2 * Copyright © 2007 Red Hat Inc.
3 * Copyright © 2007-2017 Intel Corporation
4 * Copyright © 2006 VMware, Inc.
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
14 * The above copyright notice and this permission notice (including the next
15 * paragraph) shall be included in all copies or substantial portions of the
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
21 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
23 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 * Authors: Thomas Hellström <thellstrom@vmware.com>
29 * Keith Whitwell <keithw@vmware.com>
30 * Eric Anholt <eric@anholt.net>
31 * Dave Airlie <airlied@linux.ie>
39 #include <util/u_atomic.h>
46 #include <sys/ioctl.h>
48 #include <sys/types.h>
53 #define ETIME ETIMEDOUT
55 #include "common/gen_clflush.h"
56 #include "common/gen_debug.h"
57 #include "dev/gen_device_info.h"
58 #include "libdrm_macros.h"
59 #include "main/macros.h"
60 #include "util/macros.h"
61 #include "util/hash_table.h"
62 #include "util/list.h"
63 #include "brw_bufmgr.h"
64 #include "brw_context.h"
77 /* VALGRIND_FREELIKE_BLOCK unfortunately does not actually undo the earlier
78 * VALGRIND_MALLOCLIKE_BLOCK but instead leaves vg convinced the memory is
79 * leaked. All because it does not call VG(cli_free) from its
80 * VG_USERREQ__FREELIKE_BLOCK handler. Instead of treating the memory like
81 * and allocation, we mark it available for use upon mmapping and remove
84 #define VG_DEFINED(ptr, size) VG(VALGRIND_MAKE_MEM_DEFINED(ptr, size))
85 #define VG_NOACCESS(ptr, size) VG(VALGRIND_MAKE_MEM_NOACCESS(ptr, size))
87 #define PAGE_SIZE 4096
89 #define FILE_DEBUG_FLAG DEBUG_BUFMGR
92 atomic_add_unless(int *v
, int add
, int unless
)
96 while (c
!= unless
&& (old
= p_atomic_cmpxchg(v
, c
, c
+ add
)) != c
)
101 struct bo_cache_bucket
{
102 struct list_head head
;
111 /** Array of lists of cached gem objects of power-of-two sizes */
112 struct bo_cache_bucket cache_bucket
[14 * 4];
116 struct hash_table
*name_table
;
117 struct hash_table
*handle_table
;
122 bool supports_48b_addresses
:1;
125 static int bo_set_tiling_internal(struct brw_bo
*bo
, uint32_t tiling_mode
,
128 static void bo_free(struct brw_bo
*bo
);
131 key_hash_uint(const void *key
)
133 return _mesa_hash_data(key
, 4);
137 key_uint_equal(const void *a
, const void *b
)
139 return *((unsigned *) a
) == *((unsigned *) b
);
142 static struct brw_bo
*
143 hash_find_bo(struct hash_table
*ht
, unsigned int key
)
145 struct hash_entry
*entry
= _mesa_hash_table_search(ht
, &key
);
146 return entry
? (struct brw_bo
*) entry
->data
: NULL
;
150 bo_tile_size(struct brw_bufmgr
*bufmgr
, uint64_t size
, uint32_t tiling
)
152 if (tiling
== I915_TILING_NONE
)
155 /* 965+ just need multiples of page size for tiling */
156 return ALIGN(size
, 4096);
160 * Round a given pitch up to the minimum required for X tiling on a
161 * given chip. We use 512 as the minimum to allow for a later tiling
165 bo_tile_pitch(struct brw_bufmgr
*bufmgr
, uint32_t pitch
, uint32_t tiling
)
167 unsigned long tile_width
;
169 /* If untiled, then just align it so that we can do rendering
170 * to it with the 3D engine.
172 if (tiling
== I915_TILING_NONE
)
173 return ALIGN(pitch
, 64);
175 if (tiling
== I915_TILING_X
)
180 /* 965 is flexible */
181 return ALIGN(pitch
, tile_width
);
185 * This function finds the correct bucket fit for the input size.
186 * The function works with O(1) complexity when the requested size
187 * was queried instead of iterating the size through all the buckets.
189 static struct bo_cache_bucket
*
190 bucket_for_size(struct brw_bufmgr
*bufmgr
, uint64_t size
)
192 /* Calculating the pages and rounding up to the page size. */
193 const unsigned pages
= (size
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
195 /* Row Bucket sizes clz((x-1) | 3) Row Column
196 * in pages stride size
197 * 0: 1 2 3 4 -> 30 30 30 30 4 1
198 * 1: 5 6 7 8 -> 29 29 29 29 4 1
199 * 2: 10 12 14 16 -> 28 28 28 28 8 2
200 * 3: 20 24 28 32 -> 27 27 27 27 16 4
202 const unsigned row
= 30 - __builtin_clz((pages
- 1) | 3);
203 const unsigned row_max_pages
= 4 << row
;
205 /* The '& ~2' is the special case for row 1. In row 1, max pages /
206 * 2 is 2, but the previous row maximum is zero (because there is
207 * no previous row). All row maximum sizes are power of 2, so that
208 * is the only case where that bit will be set.
210 const unsigned prev_row_max_pages
= (row_max_pages
/ 2) & ~2;
211 int col_size_log2
= row
- 1;
212 col_size_log2
+= (col_size_log2
< 0);
214 const unsigned col
= (pages
- prev_row_max_pages
+
215 ((1 << col_size_log2
) - 1)) >> col_size_log2
;
217 /* Calculating the index based on the row and column. */
218 const unsigned index
= (row
* 4) + (col
- 1);
220 return (index
< bufmgr
->num_buckets
) ?
221 &bufmgr
->cache_bucket
[index
] : NULL
;
225 brw_bo_busy(struct brw_bo
*bo
)
227 struct brw_bufmgr
*bufmgr
= bo
->bufmgr
;
228 struct drm_i915_gem_busy busy
= { .handle
= bo
->gem_handle
};
230 int ret
= drmIoctl(bufmgr
->fd
, DRM_IOCTL_I915_GEM_BUSY
, &busy
);
232 bo
->idle
= !busy
.busy
;
239 brw_bo_madvise(struct brw_bo
*bo
, int state
)
241 struct drm_i915_gem_madvise madv
= {
242 .handle
= bo
->gem_handle
,
247 drmIoctl(bo
->bufmgr
->fd
, DRM_IOCTL_I915_GEM_MADVISE
, &madv
);
249 return madv
.retained
;
252 /* drop the oldest entries that have been purged by the kernel */
254 brw_bo_cache_purge_bucket(struct brw_bufmgr
*bufmgr
,
255 struct bo_cache_bucket
*bucket
)
257 list_for_each_entry_safe(struct brw_bo
, bo
, &bucket
->head
, head
) {
258 if (brw_bo_madvise(bo
, I915_MADV_DONTNEED
))
266 static struct brw_bo
*
267 bo_alloc_internal(struct brw_bufmgr
*bufmgr
,
271 uint32_t tiling_mode
,
272 uint32_t stride
, uint64_t alignment
)
275 unsigned int page_size
= getpagesize();
277 struct bo_cache_bucket
*bucket
;
278 bool alloc_from_cache
;
283 if (flags
& BO_ALLOC_BUSY
)
286 if (flags
& BO_ALLOC_ZEROED
)
289 /* BUSY does doesn't really jive with ZEROED as we have to wait for it to
290 * be idle before we can memset. Just disallow that combination.
292 assert(!(busy
&& zeroed
));
294 /* Round the allocated size up to a power of two number of pages. */
295 bucket
= bucket_for_size(bufmgr
, size
);
297 /* If we don't have caching at this size, don't actually round the
300 if (bucket
== NULL
) {
302 if (bo_size
< page_size
)
305 bo_size
= bucket
->size
;
308 mtx_lock(&bufmgr
->lock
);
309 /* Get a buffer out of the cache if available */
311 alloc_from_cache
= false;
312 if (bucket
!= NULL
&& !list_empty(&bucket
->head
)) {
313 if (busy
&& !zeroed
) {
314 /* Allocate new render-target BOs from the tail (MRU)
315 * of the list, as it will likely be hot in the GPU
316 * cache and in the aperture for us. If the caller
317 * asked us to zero the buffer, we don't want this
318 * because we are going to mmap it.
320 bo
= LIST_ENTRY(struct brw_bo
, bucket
->head
.prev
, head
);
322 alloc_from_cache
= true;
323 bo
->align
= alignment
;
325 assert(alignment
== 0);
326 /* For non-render-target BOs (where we're probably
327 * going to map it first thing in order to fill it
328 * with data), check if the last BO in the cache is
329 * unbusy, and only reuse in that case. Otherwise,
330 * allocating a new buffer is probably faster than
331 * waiting for the GPU to finish.
333 bo
= LIST_ENTRY(struct brw_bo
, bucket
->head
.next
, head
);
334 if (!brw_bo_busy(bo
)) {
335 alloc_from_cache
= true;
340 if (alloc_from_cache
) {
341 if (!brw_bo_madvise(bo
, I915_MADV_WILLNEED
)) {
343 brw_bo_cache_purge_bucket(bufmgr
, bucket
);
347 if (bo_set_tiling_internal(bo
, tiling_mode
, stride
)) {
353 void *map
= brw_bo_map(NULL
, bo
, MAP_WRITE
| MAP_RAW
);
358 memset(map
, 0, bo_size
);
363 if (!alloc_from_cache
) {
364 bo
= calloc(1, sizeof(*bo
));
371 struct drm_i915_gem_create create
= { .size
= bo_size
};
373 /* All new BOs we get from the kernel are zeroed, so we don't need to
374 * worry about that here.
376 ret
= drmIoctl(bufmgr
->fd
, DRM_IOCTL_I915_GEM_CREATE
, &create
);
382 bo
->gem_handle
= create
.handle
;
385 bo
->align
= alignment
;
387 bo
->tiling_mode
= I915_TILING_NONE
;
388 bo
->swizzle_mode
= I915_BIT_6_SWIZZLE_NONE
;
391 if (bo_set_tiling_internal(bo
, tiling_mode
, stride
))
394 /* Calling set_domain() will allocate pages for the BO outside of the
395 * struct mutex lock in the kernel, which is more efficient than waiting
396 * to create them during the first execbuf that uses the BO.
398 struct drm_i915_gem_set_domain sd
= {
399 .handle
= bo
->gem_handle
,
400 .read_domains
= I915_GEM_DOMAIN_CPU
,
404 if (drmIoctl(bo
->bufmgr
->fd
, DRM_IOCTL_I915_GEM_SET_DOMAIN
, &sd
) != 0)
409 p_atomic_set(&bo
->refcount
, 1);
411 bo
->cache_coherent
= bufmgr
->has_llc
;
413 if (bufmgr
->supports_48b_addresses
)
414 bo
->kflags
= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
416 mtx_unlock(&bufmgr
->lock
);
418 DBG("bo_create: buf %d (%s) %llub\n", bo
->gem_handle
, bo
->name
,
419 (unsigned long long) size
);
426 mtx_unlock(&bufmgr
->lock
);
431 brw_bo_alloc(struct brw_bufmgr
*bufmgr
,
432 const char *name
, uint64_t size
, uint64_t alignment
)
434 return bo_alloc_internal(bufmgr
, name
, size
, 0, I915_TILING_NONE
, 0, 0);
438 brw_bo_alloc_tiled(struct brw_bufmgr
*bufmgr
, const char *name
,
439 uint64_t size
, uint32_t tiling_mode
, uint32_t pitch
,
442 return bo_alloc_internal(bufmgr
, name
, size
, flags
, tiling_mode
, pitch
, 0);
446 brw_bo_alloc_tiled_2d(struct brw_bufmgr
*bufmgr
, const char *name
,
447 int x
, int y
, int cpp
, uint32_t tiling
,
448 uint32_t *pitch
, unsigned flags
)
452 unsigned long aligned_y
, height_alignment
;
454 /* If we're tiled, our allocations are in 8 or 32-row blocks,
455 * so failure to align our height means that we won't allocate
458 * If we're untiled, we still have to align to 2 rows high
459 * because the data port accesses 2x2 blocks even if the
460 * bottom row isn't to be rendered, so failure to align means
461 * we could walk off the end of the GTT and fault. This is
462 * documented on 965, and may be the case on older chipsets
463 * too so we try to be careful.
466 height_alignment
= 2;
468 if (tiling
== I915_TILING_X
)
469 height_alignment
= 8;
470 else if (tiling
== I915_TILING_Y
)
471 height_alignment
= 32;
472 aligned_y
= ALIGN(y
, height_alignment
);
475 stride
= bo_tile_pitch(bufmgr
, stride
, tiling
);
476 size
= stride
* aligned_y
;
477 size
= bo_tile_size(bufmgr
, size
, tiling
);
480 if (tiling
== I915_TILING_NONE
)
483 return bo_alloc_internal(bufmgr
, name
, size
, flags
, tiling
, stride
, 0);
487 * Returns a brw_bo wrapping the given buffer object handle.
489 * This can be used when one application needs to pass a buffer object
493 brw_bo_gem_create_from_name(struct brw_bufmgr
*bufmgr
,
494 const char *name
, unsigned int handle
)
498 /* At the moment most applications only have a few named bo.
499 * For instance, in a DRI client only the render buffers passed
500 * between X and the client are named. And since X returns the
501 * alternating names for the front/back buffer a linear search
502 * provides a sufficiently fast match.
504 mtx_lock(&bufmgr
->lock
);
505 bo
= hash_find_bo(bufmgr
->name_table
, handle
);
507 brw_bo_reference(bo
);
511 struct drm_gem_open open_arg
= { .name
= handle
};
512 int ret
= drmIoctl(bufmgr
->fd
, DRM_IOCTL_GEM_OPEN
, &open_arg
);
514 DBG("Couldn't reference %s handle 0x%08x: %s\n",
515 name
, handle
, strerror(errno
));
519 /* Now see if someone has used a prime handle to get this
520 * object from the kernel before by looking through the list
521 * again for a matching gem_handle
523 bo
= hash_find_bo(bufmgr
->handle_table
, open_arg
.handle
);
525 brw_bo_reference(bo
);
529 bo
= calloc(1, sizeof(*bo
));
533 p_atomic_set(&bo
->refcount
, 1);
535 bo
->size
= open_arg
.size
;
538 bo
->gem_handle
= open_arg
.handle
;
540 bo
->global_name
= handle
;
541 bo
->reusable
= false;
544 _mesa_hash_table_insert(bufmgr
->handle_table
, &bo
->gem_handle
, bo
);
545 _mesa_hash_table_insert(bufmgr
->name_table
, &bo
->global_name
, bo
);
547 struct drm_i915_gem_get_tiling get_tiling
= { .handle
= bo
->gem_handle
};
548 ret
= drmIoctl(bufmgr
->fd
, DRM_IOCTL_I915_GEM_GET_TILING
, &get_tiling
);
552 bo
->tiling_mode
= get_tiling
.tiling_mode
;
553 bo
->swizzle_mode
= get_tiling
.swizzle_mode
;
554 /* XXX stride is unknown */
555 DBG("bo_create_from_handle: %d (%s)\n", handle
, bo
->name
);
558 mtx_unlock(&bufmgr
->lock
);
563 mtx_unlock(&bufmgr
->lock
);
568 bo_free(struct brw_bo
*bo
)
570 struct brw_bufmgr
*bufmgr
= bo
->bufmgr
;
573 VG_NOACCESS(bo
->map_cpu
, bo
->size
);
574 drm_munmap(bo
->map_cpu
, bo
->size
);
577 VG_NOACCESS(bo
->map_wc
, bo
->size
);
578 drm_munmap(bo
->map_wc
, bo
->size
);
581 VG_NOACCESS(bo
->map_gtt
, bo
->size
);
582 drm_munmap(bo
->map_gtt
, bo
->size
);
586 struct hash_entry
*entry
;
588 if (bo
->global_name
) {
589 entry
= _mesa_hash_table_search(bufmgr
->name_table
, &bo
->global_name
);
590 _mesa_hash_table_remove(bufmgr
->name_table
, entry
);
593 entry
= _mesa_hash_table_search(bufmgr
->handle_table
, &bo
->gem_handle
);
594 _mesa_hash_table_remove(bufmgr
->handle_table
, entry
);
597 /* Close this object */
598 struct drm_gem_close close
= { .handle
= bo
->gem_handle
};
599 int ret
= drmIoctl(bufmgr
->fd
, DRM_IOCTL_GEM_CLOSE
, &close
);
601 DBG("DRM_IOCTL_GEM_CLOSE %d failed (%s): %s\n",
602 bo
->gem_handle
, bo
->name
, strerror(errno
));
607 /** Frees all cached buffers significantly older than @time. */
609 cleanup_bo_cache(struct brw_bufmgr
*bufmgr
, time_t time
)
613 if (bufmgr
->time
== time
)
616 for (i
= 0; i
< bufmgr
->num_buckets
; i
++) {
617 struct bo_cache_bucket
*bucket
= &bufmgr
->cache_bucket
[i
];
619 list_for_each_entry_safe(struct brw_bo
, bo
, &bucket
->head
, head
) {
620 if (time
- bo
->free_time
<= 1)
633 bo_unreference_final(struct brw_bo
*bo
, time_t time
)
635 struct brw_bufmgr
*bufmgr
= bo
->bufmgr
;
636 struct bo_cache_bucket
*bucket
;
638 DBG("bo_unreference final: %d (%s)\n", bo
->gem_handle
, bo
->name
);
640 bucket
= bucket_for_size(bufmgr
, bo
->size
);
641 /* Put the buffer into our internal cache for reuse if we can. */
642 if (bufmgr
->bo_reuse
&& bo
->reusable
&& bucket
!= NULL
&&
643 brw_bo_madvise(bo
, I915_MADV_DONTNEED
)) {
644 bo
->free_time
= time
;
649 list_addtail(&bo
->head
, &bucket
->head
);
656 brw_bo_unreference(struct brw_bo
*bo
)
661 assert(p_atomic_read(&bo
->refcount
) > 0);
663 if (atomic_add_unless(&bo
->refcount
, -1, 1)) {
664 struct brw_bufmgr
*bufmgr
= bo
->bufmgr
;
665 struct timespec time
;
667 clock_gettime(CLOCK_MONOTONIC
, &time
);
669 mtx_lock(&bufmgr
->lock
);
671 if (p_atomic_dec_zero(&bo
->refcount
)) {
672 bo_unreference_final(bo
, time
.tv_sec
);
673 cleanup_bo_cache(bufmgr
, time
.tv_sec
);
676 mtx_unlock(&bufmgr
->lock
);
681 bo_wait_with_stall_warning(struct brw_context
*brw
,
685 bool busy
= brw
&& brw
->perf_debug
&& !bo
->idle
;
686 double elapsed
= unlikely(busy
) ? -get_time() : 0.0;
688 brw_bo_wait_rendering(bo
);
690 if (unlikely(busy
)) {
691 elapsed
+= get_time();
692 if (elapsed
> 1e-5) /* 0.01ms */
693 perf_debug("%s a busy \"%s\" BO stalled and took %.03f ms.\n",
694 action
, bo
->name
, elapsed
* 1000);
699 print_flags(unsigned flags
)
701 if (flags
& MAP_READ
)
703 if (flags
& MAP_WRITE
)
705 if (flags
& MAP_ASYNC
)
707 if (flags
& MAP_PERSISTENT
)
709 if (flags
& MAP_COHERENT
)
717 brw_bo_map_cpu(struct brw_context
*brw
, struct brw_bo
*bo
, unsigned flags
)
719 struct brw_bufmgr
*bufmgr
= bo
->bufmgr
;
721 /* We disallow CPU maps for writing to non-coherent buffers, as the
722 * CPU map can become invalidated when a batch is flushed out, which
723 * can happen at unpredictable times. You should use WC maps instead.
725 assert(bo
->cache_coherent
|| !(flags
& MAP_WRITE
));
728 DBG("brw_bo_map_cpu: %d (%s)\n", bo
->gem_handle
, bo
->name
);
730 struct drm_i915_gem_mmap mmap_arg
= {
731 .handle
= bo
->gem_handle
,
734 int ret
= drmIoctl(bufmgr
->fd
, DRM_IOCTL_I915_GEM_MMAP
, &mmap_arg
);
737 DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
738 __FILE__
, __LINE__
, bo
->gem_handle
, bo
->name
, strerror(errno
));
741 void *map
= (void *) (uintptr_t) mmap_arg
.addr_ptr
;
742 VG_DEFINED(map
, bo
->size
);
744 if (p_atomic_cmpxchg(&bo
->map_cpu
, NULL
, map
)) {
745 VG_NOACCESS(map
, bo
->size
);
746 drm_munmap(map
, bo
->size
);
751 DBG("brw_bo_map_cpu: %d (%s) -> %p, ", bo
->gem_handle
, bo
->name
,
755 if (!(flags
& MAP_ASYNC
)) {
756 bo_wait_with_stall_warning(brw
, bo
, "CPU mapping");
759 if (!bo
->cache_coherent
&& !bo
->bufmgr
->has_llc
) {
760 /* If we're reusing an existing CPU mapping, the CPU caches may
761 * contain stale data from the last time we read from that mapping.
762 * (With the BO cache, it might even be data from a previous buffer!)
763 * Even if it's a brand new mapping, the kernel may have zeroed the
764 * buffer via CPU writes.
766 * We need to invalidate those cachelines so that we see the latest
767 * contents, and so long as we only read from the CPU mmap we do not
768 * need to write those cachelines back afterwards.
770 * On LLC, the emprical evidence suggests that writes from the GPU
771 * that bypass the LLC (i.e. for scanout) do *invalidate* the CPU
772 * cachelines. (Other reads, such as the display engine, bypass the
773 * LLC entirely requiring us to keep dirty pixels for the scanout
776 gen_invalidate_range(bo
->map_cpu
, bo
->size
);
783 brw_bo_map_wc(struct brw_context
*brw
, struct brw_bo
*bo
, unsigned flags
)
785 struct brw_bufmgr
*bufmgr
= bo
->bufmgr
;
787 if (!bufmgr
->has_mmap_wc
)
791 DBG("brw_bo_map_wc: %d (%s)\n", bo
->gem_handle
, bo
->name
);
793 struct drm_i915_gem_mmap mmap_arg
= {
794 .handle
= bo
->gem_handle
,
796 .flags
= I915_MMAP_WC
,
798 int ret
= drmIoctl(bufmgr
->fd
, DRM_IOCTL_I915_GEM_MMAP
, &mmap_arg
);
801 DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
802 __FILE__
, __LINE__
, bo
->gem_handle
, bo
->name
, strerror(errno
));
806 void *map
= (void *) (uintptr_t) mmap_arg
.addr_ptr
;
807 VG_DEFINED(map
, bo
->size
);
809 if (p_atomic_cmpxchg(&bo
->map_wc
, NULL
, map
)) {
810 VG_NOACCESS(map
, bo
->size
);
811 drm_munmap(map
, bo
->size
);
816 DBG("brw_bo_map_wc: %d (%s) -> %p\n", bo
->gem_handle
, bo
->name
, bo
->map_wc
);
819 if (!(flags
& MAP_ASYNC
)) {
820 bo_wait_with_stall_warning(brw
, bo
, "WC mapping");
827 * Perform an uncached mapping via the GTT.
829 * Write access through the GTT is not quite fully coherent. On low power
830 * systems especially, like modern Atoms, we can observe reads from RAM before
831 * the write via GTT has landed. A write memory barrier that flushes the Write
832 * Combining Buffer (i.e. sfence/mfence) is not sufficient to order the later
833 * read after the write as the GTT write suffers a small delay through the GTT
834 * indirection. The kernel uses an uncached mmio read to ensure the GTT write
835 * is ordered with reads (either by the GPU, WB or WC) and unconditionally
836 * flushes prior to execbuf submission. However, if we are not informing the
837 * kernel about our GTT writes, it will not flush before earlier access, such
838 * as when using the cmdparser. Similarly, we need to be careful if we should
839 * ever issue a CPU read immediately following a GTT write.
841 * Telling the kernel about write access also has one more important
842 * side-effect. Upon receiving notification about the write, it cancels any
843 * scanout buffering for FBC/PSR and friends. Later FBC/PSR is then flushed by
844 * either SW_FINISH or DIRTYFB. The presumption is that we never write to the
845 * actual scanout via a mmaping, only to a backbuffer and so all the FBC/PSR
846 * tracking is handled on the buffer exchange instead.
849 brw_bo_map_gtt(struct brw_context
*brw
, struct brw_bo
*bo
, unsigned flags
)
851 struct brw_bufmgr
*bufmgr
= bo
->bufmgr
;
853 /* Get a mapping of the buffer if we haven't before. */
854 if (bo
->map_gtt
== NULL
) {
855 DBG("bo_map_gtt: mmap %d (%s)\n", bo
->gem_handle
, bo
->name
);
857 struct drm_i915_gem_mmap_gtt mmap_arg
= { .handle
= bo
->gem_handle
};
859 /* Get the fake offset back... */
860 int ret
= drmIoctl(bufmgr
->fd
, DRM_IOCTL_I915_GEM_MMAP_GTT
, &mmap_arg
);
862 DBG("%s:%d: Error preparing buffer map %d (%s): %s .\n",
863 __FILE__
, __LINE__
, bo
->gem_handle
, bo
->name
, strerror(errno
));
868 void *map
= drm_mmap(0, bo
->size
, PROT_READ
| PROT_WRITE
,
869 MAP_SHARED
, bufmgr
->fd
, mmap_arg
.offset
);
870 if (map
== MAP_FAILED
) {
871 DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
872 __FILE__
, __LINE__
, bo
->gem_handle
, bo
->name
, strerror(errno
));
876 /* We don't need to use VALGRIND_MALLOCLIKE_BLOCK because Valgrind will
877 * already intercept this mmap call. However, for consistency between
878 * all the mmap paths, we mark the pointer as defined now and mark it
879 * as inaccessible afterwards.
881 VG_DEFINED(map
, bo
->size
);
883 if (p_atomic_cmpxchg(&bo
->map_gtt
, NULL
, map
)) {
884 VG_NOACCESS(map
, bo
->size
);
885 drm_munmap(map
, bo
->size
);
890 DBG("bo_map_gtt: %d (%s) -> %p, ", bo
->gem_handle
, bo
->name
, bo
->map_gtt
);
893 if (!(flags
& MAP_ASYNC
)) {
894 bo_wait_with_stall_warning(brw
, bo
, "GTT mapping");
901 can_map_cpu(struct brw_bo
*bo
, unsigned flags
)
903 if (bo
->cache_coherent
)
906 /* Even if the buffer itself is not cache-coherent (such as a scanout), on
907 * an LLC platform reads always are coherent (as they are performed via the
908 * central system agent). It is just the writes that we need to take special
909 * care to ensure that land in main memory and not stick in the CPU cache.
911 if (!(flags
& MAP_WRITE
) && bo
->bufmgr
->has_llc
)
914 /* If PERSISTENT or COHERENT are set, the mmapping needs to remain valid
915 * across batch flushes where the kernel will change cache domains of the
916 * bo, invalidating continued access to the CPU mmap on non-LLC device.
918 * Similarly, ASYNC typically means that the buffer will be accessed via
919 * both the CPU and the GPU simultaneously. Batches may be executed that
920 * use the BO even while it is mapped. While OpenGL technically disallows
921 * most drawing while non-persistent mappings are active, we may still use
922 * the GPU for blits or other operations, causing batches to happen at
923 * inconvenient times.
925 if (flags
& (MAP_PERSISTENT
| MAP_COHERENT
| MAP_ASYNC
))
928 return !(flags
& MAP_WRITE
);
932 brw_bo_map(struct brw_context
*brw
, struct brw_bo
*bo
, unsigned flags
)
934 if (bo
->tiling_mode
!= I915_TILING_NONE
&& !(flags
& MAP_RAW
))
935 return brw_bo_map_gtt(brw
, bo
, flags
);
939 if (can_map_cpu(bo
, flags
))
940 map
= brw_bo_map_cpu(brw
, bo
, flags
);
942 map
= brw_bo_map_wc(brw
, bo
, flags
);
944 /* Allow the attempt to fail by falling back to the GTT where necessary.
946 * Not every buffer can be mmaped directly using the CPU (or WC), for
947 * example buffers that wrap stolen memory or are imported from other
948 * devices. For those, we have little choice but to use a GTT mmapping.
949 * However, if we use a slow GTT mmapping for reads where we expected fast
950 * access, that order of magnitude difference in throughput will be clearly
951 * expressed by angry users.
953 * We skip MAP_RAW because we want to avoid map_gtt's fence detiling.
955 if (!map
&& !(flags
& MAP_RAW
)) {
957 perf_debug("Fallback GTT mapping for %s with access flags %x\n",
960 map
= brw_bo_map_gtt(brw
, bo
, flags
);
967 brw_bo_subdata(struct brw_bo
*bo
, uint64_t offset
,
968 uint64_t size
, const void *data
)
970 struct brw_bufmgr
*bufmgr
= bo
->bufmgr
;
972 struct drm_i915_gem_pwrite pwrite
= {
973 .handle
= bo
->gem_handle
,
976 .data_ptr
= (uint64_t) (uintptr_t) data
,
979 int ret
= drmIoctl(bufmgr
->fd
, DRM_IOCTL_I915_GEM_PWRITE
, &pwrite
);
982 DBG("%s:%d: Error writing data to buffer %d: "
983 "(%"PRIu64
" %"PRIu64
") %s .\n",
984 __FILE__
, __LINE__
, bo
->gem_handle
, offset
, size
, strerror(errno
));
990 /** Waits for all GPU rendering with the object to have completed. */
992 brw_bo_wait_rendering(struct brw_bo
*bo
)
994 /* We require a kernel recent enough for WAIT_IOCTL support.
995 * See intel_init_bufmgr()
1001 * Waits on a BO for the given amount of time.
1003 * @bo: buffer object to wait for
1004 * @timeout_ns: amount of time to wait in nanoseconds.
1005 * If value is less than 0, an infinite wait will occur.
1007 * Returns 0 if the wait was successful ie. the last batch referencing the
1008 * object has completed within the allotted time. Otherwise some negative return
1009 * value describes the error. Of particular interest is -ETIME when the wait has
1010 * failed to yield the desired result.
1012 * Similar to brw_bo_wait_rendering except a timeout parameter allows
1013 * the operation to give up after a certain amount of time. Another subtle
1014 * difference is the internal locking semantics are different (this variant does
1015 * not hold the lock for the duration of the wait). This makes the wait subject
1016 * to a larger userspace race window.
1018 * The implementation shall wait until the object is no longer actively
1019 * referenced within a batch buffer at the time of the call. The wait will
1020 * not guarantee that the buffer is re-issued via another thread, or an flinked
1021 * handle. Userspace must make sure this race does not occur if such precision
1024 * Note that some kernels have broken the inifite wait for negative values
1025 * promise, upgrade to latest stable kernels if this is the case.
1028 brw_bo_wait(struct brw_bo
*bo
, int64_t timeout_ns
)
1030 struct brw_bufmgr
*bufmgr
= bo
->bufmgr
;
1032 /* If we know it's idle, don't bother with the kernel round trip */
1033 if (bo
->idle
&& !bo
->external
)
1036 struct drm_i915_gem_wait wait
= {
1037 .bo_handle
= bo
->gem_handle
,
1038 .timeout_ns
= timeout_ns
,
1040 int ret
= drmIoctl(bufmgr
->fd
, DRM_IOCTL_I915_GEM_WAIT
, &wait
);
1050 brw_bufmgr_destroy(struct brw_bufmgr
*bufmgr
)
1052 mtx_destroy(&bufmgr
->lock
);
1054 /* Free any cached buffer objects we were going to reuse */
1055 for (int i
= 0; i
< bufmgr
->num_buckets
; i
++) {
1056 struct bo_cache_bucket
*bucket
= &bufmgr
->cache_bucket
[i
];
1058 list_for_each_entry_safe(struct brw_bo
, bo
, &bucket
->head
, head
) {
1059 list_del(&bo
->head
);
1065 _mesa_hash_table_destroy(bufmgr
->name_table
, NULL
);
1066 _mesa_hash_table_destroy(bufmgr
->handle_table
, NULL
);
1072 bo_set_tiling_internal(struct brw_bo
*bo
, uint32_t tiling_mode
,
1075 struct brw_bufmgr
*bufmgr
= bo
->bufmgr
;
1076 struct drm_i915_gem_set_tiling set_tiling
;
1079 if (bo
->global_name
== 0 &&
1080 tiling_mode
== bo
->tiling_mode
&& stride
== bo
->stride
)
1083 memset(&set_tiling
, 0, sizeof(set_tiling
));
1085 /* set_tiling is slightly broken and overwrites the
1086 * input on the error path, so we have to open code
1089 set_tiling
.handle
= bo
->gem_handle
;
1090 set_tiling
.tiling_mode
= tiling_mode
;
1091 set_tiling
.stride
= stride
;
1093 ret
= ioctl(bufmgr
->fd
, DRM_IOCTL_I915_GEM_SET_TILING
, &set_tiling
);
1094 } while (ret
== -1 && (errno
== EINTR
|| errno
== EAGAIN
));
1098 bo
->tiling_mode
= set_tiling
.tiling_mode
;
1099 bo
->swizzle_mode
= set_tiling
.swizzle_mode
;
1100 bo
->stride
= set_tiling
.stride
;
1105 brw_bo_get_tiling(struct brw_bo
*bo
, uint32_t *tiling_mode
,
1106 uint32_t *swizzle_mode
)
1108 *tiling_mode
= bo
->tiling_mode
;
1109 *swizzle_mode
= bo
->swizzle_mode
;
1113 static struct brw_bo
*
1114 brw_bo_gem_create_from_prime_internal(struct brw_bufmgr
*bufmgr
, int prime_fd
,
1115 int tiling_mode
, uint32_t stride
)
1120 mtx_lock(&bufmgr
->lock
);
1121 int ret
= drmPrimeFDToHandle(bufmgr
->fd
, prime_fd
, &handle
);
1123 DBG("create_from_prime: failed to obtain handle from fd: %s\n",
1125 mtx_unlock(&bufmgr
->lock
);
1130 * See if the kernel has already returned this buffer to us. Just as
1131 * for named buffers, we must not create two bo's pointing at the same
1134 bo
= hash_find_bo(bufmgr
->handle_table
, handle
);
1136 brw_bo_reference(bo
);
1140 bo
= calloc(1, sizeof(*bo
));
1144 p_atomic_set(&bo
->refcount
, 1);
1146 /* Determine size of bo. The fd-to-handle ioctl really should
1147 * return the size, but it doesn't. If we have kernel 3.12 or
1148 * later, we can lseek on the prime fd to get the size. Older
1149 * kernels will just fail, in which case we fall back to the
1150 * provided (estimated or guess size). */
1151 ret
= lseek(prime_fd
, 0, SEEK_END
);
1155 bo
->bufmgr
= bufmgr
;
1157 bo
->gem_handle
= handle
;
1158 _mesa_hash_table_insert(bufmgr
->handle_table
, &bo
->gem_handle
, bo
);
1161 bo
->reusable
= false;
1162 bo
->external
= true;
1164 if (tiling_mode
< 0) {
1165 struct drm_i915_gem_get_tiling get_tiling
= { .handle
= bo
->gem_handle
};
1166 if (drmIoctl(bufmgr
->fd
, DRM_IOCTL_I915_GEM_GET_TILING
, &get_tiling
))
1169 bo
->tiling_mode
= get_tiling
.tiling_mode
;
1170 bo
->swizzle_mode
= get_tiling
.swizzle_mode
;
1171 /* XXX stride is unknown */
1173 bo_set_tiling_internal(bo
, tiling_mode
, stride
);
1177 mtx_unlock(&bufmgr
->lock
);
1182 mtx_unlock(&bufmgr
->lock
);
1187 brw_bo_gem_create_from_prime(struct brw_bufmgr
*bufmgr
, int prime_fd
)
1189 return brw_bo_gem_create_from_prime_internal(bufmgr
, prime_fd
, -1, 0);
1193 brw_bo_gem_create_from_prime_tiled(struct brw_bufmgr
*bufmgr
, int prime_fd
,
1194 uint32_t tiling_mode
, uint32_t stride
)
1196 assert(tiling_mode
== I915_TILING_NONE
||
1197 tiling_mode
== I915_TILING_X
||
1198 tiling_mode
== I915_TILING_Y
);
1200 return brw_bo_gem_create_from_prime_internal(bufmgr
, prime_fd
,
1201 tiling_mode
, stride
);
1205 brw_bo_make_external(struct brw_bo
*bo
)
1207 struct brw_bufmgr
*bufmgr
= bo
->bufmgr
;
1209 if (!bo
->external
) {
1210 mtx_lock(&bufmgr
->lock
);
1211 if (!bo
->external
) {
1212 _mesa_hash_table_insert(bufmgr
->handle_table
, &bo
->gem_handle
, bo
);
1213 bo
->external
= true;
1215 mtx_unlock(&bufmgr
->lock
);
1220 brw_bo_gem_export_to_prime(struct brw_bo
*bo
, int *prime_fd
)
1222 struct brw_bufmgr
*bufmgr
= bo
->bufmgr
;
1224 brw_bo_make_external(bo
);
1226 if (drmPrimeHandleToFD(bufmgr
->fd
, bo
->gem_handle
,
1227 DRM_CLOEXEC
, prime_fd
) != 0)
1230 bo
->reusable
= false;
1236 brw_bo_export_gem_handle(struct brw_bo
*bo
)
1238 brw_bo_make_external(bo
);
1240 return bo
->gem_handle
;
1244 brw_bo_flink(struct brw_bo
*bo
, uint32_t *name
)
1246 struct brw_bufmgr
*bufmgr
= bo
->bufmgr
;
1248 if (!bo
->global_name
) {
1249 struct drm_gem_flink flink
= { .handle
= bo
->gem_handle
};
1251 if (drmIoctl(bufmgr
->fd
, DRM_IOCTL_GEM_FLINK
, &flink
))
1254 brw_bo_make_external(bo
);
1255 mtx_lock(&bufmgr
->lock
);
1256 if (!bo
->global_name
) {
1257 bo
->global_name
= flink
.name
;
1258 _mesa_hash_table_insert(bufmgr
->name_table
, &bo
->global_name
, bo
);
1260 mtx_unlock(&bufmgr
->lock
);
1262 bo
->reusable
= false;
1265 *name
= bo
->global_name
;
1270 * Enables unlimited caching of buffer objects for reuse.
1272 * This is potentially very memory expensive, as the cache at each bucket
1273 * size is only bounded by how many buffers of that size we've managed to have
1274 * in flight at once.
1277 brw_bufmgr_enable_reuse(struct brw_bufmgr
*bufmgr
)
1279 bufmgr
->bo_reuse
= true;
1283 add_bucket(struct brw_bufmgr
*bufmgr
, int size
)
1285 unsigned int i
= bufmgr
->num_buckets
;
1287 assert(i
< ARRAY_SIZE(bufmgr
->cache_bucket
));
1289 list_inithead(&bufmgr
->cache_bucket
[i
].head
);
1290 bufmgr
->cache_bucket
[i
].size
= size
;
1291 bufmgr
->num_buckets
++;
1293 assert(bucket_for_size(bufmgr
, size
) == &bufmgr
->cache_bucket
[i
]);
1294 assert(bucket_for_size(bufmgr
, size
- 2048) == &bufmgr
->cache_bucket
[i
]);
1295 assert(bucket_for_size(bufmgr
, size
+ 1) != &bufmgr
->cache_bucket
[i
]);
1299 init_cache_buckets(struct brw_bufmgr
*bufmgr
)
1301 uint64_t size
, cache_max_size
= 64 * 1024 * 1024;
1303 /* OK, so power of two buckets was too wasteful of memory.
1304 * Give 3 other sizes between each power of two, to hopefully
1305 * cover things accurately enough. (The alternative is
1306 * probably to just go for exact matching of sizes, and assume
1307 * that for things like composited window resize the tiled
1308 * width/height alignment and rounding of sizes to pages will
1309 * get us useful cache hit rates anyway)
1311 add_bucket(bufmgr
, 4096);
1312 add_bucket(bufmgr
, 4096 * 2);
1313 add_bucket(bufmgr
, 4096 * 3);
1315 /* Initialize the linked lists for BO reuse cache. */
1316 for (size
= 4 * 4096; size
<= cache_max_size
; size
*= 2) {
1317 add_bucket(bufmgr
, size
);
1319 add_bucket(bufmgr
, size
+ size
* 1 / 4);
1320 add_bucket(bufmgr
, size
+ size
* 2 / 4);
1321 add_bucket(bufmgr
, size
+ size
* 3 / 4);
1326 brw_create_hw_context(struct brw_bufmgr
*bufmgr
)
1328 struct drm_i915_gem_context_create create
= { };
1329 int ret
= drmIoctl(bufmgr
->fd
, DRM_IOCTL_I915_GEM_CONTEXT_CREATE
, &create
);
1331 DBG("DRM_IOCTL_I915_GEM_CONTEXT_CREATE failed: %s\n", strerror(errno
));
1335 return create
.ctx_id
;
1339 brw_hw_context_set_priority(struct brw_bufmgr
*bufmgr
,
1343 struct drm_i915_gem_context_param p
= {
1345 .param
= I915_CONTEXT_PARAM_PRIORITY
,
1351 if (drmIoctl(bufmgr
->fd
, DRM_IOCTL_I915_GEM_CONTEXT_SETPARAM
, &p
))
1358 brw_destroy_hw_context(struct brw_bufmgr
*bufmgr
, uint32_t ctx_id
)
1360 struct drm_i915_gem_context_destroy d
= { .ctx_id
= ctx_id
};
1363 drmIoctl(bufmgr
->fd
, DRM_IOCTL_I915_GEM_CONTEXT_DESTROY
, &d
) != 0) {
1364 fprintf(stderr
, "DRM_IOCTL_I915_GEM_CONTEXT_DESTROY failed: %s\n",
1370 brw_reg_read(struct brw_bufmgr
*bufmgr
, uint32_t offset
, uint64_t *result
)
1372 struct drm_i915_reg_read reg_read
= { .offset
= offset
};
1373 int ret
= drmIoctl(bufmgr
->fd
, DRM_IOCTL_I915_REG_READ
, ®_read
);
1375 *result
= reg_read
.val
;
1380 gem_param(int fd
, int name
)
1382 int v
= -1; /* No param uses (yet) the sign bit, reserve it for errors */
1384 struct drm_i915_getparam gp
= { .param
= name
, .value
= &v
};
1385 if (drmIoctl(fd
, DRM_IOCTL_I915_GETPARAM
, &gp
))
1392 gem_supports_48b_addresses(int fd
)
1394 struct drm_i915_gem_exec_object2 obj
= {
1395 .flags
= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
,
1398 struct drm_i915_gem_execbuffer2 execbuf
= {
1399 .buffers_ptr
= (uintptr_t)&obj
,
1404 int ret
= drmIoctl(fd
, DRM_IOCTL_I915_GEM_EXECBUFFER2
, &execbuf
);
1406 return ret
== -1 && errno
== ENOENT
;
1410 * Initializes the GEM buffer manager, which uses the kernel to allocate, map,
1411 * and manage map buffer objections.
1413 * \param fd File descriptor of the opened DRM device.
1416 brw_bufmgr_init(struct gen_device_info
*devinfo
, int fd
)
1418 struct brw_bufmgr
*bufmgr
;
1420 bufmgr
= calloc(1, sizeof(*bufmgr
));
1424 /* Handles to buffer objects belong to the device fd and are not
1425 * reference counted by the kernel. If the same fd is used by
1426 * multiple parties (threads sharing the same screen bufmgr, or
1427 * even worse the same device fd passed to multiple libraries)
1428 * ownership of those handles is shared by those independent parties.
1430 * Don't do this! Ensure that each library/bufmgr has its own device
1431 * fd so that its namespace does not clash with another.
1435 if (mtx_init(&bufmgr
->lock
, mtx_plain
) != 0) {
1440 bufmgr
->has_llc
= devinfo
->has_llc
;
1441 bufmgr
->has_mmap_wc
= gem_param(fd
, I915_PARAM_MMAP_VERSION
) > 0;
1442 bufmgr
->supports_48b_addresses
=
1443 devinfo
->gen
>= 8 && gem_supports_48b_addresses(fd
);
1445 init_cache_buckets(bufmgr
);
1447 bufmgr
->name_table
=
1448 _mesa_hash_table_create(NULL
, key_hash_uint
, key_uint_equal
);
1449 bufmgr
->handle_table
=
1450 _mesa_hash_table_create(NULL
, key_hash_uint
, key_uint_equal
);