%s(name=%d) get-tiling failed, ret=%d\n", __FUNCTION__, name, errno));
		gem_close(kgem->fd, args.handle);
		return NULL;
	}

	/* Query actual size, overriding specified if available */
	seek = lseek(args.fd, 0, SEEK_END);
	DBG(("%s: estimated size=%ld, actual=%lld\n",
	     __FUNCTION__, (long)size, (long long)seek));
	if (seek != -1) {
		if (size > seek) {
			DBG(("%s(name=%d) estimated required size [%d] is larger than actual [%ld]\n", __FUNCTION__, name, size, (long)seek));
			gem_close(kgem->fd, args.handle);
			return NULL;
		}
		size = seek;
	}

	DBG(("%s: new handle=%d, tiling=%d\n", __FUNCTION__,
	     args.handle, tiling.tiling_mode));
	bo = __kgem_bo_alloc(args.handle, NUM_PAGES(size));
	if (bo == NULL) {
		gem_close(kgem->fd, args.handle);
		return NULL;
	}

	bo->unique_id = kgem_get_unique_id(kgem);
	bo->tiling = tiling.tiling_mode;
	bo->reusable = false;
	bo->prime = true;
	bo->domain = DOMAIN_NONE;

	/* is this a special bo (e.g. scanout or CPU coherent)? */

	VG_CLEAR(caching);
	caching.handle = args.handle;
	caching.caching = kgem->has_llc;
	(void)drmIoctl(kgem->fd, LOCAL_IOCTL_I915_GEM_GET_CACHING, &caching);
	DBG(("%s: imported handle=%d has caching %d\n", __FUNCTION__, args.handle, caching.caching));
	switch (caching.caching) {
	case 0:
		if (kgem->has_llc) {
			DBG(("%s: interpreting handle=%d as a foreign scanout\n",
			     __FUNCTION__, args.handle));
			bo->scanout = true;
		}
		break;
	case 1:
		if (!kgem->has_llc) {
			DBG(("%s: interpreting handle=%d as a foreign snooped buffer\n",
			     __FUNCTION__, args.handle));
			bo->snoop = true;
			if (bo->tiling) {
				DBG(("%s: illegal snooped tiled buffer\n", __FUNCTION__));
				kgem_bo_free(kgem, bo);
				return NULL;
			}
		}
		break;
	case 2:
		DBG(("%s: interpreting handle=%d as a foreign scanout\n",
		     __FUNCTION__, args.handle));
		bo->scanout = true;
		break;
	}

	debug_alloc__bo(kgem, bo);
	return bo;
#else
	return NULL;
#endif
}

int kgem_bo_export_to_prime(struct kgem *kgem, struct kgem_bo *bo)
{
#if defined(DRM_IOCTL_PRIME_HANDLE_TO_FD) && defined(O_CLOEXEC)
	struct drm_prime_handle args;

	assert(kgem_bo_is_fenced(kgem, bo));

	VG_CLEAR(args);
	args.handle = bo->handle;
	args.flags = O_CLOEXEC;

	if (do_ioctl(kgem->fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args))
		return -1;

	bo->reusable = false;
	return args.fd;
#else
	return -1;
#endif
}

struct kgem_bo *kgem_create_linear(struct kgem *kgem, int size, unsigned flags)
{
	struct kgem_bo *bo;
	uint32_t handle;

	DBG(("%s(%d)\n", __FUNCTION__, size));
	assert(size);

	if (flags & CREATE_GTT_MAP && kgem->has_llc) {
		flags &= ~CREATE_GTT_MAP;
		flags |= CREATE_CPU_MAP;
	}

	size = NUM_PAGES(size);
	if ((flags & CREATE_UNCACHED) == 0) {
		bo = search_linear_cache(kgem, size, CREATE_INACTIVE | flags);
		if (bo) {
			assert(!bo->purged);
			assert(!bo->delta);
			assert(bo->domain != DOMAIN_GPU);
			ASSERT_IDLE(kgem, bo->handle);
			bo->refcnt = 1;
			return bo;
		}

		if (flags & CREATE_CACHED)
			return NULL;
	}

	handle = gem_create(kgem->fd, size);
	if (handle == 0)
		return NULL;

	DBG(("%s: new handle=%d, num_pages=%d\n", __FUNCTION__, handle, size));
	bo = __kgem_bo_alloc(handle, size);
	if (bo == NULL) {
		gem_close(kgem->fd, handle);
		return NULL;
	}

	debug_alloc__bo(kgem, bo);
	return bo;
}

int kgem_choose_tiling(struct kgem *kgem, int tiling, int width, int height, int bpp)
{
	if (DBG_NO_TILING)
		return tiling < 0 ? tiling : I915_TILING_NONE;

	if (kgem->gen < 040) {
		if (tiling && width * bpp > 8192 * 8) {
			DBG(("%s: pitch too large for tliing [%d]\n",
			     __FUNCTION__, width*bpp/8));
			tiling = I915_TILING_NONE;
			goto done;
		}
	} else {
		if (width*bpp > (MAXSHORT-512) * 8) {
			if (tiling > 0)
				tiling = -tiling;
			else if (tiling == 0)
				tiling = -I915_TILING_X;
			DBG(("%s: large pitch [%d], forcing TILING [%d]\n",
			     __FUNCTION__, width*bpp/8, tiling));
		} else if (tiling && (width|height) > 8192) {
			DBG(("%s: large tiled buffer [%dx%d], forcing TILING_X\n",
			     __FUNCTION__, width, height));
			tiling = -I915_TILING_X;
		}

		/* fences limited to 128k (256k on ivb) */
		assert(width * bpp <= 128 * 1024 * 8);
	}

	if (tiling < 0)
		return tiling;

	if (tiling == I915_TILING_Y && !kgem->can_render_y)
		tiling = I915_TILING_X;

	if (tiling && (height == 1 || width == 1)) {
		DBG(("%s: disabling tiling [%dx%d] for single row/col\n",
		     __FUNCTION__,width, height));
		tiling = I915_TILING_NONE;
		goto done;
	}
	if (tiling == I915_TILING_Y && height <= 16) {
		DBG(("%s: too short [%d] for TILING_Y\n",
		     __FUNCTION__,height));
		tiling = I915_TILING_X;
	}
	if (tiling && width * bpp > 8 * (4096 - 64)) {
		DBG(("%s: TLB miss between lines %dx%d (pitch=%d), forcing tiling %d\n",
		     __FUNCTION__,
		     width, height, width*bpp/8,
		     tiling));
		return -tiling;
	}
	if (tiling == I915_TILING_X && height < 4) {
		DBG(("%s: too short [%d] for TILING_X\n",
		     __FUNCTION__, height));
		tiling = I915_TILING_NONE;
		goto done;
	}

	if (tiling == I915_TILING_X && width * bpp <= 8*512) {
		DBG(("%s: too thin [width %d, %d bpp] for TILING_X\n",
		     __FUNCTION__, width, bpp));
		tiling = I915_TILING_NONE;
		goto done;
	}
	if (tiling == I915_TILING_Y && width * bpp < 8*128) {
		DBG(("%s: too thin [%d] for TILING_Y\n",
		     __FUNCTION__, width));
		tiling = I915_TILING_NONE;
		goto done;
	}

	if (tiling && ALIGN(height, 2) * ALIGN(width*bpp, 8*64) <= 4096 * 8) {
		DBG(("%s: too small [%d bytes] for TILING_%c\n", __FUNCTION__,
		     ALIGN(height, 2) * ALIGN(width*bpp, 8*64) / 8,
		     tiling == I915_TILING_X ? 'X' : 'Y'));
		tiling = I915_TILING_NONE;
		goto done;
	}

	if (tiling && width * bpp >= 8 * 4096 / 2) {
		DBG(("%s: TLB near-miss between lines %dx%d (pitch=%d), forcing tiling %d\n",
		     __FUNCTION__,
		     width, height, width*bpp/8,
		     tiling));
		return -tiling;
	}

done:
	DBG(("%s: %dx%d -> %d\n", __FUNCTION__, width, height, tiling));
	return tiling;
}

static int bits_per_pixel(int depth)
{
	switch (depth) {
	case 8: return 8;
	case 15:
	case 16: return 16;
	case 24:
	case 30:
	case 32: return 32;
	default: return 0;
	}
}

unsigned kgem_can_create_2d(struct kgem *kgem,
			    int width, int height, int depth)
{
	uint32_t pitch, size;
	unsigned flags = 0;
	int tiling;
	int bpp;

	DBG(("%s: %dx%d @ %d\n", __FUNCTION__, width, height, depth));

	bpp = bits_per_pixel(depth);
	if (bpp == 0) {
		DBG(("%s: unhandled depth %d\n", __FUNCTION__, depth));
		return 0;
	}

	if (width > MAXSHORT || height > MAXSHORT) {
		DBG(("%s: unhandled size %dx%d\n",
		     __FUNCTION__, width, height));
		return 0;
	}

	size = kgem_surface_size(kgem, false, 0,
				 width, height, bpp,
				 I915_TILING_NONE, &pitch);
	DBG(("%s: untiled size=%d\n", __FUNCTION__, size));
	if (size > 0) {
		if (size <= kgem->max_cpu_size)
			flags |= KGEM_CAN_CREATE_CPU;
		if (size > 4096 && size <= kgem->max_gpu_size)
			flags |= KGEM_CAN_CREATE_GPU;
		if (size <= PAGE_SIZE*kgem->aperture_mappable/4 || kgem->has_wc_mmap)
			flags |= KGEM_CAN_CREATE_GTT;
		if (size > kgem->large_object_size)
			flags |= KGEM_CAN_CREATE_LARGE;
		if (size > kgem->max_object_size) {
			DBG(("%s: too large (untiled) %d > %d\n",
			     __FUNCTION__, size, kgem->max_object_size));
			return 0;
		}
	}

	tiling = kgem_choose_tiling(kgem, I915_TILING_X,
				    width, height, bpp);
	if (tiling != I915_TILING_NONE) {
		size = kgem_surface_size(kgem, false, 0,
					 width, height, bpp, tiling,
					 &pitch);
		DBG(("%s: tiled[%d] size=%d\n", __FUNCTION__, tiling, size));
		if (size > 0 && size <= kgem->max_gpu_size)
			flags |= KGEM_CAN_CREATE_GPU | KGEM_CAN_CREATE_TILED;
		if (size > 0 && size <= PAGE_SIZE*kgem->aperture_mappable/4)
			flags |= KGEM_CAN_CREATE_GTT;
		if (size > PAGE_SIZE*kgem->aperture_mappable/4)
			flags &= ~KGEM_CAN_CREATE_GTT;
		if (size > kgem->large_object_size)
			flags |= KGEM_CAN_CREATE_LARGE;
		if (size > kgem->max_object_size) {
			DBG(("%s: too large (tiled) %d > %d\n",
			     __FUNCTION__, size, kgem->max_object_size));
			return 0;
		}
		if (kgem->gen < 040) {
			int fence_size = 1024 * 1024;
			while (fence_size < size)
				fence_size <<= 1;
			if (fence_size > kgem->max_gpu_size)
				flags &= ~KGEM_CAN_CREATE_GPU | KGEM_CAN_CREATE_TILED;
			if (fence_size > PAGE_SIZE*kgem->aperture_fenceable/4)
				flags &= ~KGEM_CAN_CREATE_GTT;
		}
	}

	return flags;
}

inline int kgem_bo_fenced_size(struct kgem *kgem, struct kgem_bo *bo)
{
	unsigned int size;

	assert(bo->tiling);
	assert_tiling(kgem, bo);
	assert(kgem->gen < 040);

	if (kgem->gen < 030)
		size = 512 * 1024 / PAGE_SIZE;
	else
		size = 1024 * 1024 / PAGE_SIZE;
	while (size < num_pages(bo))
		size <<= 1;

	return size;
}

static struct kgem_bo *
__kgem_bo_create_as_display(struct kgem *kgem, int size, int tiling, int pitch)
{
	struct local_i915_gem_create2 args;
	struct kgem_bo *bo;

	if (!kgem->has_create2)
		return NULL;

	memset(&args, 0, sizeof(args));
	args.size = size * PAGE_SIZE;
	args.placement = LOCAL_I915_CREATE_PLACEMENT_STOLEN;
	args.caching = DISPLAY;
	args.tiling_mode = tiling;
	args.stride = pitch;

	if (do_ioctl(kgem->fd, LOCAL_IOCTL_I915_GEM_CREATE2, &args)) {
		args.placement = LOCAL_I915_CREATE_PLACEMENT_SYSTEM;
		if (do_ioctl(kgem->fd, LOCAL_IOCTL_I915_GEM_CREATE2, &args))
			return NULL;
	}

	bo = __kgem_bo_alloc(args.handle, size);
	if (bo == NULL) {
		gem_close(kgem->fd, args.handle);
		return NULL;
	}

	bo->unique_id = kgem_get_unique_id(kgem);
	bo->tiling = tiling;
	bo->pitch = pitch;
	if (args.placement == LOCAL_I915_CREATE_PLACEMENT_STOLEN) {
		bo->purged = true; /* for asserts against CPU access */
	}
	bo->reusable = false; /* so that unclaimed scanouts are freed */
	bo->domain = DOMAIN_NONE;

	if (__kgem_busy(kgem, bo->handle)) {
		assert(bo->exec == NULL);
		list_add(&bo->request, &kgem->flushing);
		bo->rq = (void *)kgem;
		kgem->need_retire = true;
	}

	assert_tiling(kgem, bo);
	debug_alloc__bo(kgem, bo);

	return bo;
}

static void __kgem_bo_make_scanout(struct kgem *kgem,
				   struct kgem_bo *bo,
				   int width, int height)
{
	ScrnInfoPtr scrn = __to_sna(kgem)->scrn;
	struct drm_mode_fb_cmd arg;

	assert(bo->proxy == NULL);

	if (!scrn->vtSema)
		return;

	DBG(("%s: create fb %dx%d@%d/%d\n",
	     __FUNCTION__, width, height, scrn->depth, scrn->bitsPerPixel));

	VG_CLEAR(arg);
	arg.width = width;
	arg.height = height;
	arg.pitch = bo->pitch;
	arg.bpp = scrn->bitsPerPixel;
	arg.depth = scrn->depth;
	arg.handle = bo->handle;

	/* First move the scanout out of cached memory */
	if (kgem->has_llc) {
		if (!gem_set_caching(kgem->fd, bo->handle, DISPLAY) &&
		    !gem_set_caching(kgem->fd, bo->handle, UNCACHED))
			return;
	}

	bo->scanout = true;

	/* Then pre-emptively move the object into the mappable
	 * portion to avoid rebinding later when busy.
	 */
	if (bo->map__gtt == NULL)
		bo->map__gtt = __kgem_bo_map__gtt(kgem, bo);
	if (bo->map__gtt) {
		if (sigtrap_get() == 0) {
			*(uint32_t *)bo->map__gtt = 0;
			sigtrap_put();
		}
		bo->domain = DOMAIN_GTT;
	}

	if (do_ioctl(kgem->fd, DRM_IOCTL_MODE_ADDFB, &arg) == 0) {
		DBG(("%s: attached fb=%d to handle=%d\n",
		     __FUNCTION__, arg.fb_id, arg.handle));
		bo->delta = arg.fb_id;
	}
}

static bool tiling_changed(struct kgem_bo *bo, int tiling, int pitch)
{
	if (tiling != bo->tiling)
		return true;

	return tiling != I915_TILING_NONE && pitch != bo->pitch;
}

static void set_gpu_tiling(struct kgem *kgem,
			   struct kgem_bo *bo,
			   int tiling, int pitch)
{
	DBG(("%s: handle=%d, tiling=%d, pitch=%d\n",
	     __FUNCTION__, bo->handle, tiling, pitch));

	if (tiling_changed(bo, tiling, pitch) && bo->map__gtt) {
		if (!list_is_empty(&bo->vma)) {
			list_del(&bo->vma);
			kgem->vma[0].count--;
		}
		munmap(bo->map__gtt, bytes(bo));
		bo->map__gtt = NULL;
	}

	bo->tiling = tiling;
	bo->pitch = pitch;
}

bool kgem_bo_is_fenced(struct kgem *kgem, struct kgem_bo *bo)
{
	struct drm_i915_gem_get_tiling tiling;

	assert(kgem);
	assert(bo);

	VG_CLEAR(tiling);
	tiling.handle = bo->handle;
	tiling.tiling_mode = bo->tiling;
	(void)do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_GET_TILING, &tiling);
	return tiling.tiling_mode == bo->tiling; /* assume pitch is fine! */
}

struct kgem_bo *kgem_create_2d(struct kgem *kgem,
			       int width,
			       int height,
			       int bpp,
			       int tiling,
			       uint32_t flags)
{
	struct list *cache;
	struct kgem_bo *bo;
	uint32_t pitch, tiled_height, size;
	uint32_t handle;
	int i, bucket, retry;
	bool exact = flags & (CREATE_EXACT | CREATE_SCANOUT);

	if (tiling < 0)
		exact = true, tiling = -tiling;

	DBG(("%s(%dx%d, bpp=%d, tiling=%d, exact=%d, inactive=%d, cpu-mapping=%d, gtt-mapping=%d, scanout?=%d, prime?=%d, temp?=%d)\n", __FUNCTION__,
	     width, height, bpp, tiling, exact,
	     !!(flags & CREATE_INACTIVE),
	     !!(flags & CREATE_CPU_MAP),
	     !!(flags & CREATE_GTT_MAP),
	     !!(flags & CREATE_SCANOUT),
	     !!(flags & CREATE_PRIME),
	     !!(flags & CREATE_TEMPORARY)));

	size = kgem_surface_size(kgem, kgem->has_relaxed_fencing, flags,
				 width, height, bpp, tiling, &pitch);
	if (size == 0) {
		DBG(("%s: invalid surface size (too large?)\n", __FUNCTION__));
		return NULL;
	}

	size /= PAGE_SIZE;
	bucket = cache_bucket(size);

	if (flags & CREATE_SCANOUT) {
		struct kgem_bo *last = NULL;

		list_for_each_entry_reverse(bo, &kgem->scanout, list) {
			assert(bo->scanout);
			assert(!bo->flush);
			assert(!bo->refcnt);
			assert_tiling(kgem, bo);

			if (size > num_pages(bo) || num_pages(bo) > 2*size)
				continue;

			if (bo->tiling != tiling || bo->pitch != pitch)
				/* No tiling/pitch without recreating fb */
				continue;

			if (bo->delta && !check_scanout_size(kgem, bo, width, height))
				kgem_bo_rmfb(kgem, bo);

			if (flags & CREATE_INACTIVE && bo->rq) {
				last = bo;
				continue;
			}

			list_del(&bo->list);

			bo->unique_id = kgem_get_unique_id(kgem);
			DBG(("  1:from scanout: pitch=%d, tiling=%d, handle=%d, id=%d\n",
			     bo->pitch, bo->tiling, bo->handle, bo->unique_id));
			assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo));
			assert_tiling(kgem, bo);
			bo->refcnt = 1;
			return bo;
		}

		if (last) {
			list_del(&last->list);

			last->unique_id = kgem_get_unique_id(kgem);
			DBG(("  1:from scanout: pitch=%d, tiling=%d, handle=%d, id=%d\n",
			     last->pitch, last->tiling, last->handle, last->unique_id));
			assert(last->pitch*kgem_aligned_height(kgem, height, last->tiling) <= kgem_bo_size(last));
			assert_tiling(kgem, last);
			last->refcnt = 1;
			return last;
		}

		if (__to_sna(kgem)->scrn->vtSema) {
			ScrnInfoPtr scrn = __to_sna(kgem)->scrn;

			list_for_each_entry_reverse(bo, &kgem->scanout, list) {
				struct drm_mode_fb_cmd arg;

				assert(bo->scanout);
				assert(!bo->refcnt);

				if (size > num_pages(bo) || num_pages(bo) > 2*size)
					continue;

				if (flags & CREATE_INACTIVE && bo->rq)
					continue;

				list_del(&bo->list);

				if (bo->tiling != tiling || bo->pitch != pitch) {
					if (bo->delta) {
						kgem_bo_rmfb(kgem, bo);
						bo->delta = 0;
					}

					if (!kgem_set_tiling(kgem, bo,
							     tiling, pitch)) {
						bo->scanout = false;
						__kgem_bo_destroy(kgem, bo);
						break;
					}
				}

				VG_CLEAR(arg);
				arg.width = width;
				arg.height = height;
				arg.pitch = bo->pitch;
				arg.bpp = scrn->bitsPerPixel;
				arg.depth = scrn->depth;
				arg.handle = bo->handle;

				if (do_ioctl(kgem->fd, DRM_IOCTL_MODE_ADDFB, &arg)) {
					bo->scanout = false;
					__kgem_bo_destroy(kgem, bo);
					break;
				}

				bo->delta = arg.fb_id;
				bo->unique_id = kgem_get_unique_id(kgem);

				DBG(("  2:from scanout: pitch=%d, tiling=%d, handle=%d, id=%d\n",
				     bo->pitch, bo->tiling, bo->handle, bo->unique_id));
				assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo));
				assert_tiling(kgem, bo);
				bo->refcnt = 1;
				return bo;
			}
		}

		if (flags & CREATE_CACHED)
			return NULL;

		bo = __kgem_bo_create_as_display(kgem, size, tiling, pitch);
		if (bo)
			return bo;

		flags |= CREATE_INACTIVE;
	}

	if (bucket >= NUM_CACHE_BUCKETS) {
		DBG(("%s: large bo num pages=%d, bucket=%d\n",
		     __FUNCTION__, size, bucket));

		if (flags & CREATE_INACTIVE)
			goto large_inactive;

		tiled_height = kgem_aligned_height(kgem, height, tiling);

		list_for_each_entry(bo, &kgem->large, list) {
			assert(!bo->purged);
			assert(!bo->scanout);
			assert(bo->refcnt == 0);
			assert(bo->reusable);
			assert_tiling(kgem, bo);

			if (kgem->gen < 040) {
				if (bo->pitch < pitch) {
					DBG(("tiled and pitch too small: tiling=%d, (want %d), pitch=%d, need %d\n",
					     bo->tiling, tiling,
					     bo->pitch, pitch));
					continue;
				}

				if (bo->pitch * tiled_height > bytes(bo))
					continue;
			} else {
				if (num_pages(bo) < size)
					continue;

				if (!kgem_set_tiling(kgem, bo, tiling, pitch)) {
					if (exact) {
						DBG(("tiled and pitch not exact: tiling=%d, (want %d), pitch=%d, need %d\n",
						     bo->tiling, tiling,
						     bo->pitch, pitch));
						continue;
					}

					set_gpu_tiling(kgem, bo, tiling, pitch);
				}
			}

			kgem_bo_remove_from_active(kgem, bo);

			bo->unique_id = kgem_get_unique_id(kgem);
			bo->delta = 0;
			DBG(("  1:from active: pitch=%d, tiling=%d, handle=%d, id=%d\n",
			     bo->pitch, bo->tiling, bo->handle, bo->unique_id));
			assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo));
			assert_tiling(kgem, bo);
			bo->refcnt = 1;
			return bo;
		}

large_inactive:
		__kgem_throttle_retire(kgem, flags);
		list_for_each_entry(bo, &kgem->large_inactive, list) {
			assert(bo->refcnt == 0);
			assert(bo->reusable);
			assert(!bo->scanout);
			assert_tiling(kgem, bo);

			if (size > num_pages(bo))
				continue;

			if (!kgem_set_tiling(kgem, bo, tiling, pitch)) {
				if (kgem->gen >= 040 && !exact)
					set_gpu_tiling(kgem, bo, tiling, pitch);
				else
					continue;
			}

			if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo)) {
				kgem_bo_free(kgem, bo);
				break;
			}

			list_del(&bo->list);

			assert(bo->domain != DOMAIN_GPU);
			bo->unique_id = kgem_get_unique_id(kgem);
			bo->delta = 0;
			DBG(("  1:from large inactive: pitch=%d, tiling=%d, handle=%d, id=%d\n",
			     bo->pitch, bo->tiling, bo->handle, bo->unique_id));
			assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo));
			assert_tiling(kgem, bo);
			bo->refcnt = 1;

			if (flags & CREATE_SCANOUT)
				__kgem_bo_make_scanout(kgem, bo, width, height);

			return bo;
		}

		goto create;
	}

	if (flags & (CREATE_CPU_MAP | CREATE_GTT_MAP)) {
		int for_cpu = !!(flags & CREATE_CPU_MAP);
		if (kgem->has_llc && tiling == I915_TILING_NONE)
			for_cpu = 1;
		/* We presume that we will need to upload to this bo,
		 * and so would prefer to have an active VMA.
		 */
		cache = &kgem->vma[for_cpu].inactive[bucket];
		do {
			list_for_each_entry(bo, cache, vma) {
				assert(bucket(bo) == bucket);
				assert(bo->refcnt == 0);
				assert(!bo->scanout);
				assert(for_cpu ? !!bo->map__cpu : (bo->map__gtt || bo->map__wc));
				assert(bo->rq == NULL);
				assert(bo->exec == NULL);
				assert(list_is_empty(&bo->request));
				assert(bo->flush == false);
				assert_tiling(kgem, bo);

				if (size > num_pages(bo)) {
					DBG(("inactive too small: %d < %d\n",
					     num_pages(bo), size));
					continue;
				}

				if (flags & UNCACHED && !kgem->has_llc && bo->domain != DOMAIN_CPU)
					continue;

				if (bo->tiling != tiling ||
				    (tiling != I915_TILING_NONE && bo->pitch != pitch)) {
					if (bo->map__gtt ||
					    !kgem_set_tiling(kgem, bo,
							     tiling, pitch)) {
						DBG(("inactive GTT vma with wrong tiling: %d < %d\n",
						     bo->tiling, tiling));
						kgem_bo_free(kgem, bo);
						break;
					}
				}

				if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo)) {
					kgem_bo_free(kgem, bo);
					break;
				}

				if (tiling == I915_TILING_NONE)
					bo->pitch = pitch;

				assert(bo->tiling == tiling);
				assert(bo->pitch >= pitch);
				bo->delta = 0;
				bo->unique_id = kgem_get_unique_id(kgem);

				kgem_bo_remove_from_inactive(kgem, bo);
				assert(list_is_empty(&bo->list));
				assert(list_is_empty(&bo->vma));

				DBG(("  from inactive vma: pitch=%d, tiling=%d: handle=%d, id=%d\n",
				     bo->pitch, bo->tiling, bo->handle, bo->unique_id));
				assert(bo->reusable);
				assert(bo->domain != DOMAIN_GPU);
				ASSERT_IDLE(kgem, bo->handle);
				assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo));
				assert_tiling(kgem, bo);
				bo->refcnt = 1;
				return bo;
			}
		} while (!list_is_empty(cache) &&
			 __kgem_throttle_retire(kgem, flags));

		if (flags & CREATE_CPU_MAP && !kgem->has_llc) {
			if (list_is_empty(&kgem->active[bucket][tiling]) &&
			    list_is_empty(&kgem->inactive[bucket]))
				flags &= ~CREATE_CACHED;

			goto create;
		}
	}

	if (flags & CREATE_INACTIVE)
		goto skip_active_search;

	/* Best active match */
	retry = NUM_CACHE_BUCKETS - bucket;
	if (retry > 3 && (flags & CREATE_TEMPORARY) == 0)
		retry = 3;
search_active:
	assert(bucket < NUM_CACHE_BUCKETS);
	cache = &kgem->active[bucket][tiling];
	if (tiling) {
		tiled_height = kgem_aligned_height(kgem, height, tiling);
		list_for_each_entry(bo, cache, list) {
			assert(!bo->purged);
			assert(bo->refcnt == 0);
			assert(bucket(bo) == bucket);
			assert(bo->reusable);
			assert(bo->tiling == tiling);
			assert(bo->flush == false);
			assert(!bo->scanout);
			assert_tiling(kgem, bo);

			if (kgem->gen < 040) {
				if (bo->pitch < pitch) {
					DBG(("tiled and pitch too small: tiling=%d, (want %d), pitch=%d, need %d\n",
					     bo->tiling, tiling,
					     bo->pitch, pitch));
					continue;
				}

				if (bo->pitch * tiled_height > bytes(bo))
					continue;
			} else {
				if (num_pages(bo) < size)
					continue;

				if (!kgem_set_tiling(kgem, bo, tiling, pitch)) {
					if (exact) {
						DBG(("tiled and pitch not exact: tiling=%d, (want %d), pitch=%d, need %d\n",
						     bo->tiling, tiling,
						     bo->pitch, pitch));
						continue;
					}

					set_gpu_tiling(kgem, bo, tiling, pitch);
				}
			}
			assert(bo->tiling == tiling);
			assert(bo->pitch >= pitch);

			kgem_bo_remove_from_active(kgem, bo);

			bo->unique_id = kgem_get_unique_id(kgem);
			bo->delta = 0;
			DBG(("  1:from active: pitch=%d, tiling=%d, handle=%d, id=%d\n",
			     bo->pitch, bo->tiling, bo->handle, bo->unique_id));
			assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo));
			assert_tiling(kgem, bo);
			bo->refcnt = 1;
			return bo;
		}
	} else {
		list_for_each_entry(bo, cache, list) {
			assert(bucket(bo) == bucket);
			assert(!bo->purged);
			assert(bo->refcnt == 0);
			assert(bo->reusable);
			assert(!bo->scanout);
			assert(bo->tiling == tiling);
			assert(bo->flush == false);
			assert_tiling(kgem, bo);

			if (num_pages(bo) < size)
				continue;

			kgem_bo_remove_from_active(kgem, bo);

			bo->pitch = pitch;
			bo->unique_id = kgem_get_unique_id(kgem);
			bo->delta = 0;
			DBG(("  1:from active: pitch=%d, tiling=%d, handle=%d, id=%d\n",
			     bo->pitch, bo->tiling, bo->handle, bo->unique_id));
			assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo));
			assert_tiling(kgem, bo);
			bo->refcnt = 1;
			return bo;
		}
	}

	if (kgem->gen >= 040) {
		for (i = I915_TILING_Y; i >= I915_TILING_NONE; i--) {
			cache = &kgem->active[bucket][i];
			list_for_each_entry(bo, cache, list) {
				assert(!bo->purged);
				assert(bo->refcnt == 0);
				assert(bo->reusable);
				assert(!bo->scanout);
				assert(bo->flush == false);
				assert_tiling(kgem, bo);

				if (num_pages(bo) < size)
					continue;

				if (!kgem_set_tiling(kgem, bo, tiling, pitch)) {
					if (exact || kgem->gen < 040)
						continue;

					set_gpu_tiling(kgem, bo, tiling, pitch);
				}
				assert(bo->tiling == tiling);
				assert(bo->pitch >= pitch);

				kgem_bo_remove_from_active(kgem, bo);

				bo->unique_id = kgem_get_unique_id(kgem);
				bo->delta = 0;
				DBG(("  1:from active: pitch=%d, tiling=%d, handle=%d, id=%d\n",
				     bo->pitch, bo->tiling, bo->handle, bo->unique_id));
				assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo));
				assert_tiling(kgem, bo);
				bo->refcnt = 1;
				return bo;
			}
		}
	} else if (!exact) { /* allow an active near-miss? */
		for (i = tiling; i >= I915_TILING_NONE; i--) {
			tiled_height = kgem_surface_size(kgem, kgem->has_relaxed_fencing, flags,
							 width, height, bpp, tiling, &pitch);
			cache = active(kgem, tiled_height / PAGE_SIZE, i);
			tiled_height = kgem_aligned_height(kgem, height, i);
			list_for_each_entry(bo, cache, list) {
				assert(!bo->purged);
				assert(bo->refcnt == 0);
				assert(bo->reusable);
				assert(!bo->scanout);
				assert(bo->flush == false);
				assert_tiling(kgem, bo);

				if (bo->tiling) {
					if (bo->pitch < pitch) {
						DBG(("tiled and pitch too small: tiling=%d, (want %d), pitch=%d, need %d\n",
						     bo->tiling, tiling,
						     bo->pitch, pitch));
						continue;
					}
				} else
					bo->pitch = pitch;

				if (bo->pitch * tiled_height > bytes(bo))
					continue;

				kgem_bo_remove_from_active(kgem, bo);

				bo->unique_id = kgem_get_unique_id(kgem);
				bo->delta = 0;
				DBG(("  1:from active: pitch=%d, tiling=%d, handle=%d, id=%d\n",
				     bo->pitch, bo->tiling, bo->handle, bo->unique_id));
				assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo));
				assert_tiling(kgem, bo);
				bo->refcnt = 1;
				return bo;
			}
		}
	}

	if (--retry) {
		bucket++;
		goto search_active;
	}

skip_active_search:
	bucket = cache_bucket(size);
	retry = NUM_CACHE_BUCKETS - bucket;
	if (retry > 3)
		retry = 3;
search_inactive:
	/* Now just look for a close match and prefer any currently active */
	assert(bucket < NUM_CACHE_BUCKETS);
	cache = &kgem->inactive[bucket];
	list_for_each_entry(bo, cache, list) {
		assert(bucket(bo) == bucket);
		assert(bo->reusable);
		assert(!bo->scanout);
		assert(bo->flush == false);
		assert_tiling(kgem, bo);

		if (size > num_pages(bo)) {
			DBG(("inactive too small: %d < %d\n",
			     num_pages(bo), size));
			continue;
		}

		if (!kgem_set_tiling(kgem, bo, tiling, pitch)) {
			if (exact || kgem->gen < 040) {
				kgem_bo_free(kgem, bo);
				break;
			}

			set_gpu_tiling(kgem, bo, tiling, pitch);
		}

		if (bo->purged && !kgem_bo_clear_purgeable(kgem, bo)) {
			kgem_bo_free(kgem, bo);
			break;
		}

		kgem_bo_remove_from_inactive(kgem, bo);
		assert(list_is_empty(&bo->list));
		assert(list_is_empty(&bo->vma));
		assert(bo->tiling == tiling);
		assert(bo->pitch >= pitch);

		bo->delta = 0;
		bo->unique_id = kgem_get_unique_id(kgem);
		assert(bo->pitch);
		DBG(("  from inactive: pitch=%d, tiling=%d: handle=%d, id=%d\n",
		     bo->pitch, bo->tiling, bo->handle, bo->unique_id));
		assert(bo->refcnt == 0);
		assert(bo->reusable);
		assert((flags & CREATE_INACTIVE) == 0 || bo->domain != DOMAIN_GPU);
		ASSERT_MAYBE_IDLE(kgem, bo->handle, flags & CREATE_INACTIVE);
		assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo));
		assert_tiling(kgem, bo);
		bo->refcnt = 1;

		if (flags & CREATE_SCANOUT)
			__kgem_bo_make_scanout(kgem, bo, width, height);

		return bo;
	}

	if ((flags & CREATE_NO_RETIRE) == 0) {
		list_for_each_entry_reverse(bo, &kgem->active[bucket][tiling], list) {
			if (bo->exec)
				break;

			if (size > num_pages(bo))
				continue;

			if (__kgem_busy(kgem, bo->handle)) {
				if (flags & CREATE_NO_THROTTLE)
					goto no_retire;

				do {
					if (!kgem->need_throttle) {
						DBG(("%s: not throttling for active handle=%d\n", __FUNCTION__, bo->handle));
						goto no_retire;
					}

					__kgem_throttle(kgem, false);
				} while (__kgem_busy(kgem, bo->handle));
			}

			DBG(("%s: flushed active handle=%d\n", __FUNCTION__, bo->handle));

			kgem_bo_remove_from_active(kgem, bo);
			__kgem_bo_clear_busy(bo);

			if (!kgem_set_tiling(kgem, bo, tiling, pitch)) {
				if (exact || kgem->gen < 040)
					goto no_retire;

				set_gpu_tiling(kgem, bo, tiling, pitch);
			}
			assert(bo->tiling == tiling);
			assert(bo->pitch >= pitch);

			bo->unique_id = kgem_get_unique_id(kgem);
			bo->delta = 0;
			DBG(("  2:from active: pitch=%d, tiling=%d, handle=%d, id=%d\n",
			     bo->pitch, bo->tiling, bo->handle, bo->unique_id));
			assert(bo->pitch*kgem_aligned_height(kgem, height, bo->tiling) <= kgem_bo_size(bo));
			assert_tiling(kgem, bo);
			bo->refcnt = 1;

			if (flags & CREATE_SCANOUT)
				__kgem_bo_make_scanout(kgem, bo, width, height);

			return bo;
		}
no_retire:
		flags |= CREATE_NO_RETIRE;
	}

	if (--retry) {
		bucket++;
		goto search_inactive;
	}

create:
	if (flags & CREATE_CACHED) {
		DBG(("%s: no cached bo found, requested not to create a new bo\n", __FUNCTION__));
		return NULL;
	}

	if (bucket >= NUM_CACHE_BUCKETS)
		size = ALIGN(size, 1024);
	handle = gem_create(kgem->fd, size);
	if (handle == 0) {
		DBG(("%s: kernel allocation (gem_create) failure\n", __FUNCTION__));
		return NULL;
	}

	bo = __kgem_bo_alloc(handle, size);
	if (!bo) {
		DBG(("%s: malloc failed\n", __FUNCTION__));
		gem_close(kgem->fd, handle);
		return NULL;
	}

	bo->unique_id = kgem_get_unique_id(kgem);
	if (kgem_set_tiling(kgem, bo, tiling, pitch)) {
		if (flags & CREATE_SCANOUT)
			__kgem_bo_make_scanout(kgem, bo, width, height);
	} else {
		if (kgem->gen >= 040) {
			assert(!kgem->can_fence);
			bo->tiling = tiling;
			bo->pitch = pitch;
		} else {
			if (flags & CREATE_EXACT) {
				DBG(("%s: failed to set exact tiling (gem_set_tiling)\n", __FUNCTION__));
				gem_close(kgem->fd, handle);
				free(bo);
				return NULL;
			}
		}
	}

	assert(bytes(bo) >= bo->pitch * kgem_aligned_height(kgem, height, bo->tiling));
	assert_tiling(kgem, bo);

	debug_alloc__bo(kgem, bo);

	DBG(("  new pitch=%d, tiling=%d, handle=%d, id=%d, num_pages=%d [%d], bucket=%d\n",
	     bo->pitch, bo->tiling, bo->handle, bo->unique_id,
	     size, num_pages(bo), bucket(bo)));
	return bo;
}

struct kgem_bo *kgem_create_cpu_2d(struct kgem *kgem,
				   int width,
				   int height,
				   int bpp,
				   uint32_t flags)
{
	struct kgem_bo *bo;
	int stride, size;

	if (DBG_NO_CPU)
		return NULL;

	DBG(("%s(%dx%d, bpp=%d)\n", __FUNCTION__, width, height, bpp));

	if (kgem->has_llc) {
		bo = kgem_create_2d(kgem, width, height, bpp,
				    I915_TILING_NONE, flags);
		if (bo == NULL)
			return bo;

		assert(bo->tiling == I915_TILING_NONE);
		assert_tiling(kgem, bo);

		if (kgem_bo_map__cpu(kgem, bo) == NULL) {
			kgem_bo_destroy(kgem, bo);
			return NULL;
		}

		return bo;
	}

	assert(width > 0 && height > 0);
	stride = ALIGN(width, 2) * bpp >> 3;
	stride = ALIGN(stride, 4);
	size = stride * ALIGN(height, 2);
	assert(size >= PAGE_SIZE);

	DBG(("%s: %dx%d, %d bpp, stride=%d\n",
	     __FUNCTION__, width, height, bpp, stride));

	bo = search_snoop_cache(kgem, NUM_PAGES(size), 0);
	if (bo) {
		assert(bo->tiling == I915_TILING_NONE);
		assert_tiling(kgem, bo);
		assert(bo->snoop);
		bo->refcnt = 1;
		bo->pitch = stride;
		bo->unique_id = kgem_get_unique_id(kgem);
		return bo;
	}

	if (kgem->has_caching) {
		bo = kgem_create_linear(kgem, size, flags);
		if (bo == NULL)
			return NULL;

		assert(bo->tiling == I915_TILING_NONE);
		assert_tiling(kgem, bo);

		assert(!__kgem_busy(kgem, bo->handle));
		if (!gem_set_caching(kgem->fd, bo->handle, SNOOPED)) {
			kgem_bo_destroy(kgem, bo);
			return NULL;
		}
		bo->snoop = true;

		if (kgem_bo_map__cpu(kgem, bo) == NULL) {
			kgem_bo_destroy(kgem, bo);
			return NULL;
		}

		bo->pitch = stride;
		bo->unique_id = kgem_get_unique_id(kgem);
		return bo;
	}

	if (kgem->has_userptr) {
		void *ptr;

		/* XXX */
		//if (posix_memalign(&ptr, 64, ALIGN(size, 64)))
		if (posix_memalign(&ptr, PAGE_SIZE, ALIGN(size, PAGE_SIZE)))
			return NULL;

		bo = kgem_create_map(kgem, ptr, size, false);
		if (bo == NULL) {
			free(ptr);
			return NULL;
		}

		bo->pitch = stride;
		bo->unique_id = kgem_get_unique_id(kgem);
		return bo;
	}

	return NULL;
}

void _kgem_bo_destroy(struct kgem *kgem, struct kgem_bo *bo)
{
	DBG(("%s: handle=%d, proxy? %d\n",
	     __FUNCTION__, bo->handle, bo->proxy != NULL));

	if (bo->proxy) {
		assert(!bo->reusable);
		kgem_bo_binding_free(kgem, bo);

		assert(list_is_empty(&bo->list));
		_list_del(&bo->vma);
		_list_del(&bo->request);

		if (bo->io && bo->domain == DOMAIN_CPU)
			_kgem_bo_delete_buffer(kgem, bo);

		kgem_bo_unref(kgem, bo->proxy);

		if (DBG_NO_MALLOC_CACHE) {
			free(bo);
		} else {
			*(struct kgem_bo **)bo = __kgem_freed_bo;
			__kgem_freed_bo = bo;
		}
	} else
		__kgem_bo_destroy(kgem, bo);
}

static void __kgem_flush(struct kgem *kgem, struct kgem_bo *bo)
{
	assert(bo->rq);
	assert(bo->exec == NULL);
	assert(bo->needs_flush);

	/* The kernel will emit a flush *and* update its own flushing lists. */
	if (!__kgem_busy(kgem, bo->handle))
		__kgem_bo_clear_busy(bo);

	DBG(("%s: handle=%d, busy?=%d\n",
	     __FUNCTION__, bo->handle, bo->rq != NULL));
}

void kgem_scanout_flush(struct kgem *kgem, struct kgem_bo *bo)
{
	if (!bo->needs_flush && !bo->gtt_dirty)
		return;

	kgem_bo_submit(kgem, bo);

	/* If the kernel fails to emit the flush, then it will be forced when
	 * we assume direct access. And as the usual failure is EIO, we do
	 * not actually care.
	 */
	assert(bo->exec == NULL);
	if (bo->rq)
		__kgem_flush(kgem, bo);

	if (bo->scanout && kgem->needs_dirtyfb) {
		struct drm_mode_fb_dirty_cmd cmd;
		memset(&cmd, 0, sizeof(cmd));
		cmd.fb_id = bo->delta;
		(void)drmIoctl(kgem->fd, DRM_IOCTL_MODE_DIRTYFB, &cmd);
	}

	/* Whatever actually happens, we can regard the GTT write domain
	 * as being flushed.
	 */
	__kgem_bo_clear_dirty(bo);
}

inline static bool nearly_idle(struct kgem *kgem)
{
	int ring = kgem->ring == KGEM_BLT;

	assert(ring < ARRAY_SIZE(kgem->requests));
	if (list_is_singular(&kgem->requests[ring]))
		return true;

	return __kgem_ring_is_idle(kgem, ring);
}

inline static bool needs_semaphore(struct kgem *kgem, struct kgem_bo *bo)
{
	if (kgem->needs_semaphore)
		return false;

	if (bo->rq == NULL || RQ_RING(bo->rq) == kgem->ring)
		return false;

	kgem->needs_semaphore = true;
	return true;
}

inline static bool needs_reservation(struct kgem *kgem, struct kgem_bo *bo)
{
	if (kgem->needs_reservation)
		return false;

	if (bo->presumed_offset)
		return false;

	kgem->needs_reservation = true;
	return nearly_idle(kgem);
}

inline static bool needs_batch_flush(struct kgem *kgem, struct kgem_bo *bo)
{
	bool flush = false;

	if (needs_semaphore(kgem, bo)) {
		DBG(("%s: flushing before handle=%d for required semaphore\n", __FUNCTION__, bo->handle));
		flush = true;
	}

	if (needs_reservation(kgem, bo)) {
		DBG(("%s: flushing before handle=%d for new reservation\n", __FUNCTION__, bo->handle));
		flush = true;
	}

	return kgem->nreloc ? flush : false;
}

static bool aperture_check(struct kgem *kgem, unsigned num_pages)
{
	struct drm_i915_gem_get_aperture aperture;
	int reserve;

	if (kgem->aperture)
		return false;

	/* Leave some space in case of alignment issues */
	reserve = kgem->aperture_mappable / 2;
	if (kgem->gen < 033 && reserve < kgem->aperture_max_fence)
		reserve = kgem->aperture_max_fence;
	if (!kgem->has_llc)
		reserve += kgem->nexec * PAGE_SIZE * 2;

	DBG(("%s: num_pages=%d, holding %d pages in reserve, total aperture %d\n",
	     __FUNCTION__, num_pages, reserve, kgem->aperture_total));
	num_pages += reserve;

	VG_CLEAR(aperture);
	aperture.aper_available_size = kgem->aperture_total;
	aperture.aper_available_size *= PAGE_SIZE;
	(void)do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_GET_APERTURE, &aperture);

	DBG(("%s: aperture required %ld bytes, available %ld bytes\n",
	     __FUNCTION__,
	     (long)num_pages * PAGE_SIZE,
	     (long)aperture.aper_available_size));

	return num_pages <= aperture.aper_available_size / PAGE_SIZE;
}

static inline bool kgem_flush(struct kgem *kgem, bool flush)
{
	if (unlikely(kgem->wedged))
		return false;

	if (kgem->nreloc == 0)
		return true;

	if (__to_sna(kgem)->flags & SNA_POWERSAVE)
		return true;

	if (kgem->flush == flush && kgem->aperture < kgem->aperture_low)
		return true;

	DBG(("%s: opportunistic flushing? flush=%d,%d, aperture=%d/%d, idle?=%d\n",
	     __FUNCTION__, kgem->flush, flush, kgem->aperture, kgem->aperture_low, kgem_ring_is_idle(kgem, kgem->ring)));
	return !kgem_ring_is_idle(kgem, kgem->ring);
}

bool kgem_check_bo(struct kgem *kgem, ...)
{
	va_list ap;
	struct kgem_bo *bo;
	int num_exec = 0;
	int num_pages = 0;
	bool flush = false;
	bool busy = true;

	va_start(ap, kgem);
	while ((bo = va_arg(ap, struct kgem_bo *))) {
		while (bo->proxy)
			bo = bo->proxy;
		if (bo->exec)
			continue;

		if (needs_batch_flush(kgem, bo)) {
			va_end(ap);
			return false;
		}

		num_pages += num_pages(bo);
		num_exec++;

		flush |= bo->flush;
		busy &= bo->rq != NULL;
	}
	va_end(ap);

	DBG(("%s: num_pages=+%d, num_exec=+%d\n",
	     __FUNCTION__, num_pages, num_exec));

	if (!num_pages)
		return true;

	if (kgem->nexec + num_exec >= KGEM_EXEC_SIZE(kgem)) {
		DBG(("%s: out of exec slots (%d + %d / %d)\n", __FUNCTION__,
		     kgem->nexec, num_exec, KGEM_EXEC_SIZE(kgem)));
		return false;
	}

	if (num_pages + kgem->aperture > kgem->aperture_high) {
		DBG(("%s: final aperture usage (%d + %d) is greater than high water mark (%d)\n",
		     __FUNCTION__, kgem->aperture, num_pages, kgem->aperture_high));
		return aperture_check(kgem, num_pages);
	}

	if (busy)
		return true;

	return kgem_flush(kgem, flush);
}

bool kgem_check_bo_fenced(struct kgem *kgem, struct kgem_bo *bo)
{
	assert(bo->refcnt);
	while (bo->proxy)
		bo = bo->proxy;
	assert(bo->refcnt);

	if (bo->exec) {
		if (kgem->gen < 040 &&
		    bo->tiling != I915_TILING_NONE &&
		    (bo->exec->flags & EXEC_OBJECT_NEEDS_FENCE) == 0) {
			uint32_t size;

			assert(bo->tiling == I915_TILING_X);

			if (kgem->nfence >= kgem->fence_max)
				return false;

			if (kgem->aperture_fenced) {
				size = 3*kgem->aperture_fenced;
				if (kgem->aperture_total == kgem->aperture_mappable)
					size += kgem->aperture;
				if (size > kgem->aperture_fenceable &&
				    kgem_ring_is_idle(kgem, kgem->ring)) {
					DBG(("%s: opportunistic fence flush\n", __FUNCTION__));
					return false;
				}
			}

			size = kgem_bo_fenced_size(kgem, bo);
			if (size > kgem->aperture_max_fence)
				kgem->aperture_max_fence = size;
			size += kgem->aperture_fenced;
			if (kgem->gen < 033 && size < 2 * kgem->aperture_max_fence)
				size = 2 * kgem->aperture_max_fence;
			if (kgem->aperture_total == kgem->aperture_mappable)
				size += kgem->aperture;
			if (size > kgem->aperture_fenceable) {
				DBG(("%s: estimated fence space required %d (fenced=%d, max_fence=%d, aperture=%d) exceeds fenceable aperture %d\n",
				     __FUNCTION__, size, kgem->aperture_fenced, kgem->aperture_max_fence, kgem->aperture, kgem->aperture_fenceable));
				return false;
			}
		}

		return true;
	}

	if (kgem->nexec >= KGEM_EXEC_SIZE(kgem) - 1)
		return false;

	if (needs_batch_flush(kgem, bo))
		return false;

	assert_tiling(kgem, bo);
	if (kgem->gen < 040 && bo->tiling != I915_TILING_NONE) {
		uint32_t size;

		assert(bo->tiling == I915_TILING_X);

		if (kgem->nfence >= kgem->fence_max)
			return false;

		if (kgem->aperture_fenced) {
			size = 3*kgem->aperture_fenced;
			if (kgem->aperture_total == kgem->aperture_mappable)
				size += kgem->aperture;
			if (size > kgem->aperture_fenceable &&
			    kgem_ring_is_idle(kgem, kgem->ring)) {
				DBG(("%s: opportunistic fence flush\n", __FUNCTION__));
				return false;
			}
		}

		size = kgem_bo_fenced_size(kgem, bo);
		if (size > kgem->aperture_max_fence)
			kgem->aperture_max_fence = size;
		size += kgem->aperture_fenced;
		if (kgem->gen < 033 && size < 2 * kgem->aperture_max_fence)
			size = 2 * kgem->aperture_max_fence;
		if (kgem->aperture_total == kgem->aperture_mappable)
			size += kgem->aperture;
		if (size > kgem->aperture_fenceable) {
			DBG(("%s: estimated fence space required %d (fenced=%d, max_fence=%d, aperture=%d) exceeds fenceable aperture %d\n",
			     __FUNCTION__, size, kgem->aperture_fenced, kgem->aperture_max_fence, kgem->aperture, kgem->aperture_fenceable));
			return false;
		}
	}

	if (kgem->aperture + kgem->aperture_fenced + num_pages(bo) > kgem->aperture_high) {
		DBG(("%s: final aperture usage (%d + %d) is greater than high water mark (%d)\n",
		     __FUNCTION__, kgem->aperture, num_pages(bo), kgem->aperture_high));
		return aperture_check(kgem, num_pages(bo));
	}

	if (bo->rq)
		return true;

	return kgem_flush(kgem, bo->flush);
}

bool kgem_check_many_bo_fenced(struct kgem *kgem, ...)
{
	va_list ap;
	struct kgem_bo *bo;
	int num_fence = 0;
	int num_exec = 0;
	int num_pages = 0;
	int fenced_size = 0;
	bool flush = false;
	bool busy = true;

	va_start(ap, kgem);
	while ((bo = va_arg(ap, struct kgem_bo *))) {
		assert(bo->refcnt);
		while (bo->proxy)
			bo = bo->proxy;
		assert(bo->refcnt);
		if (bo->exec) {
			if (kgem->gen >= 040 || bo->tiling == I915_TILING_NONE)
				continue;

			if ((bo->exec->flags & EXEC_OBJECT_NEEDS_FENCE) == 0) {
				fenced_size += kgem_bo_fenced_size(kgem, bo);
				num_fence++;
			}

			continue;
		}

		if (needs_batch_flush(kgem, bo)) {
			va_end(ap);
			return false;
		}

		assert_tiling(kgem, bo);
		num_pages += num_pages(bo);
		num_exec++;
		if (kgem->gen < 040 && bo->tiling) {
			uint32_t size = kgem_bo_fenced_size(kgem, bo);
			if (size > kgem->aperture_max_fence)
				kgem->aperture_max_fence = size;
			fenced_size += size;
			num_fence++;
		}

		flush |= bo->flush;
		busy &= bo->rq != NULL;
	}
	va_end(ap);

	if (num_fence) {
		uint32_t size;

		if (kgem->nfence + num_fence > kgem->fence_max)
			return false;

		if (kgem->aperture_fenced) {
			size = 3*kgem->aperture_fenced;
			if (kgem->aperture_total == kgem->aperture_mappable)
				size += kgem->aperture;
			if (size > kgem->aperture_fenceable &&
			    kgem_ring_is_idle(kgem, kgem->ring)) {
				DBG(("%s: opportunistic fence flush\n", __FUNCTION__));
				return false;
			}
		}

		size = kgem->aperture_fenced;
		size += fenced_size;
		if (kgem->gen < 033 && size < 2 * kgem->aperture_max_fence)
			size = 2 * kgem->aperture_max_fence;
		if (kgem->aperture_total == kgem->aperture_mappable)
			size += kgem->aperture;
		if (size > kgem->aperture_fenceable) {
			DBG(("%s: estimated fence space required %d (fenced=%d, max_fence=%d, aperture=%d) exceeds fenceable aperture %d\n",
			     __FUNCTION__, size, kgem->aperture_fenced, kgem->aperture_max_fence, kgem->aperture, kgem->aperture_fenceable));
			return false;
		}
	}

	if (num_pages == 0)
		return true;

	if (kgem->nexec + num_exec >= KGEM_EXEC_SIZE(kgem))
		return false;

	if (num_pages + kgem->aperture > kgem->aperture_high - kgem->aperture_fenced) {
		DBG(("%s: final aperture usage (%d + %d + %d) is greater than high water mark (%d)\n",
		     __FUNCTION__, kgem->aperture, kgem->aperture_fenced, num_pages, kgem->aperture_high));
		return aperture_check(kgem, num_pages);
	}

	if (busy)
		return true;

	return kgem_flush(kgem, flush);
}

void __kgem_bcs_set_tiling(struct kgem *kgem,
			   struct kgem_bo *src,
			   struct kgem_bo *dst)
{
	uint32_t state, *b;

	DBG(("%s: src handle=%d:tiling=%d, dst handle=%d:tiling=%d\n",
	     __FUNCTION__,
	     src ? src->handle : 0, src ? src->tiling : 0,
	     dst ? dst->handle : 0, dst ? dst->tiling : 0));
	assert(kgem->mode == KGEM_BLT);
	assert(dst == NULL || kgem_bo_can_blt(kgem, dst));
	assert(src == NULL || kgem_bo_can_blt(kgem, src));

	state = 0;
	if (dst && dst->tiling == I915_TILING_Y)
		state |= BCS_DST_Y;
	if (src && src->tiling == I915_TILING_Y)
		state |= BCS_SRC_Y;

	if (kgem->bcs_state == state)
		return;

	DBG(("%s: updating SWCTRL %x -> %x\n", __FUNCTION__,
	     kgem->bcs_state, state));

	/* Over-estimate space in case we need to re-emit the cmd packet */
	if (!kgem_check_batch(kgem, 24)) {
		_kgem_submit(kgem);
		_kgem_set_mode(kgem, KGEM_BLT);
		if (state == 0)
			return;
	}

	b = kgem->batch + kgem->nbatch;
	if (kgem->nbatch) {
		*b++ = MI_FLUSH_DW;
		*b++ = 0;
		*b++ = 0;
		*b++ = 0;
	}
	*b++ = MI_LOAD_REGISTER_IMM;
	*b++ = BCS_SWCTRL;
	*b++ = (BCS_SRC_Y | BCS_DST_Y) << 16 | state;
	kgem->nbatch = b - kgem->batch;

	kgem->bcs_state = state;
}

uint32_t kgem_add_reloc(struct kgem *kgem,
			uint32_t pos,
			struct kgem_bo *bo,
			uint32_t read_write_domain,
			uint32_t delta)
{
	int index;

	DBG(("%s: handle=%d, pos=%d, delta=%d, domains=%08x\n",
	     __FUNCTION__, bo ? bo->handle : 0, pos, delta, read_write_domain));

	assert(kgem->gen < 0100);
	assert((read_write_domain & 0x7fff) == 0 || bo != NULL);

	index = kgem->nreloc++;
	assert(index < ARRAY_SIZE(kgem->reloc));
	kgem->reloc[index].offset = pos * sizeof(kgem->batch[0]);
	if (bo) {
		assert(kgem->mode != KGEM_NONE);
		assert(bo->refcnt);
		while (bo->proxy) {
			DBG(("%s: adding proxy [delta=%d] for handle=%d\n",
			     __FUNCTION__, bo->delta, bo->handle));
			delta += bo->delta;
			assert(bo->handle == bo->proxy->handle);
			/* need to release the cache upon batch submit */
			if (bo->exec == NULL) {
				list_move_tail(&bo->request,
					       &kgem->next_request->buffers);
				bo->rq = MAKE_REQUEST(kgem->next_request,
						      kgem->ring);
				bo->exec = &_kgem_dummy_exec;
				bo->domain = DOMAIN_GPU;
			}

			if (read_write_domain & 0x7fff && !bo->gpu_dirty)
				__kgem_bo_mark_dirty(bo);

			bo = bo->proxy;
			assert(bo->refcnt);
		}
		assert(bo->refcnt);

		if (bo->exec == NULL)
			kgem_add_bo(kgem, bo);
		assert(bo->rq == MAKE_REQUEST(kgem->next_request, kgem->ring));
		assert(RQ_RING(bo->rq) == kgem->ring);

		if (kgem->gen < 040 && read_write_domain & KGEM_RELOC_FENCED) {
			if (bo->tiling &&
			    (bo->exec->flags & EXEC_OBJECT_NEEDS_FENCE) == 0) {
				assert(bo->tiling == I915_TILING_X);
				assert(kgem->nfence < kgem->fence_max);
				kgem->aperture_fenced +=
					kgem_bo_fenced_size(kgem, bo);
				kgem->nfence++;
			}
			bo->exec->flags |= EXEC_OBJECT_NEEDS_FENCE;
		}

		kgem->reloc[index].delta = delta;
		kgem->reloc[index].target_handle = bo->target_handle;
		kgem->reloc[index].presumed_offset = bo->presumed_offset;

		if (read_write_domain & 0x7fff && !bo->gpu_dirty) {
			assert(!bo->snoop || kgem->can_blt_cpu);
			__kgem_bo_mark_dirty(bo);
		}

		delta += bo->presumed_offset;
	} else {
		kgem->reloc[index].delta = delta;
		kgem->reloc[index].target_handle = ~0U;
		kgem->reloc[index].presumed_offset = 0;
		if (kgem->nreloc__self < 256)
			kgem->reloc__self[kgem->nreloc__self++] = index;
	}
	kgem->reloc[index].read_domains = read_write_domain >> 16;
	kgem->reloc[index].write_domain = read_write_domain & 0x7fff;

	return delta;
}

uint64_t kgem_add_reloc64(struct kgem *kgem,
			  uint32_t pos,
			  struct kgem_bo *bo,
			  uint32_t read_write_domain,
			  uint64_t delta)
{
	int index;

	DBG(("%s: handle=%d, pos=%d, delta=%ld, domains=%08x\n",
	     __FUNCTION__, bo ? bo->handle : 0, pos, (long)delta, read_write_domain));

	assert(kgem->gen >= 0100);
	assert((read_write_domain & 0x7fff) == 0 || bo != NULL);

	index = kgem->nreloc++;
	assert(index < ARRAY_SIZE(kgem->reloc));
	kgem->reloc[index].offset = pos * sizeof(kgem->batch[0]);
	if (bo) {
		assert(kgem->mode != KGEM_NONE);
		assert(bo->refcnt);
		while (bo->proxy) {
			DBG(("%s: adding proxy [delta=%ld] for handle=%d\n",
			     __FUNCTION__, (long)bo->delta, bo->handle));
			delta += bo->delta;
			assert(bo->handle == bo->proxy->handle);
			/* need to release the cache upon batch submit */
			if (bo->exec == NULL) {
				list_move_tail(&bo->request,
					       &kgem->next_request->buffers);
				bo->rq = MAKE_REQUEST(kgem->next_request,
						      kgem->ring);
				bo->exec = &_kgem_dummy_exec;
				bo->domain = DOMAIN_GPU;
			}

			if (read_write_domain & 0x7fff && !bo->gpu_dirty)
				__kgem_bo_mark_dirty(bo);

			bo = bo->proxy;
			assert(bo->refcnt);
		}
		assert(bo->refcnt);

		if (bo->exec == NULL)
			kgem_add_bo(kgem, bo);
		assert(bo->rq == MAKE_REQUEST(kgem->next_request, kgem->ring));
		assert(RQ_RING(bo->rq) == kgem->ring);

		DBG(("%s[%d] = (delta=%d, target handle=%d, presumed=%llx)\n",
					__FUNCTION__, index, delta, bo->target_handle, (long long)bo->presumed_offset));
		kgem->reloc[index].delta = delta;
		kgem->reloc[index].target_handle = bo->target_handle;
		kgem->reloc[index].presumed_offset = bo->presumed_offset;

		if (read_write_domain & 0x7fff && !bo->gpu_dirty) {
			assert(!bo->snoop || kgem->can_blt_cpu);
			__kgem_bo_mark_dirty(bo);
		}

		delta += bo->presumed_offset;
	} else {
		DBG(("%s[%d] = (delta=%d, target handle=batch)\n",
					__FUNCTION__, index, delta));
		kgem->reloc[index].delta = delta;
		kgem->reloc[index].target_handle = ~0U;
		kgem->reloc[index].presumed_offset = 0;
		if (kgem->nreloc__self < 256)
			kgem->reloc__self[kgem->nreloc__self++] = index;
	}
	kgem->reloc[index].read_domains = read_write_domain >> 16;
	kgem->reloc[index].write_domain = read_write_domain & 0x7fff;

	return delta;
}

static void kgem_trim_vma_cache(struct kgem *kgem, int type, int bucket)
{
	int i, j;

	DBG(("%s: type=%d, count=%d (bucket: %d)\n",
	     __FUNCTION__, type, kgem->vma[type].count, bucket));
	if (kgem->vma[type].count <= 0)
	       return;

	if (kgem->need_purge)
		kgem_purge_cache(kgem);

	/* vma are limited on a per-process basis to around 64k.
	 * This includes all malloc arenas as well as other file
	 * mappings. In order to be fair and not hog the cache,
	 * and more importantly not to exhaust that limit and to
	 * start failing mappings, we keep our own number of open
	 * vma to within a conservative value.
	 */
	i = 0;
	while (kgem->vma[type].count > 0) {
		struct kgem_bo *bo = NULL;

		for (j = 0;
		     bo == NULL && j < ARRAY_SIZE(kgem->vma[type].inactive);
		     j++) {
			struct list *head = &kgem->vma[type].inactive[i++%ARRAY_SIZE(kgem->vma[type].inactive)];
			if (!list_is_empty(head))
				bo = list_last_entry(head, struct kgem_bo, vma);
		}
		if (bo == NULL)
			break;

		DBG(("%s: discarding inactive %s vma cache for %d\n",
		     __FUNCTION__, type ? "CPU" : "GTT", bo->handle));

		assert(bo->rq == NULL);
		if (type) {
			VG(VALGRIND_MAKE_MEM_NOACCESS(MAP(bo->map__cpu), bytes(bo)));
			munmap(MAP(bo->map__cpu), bytes(bo));
			bo->map__cpu = NULL;
		} else {
			if (bo->map__wc) {
				VG(VALGRIND_MAKE_MEM_NOACCESS(bo->map__wc, bytes(bo)));
				munmap(bo->map__wc, bytes(bo));
				bo->map__wc = NULL;
			}
			if (bo->map__gtt) {
				munmap(bo->map__gtt, bytes(bo));
				bo->map__gtt = NULL;
			}
		}

		list_del(&bo->vma);
		kgem->vma[type].count--;
	}
}

static void *__kgem_bo_map__gtt_or_wc(struct kgem *kgem, struct kgem_bo *bo)
{
	void *ptr;

	DBG(("%s: handle=%d\n", __FUNCTION__, bo->handle));

	assert(bo->proxy == NULL);
	assert(!bo->snoop);

	kgem_trim_vma_cache(kgem, MAP_GTT, bucket(bo));

	if (bo->tiling || !kgem->has_wc_mmap) {
		assert(kgem->gen != 021 || bo->tiling != I915_TILING_Y);
		warn_unless(num_pages(bo) <= kgem->aperture_mappable / 2);

		ptr = bo->map__gtt;
		if (ptr == NULL)
			ptr = __kgem_bo_map__gtt(kgem, bo);
	} else {
		ptr = bo->map__wc;
		if (ptr == NULL)
			ptr = __kgem_bo_map__wc(kgem, bo);
	}

	return ptr;
}

void *kgem_bo_map__async(struct kgem *kgem, struct kgem_bo *bo)
{
	DBG(("%s: handle=%d, offset=%ld, tiling=%d, map=%p:%p, domain=%d\n", __FUNCTION__,
	     bo->handle, (long)bo->presumed_offset, bo->tiling, bo->map__gtt, bo->map__cpu, bo->domain));

	assert(bo->proxy == NULL);
	assert(list_is_empty(&bo->list));
	assert_tiling(kgem, bo);
	assert(!bo->purged || bo->reusable);

	if (bo->tiling == I915_TILING_NONE && !bo->scanout && kgem->has_llc) {
		DBG(("%s: converting request for GTT map into CPU map\n",
		     __FUNCTION__));
		return kgem_bo_map__cpu(kgem, bo);
	}

	return __kgem_bo_map__gtt_or_wc(kgem, bo);
}

void *kgem_bo_map(struct kgem *kgem, struct kgem_bo *bo)
{
	void *ptr;

	DBG(("%s: handle=%d, offset=%ld, tiling=%d, map=%p:%p, domain=%d\n", __FUNCTION__,
	     bo->handle, (long)bo->presumed_offset, bo->tiling, bo->map__gtt, bo->map__cpu, bo->domain));

	assert(bo->proxy == NULL);
	assert(list_is_empty(&bo->list));
	assert(bo->exec == NULL);
	assert_tiling(kgem, bo);
	assert(!bo->purged || bo->reusable);

	if (bo->tiling == I915_TILING_NONE && !bo->scanout &&
	    (kgem->has_llc || bo->domain == DOMAIN_CPU)) {
		DBG(("%s: converting request for GTT map into CPU map\n",
		     __FUNCTION__));
		ptr = kgem_bo_map__cpu(kgem, bo);
		if (ptr)
			kgem_bo_sync__cpu(kgem, bo);
		return ptr;
	}

	ptr = __kgem_bo_map__gtt_or_wc(kgem, bo);

	if (bo->domain != DOMAIN_GTT || FORCE_MMAP_SYNC & (1 << DOMAIN_GTT)) {
		struct drm_i915_gem_set_domain set_domain;

		DBG(("%s: sync: needs_flush? %d, domain? %d, busy? %d\n", __FUNCTION__,
		     bo->needs_flush, bo->domain, __kgem_busy(kgem, bo->handle)));

		/* XXX use PROT_READ to avoid the write flush? */

		VG_CLEAR(set_domain);
		set_domain.handle = bo->handle;
		set_domain.read_domains = I915_GEM_DOMAIN_GTT;
		set_domain.write_domain = I915_GEM_DOMAIN_GTT;
		if (do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain)) {
			DBG(("%s: sync: GPU hang detected\n", __FUNCTION__));
			kgem_throttle(kgem);
		}
		bo->needs_flush = false;
		kgem_bo_retire(kgem, bo);
		bo->domain = DOMAIN_GTT;
		bo->gtt_dirty = true;
	}

	return ptr;
}

void *kgem_bo_map__gtt(struct kgem *kgem, struct kgem_bo *bo)
{
	DBG(("%s: handle=%d, offset=%ld, tiling=%d, map=%p:%p, domain=%d\n", __FUNCTION__,
	     bo->handle, (long)bo->presumed_offset, bo->tiling, bo->map__gtt, bo->map__cpu, bo->domain));

	assert(bo->proxy == NULL);
	assert(bo->exec == NULL);
	assert(list_is_empty(&bo->list));
	assert_tiling(kgem, bo);
	assert(!bo->purged || bo->reusable);

	return __kgem_bo_map__gtt_or_wc(kgem, bo);
}

void *kgem_bo_map__wc(struct kgem *kgem, struct kgem_bo *bo)
{
	DBG(("%s: handle=%d, offset=%ld, tiling=%d, map=%p:%p, domain=%d\n", __FUNCTION__,
	     bo->handle, (long)bo->presumed_offset, bo->tiling, bo->map__gtt, bo->map__cpu, bo->domain));

	assert(bo->proxy == NULL);
	assert(list_is_empty(&bo->list));
	assert_tiling(kgem, bo);
	assert(!bo->purged || bo->reusable);

	if (bo->map__wc)
		return bo->map__wc;
	if (!kgem->has_wc_mmap)
		return NULL;

	kgem_trim_vma_cache(kgem, MAP_GTT, bucket(bo));
	return __kgem_bo_map__wc(kgem, bo);
}

void *kgem_bo_map__cpu(struct kgem *kgem, struct kgem_bo *bo)
{
	DBG(("%s(handle=%d, size=%d, map=%p:%p)\n",
	     __FUNCTION__, bo->handle, bytes(bo), bo->map__gtt, bo->map__cpu));
	assert(!bo->purged);
	assert(list_is_empty(&bo->list));
	assert(bo->proxy == NULL);
	assert_tiling(kgem, bo);

	if (bo->map__cpu)
		return MAP(bo->map__cpu);

	kgem_trim_vma_cache(kgem, MAP_CPU, bucket(bo));

	return __kgem_bo_map__cpu(kgem, bo);
}

void *kgem_bo_map__debug(struct kgem *kgem, struct kgem_bo *bo)
{
	void *ptr;

	if (bo->tiling == I915_TILING_NONE && kgem->has_llc) {
		ptr = MAP(bo->map__cpu);
		if (ptr == NULL)
			ptr = __kgem_bo_map__cpu(kgem, bo);
	} else if (bo->tiling || !kgem->has_wc_mmap) {
		ptr = bo->map__gtt;
		if (ptr == NULL)
			ptr = __kgem_bo_map__gtt(kgem, bo);
	} else {
		ptr = bo->map__wc;
		if (ptr == NULL)
			ptr = __kgem_bo_map__wc(kgem, bo);
	}

	return ptr;
}


uint32_t kgem_bo_flink(struct kgem *kgem, struct kgem_bo *bo)
{
	struct drm_gem_flink flink;

	assert(kgem_bo_is_fenced(kgem, bo));

	VG_CLEAR(flink);
	flink.handle = bo->handle;
	if (do_ioctl(kgem->fd, DRM_IOCTL_GEM_FLINK, &flink))
		return 0;

	DBG(("%s: flinked handle=%d to name=%d, marking non-reusable\n",
	     __FUNCTION__, flink.handle, flink.name));

	/* Ordinarily giving the name aware makes the buffer non-reusable.
	 * However, we track the lifetime of all clients and their hold
	 * on the buffer, and *presuming* they do not pass it on to a third
	 * party, we track the lifetime accurately.
	 */
	bo->reusable = false;
	kgem_bo_unclean(kgem, bo);

	return flink.name;
}

struct kgem_bo *kgem_create_map(struct kgem *kgem,
				void *ptr, uint32_t size,
				bool read_only)
{
	struct kgem_bo *bo;
	uintptr_t first_page, last_page;
	uint32_t handle;

	assert(MAP(ptr) == ptr);

	DBG(("%s(%p size=%d, read-only?=%d) - has_userptr?=%d\n", __FUNCTION__,
	     ptr, size, read_only, kgem->has_userptr));
	if (!kgem->has_userptr)
		return NULL;

	first_page = (uintptr_t)ptr;
	last_page = first_page + size + PAGE_SIZE - 1;

	first_page &= ~(uintptr_t)(PAGE_SIZE-1);
	last_page &= ~(uintptr_t)(PAGE_SIZE-1);
	assert(last_page > first_page);

	handle = gem_userptr(kgem->fd,
			     (void *)first_page, last_page-first_page,
			     read_only);
	if (handle == 0) {
		if (read_only && kgem->has_wc_mmap) {
			struct drm_i915_gem_set_domain set_domain;

			handle = gem_userptr(kgem->fd,
					     (void *)first_page, last_page-first_page,
					     false);

			VG_CLEAR(set_domain);
			set_domain.handle = handle;
			set_domain.read_domains = I915_GEM_DOMAIN_GTT;
			set_domain.write_domain = 0;
			if (do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain)) {
				gem_close(kgem->fd, handle);
				handle = 0;
			}
		}
		if (handle == 0) {
			DBG(("%s: import failed, errno=%d\n", __FUNCTION__, errno));
			return NULL;
		}
	}

	bo = __kgem_bo_alloc(handle, (last_page - first_page) / PAGE_SIZE);
	if (bo == NULL) {
		gem_close(kgem->fd, handle);
		return NULL;
	}

	bo->unique_id = kgem_get_unique_id(kgem);
	bo->snoop = !kgem->has_llc;
	debug_alloc__bo(kgem, bo);

	if (first_page != (uintptr_t)ptr) {
		struct kgem_bo *proxy;

		proxy = kgem_create_proxy(kgem, bo,
					  (uintptr_t)ptr - first_page, size);
		kgem_bo_destroy(kgem, bo);
		if (proxy == NULL)
			return NULL;

		bo = proxy;
	}

	bo->map__cpu = MAKE_USER_MAP(ptr);

	DBG(("%s(ptr=%p, size=%d, pages=%d, read_only=%d) => handle=%d (proxy? %d)\n",
	     __FUNCTION__, ptr, size, NUM_PAGES(size), read_only, handle, bo->proxy != NULL));
	return bo;
}

void kgem_bo_sync__cpu(struct kgem *kgem, struct kgem_bo *bo)
{
	DBG(("%s: handle=%d\n", __FUNCTION__, bo->handle));
	assert(!bo->scanout);
	assert_tiling(kgem, bo);

	kgem_bo_submit(kgem, bo);

	/* SHM pixmaps use proxies for subpage offsets */
	assert(!bo->purged);
	while (bo->proxy)
		bo = bo->proxy;
	assert(!bo->purged);

	if (bo->domain != DOMAIN_CPU || FORCE_MMAP_SYNC & (1 << DOMAIN_CPU)) {
		struct drm_i915_gem_set_domain set_domain;

		DBG(("%s: SYNC: handle=%d, needs_flush? %d, domain? %d, busy? %d\n",
		     __FUNCTION__, bo->handle,
		     bo->needs_flush, bo->domain,
		     __kgem_busy(kgem, bo->handle)));

		VG_CLEAR(set_domain);
		set_domain.handle = bo->handle;
		set_domain.read_domains = I915_GEM_DOMAIN_CPU;
		set_domain.write_domain = I915_GEM_DOMAIN_CPU;

		if (do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain)) {
			DBG(("%s: sync: GPU hang detected\n", __FUNCTION__));
			kgem_throttle(kgem);
		}
		bo->needs_flush = false;
		kgem_bo_retire(kgem, bo);
		bo->domain = DOMAIN_CPU;
		bo->gtt_dirty = true;
	}
}

void kgem_bo_sync__cpu_full(struct kgem *kgem, struct kgem_bo *bo, bool write)
{
	DBG(("%s: handle=%d\n", __FUNCTION__, bo->handle));
	assert(!bo->scanout || !write);
	assert_tiling(kgem, bo);

	if (write || bo->needs_flush)
		kgem_bo_submit(kgem, bo);

	/* SHM pixmaps use proxies for subpage offsets */
	assert(!bo->purged);
	assert(bo->refcnt);
	while (bo->proxy)
		bo = bo->proxy;
	assert(bo->refcnt);
	assert(!bo->purged);

	if (bo->rq == NULL && (kgem->has_llc || bo->snoop) && !write)
		return;

	if (bo->domain != DOMAIN_CPU || FORCE_MMAP_SYNC & (1 << DOMAIN_CPU)) {
		struct drm_i915_gem_set_domain set_domain;

		DBG(("%s: SYNC: handle=%d, needs_flush? %d, domain? %d, busy? %d\n",
		     __FUNCTION__, bo->handle,
		     bo->needs_flush, bo->domain,
		     __kgem_busy(kgem, bo->handle)));

		VG_CLEAR(set_domain);
		set_domain.handle = bo->handle;
		set_domain.read_domains = I915_GEM_DOMAIN_CPU;
		set_domain.write_domain = write ? I915_GEM_DOMAIN_CPU : 0;

		if (do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain)) {
			DBG(("%s: sync: GPU hang detected\n", __FUNCTION__));
			kgem_throttle(kgem);
		}
		bo->needs_flush = false;
		if (write) {
			kgem_bo_retire(kgem, bo);
			bo->domain = DOMAIN_CPU;
			bo->gtt_dirty = true;
		} else {
			if (bo->exec == NULL)
				kgem_bo_maybe_retire(kgem, bo);
			bo->domain = DOMAIN_NONE;
		}
	}
}

void kgem_bo_sync__gtt(struct kgem *kgem, struct kgem_bo *bo)
{
	DBG(("%s: handle=%d\n", __FUNCTION__, bo->handle));
	assert(bo->refcnt);
	assert(bo->proxy == NULL);
	assert_tiling(kgem, bo);
	assert(!bo->snoop);

	kgem_bo_submit(kgem, bo);

	if (bo->domain != DOMAIN_GTT ||
	    !kgem->has_coherent_mmap_gtt ||
	    FORCE_MMAP_SYNC & (1 << DOMAIN_GTT)) {
		struct drm_i915_gem_set_domain set_domain;

		DBG(("%s: SYNC: handle=%d, needs_flush? %d, domain? %d, busy? %d\n",
		     __FUNCTION__, bo->handle,
		     bo->needs_flush, bo->domain,
		     __kgem_busy(kgem, bo->handle)));

		VG_CLEAR(set_domain);
		set_domain.handle = bo->handle;
		set_domain.read_domains = I915_GEM_DOMAIN_GTT;
		set_domain.write_domain = I915_GEM_DOMAIN_GTT;

		if (do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain)) {
			DBG(("%s: sync: GPU hang detected\n", __FUNCTION__));
			kgem_throttle(kgem);
		}
		bo->needs_flush = false;
		kgem_bo_retire(kgem, bo);
		bo->domain = DOMAIN_GTT;
		bo->gtt_dirty = true;
	}
}

void kgem_clear_dirty(struct kgem *kgem)
{
	struct list * const buffers = &kgem->next_request->buffers;
	struct kgem_bo *bo;

	list_for_each_entry(bo, buffers, request) {
		if (!bo->gpu_dirty)
			break;

		bo->gpu_dirty = false;
	}
}

struct kgem_bo *kgem_create_proxy(struct kgem *kgem,
				  struct kgem_bo *target,
				  int offset, int length)
{
	struct kgem_bo *bo;

	DBG(("%s: target handle=%d [proxy? %d], offset=%d, length=%d, io=%d\n",
	     __FUNCTION__, target->handle, target->proxy ? target->proxy->delta : -1,
	     offset, length, target->io));

	bo = __kgem_bo_alloc(target->handle, length);
	if (bo == NULL)
		return NULL;

	bo->unique_id = kgem_get_unique_id(kgem);
	bo->reusable = false;
	bo->size.bytes = length;

	bo->io = target->io && target->proxy == NULL;
	bo->gpu_dirty = target->gpu_dirty;
	bo->tiling = target->tiling;
	bo->pitch = target->pitch;
	bo->flush = target->flush;
	bo->snoop = target->snoop;

	assert(!bo->scanout);
	bo->proxy = kgem_bo_reference(target);
	bo->delta = offset;

	/* Proxies are only tracked for busyness on the current rq */
	if (target->exec && !bo->io) {
		assert(RQ(target->rq) == kgem->next_request);
		list_move_tail(&bo->request, &kgem->next_request->buffers);
		bo->exec = &_kgem_dummy_exec;
		bo->rq = target->rq;
	}

	return bo;
}

static struct kgem_buffer *
buffer_alloc(void)
{
	struct kgem_buffer *bo;

	bo = malloc(sizeof(*bo));
	if (bo == NULL)
		return NULL;

	bo->mem = NULL;
	bo->need_io = false;
	bo->mmapped = MMAPPED_CPU;

	return bo;
}

static struct kgem_buffer *
buffer_alloc_with_data(int num_pages)
{
	struct kgem_buffer *bo;

	bo = malloc(sizeof(*bo) + 2*UPLOAD_ALIGNMENT + num_pages * PAGE_SIZE);
	if (bo == NULL)
		return NULL;

	bo->mem = (void *)ALIGN((uintptr_t)bo + sizeof(*bo), UPLOAD_ALIGNMENT);
	bo->mmapped = false;
	return bo;
}

static inline bool
use_snoopable_buffer(struct kgem *kgem, uint32_t flags)
{
	if ((flags & KGEM_BUFFER_WRITE) == 0)
		return kgem->gen >= 030;

	return true;
}

static void
init_buffer_from_bo(struct kgem_buffer *bo, struct kgem_bo *old)
{
	DBG(("%s: reusing handle=%d for buffer\n",
	     __FUNCTION__, old->handle));

	assert(old->proxy == NULL);
	assert(list_is_empty(&old->list));

	memcpy(&bo->base, old, sizeof(*old));
	if (old->rq)
		list_replace(&old->request, &bo->base.request);
	else
		list_init(&bo->base.request);
	list_replace(&old->vma, &bo->base.vma);
	list_init(&bo->base.list);
	free(old);

	assert(bo->base.tiling == I915_TILING_NONE);

	bo->base.refcnt = 1;
}

static struct kgem_buffer *
search_snoopable_buffer(struct kgem *kgem, unsigned alloc)
{
	struct kgem_buffer *bo;
	struct kgem_bo *old;

	old = search_snoop_cache(kgem, alloc, 0);
	if (old) {
		if (!old->io) {
			bo = buffer_alloc();
			if (bo == NULL)
				return NULL;

			init_buffer_from_bo(bo, old);
		} else {
			bo = (struct kgem_buffer *)old;
			bo->base.refcnt = 1;
		}

		DBG(("%s: created CPU handle=%d for buffer, size %d\n",
		     __FUNCTION__, bo->base.handle, num_pages(&bo->base)));

		assert(bo->base.snoop);
		assert(bo->base.tiling == I915_TILING_NONE);
		assert(num_pages(&bo->base) >= alloc);
		assert(bo->mmapped == MMAPPED_CPU);
		assert(bo->need_io == false);

		bo->mem = kgem_bo_map__cpu(kgem, &bo->base);
		if (bo->mem == NULL) {
			bo->base.refcnt = 0;
			kgem_bo_free(kgem, &bo->base);
			bo = NULL;
		}

		return bo;
	}

	return NULL;
}

static struct kgem_buffer *
create_snoopable_buffer(struct kgem *kgem, unsigned alloc)
{
	struct kgem_buffer *bo;
	uint32_t handle;

	if (kgem->has_llc) {
		struct kgem_bo *old;

		bo = buffer_alloc();
		if (bo == NULL)
			return NULL;

		old = search_linear_cache(kgem, alloc,
					 CREATE_INACTIVE | CREATE_CPU_MAP | CREATE_EXACT);
		if (old) {
			init_buffer_from_bo(bo, old);
		} else {
			handle = gem_create(kgem->fd, alloc);
			if (handle == 0) {
				free(bo);
				return NULL;
			}

			__kgem_bo_init(&bo->base, handle, alloc);
			debug_alloc__bo(kgem, &bo->base);
			DBG(("%s: created CPU (LLC) handle=%d for buffer, size %d\n",
			     __FUNCTION__, bo->base.handle, alloc));
		}

		assert(bo->base.refcnt == 1);
		assert(bo->mmapped == MMAPPED_CPU);
		assert(bo->need_io == false);

		bo->mem = kgem_bo_map__cpu(kgem, &bo->base);
		if (bo->mem != NULL)
			return bo;

		bo->base.refcnt = 0; /* for valgrind */
		kgem_bo_free(kgem, &bo->base);
	}

	if (kgem->has_caching) {
		struct kgem_bo *old;

		bo = buffer_alloc();
		if (bo == NULL)
			return NULL;

		old = search_linear_cache(kgem, alloc,
					 CREATE_INACTIVE | CREATE_CPU_MAP | CREATE_EXACT);
		if (old) {
			init_buffer_from_bo(bo, old);
		} else {
			handle = gem_create(kgem->fd, alloc);
			if (handle == 0) {
				free(bo);
				return NULL;
			}

			__kgem_bo_init(&bo->base, handle, alloc);
			debug_alloc__bo(kgem, &bo->base);
			DBG(("%s: created CPU handle=%d for buffer, size %d\n",
			     __FUNCTION__, bo->base.handle, alloc));
		}

		assert(bo->base.refcnt == 1);
		assert(bo->mmapped == MMAPPED_CPU);
		assert(bo->need_io == false);
		assert(!__kgem_busy(kgem, bo->base.handle));

		if (!gem_set_caching(kgem->fd, bo->base.handle, SNOOPED))
			goto free_caching;

		bo->base.snoop = true;

		bo->mem = kgem_bo_map__cpu(kgem, &bo->base);
		if (bo->mem == NULL)
			goto free_caching;

		return bo;

free_caching:
		bo->base.refcnt = 0; /* for valgrind */
		kgem_bo_free(kgem, &bo->base);
	}

	if (kgem->has_userptr) {
		bo = buffer_alloc();
		if (bo == NULL)
			return NULL;

		//if (posix_memalign(&ptr, 64, ALIGN(size, 64)))
		if (posix_memalign(&bo->mem, PAGE_SIZE, alloc * PAGE_SIZE)) {
			free(bo);
			return NULL;
		}

		handle = gem_userptr(kgem->fd, bo->mem, alloc * PAGE_SIZE, false);
		if (handle == 0) {
			free(bo->mem);
			free(bo);
			return NULL;
		}

		__kgem_bo_init(&bo->base, handle, alloc);
		debug_alloc__bo(kgem, &bo->base);
		DBG(("%s: created snoop handle=%d for buffer\n",
		     __FUNCTION__, bo->base.handle));

		assert(bo->mmapped == MMAPPED_CPU);
		assert(bo->need_io == false);

		bo->base.refcnt = 1;
		bo->base.snoop = true;
		bo->base.map__cpu = MAKE_USER_MAP(bo->mem);

		return bo;
	}

	return NULL;
}

struct kgem_bo *kgem_create_buffer(struct kgem *kgem,
				   uint32_t size, uint32_t flags,
				   void **ret)
{
	struct kgem_buffer *bo;
	unsigned offset, alloc;
	struct kgem_bo *old;

	DBG(("%s: size=%d, flags=%x [write?=%d, inplace?=%d, last?=%d]\n",
	     __FUNCTION__, size, flags,
	     !!(flags & KGEM_BUFFER_WRITE),
	     !!(flags & KGEM_BUFFER_INPLACE),
	     !!(flags & KGEM_BUFFER_LAST)));
	assert(size);
	/* we should never be asked to create anything TOO large */
	assert(size <= kgem->max_object_size);

#if !DBG_NO_UPLOAD_CACHE
	list_for_each_entry(bo, &kgem->batch_buffers, base.list) {
		assert(bo->base.io);
		assert(bo->base.refcnt >= 1);

		/* We can reuse any write buffer which we can fit */
		if (flags == KGEM_BUFFER_LAST &&
		    bo->write == KGEM_BUFFER_WRITE &&
		    bo->base.refcnt == 1 &&
		    bo->mmapped == MMAPPED_NONE &&
		    size <= bytes(&bo->base)) {
			DBG(("%s: reusing write buffer for read of %d bytes? used=%d, total=%d\n",
			     __FUNCTION__, size, bo->used, bytes(&bo->base)));
			gem_write__cachealigned(kgem->fd, bo->base.handle,
						0, bo->used, bo->mem);
			assert(list_is_empty(&bo->base.vma));
			bo->need_io = 0;
			bo->write = 0;
			offset = 0;
			bo->used = size;
			goto done;
		}

		if (flags & KGEM_BUFFER_WRITE) {
			if ((bo->write & KGEM_BUFFER_WRITE) == 0 ||
			    (((bo->write & ~flags) & KGEM_BUFFER_INPLACE) &&
			     !bo->base.snoop)) {
				DBG(("%s: skip write %x buffer, need %x\n",
				     __FUNCTION__, bo->write, flags));
				continue;
			}
			assert(bo->mmapped || bo->need_io);
		} else {
			if (bo->write & KGEM_BUFFER_WRITE) {
				DBG(("%s: skip write %x buffer, need %x\n",
				     __FUNCTION__, bo->write, flags));
				continue;
			}
		}

		if (bo->used + size <= bytes(&bo->base)) {
			DBG(("%s: reusing buffer? used=%d + size=%d, total=%d\n",
			     __FUNCTION__, bo->used, size, bytes(&bo->base)));
			offset = bo->used;
			bo->used += size;
			goto done;
		}
	}

	if (flags & KGEM_BUFFER_WRITE) {
		list_for_each_entry(bo, &kgem->active_buffers, base.list) {
			assert(bo->base.io);
			assert(bo->base.refcnt >= 1);
			assert(bo->base.exec == NULL);
			assert(bo->mmapped);
			assert(bo->mmapped == MMAPPED_GTT || kgem->has_llc || bo->base.snoop);

			if ((bo->write & ~flags) & KGEM_BUFFER_INPLACE && !bo->base.snoop) {
				DBG(("%s: skip write %x buffer, need %x\n",
				     __FUNCTION__, bo->write, flags));
				continue;
			}

			if (bo->used + size <= bytes(&bo->base)) {
				DBG(("%s: reusing buffer? used=%d + size=%d, total=%d\n",
				     __FUNCTION__, bo->used, size, bytes(&bo->base)));
				offset = bo->used;
				bo->used += size;
				list_move(&bo->base.list, &kgem->batch_buffers);
				goto done;
			}

			if (bo->base.refcnt == 1 &&
			    size <= bytes(&bo->base) &&
			    (bo->base.rq == NULL ||
			     !__kgem_busy(kgem, bo->base.handle))) {
				DBG(("%s: reusing whole buffer? size=%d, total=%d\n",
				     __FUNCTION__, size, bytes(&bo->base)));
				__kgem_bo_clear_busy(&bo->base);
				assert(list_is_empty(&bo->base.vma));

				switch (bo->mmapped) {
				case MMAPPED_CPU:
					kgem_bo_sync__cpu(kgem, &bo->base);
					break;
				case MMAPPED_GTT:
					kgem_bo_sync__gtt(kgem, &bo->base);
					break;
				}

				offset = 0;
				bo->used = size;
				list_move(&bo->base.list, &kgem->batch_buffers);
				goto done;
			}
		}
	}
#endif

#if !DBG_NO_MAP_UPLOAD
	/* Be a little more generous and hope to hold fewer mmappings */
	alloc = ALIGN(2*size, kgem->buffer_size);
	if (alloc > MAX_CACHE_SIZE)
		alloc = ALIGN(size, kgem->buffer_size);
	if (alloc > MAX_CACHE_SIZE)
		alloc = PAGE_ALIGN(size);
	assert(alloc);

	alloc /= PAGE_SIZE;
	if (alloc > kgem->aperture_mappable / 4 && !kgem->has_wc_mmap)
		flags &= ~KGEM_BUFFER_INPLACE;

	if (kgem->has_llc &&
	    (flags & KGEM_BUFFER_WRITE_INPLACE) != KGEM_BUFFER_WRITE_INPLACE) {
		bo = buffer_alloc();
		if (bo == NULL)
			goto skip_llc;

		old = NULL;
		if ((flags & KGEM_BUFFER_WRITE) == 0)
			old = search_linear_cache(kgem, alloc, CREATE_CPU_MAP);
		if (old == NULL)
			old = search_linear_cache(kgem, alloc, CREATE_INACTIVE | CREATE_CPU_MAP);
		if (old == NULL)
			old = search_linear_cache(kgem, NUM_PAGES(size), CREATE_INACTIVE | CREATE_CPU_MAP);
		if (old) {
			DBG(("%s: found LLC handle=%d for buffer\n",
			     __FUNCTION__, old->handle));

			init_buffer_from_bo(bo, old);
		} else {
			uint32_t handle = gem_create(kgem->fd, alloc);
			if (handle == 0) {
				free(bo);
				goto skip_llc;
			}
			__kgem_bo_init(&bo->base, handle, alloc);
			debug_alloc__bo(kgem, &bo->base);
			DBG(("%s: created LLC handle=%d for buffer\n",
			     __FUNCTION__, bo->base.handle));
		}

		assert(bo->mmapped);
		assert(!bo->need_io);

		bo->mem = kgem_bo_map__cpu(kgem, &bo->base);
		if (bo->mem) {
			if (flags & KGEM_BUFFER_WRITE)
				kgem_bo_sync__cpu(kgem, &bo->base);
			flags &= ~KGEM_BUFFER_INPLACE;
			goto init;
		} else {
			bo->base.refcnt = 0; /* for valgrind */
			kgem_bo_free(kgem, &bo->base);
		}
	}
skip_llc:

	if ((flags & KGEM_BUFFER_WRITE_INPLACE) == KGEM_BUFFER_WRITE_INPLACE) {
		/* The issue with using a GTT upload buffer is that we may
		 * cause eviction-stalls in order to free up some GTT space.
		 * An is-mappable? ioctl could help us detect when we are
		 * about to block, or some per-page magic in the kernel.
		 *
		 * XXX This is especially noticeable on memory constrained
		 * devices like gen2 or with relatively slow gpu like i3.
		 */
		DBG(("%s: searching for an inactive GTT map for upload\n",
		     __FUNCTION__));
		old = search_linear_cache(kgem, alloc,
					  CREATE_EXACT | CREATE_INACTIVE | CREATE_GTT_MAP);
#if HAVE_I915_GEM_BUFFER_INFO
		if (old) {
			struct drm_i915_gem_buffer_info info;

			/* An example of such a non-blocking ioctl might work */

			VG_CLEAR(info);
			info.handle = handle;
			if (do_ioctl(kgem->fd,
				     DRM_IOCTL_I915_GEM_BUFFER_INFO,
				     &fino) == 0) {
				old->presumed_offset = info.addr;
				if ((info.flags & I915_GEM_MAPPABLE) == 0) {
					kgem_bo_move_to_inactive(kgem, old);
					old = NULL;
				}
			}
		}
#endif
		if (old == NULL)
			old = search_linear_cache(kgem, NUM_PAGES(size),
						  CREATE_EXACT | CREATE_INACTIVE | CREATE_GTT_MAP);
		if (old == NULL) {
			old = search_linear_cache(kgem, alloc, CREATE_INACTIVE);
			if (old && !kgem_bo_can_map(kgem, old)) {
				_kgem_bo_destroy(kgem, old);
				old = NULL;
			}
		}
		if (old) {
			DBG(("%s: reusing handle=%d for buffer\n",
			     __FUNCTION__, old->handle));
			assert(kgem_bo_can_map(kgem, old));
			assert(!old->snoop);
			assert(old->rq == NULL);

			bo = buffer_alloc();
			if (bo == NULL)
				return NULL;

			init_buffer_from_bo(bo, old);
			assert(num_pages(&bo->base) >= NUM_PAGES(size));

			assert(bo->mmapped);
			assert(bo->base.refcnt == 1);

			bo->mem = kgem_bo_map(kgem, &bo->base);
			if (bo->mem) {
				if (bo->mem == MAP(bo->base.map__cpu))
					flags &= ~KGEM_BUFFER_INPLACE;
				else
					bo->mmapped = MMAPPED_GTT;
				goto init;
			} else {
				bo->base.refcnt = 0;
				kgem_bo_free(kgem, &bo->base);
			}
		}
	}
#else
	flags &= ~KGEM_BUFFER_INPLACE;
#endif
	/* Be more parsimonious with pwrite/pread/cacheable buffers */
	if ((flags & KGEM_BUFFER_INPLACE) == 0)
		alloc = NUM_PAGES(size);

	if (use_snoopable_buffer(kgem, flags)) {
		bo = search_snoopable_buffer(kgem, alloc);
		if (bo) {
			if (flags & KGEM_BUFFER_WRITE)
				kgem_bo_sync__cpu(kgem, &bo->base);
			flags &= ~KGEM_BUFFER_INPLACE;
			goto init;
		}

		if ((flags & KGEM_BUFFER_INPLACE) == 0) {
			bo = create_snoopable_buffer(kgem, alloc);
			if (bo)
				goto init;
		}
	}

	flags &= ~KGEM_BUFFER_INPLACE;

	old = NULL;
	if ((flags & KGEM_BUFFER_WRITE) == 0)
		old = search_linear_cache(kgem, alloc, 0);
	if (old == NULL)
		old = search_linear_cache(kgem, alloc, CREATE_INACTIVE);
	if (old) {
		DBG(("%s: reusing ordinary handle %d for io\n",
		     __FUNCTION__, old->handle));
		bo = buffer_alloc_with_data(num_pages(old));
		if (bo == NULL)
			return NULL;

		init_buffer_from_bo(bo, old);
		bo->need_io = flags & KGEM_BUFFER_WRITE;
	} else {
		unsigned hint;

		if (use_snoopable_buffer(kgem, flags)) {
			bo = create_snoopable_buffer(kgem, alloc);
			if (bo)
				goto init;
		}

		bo = buffer_alloc();
		if (bo == NULL)
			return NULL;

		hint = CREATE_INACTIVE;
		if (flags & KGEM_BUFFER_WRITE)
			hint |= CREATE_CPU_MAP;
		old = search_linear_cache(kgem, alloc, hint);
		if (old) {
			DBG(("%s: reusing handle=%d for buffer\n",
			     __FUNCTION__, old->handle));

			init_buffer_from_bo(bo, old);
		} else {
			uint32_t handle = gem_create(kgem->fd, alloc);
			if (handle == 0) {
				free(bo);
				return NULL;
			}

			DBG(("%s: created handle=%d for buffer\n",
			     __FUNCTION__, handle));

			__kgem_bo_init(&bo->base, handle, alloc);
			debug_alloc__bo(kgem, &bo->base);
		}

		assert(bo->mmapped);
		assert(!bo->need_io);
		assert(bo->base.refcnt == 1);

		if (flags & KGEM_BUFFER_WRITE) {
			bo->mem = kgem_bo_map__cpu(kgem, &bo->base);
			if (bo->mem != NULL) {
				kgem_bo_sync__cpu(kgem, &bo->base);
				goto init;
			}
		}

		DBG(("%s: failing back to new pwrite buffer\n", __FUNCTION__));
		old = &bo->base;
		bo = buffer_alloc_with_data(num_pages(old));
		if (bo == NULL) {
			old->refcnt= 0;
			kgem_bo_free(kgem, old);
			return NULL;
		}

		init_buffer_from_bo(bo, old);

		assert(bo->mem);
		assert(!bo->mmapped);
		assert(bo->base.refcnt == 1);

		bo->need_io = flags & KGEM_BUFFER_WRITE;
	}
init:
	bo->base.io = true;
	assert(bo->base.refcnt == 1);
	assert(num_pages(&bo->base) >= NUM_PAGES(size));
	assert(!bo->need_io || !bo->base.needs_flush);
	assert(!bo->need_io || bo->base.domain != DOMAIN_GPU);
	assert(bo->mem);
	assert(bo->mmapped != MMAPPED_GTT || bo->base.map__gtt == bo->mem || bo->base.map__wc == bo->mem);
	assert(bo->mmapped != MMAPPED_CPU || MAP(bo->base.map__cpu) == bo->mem);

	bo->used = size;
	bo->write = flags & KGEM_BUFFER_WRITE_INPLACE;
	offset = 0;

	assert(list_is_empty(&bo->base.list));
	list_add(&bo->base.list, &kgem->batch_buffers);

	DBG(("%s(pages=%d [%d]) new handle=%d, used=%d, write=%d\n",
	     __FUNCTION__, num_pages(&bo->base), alloc, bo->base.handle, bo->used, bo->write));

done:
	bo->used = ALIGN(bo->used, UPLOAD_ALIGNMENT);
	assert(bo->used && bo->used <= bytes(&bo->base));
	assert(bo->mem);
	*ret = (char *)bo->mem + offset;
	return kgem_create_proxy(kgem, &bo->base, offset, size);
}

bool kgem_buffer_is_inplace(struct kgem_bo *_bo)
{
	struct kgem_buffer *bo = (struct kgem_buffer *)_bo->proxy;
	return bo->write & KGEM_BUFFER_WRITE_INPLACE;
}

struct kgem_bo *kgem_create_buffer_2d(struct kgem *kgem,
				      int width, int height, int bpp,
				      uint32_t flags,
				      void **ret)
{
	struct kgem_bo *bo;
	int stride;

	assert(width > 0 && height > 0);
	assert(ret != NULL);
	stride = ALIGN(width, 2) * bpp >> 3;
	stride = ALIGN(stride, kgem->gen >= 0100 ? 32 : 4);

	DBG(("%s: %dx%d, %d bpp, stride=%d\n",
	     __FUNCTION__, width, height, bpp, stride));

	bo = kgem_create_buffer(kgem, stride * ALIGN(height, 2), flags, ret);
	if (bo == NULL) {
		DBG(("%s: allocation failure for upload buffer\n",
		     __FUNCTION__));
		return NULL;
	}
	assert(*ret != NULL);
	assert(bo->proxy != NULL);

	if (height & 1) {
		struct kgem_buffer *io = (struct kgem_buffer *)bo->proxy;
		int min;

		assert(io->used);

		/* Having padded this surface to ensure that accesses to
		 * the last pair of rows is valid, remove the padding so
		 * that it can be allocated to other pixmaps.
		 */
		min = bo->delta + height * stride;
		min = ALIGN(min, UPLOAD_ALIGNMENT);
		if (io->used != min) {
			DBG(("%s: trimming buffer from %d to %d\n",
			     __FUNCTION__, io->used, min));
			io->used = min;
		}
		bo->size.bytes -= stride;
	}

	bo->map__cpu = *ret;
	bo->pitch = stride;
	bo->unique_id = kgem_get_unique_id(kgem);
	return bo;
}

struct kgem_bo *kgem_upload_source_image(struct kgem *kgem,
					 const void *data,
					 const BoxRec *box,
					 int stride, int bpp)
{
	int width  = box->x2 - box->x1;
	int height = box->y2 - box->y1;
	struct kgem_bo *bo;
	void *dst;

	if (!kgem_can_create_2d(kgem, width, height, bpp))
		return NULL;

	DBG(("%s : (%d, %d), (%d, %d), stride=%d, bpp=%d\n",
	     __FUNCTION__, box->x1, box->y1, box->x2, box->y2, stride, bpp));

	assert(data);
	assert(width > 0);
	assert(height > 0);
	assert(stride);
	assert(bpp);

	bo = kgem_create_buffer_2d(kgem,
				   width, height, bpp,
				   KGEM_BUFFER_WRITE_INPLACE, &dst);
	if (bo == NULL)
		return NULL;

	if (sigtrap_get()) {
		kgem_bo_destroy(kgem, bo);
		return NULL;
	}

	memcpy_blt(data, dst, bpp,
		   stride, bo->pitch,
		   box->x1, box->y1,
		   0, 0,
		   width, height);

	sigtrap_put();
	return bo;
}

void kgem_proxy_bo_attach(struct kgem_bo *bo,
			  struct kgem_bo **ptr)
{
	DBG(("%s: handle=%d\n", __FUNCTION__, bo->handle));
	assert(bo->map__gtt == NULL);
	assert(bo->proxy);
	list_add(&bo->vma, &bo->proxy->vma);
	bo->map__gtt = ptr;
	*ptr = kgem_bo_reference(bo);
}

void kgem_buffer_read_sync(struct kgem *kgem, struct kgem_bo *_bo)
{
	struct kgem_buffer *bo;
	uint32_t offset = _bo->delta, length = _bo->size.bytes;

	/* We expect the caller to have already submitted the batch */
	assert(_bo->io);
	assert(_bo->exec == NULL);
	assert(_bo->rq == NULL);
	assert(_bo->proxy);

	_bo = _bo->proxy;
	assert(_bo->proxy == NULL);
	assert(_bo->exec == NULL);

	bo = (struct kgem_buffer *)_bo;

	DBG(("%s(offset=%d, length=%d, snooped=%d)\n", __FUNCTION__,
	     offset, length, bo->base.snoop));

	if (bo->mmapped) {
		struct drm_i915_gem_set_domain set_domain;

		DBG(("%s: sync: needs_flush? %d, domain? %d, busy? %d\n",
		     __FUNCTION__,
		     bo->base.needs_flush,
		     bo->base.domain,
		     __kgem_busy(kgem, bo->base.handle)));

		assert(bo->mmapped == MMAPPED_GTT || bo->base.snoop || kgem->has_llc);

		VG_CLEAR(set_domain);
		set_domain.handle = bo->base.handle;
		set_domain.write_domain = 0;
		set_domain.read_domains =
			bo->mmapped == MMAPPED_CPU ? I915_GEM_DOMAIN_CPU : I915_GEM_DOMAIN_GTT;

		if (do_ioctl(kgem->fd, DRM_IOCTL_I915_GEM_SET_DOMAIN, &set_domain)) {
			DBG(("%s: sync: GPU hang detected\n", __FUNCTION__));
			kgem_throttle(kgem);
		}
	} else {
		if (gem_read(kgem->fd,
			     bo->base.handle, (char *)bo->mem+offset,
			     offset, length))
			return;
	}
	kgem_bo_maybe_retire(kgem, &bo->base);
	bo->base.domain = DOMAIN_NONE;
}

uint32_t kgem_bo_get_binding(struct kgem_bo *bo, uint32_t format)
{
	struct kgem_bo_binding *b;

	assert(bo->refcnt);

	for (b = &bo->binding; b && b->offset; b = b->next)
		if (format == b->format)
			return b->offset;

	return 0;
}

void kgem_bo_set_binding(struct kgem_bo *bo, uint32_t format, uint16_t offset)
{
	struct kgem_bo_binding *b;

	assert(bo->refcnt);

	for (b = &bo->binding; b; b = b->next) {
		if (b->offset)
			continue;

		b->offset = offset;
		b->format = format;

		if (b->next)
			b->next->offset = 0;

		return;
	}

	b = malloc(sizeof(*b));
	if (b) {
		b->next = bo->binding.next;
		b->format = format;
		b->offset = offset;
		bo->binding.next = b;
	}
}

struct kgem_bo *
kgem_replace_bo(struct kgem *kgem,
		struct kgem_bo *src,
		uint32_t width,
		uint32_t height,
		uint32_t pitch,
		uint32_t bpp)
{
	struct kgem_bo *dst;
	uint32_t br00, br13;
	uint32_t handle;
	uint32_t size;
	uint32_t *b;

	DBG(("%s: replacing bo handle=%d, size=%dx%d pitch=%d, with pitch=%d\n",
	     __FUNCTION__, src->handle,  width, height, src->pitch, pitch));

	/* We only expect to be called to fixup small buffers, hence why
	 * we only attempt to allocate a linear bo.
	 */
	assert(src->tiling == I915_TILING_NONE);
	assert(kgem_bo_can_blt(kgem, src));

	size = height * pitch;
	size = NUM_PAGES(size);

	dst = search_linear_cache(kgem, size, 0);
	if (dst == NULL)
		dst = search_linear_cache(kgem, size, CREATE_INACTIVE);
	if (dst == NULL) {
		handle = gem_create(kgem->fd, size);
		if (handle == 0)
			return NULL;

		dst = __kgem_bo_alloc(handle, size);
		if (dst == NULL) {
			gem_close(kgem->fd, handle);
			return NULL;
		}

		debug_alloc__bo(kgem, dst);
	}
	dst->pitch = pitch;
	dst->unique_id = kgem_get_unique_id(kgem);
	dst->refcnt = 1;
	assert(dst->tiling == I915_TILING_NONE);
	assert(kgem_bo_can_blt(kgem, dst));

	kgem_set_mode(kgem, KGEM_BLT, dst);
	if (!kgem_check_batch(kgem, 10) ||
	    !kgem_check_reloc(kgem, 2) ||
	    !kgem_check_many_bo_fenced(kgem, src, dst, NULL)) {
		kgem_submit(kgem);
		if (!kgem_check_many_bo_fenced(kgem, src, dst, NULL)) {
			kgem_bo_destroy(kgem, dst);
			return NULL;
		}
		_kgem_set_mode(kgem, KGEM_BLT);
	}
	kgem_bcs_set_tiling(kgem, src, dst);

	br00 = XY_SRC_COPY_BLT_CMD;
	br13 = pitch;
	pitch = src->pitch;
	if (kgem->gen >= 040 && src->tiling) {
		br00 |= BLT_SRC_TILED;
		pitch >>= 2;
	}

	br13 |= 0xcc << 16;
	switch (bpp) {
	default:
	case 32: br00 |= BLT_WRITE_ALPHA | BLT_WRITE_RGB;
		 br13 |= 1 << 25; /* RGB8888 */
	case 16: br13 |= 1 << 24; /* RGB565 */
	case 8: break;
	}

	b = kgem->batch + kgem->nbatch;
	if (kgem->gen >= 0100) {
		b[0] = br00 | 8;
		b[1] = br13;
		b[2] = 0;
		b[3] = height << 16 | width;
		*(uint64_t *)(b+4) =
			kgem_add_reloc64(kgem, kgem->nbatch + 4, dst,
					 I915_GEM_DOMAIN_RENDER << 16 |
					 I915_GEM_DOMAIN_RENDER |
					 KGEM_RELOC_FENCED,
					 0);
		b[6] = 0;
		b[7] = pitch;
		*(uint64_t *)(b+8) =
			kgem_add_reloc64(kgem, kgem->nbatch + 8, src,
					 I915_GEM_DOMAIN_RENDER << 16 |
					 KGEM_RELOC_FENCED,
					 0);
		kgem->nbatch += 10;
	} else {
		b[0] = br00 | 6;
		b[1] = br13;
		b[2] = 0;
		b[3] = height << 16 | width;
		b[4] = kgem_add_reloc(kgem, kgem->nbatch + 4, dst,
				      I915_GEM_DOMAIN_RENDER << 16 |
				      I915_GEM_DOMAIN_RENDER |
				      KGEM_RELOC_FENCED,
				      0);
		b[5] = 0;
		b[6] = pitch;
		b[7] = kgem_add_reloc(kgem, kgem->nbatch + 7, src,
				      I915_GEM_DOMAIN_RENDER << 16 |
				      KGEM_RELOC_FENCED,
				      0);
		kgem->nbatch += 8;
	}

	return dst;
}

bool kgem_bo_convert_to_gpu(struct kgem *kgem,
			    struct kgem_bo *bo,
			    unsigned flags)
{
	DBG(("%s: converting handle=%d from CPU to GPU, flags=%x, busy?=%d\n",
	     __FUNCTION__, bo->handle, flags, __kgem_bo_is_busy(kgem, bo)));
	assert(bo->tiling == I915_TILING_NONE);

	if (flags & (__MOVE_PRIME | __MOVE_SCANOUT))
		return false;

	if (kgem->has_llc)
		return true;

	if (flags & MOVE_ASYNC_HINT && __kgem_bo_is_busy(kgem, bo))
		return false;

	assert(bo->snoop);

	kgem_bo_submit(kgem, bo);

	if (!gem_set_caching(kgem->fd, bo->handle, UNCACHED))
		return false;

	bo->snoop = false;
	return true;
}