2 * Copyright © 2008 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Eric Anholt <eric@anholt.net>
28 #include <linux/acpi.h>
29 #include <linux/pnp.h>
30 #include "linux/string.h"
31 #include "linux/bitops.h"
37 /** @file i915_gem_tiling.c
39 * Support for managing tiling state of buffer objects.
41 * The idea behind tiling is to increase cache hit rates by rearranging
42 * pixel data so that a group of pixel accesses are in the same cacheline.
43 * Performance improvement from doing this on the back/depth buffer are on
46 * Intel architectures make this somewhat more complicated, though, by
47 * adjustments made to addressing of data when the memory is in interleaved
48 * mode (matched pairs of DIMMS) to improve memory bandwidth.
49 * For interleaved memory, the CPU sends every sequential 64 bytes
50 * to an alternate memory channel so it can get the bandwidth from both.
52 * The GPU also rearranges its accesses for increased bandwidth to interleaved
53 * memory, and it matches what the CPU does for non-tiled. However, when tiled
54 * it does it a little differently, since one walks addresses not just in the
55 * X direction but also Y. So, along with alternating channels when bit
56 * 6 of the address flips, it also alternates when other bits flip -- Bits 9
57 * (every 512 bytes, an X tile scanline) and 10 (every two X tile scanlines)
58 * are common to both the 915 and 965-class hardware.
60 * The CPU also sometimes XORs in higher bits as well, to improve
61 * bandwidth doing strided access like we do so frequently in graphics. This
62 * is called "Channel XOR Randomization" in the MCH documentation. The result
63 * is that the CPU is XORing in either bit 11 or bit 17 to bit 6 of its address
66 * All of this bit 6 XORing has an effect on our memory management,
67 * as we need to make sure that the 3d driver can correctly address object
70 * If we don't have interleaved memory, all tiling is safe and no swizzling is
73 * When bit 17 is XORed in, we simply refuse to tile at all. Bit
74 * 17 is not just a page offset, so as we page an objet out and back in,
75 * individual pages in it will have different bit 17 addresses, resulting in
76 * each 64 bytes being swapped with its neighbor!
78 * Otherwise, if interleaved, we have to tell the 3d driver what the address
79 * swizzling it needs to do is, since it's writing with the CPU to the pages
80 * (bit 6 and potentially bit 11 XORed in), and the GPU is reading from the
81 * pages (bit 6, 9, and 10 XORed in), resulting in a cumulative bit swizzling
82 * required by the CPU of XORing in bit 6, 9, 10, and potentially 11, in order
83 * to match what the GPU expects.
86 #define MCHBAR_I915 0x44
87 #define MCHBAR_I965 0x48
88 #define MCHBAR_SIZE (4*4096)
90 #define DEVEN_REG 0x54
91 #define DEVEN_MCHBAR_EN (1 << 28)
93 /* Allocate space for the MCH regs if needed, return nonzero on error */
95 intel_alloc_mchbar_resource(struct drm_device
*dev
)
97 struct pci_dev
*bridge_dev
;
98 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
99 int reg
= IS_I965G(dev
) ? MCHBAR_I965
: MCHBAR_I915
;
100 u32 temp_lo
, temp_hi
= 0;
104 bridge_dev
= pci_get_bus_and_slot(0, PCI_DEVFN(0,0));
106 DRM_DEBUG("no bridge dev?!\n");
112 pci_read_config_dword(bridge_dev
, reg
+ 4, &temp_hi
);
113 pci_read_config_dword(bridge_dev
, reg
, &temp_lo
);
114 mchbar_addr
= ((u64
)temp_hi
<< 32) | temp_lo
;
116 /* If ACPI doesn't have it, assume we need to allocate it ourselves */
119 pnp_range_reserved(mchbar_addr
, mchbar_addr
+ MCHBAR_SIZE
)) {
125 /* Get some space for it */
126 ret
= pci_bus_alloc_resource(bridge_dev
->bus
, &dev_priv
->mch_res
,
127 MCHBAR_SIZE
, MCHBAR_SIZE
,
129 0, pcibios_align_resource
,
132 DRM_DEBUG("failed bus alloc: %d\n", ret
);
133 dev_priv
->mch_res
.start
= 0;
138 pci_write_config_dword(bridge_dev
, reg
+ 4,
139 upper_32_bits(dev_priv
->mch_res
.start
));
141 pci_write_config_dword(bridge_dev
, reg
,
142 lower_32_bits(dev_priv
->mch_res
.start
));
144 pci_dev_put(bridge_dev
);
149 /* Setup MCHBAR if possible, return true if we should disable it again */
151 intel_setup_mchbar(struct drm_device
*dev
)
153 struct pci_dev
*bridge_dev
;
154 int mchbar_reg
= IS_I965G(dev
) ? MCHBAR_I965
: MCHBAR_I915
;
156 bool need_disable
= false, enabled
;
158 bridge_dev
= pci_get_bus_and_slot(0, PCI_DEVFN(0,0));
160 DRM_DEBUG("no bridge dev?!\n");
164 if (IS_I915G(dev
) || IS_I915GM(dev
)) {
165 pci_read_config_dword(bridge_dev
, DEVEN_REG
, &temp
);
166 enabled
= !!(temp
& DEVEN_MCHBAR_EN
);
168 pci_read_config_dword(bridge_dev
, mchbar_reg
, &temp
);
172 /* If it's already enabled, don't have to do anything */
176 if (intel_alloc_mchbar_resource(dev
))
181 /* Space is allocated or reserved, so enable it. */
182 if (IS_I915G(dev
) || IS_I915GM(dev
)) {
183 pci_write_config_dword(bridge_dev
, DEVEN_REG
,
184 temp
| DEVEN_MCHBAR_EN
);
186 pci_read_config_dword(bridge_dev
, mchbar_reg
, &temp
);
187 pci_write_config_dword(bridge_dev
, mchbar_reg
, temp
| 1);
190 pci_dev_put(bridge_dev
);
196 intel_teardown_mchbar(struct drm_device
*dev
, bool disable
)
198 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
199 struct pci_dev
*bridge_dev
;
200 int mchbar_reg
= IS_I965G(dev
) ? MCHBAR_I965
: MCHBAR_I915
;
203 bridge_dev
= pci_get_bus_and_slot(0, PCI_DEVFN(0,0));
205 DRM_DEBUG("no bridge dev?!\n");
210 if (IS_I915G(dev
) || IS_I915GM(dev
)) {
211 pci_read_config_dword(bridge_dev
, DEVEN_REG
, &temp
);
212 temp
&= ~DEVEN_MCHBAR_EN
;
213 pci_write_config_dword(bridge_dev
, DEVEN_REG
, temp
);
215 pci_read_config_dword(bridge_dev
, mchbar_reg
, &temp
);
217 pci_write_config_dword(bridge_dev
, mchbar_reg
, temp
);
221 if (dev_priv
->mch_res
.start
)
222 release_resource(&dev_priv
->mch_res
);
226 * Detects bit 6 swizzling of address lookup between IGD access and CPU
227 * access through main memory.
230 i915_gem_detect_bit_6_swizzle(struct drm_device
*dev
)
232 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
233 uint32_t swizzle_x
= I915_BIT_6_SWIZZLE_UNKNOWN
;
234 uint32_t swizzle_y
= I915_BIT_6_SWIZZLE_UNKNOWN
;
238 /* As far as we know, the 865 doesn't have these bit 6
241 swizzle_x
= I915_BIT_6_SWIZZLE_NONE
;
242 swizzle_y
= I915_BIT_6_SWIZZLE_NONE
;
243 } else if (IS_MOBILE(dev
)) {
246 /* Try to make sure MCHBAR is enabled before poking at it */
247 need_disable
= intel_setup_mchbar(dev
);
249 /* On mobile 9xx chipsets, channel interleave by the CPU is
250 * determined by DCC. For single-channel, neither the CPU
251 * nor the GPU do swizzling. For dual channel interleaved,
252 * the GPU's interleave is bit 9 and 10 for X tiled, and bit
253 * 9 for Y tiled. The CPU's interleave is independent, and
254 * can be based on either bit 11 (haven't seen this yet) or
257 dcc
= I915_READ(DCC
);
258 switch (dcc
& DCC_ADDRESSING_MODE_MASK
) {
259 case DCC_ADDRESSING_MODE_SINGLE_CHANNEL
:
260 case DCC_ADDRESSING_MODE_DUAL_CHANNEL_ASYMMETRIC
:
261 swizzle_x
= I915_BIT_6_SWIZZLE_NONE
;
262 swizzle_y
= I915_BIT_6_SWIZZLE_NONE
;
264 case DCC_ADDRESSING_MODE_DUAL_CHANNEL_INTERLEAVED
:
265 if (dcc
& DCC_CHANNEL_XOR_DISABLE
) {
266 /* This is the base swizzling by the GPU for
269 swizzle_x
= I915_BIT_6_SWIZZLE_9_10
;
270 swizzle_y
= I915_BIT_6_SWIZZLE_9
;
271 } else if ((dcc
& DCC_CHANNEL_XOR_BIT_17
) == 0) {
272 /* Bit 11 swizzling by the CPU in addition. */
273 swizzle_x
= I915_BIT_6_SWIZZLE_9_10_11
;
274 swizzle_y
= I915_BIT_6_SWIZZLE_9_11
;
276 /* Bit 17 swizzling by the CPU in addition. */
277 swizzle_x
= I915_BIT_6_SWIZZLE_9_10_17
;
278 swizzle_y
= I915_BIT_6_SWIZZLE_9_17
;
282 if (dcc
== 0xffffffff) {
283 DRM_ERROR("Couldn't read from MCHBAR. "
284 "Disabling tiling.\n");
285 swizzle_x
= I915_BIT_6_SWIZZLE_UNKNOWN
;
286 swizzle_y
= I915_BIT_6_SWIZZLE_UNKNOWN
;
289 intel_teardown_mchbar(dev
, need_disable
);
291 /* The 965, G33, and newer, have a very flexible memory
292 * configuration. It will enable dual-channel mode
293 * (interleaving) on as much memory as it can, and the GPU
294 * will additionally sometimes enable different bit 6
295 * swizzling for tiled objects from the CPU.
297 * Here's what I found on the G965:
298 * slot fill memory size swizzling
299 * 0A 0B 1A 1B 1-ch 2-ch
301 * 512 0 512 0 16 1008 X
302 * 512 0 0 512 16 1008 X
303 * 0 512 0 512 16 1008 X
304 * 1024 1024 1024 0 2048 1024 O
306 * We could probably detect this based on either the DRB
307 * matching, which was the case for the swizzling required in
308 * the table above, or from the 1-ch value being less than
309 * the minimum size of a rank.
311 if (I915_READ16(C0DRB3
) != I915_READ16(C1DRB3
)) {
312 swizzle_x
= I915_BIT_6_SWIZZLE_NONE
;
313 swizzle_y
= I915_BIT_6_SWIZZLE_NONE
;
315 swizzle_x
= I915_BIT_6_SWIZZLE_9_10
;
316 swizzle_y
= I915_BIT_6_SWIZZLE_9
;
320 /* FIXME: check with memory config on IGDNG */
322 DRM_ERROR("disable tiling on IGDNG...\n");
323 swizzle_x
= I915_BIT_6_SWIZZLE_UNKNOWN
;
324 swizzle_y
= I915_BIT_6_SWIZZLE_UNKNOWN
;
327 dev_priv
->mm
.bit_6_swizzle_x
= swizzle_x
;
328 dev_priv
->mm
.bit_6_swizzle_y
= swizzle_y
;
333 * Returns the size of the fence for a tiled object of the given size.
336 i915_get_fence_size(struct drm_device
*dev
, int size
)
342 /* The 965 can have fences at any page boundary. */
343 return ALIGN(size
, 4096);
345 /* Align the size to a power of two greater than the smallest
353 for (i
= start
; i
< size
; i
<<= 1)
360 /* Check pitch constriants for all chips & tiling formats */
362 i915_tiling_ok(struct drm_device
*dev
, int stride
, int size
, int tiling_mode
)
366 /* Linear is always fine */
367 if (tiling_mode
== I915_TILING_NONE
)
371 (tiling_mode
== I915_TILING_Y
&& HAS_128_BYTE_Y_TILING(dev
)))
376 /* check maximum stride & object size */
378 /* i965 stores the end address of the gtt mapping in the fence
379 * reg, so dont bother to check the size */
380 if (stride
/ 128 > I965_FENCE_MAX_PITCH_VAL
)
382 } else if (IS_I9XX(dev
)) {
383 uint32_t pitch_val
= ffs(stride
/ tile_width
) - 1;
385 /* XXX: For Y tiling, FENCE_MAX_PITCH_VAL is actually 6 (8KB)
386 * instead of 4 (2KB) on 945s.
388 if (pitch_val
> I915_FENCE_MAX_PITCH_VAL
||
389 size
> (I830_FENCE_MAX_SIZE_VAL
<< 20))
392 uint32_t pitch_val
= ffs(stride
/ tile_width
) - 1;
394 if (pitch_val
> I830_FENCE_MAX_PITCH_VAL
||
395 size
> (I830_FENCE_MAX_SIZE_VAL
<< 19))
399 /* 965+ just needs multiples of tile width */
401 if (stride
& (tile_width
- 1))
406 /* Pre-965 needs power of two tile widths */
407 if (stride
< tile_width
)
410 if (stride
& (stride
- 1))
413 /* We don't 0handle the aperture area covered by the fence being bigger
414 * than the object size.
416 if (i915_get_fence_size(dev
, size
) != size
)
423 i915_gem_object_fence_offset_ok(struct drm_gem_object
*obj
, int tiling_mode
)
425 struct drm_device
*dev
= obj
->dev
;
426 struct drm_i915_gem_object
*obj_priv
= obj
->driver_private
;
428 if (obj_priv
->gtt_space
== NULL
)
431 if (tiling_mode
== I915_TILING_NONE
)
434 if (!IS_I965G(dev
)) {
435 if (obj_priv
->gtt_offset
& (obj
->size
- 1))
438 if (obj_priv
->gtt_offset
& ~I915_FENCE_START_MASK
)
441 if (obj_priv
->gtt_offset
& ~I830_FENCE_START_MASK
)
450 * Sets the tiling mode of an object, returning the required swizzling of
451 * bit 6 of addresses in the object.
454 i915_gem_set_tiling(struct drm_device
*dev
, void *data
,
455 struct drm_file
*file_priv
)
457 struct drm_i915_gem_set_tiling
*args
= data
;
458 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
459 struct drm_gem_object
*obj
;
460 struct drm_i915_gem_object
*obj_priv
;
463 obj
= drm_gem_object_lookup(dev
, file_priv
, args
->handle
);
466 obj_priv
= obj
->driver_private
;
468 if (!i915_tiling_ok(dev
, args
->stride
, obj
->size
, args
->tiling_mode
)) {
469 mutex_lock(&dev
->struct_mutex
);
470 drm_gem_object_unreference(obj
);
471 mutex_unlock(&dev
->struct_mutex
);
475 if (args
->tiling_mode
== I915_TILING_NONE
) {
476 args
->swizzle_mode
= I915_BIT_6_SWIZZLE_NONE
;
479 if (args
->tiling_mode
== I915_TILING_X
)
480 args
->swizzle_mode
= dev_priv
->mm
.bit_6_swizzle_x
;
482 args
->swizzle_mode
= dev_priv
->mm
.bit_6_swizzle_y
;
484 /* Hide bit 17 swizzling from the user. This prevents old Mesa
485 * from aborting the application on sw fallbacks to bit 17,
486 * and we use the pread/pwrite bit17 paths to swizzle for it.
487 * If there was a user that was relying on the swizzle
488 * information for drm_intel_bo_map()ed reads/writes this would
489 * break it, but we don't have any of those.
491 if (args
->swizzle_mode
== I915_BIT_6_SWIZZLE_9_17
)
492 args
->swizzle_mode
= I915_BIT_6_SWIZZLE_9
;
493 if (args
->swizzle_mode
== I915_BIT_6_SWIZZLE_9_10_17
)
494 args
->swizzle_mode
= I915_BIT_6_SWIZZLE_9_10
;
496 /* If we can't handle the swizzling, make it untiled. */
497 if (args
->swizzle_mode
== I915_BIT_6_SWIZZLE_UNKNOWN
) {
498 args
->tiling_mode
= I915_TILING_NONE
;
499 args
->swizzle_mode
= I915_BIT_6_SWIZZLE_NONE
;
504 mutex_lock(&dev
->struct_mutex
);
505 if (args
->tiling_mode
!= obj_priv
->tiling_mode
||
506 args
->stride
!= obj_priv
->stride
) {
507 /* We need to rebind the object if its current allocation
508 * no longer meets the alignment restrictions for its new
509 * tiling mode. Otherwise we can just leave it alone, but
510 * need to ensure that any fence register is cleared.
512 if (!i915_gem_object_fence_offset_ok(obj
, args
->tiling_mode
))
513 ret
= i915_gem_object_unbind(obj
);
515 ret
= i915_gem_object_put_fence_reg(obj
);
517 WARN(ret
!= -ERESTARTSYS
,
518 "failed to reset object for tiling switch");
519 args
->tiling_mode
= obj_priv
->tiling_mode
;
520 args
->stride
= obj_priv
->stride
;
524 /* If we've changed tiling, GTT-mappings of the object
525 * need to re-fault to ensure that the correct fence register
528 i915_gem_release_mmap(obj
);
530 obj_priv
->tiling_mode
= args
->tiling_mode
;
531 obj_priv
->stride
= args
->stride
;
534 drm_gem_object_unreference(obj
);
535 mutex_unlock(&dev
->struct_mutex
);
541 * Returns the current tiling mode and required bit 6 swizzling for the object.
544 i915_gem_get_tiling(struct drm_device
*dev
, void *data
,
545 struct drm_file
*file_priv
)
547 struct drm_i915_gem_get_tiling
*args
= data
;
548 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
549 struct drm_gem_object
*obj
;
550 struct drm_i915_gem_object
*obj_priv
;
552 obj
= drm_gem_object_lookup(dev
, file_priv
, args
->handle
);
555 obj_priv
= obj
->driver_private
;
557 mutex_lock(&dev
->struct_mutex
);
559 args
->tiling_mode
= obj_priv
->tiling_mode
;
560 switch (obj_priv
->tiling_mode
) {
562 args
->swizzle_mode
= dev_priv
->mm
.bit_6_swizzle_x
;
565 args
->swizzle_mode
= dev_priv
->mm
.bit_6_swizzle_y
;
567 case I915_TILING_NONE
:
568 args
->swizzle_mode
= I915_BIT_6_SWIZZLE_NONE
;
571 DRM_ERROR("unknown tiling mode\n");
574 /* Hide bit 17 from the user -- see comment in i915_gem_set_tiling */
575 if (args
->swizzle_mode
== I915_BIT_6_SWIZZLE_9_17
)
576 args
->swizzle_mode
= I915_BIT_6_SWIZZLE_9
;
577 if (args
->swizzle_mode
== I915_BIT_6_SWIZZLE_9_10_17
)
578 args
->swizzle_mode
= I915_BIT_6_SWIZZLE_9_10
;
580 drm_gem_object_unreference(obj
);
581 mutex_unlock(&dev
->struct_mutex
);
587 * Swap every 64 bytes of this page around, to account for it having a new
588 * bit 17 of its physical address and therefore being interpreted differently
592 i915_gem_swizzle_page(struct page
*page
)
602 for (i
= 0; i
< PAGE_SIZE
; i
+= 128) {
603 memcpy(temp
, &vaddr
[i
], 64);
604 memcpy(&vaddr
[i
], &vaddr
[i
+ 64], 64);
605 memcpy(&vaddr
[i
+ 64], temp
, 64);
614 i915_gem_object_do_bit_17_swizzle(struct drm_gem_object
*obj
)
616 struct drm_device
*dev
= obj
->dev
;
617 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
618 struct drm_i915_gem_object
*obj_priv
= obj
->driver_private
;
619 int page_count
= obj
->size
>> PAGE_SHIFT
;
622 if (dev_priv
->mm
.bit_6_swizzle_x
!= I915_BIT_6_SWIZZLE_9_10_17
)
625 if (obj_priv
->bit_17
== NULL
)
628 for (i
= 0; i
< page_count
; i
++) {
629 char new_bit_17
= page_to_phys(obj_priv
->pages
[i
]) >> 17;
630 if ((new_bit_17
& 0x1) !=
631 (test_bit(i
, obj_priv
->bit_17
) != 0)) {
632 int ret
= i915_gem_swizzle_page(obj_priv
->pages
[i
]);
634 DRM_ERROR("Failed to swizzle page\n");
637 set_page_dirty(obj_priv
->pages
[i
]);
643 i915_gem_object_save_bit_17_swizzle(struct drm_gem_object
*obj
)
645 struct drm_device
*dev
= obj
->dev
;
646 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
647 struct drm_i915_gem_object
*obj_priv
= obj
->driver_private
;
648 int page_count
= obj
->size
>> PAGE_SHIFT
;
651 if (dev_priv
->mm
.bit_6_swizzle_x
!= I915_BIT_6_SWIZZLE_9_10_17
)
654 if (obj_priv
->bit_17
== NULL
) {
655 obj_priv
->bit_17
= kmalloc(BITS_TO_LONGS(page_count
) *
656 sizeof(long), GFP_KERNEL
);
657 if (obj_priv
->bit_17
== NULL
) {
658 DRM_ERROR("Failed to allocate memory for bit 17 "
664 for (i
= 0; i
< page_count
; i
++) {
665 if (page_to_phys(obj_priv
->pages
[i
]) & (1 << 17))
666 __set_bit(i
, obj_priv
->bit_17
);
668 __clear_bit(i
, obj_priv
->bit_17
);