]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blame - drivers/gpu/drm/i915/intel_overlay.c
drm/i915: use pointer = k[cmz...]alloc(sizeof(*pointer), ...) pattern
[mirror_ubuntu-bionic-kernel.git] / drivers / gpu / drm / i915 / intel_overlay.c
CommitLineData
02e792fb
DV
1/*
2 * Copyright © 2009
3 *
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:
10 *
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
13 * Software.
14 *
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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 *
23 * Authors:
24 * Daniel Vetter <daniel@ffwll.ch>
25 *
26 * Derived from Xorg ddx, xf86-video-intel, src/i830_video.c
27 */
760285e7
DH
28#include <drm/drmP.h>
29#include <drm/i915_drm.h>
02e792fb
DV
30#include "i915_drv.h"
31#include "i915_reg.h"
32#include "intel_drv.h"
33
34/* Limits for overlay size. According to intel doc, the real limits are:
35 * Y width: 4095, UV width (planar): 2047, Y height: 2047,
36 * UV width (planar): * 1023. But the xorg thinks 2048 for height and width. Use
37 * the mininum of both. */
38#define IMAGE_MAX_WIDTH 2048
39#define IMAGE_MAX_HEIGHT 2046 /* 2 * 1023 */
40/* on 830 and 845 these large limits result in the card hanging */
41#define IMAGE_MAX_WIDTH_LEGACY 1024
42#define IMAGE_MAX_HEIGHT_LEGACY 1088
43
44/* overlay register definitions */
45/* OCMD register */
46#define OCMD_TILED_SURFACE (0x1<<19)
47#define OCMD_MIRROR_MASK (0x3<<17)
48#define OCMD_MIRROR_MODE (0x3<<17)
49#define OCMD_MIRROR_HORIZONTAL (0x1<<17)
50#define OCMD_MIRROR_VERTICAL (0x2<<17)
51#define OCMD_MIRROR_BOTH (0x3<<17)
52#define OCMD_BYTEORDER_MASK (0x3<<14) /* zero for YUYV or FOURCC YUY2 */
53#define OCMD_UV_SWAP (0x1<<14) /* YVYU */
54#define OCMD_Y_SWAP (0x2<<14) /* UYVY or FOURCC UYVY */
55#define OCMD_Y_AND_UV_SWAP (0x3<<14) /* VYUY */
56#define OCMD_SOURCE_FORMAT_MASK (0xf<<10)
57#define OCMD_RGB_888 (0x1<<10) /* not in i965 Intel docs */
58#define OCMD_RGB_555 (0x2<<10) /* not in i965 Intel docs */
59#define OCMD_RGB_565 (0x3<<10) /* not in i965 Intel docs */
60#define OCMD_YUV_422_PACKED (0x8<<10)
61#define OCMD_YUV_411_PACKED (0x9<<10) /* not in i965 Intel docs */
62#define OCMD_YUV_420_PLANAR (0xc<<10)
63#define OCMD_YUV_422_PLANAR (0xd<<10)
64#define OCMD_YUV_410_PLANAR (0xe<<10) /* also 411 */
65#define OCMD_TVSYNCFLIP_PARITY (0x1<<9)
66#define OCMD_TVSYNCFLIP_ENABLE (0x1<<7)
d7961364 67#define OCMD_BUF_TYPE_MASK (0x1<<5)
02e792fb
DV
68#define OCMD_BUF_TYPE_FRAME (0x0<<5)
69#define OCMD_BUF_TYPE_FIELD (0x1<<5)
70#define OCMD_TEST_MODE (0x1<<4)
71#define OCMD_BUFFER_SELECT (0x3<<2)
72#define OCMD_BUFFER0 (0x0<<2)
73#define OCMD_BUFFER1 (0x1<<2)
74#define OCMD_FIELD_SELECT (0x1<<2)
75#define OCMD_FIELD0 (0x0<<1)
76#define OCMD_FIELD1 (0x1<<1)
77#define OCMD_ENABLE (0x1<<0)
78
79/* OCONFIG register */
80#define OCONF_PIPE_MASK (0x1<<18)
81#define OCONF_PIPE_A (0x0<<18)
82#define OCONF_PIPE_B (0x1<<18)
83#define OCONF_GAMMA2_ENABLE (0x1<<16)
84#define OCONF_CSC_MODE_BT601 (0x0<<5)
85#define OCONF_CSC_MODE_BT709 (0x1<<5)
86#define OCONF_CSC_BYPASS (0x1<<4)
87#define OCONF_CC_OUT_8BIT (0x1<<3)
88#define OCONF_TEST_MODE (0x1<<2)
89#define OCONF_THREE_LINE_BUFFER (0x1<<0)
90#define OCONF_TWO_LINE_BUFFER (0x0<<0)
91
92/* DCLRKM (dst-key) register */
93#define DST_KEY_ENABLE (0x1<<31)
94#define CLK_RGB24_MASK 0x0
95#define CLK_RGB16_MASK 0x070307
96#define CLK_RGB15_MASK 0x070707
97#define CLK_RGB8I_MASK 0xffffff
98
99#define RGB16_TO_COLORKEY(c) \
100 (((c & 0xF800) << 8) | ((c & 0x07E0) << 5) | ((c & 0x001F) << 3))
101#define RGB15_TO_COLORKEY(c) \
102 (((c & 0x7c00) << 9) | ((c & 0x03E0) << 6) | ((c & 0x001F) << 3))
103
104/* overlay flip addr flag */
105#define OFC_UPDATE 0x1
106
107/* polyphase filter coefficients */
108#define N_HORIZ_Y_TAPS 5
109#define N_VERT_Y_TAPS 3
110#define N_HORIZ_UV_TAPS 3
111#define N_VERT_UV_TAPS 3
112#define N_PHASES 17
113#define MAX_TAPS 5
114
115/* memory bufferd overlay registers */
116struct overlay_registers {
0206e353
AJ
117 u32 OBUF_0Y;
118 u32 OBUF_1Y;
119 u32 OBUF_0U;
120 u32 OBUF_0V;
121 u32 OBUF_1U;
122 u32 OBUF_1V;
123 u32 OSTRIDE;
124 u32 YRGB_VPH;
125 u32 UV_VPH;
126 u32 HORZ_PH;
127 u32 INIT_PHS;
128 u32 DWINPOS;
129 u32 DWINSZ;
130 u32 SWIDTH;
131 u32 SWIDTHSW;
132 u32 SHEIGHT;
133 u32 YRGBSCALE;
134 u32 UVSCALE;
135 u32 OCLRC0;
136 u32 OCLRC1;
137 u32 DCLRKV;
138 u32 DCLRKM;
139 u32 SCLRKVH;
140 u32 SCLRKVL;
141 u32 SCLRKEN;
142 u32 OCONFIG;
143 u32 OCMD;
144 u32 RESERVED1; /* 0x6C */
145 u32 OSTART_0Y;
146 u32 OSTART_1Y;
147 u32 OSTART_0U;
148 u32 OSTART_0V;
149 u32 OSTART_1U;
150 u32 OSTART_1V;
151 u32 OTILEOFF_0Y;
152 u32 OTILEOFF_1Y;
153 u32 OTILEOFF_0U;
154 u32 OTILEOFF_0V;
155 u32 OTILEOFF_1U;
156 u32 OTILEOFF_1V;
157 u32 FASTHSCALE; /* 0xA0 */
158 u32 UVSCALEV; /* 0xA4 */
159 u32 RESERVEDC[(0x200 - 0xA8) / 4]; /* 0xA8 - 0x1FC */
160 u16 Y_VCOEFS[N_VERT_Y_TAPS * N_PHASES]; /* 0x200 */
161 u16 RESERVEDD[0x100 / 2 - N_VERT_Y_TAPS * N_PHASES];
162 u16 Y_HCOEFS[N_HORIZ_Y_TAPS * N_PHASES]; /* 0x300 */
163 u16 RESERVEDE[0x200 / 2 - N_HORIZ_Y_TAPS * N_PHASES];
164 u16 UV_VCOEFS[N_VERT_UV_TAPS * N_PHASES]; /* 0x500 */
165 u16 RESERVEDF[0x100 / 2 - N_VERT_UV_TAPS * N_PHASES];
166 u16 UV_HCOEFS[N_HORIZ_UV_TAPS * N_PHASES]; /* 0x600 */
167 u16 RESERVEDG[0x100 / 2 - N_HORIZ_UV_TAPS * N_PHASES];
02e792fb
DV
168};
169
23f09ce3
CW
170struct intel_overlay {
171 struct drm_device *dev;
172 struct intel_crtc *crtc;
173 struct drm_i915_gem_object *vid_bo;
174 struct drm_i915_gem_object *old_vid_bo;
175 int active;
176 int pfit_active;
177 u32 pfit_vscale_ratio; /* shifted-point number, (1<<12) == 1.0 */
178 u32 color_key;
179 u32 brightness, contrast, saturation;
180 u32 old_xscale, old_yscale;
181 /* register access */
182 u32 flip_addr;
183 struct drm_i915_gem_object *reg_bo;
184 /* flip handling */
185 uint32_t last_flip_req;
b303cf95 186 void (*flip_tail)(struct intel_overlay *);
23f09ce3 187};
02e792fb 188
75020bc1 189static struct overlay_registers __iomem *
8d74f656 190intel_overlay_map_regs(struct intel_overlay *overlay)
02e792fb 191{
0206e353 192 drm_i915_private_t *dev_priv = overlay->dev->dev_private;
75020bc1 193 struct overlay_registers __iomem *regs;
02e792fb 194
9bb2ff73 195 if (OVERLAY_NEEDS_PHYSICAL(overlay->dev))
75020bc1 196 regs = (struct overlay_registers __iomem *)overlay->reg_bo->phys_obj->handle->vaddr;
9bb2ff73 197 else
5d4545ae 198 regs = io_mapping_map_wc(dev_priv->gtt.mappable,
f343c5f6 199 i915_gem_obj_ggtt_offset(overlay->reg_bo));
02e792fb 200
9bb2ff73 201 return regs;
8d74f656 202}
02e792fb 203
9bb2ff73 204static void intel_overlay_unmap_regs(struct intel_overlay *overlay,
75020bc1 205 struct overlay_registers __iomem *regs)
8d74f656
CW
206{
207 if (!OVERLAY_NEEDS_PHYSICAL(overlay->dev))
9bb2ff73 208 io_mapping_unmap(regs);
02e792fb
DV
209}
210
b6c028e0 211static int intel_overlay_do_wait_request(struct intel_overlay *overlay,
b303cf95 212 void (*tail)(struct intel_overlay *))
02e792fb
DV
213{
214 struct drm_device *dev = overlay->dev;
852835f3 215 drm_i915_private_t *dev_priv = dev->dev_private;
6d90c952 216 struct intel_ring_buffer *ring = &dev_priv->ring[RCS];
b6c028e0 217 int ret;
02e792fb 218
b303cf95 219 BUG_ON(overlay->last_flip_req);
0025c077 220 ret = i915_add_request(ring, &overlay->last_flip_req);
acb868d3
CW
221 if (ret)
222 return ret;
223
b303cf95 224 overlay->flip_tail = tail;
199b2bc2 225 ret = i915_wait_seqno(ring, overlay->last_flip_req);
b6c028e0 226 if (ret)
03f77ea5 227 return ret;
b2da9fe5 228 i915_gem_retire_requests(dev);
02e792fb 229
03f77ea5 230 overlay->last_flip_req = 0;
02e792fb 231 return 0;
02e792fb
DV
232}
233
02e792fb
DV
234/* overlay needs to be disable in OCMD reg */
235static int intel_overlay_on(struct intel_overlay *overlay)
236{
237 struct drm_device *dev = overlay->dev;
e1f99ce6 238 struct drm_i915_private *dev_priv = dev->dev_private;
6d90c952 239 struct intel_ring_buffer *ring = &dev_priv->ring[RCS];
02e792fb 240 int ret;
02e792fb
DV
241
242 BUG_ON(overlay->active);
03f77ea5 243 overlay->active = 1;
b6c028e0 244
6306cb4f 245 WARN_ON(IS_I830(dev) && !(dev_priv->quirks & QUIRK_PIPEA_FORCE));
106dadac 246
6d90c952 247 ret = intel_ring_begin(ring, 4);
acb868d3
CW
248 if (ret)
249 return ret;
e1f99ce6 250
6d90c952
DV
251 intel_ring_emit(ring, MI_OVERLAY_FLIP | MI_OVERLAY_ON);
252 intel_ring_emit(ring, overlay->flip_addr | OFC_UPDATE);
253 intel_ring_emit(ring, MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP);
254 intel_ring_emit(ring, MI_NOOP);
255 intel_ring_advance(ring);
02e792fb 256
acb868d3 257 return intel_overlay_do_wait_request(overlay, NULL);
02e792fb
DV
258}
259
260/* overlay needs to be enabled in OCMD reg */
8dc5d147
CW
261static int intel_overlay_continue(struct intel_overlay *overlay,
262 bool load_polyphase_filter)
02e792fb
DV
263{
264 struct drm_device *dev = overlay->dev;
0206e353 265 drm_i915_private_t *dev_priv = dev->dev_private;
6d90c952 266 struct intel_ring_buffer *ring = &dev_priv->ring[RCS];
02e792fb
DV
267 u32 flip_addr = overlay->flip_addr;
268 u32 tmp;
e1f99ce6 269 int ret;
02e792fb
DV
270
271 BUG_ON(!overlay->active);
272
273 if (load_polyphase_filter)
274 flip_addr |= OFC_UPDATE;
275
276 /* check for underruns */
277 tmp = I915_READ(DOVSTA);
278 if (tmp & (1 << 17))
279 DRM_DEBUG("overlay underrun, DOVSTA: %x\n", tmp);
280
6d90c952 281 ret = intel_ring_begin(ring, 2);
acb868d3 282 if (ret)
e1f99ce6 283 return ret;
acb868d3 284
6d90c952
DV
285 intel_ring_emit(ring, MI_OVERLAY_FLIP | MI_OVERLAY_CONTINUE);
286 intel_ring_emit(ring, flip_addr);
287 intel_ring_advance(ring);
5a5a0c64 288
0025c077 289 return i915_add_request(ring, &overlay->last_flip_req);
5a5a0c64
DV
290}
291
b303cf95 292static void intel_overlay_release_old_vid_tail(struct intel_overlay *overlay)
5a5a0c64 293{
05394f39 294 struct drm_i915_gem_object *obj = overlay->old_vid_bo;
5a5a0c64 295
b303cf95 296 i915_gem_object_unpin(obj);
05394f39 297 drm_gem_object_unreference(&obj->base);
5a5a0c64 298
b303cf95
CW
299 overlay->old_vid_bo = NULL;
300}
03f77ea5 301
b303cf95
CW
302static void intel_overlay_off_tail(struct intel_overlay *overlay)
303{
05394f39 304 struct drm_i915_gem_object *obj = overlay->vid_bo;
02e792fb 305
b303cf95
CW
306 /* never have the overlay hw on without showing a frame */
307 BUG_ON(!overlay->vid_bo);
02e792fb 308
b303cf95 309 i915_gem_object_unpin(obj);
05394f39 310 drm_gem_object_unreference(&obj->base);
b303cf95 311 overlay->vid_bo = NULL;
03f77ea5 312
b303cf95
CW
313 overlay->crtc->overlay = NULL;
314 overlay->crtc = NULL;
315 overlay->active = 0;
02e792fb
DV
316}
317
318/* overlay needs to be disabled in OCMD reg */
ce453d81 319static int intel_overlay_off(struct intel_overlay *overlay)
02e792fb 320{
02e792fb 321 struct drm_device *dev = overlay->dev;
e1f99ce6 322 struct drm_i915_private *dev_priv = dev->dev_private;
6d90c952 323 struct intel_ring_buffer *ring = &dev_priv->ring[RCS];
8dc5d147 324 u32 flip_addr = overlay->flip_addr;
e1f99ce6 325 int ret;
02e792fb
DV
326
327 BUG_ON(!overlay->active);
328
329 /* According to intel docs the overlay hw may hang (when switching
330 * off) without loading the filter coeffs. It is however unclear whether
331 * this applies to the disabling of the overlay or to the switching off
332 * of the hw. Do it in both cases */
333 flip_addr |= OFC_UPDATE;
334
6d90c952 335 ret = intel_ring_begin(ring, 6);
acb868d3 336 if (ret)
e1f99ce6 337 return ret;
acb868d3 338
02e792fb 339 /* wait for overlay to go idle */
6d90c952
DV
340 intel_ring_emit(ring, MI_OVERLAY_FLIP | MI_OVERLAY_CONTINUE);
341 intel_ring_emit(ring, flip_addr);
342 intel_ring_emit(ring, MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP);
02e792fb 343 /* turn overlay off */
a9193983
DV
344 if (IS_I830(dev)) {
345 /* Workaround: Don't disable the overlay fully, since otherwise
346 * it dies on the next OVERLAY_ON cmd. */
347 intel_ring_emit(ring, MI_NOOP);
348 intel_ring_emit(ring, MI_NOOP);
349 intel_ring_emit(ring, MI_NOOP);
350 } else {
351 intel_ring_emit(ring, MI_OVERLAY_FLIP | MI_OVERLAY_OFF);
352 intel_ring_emit(ring, flip_addr);
353 intel_ring_emit(ring, MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP);
354 }
6d90c952 355 intel_ring_advance(ring);
02e792fb 356
acb868d3 357 return intel_overlay_do_wait_request(overlay, intel_overlay_off_tail);
12ca45fe
DV
358}
359
03f77ea5
DV
360/* recover from an interruption due to a signal
361 * We have to be careful not to repeat work forever an make forward progess. */
ce453d81 362static int intel_overlay_recover_from_interrupt(struct intel_overlay *overlay)
03f77ea5
DV
363{
364 struct drm_device *dev = overlay->dev;
852835f3 365 drm_i915_private_t *dev_priv = dev->dev_private;
6d90c952 366 struct intel_ring_buffer *ring = &dev_priv->ring[RCS];
03f77ea5 367 int ret;
03f77ea5 368
b303cf95
CW
369 if (overlay->last_flip_req == 0)
370 return 0;
03f77ea5 371
199b2bc2 372 ret = i915_wait_seqno(ring, overlay->last_flip_req);
b6c028e0 373 if (ret)
03f77ea5 374 return ret;
b2da9fe5 375 i915_gem_retire_requests(dev);
03f77ea5 376
b303cf95
CW
377 if (overlay->flip_tail)
378 overlay->flip_tail(overlay);
03f77ea5 379
03f77ea5
DV
380 overlay->last_flip_req = 0;
381 return 0;
382}
383
5a5a0c64
DV
384/* Wait for pending overlay flip and release old frame.
385 * Needs to be called before the overlay register are changed
8d74f656
CW
386 * via intel_overlay_(un)map_regs
387 */
02e792fb
DV
388static int intel_overlay_release_old_vid(struct intel_overlay *overlay)
389{
5cd68c98
CW
390 struct drm_device *dev = overlay->dev;
391 drm_i915_private_t *dev_priv = dev->dev_private;
6d90c952 392 struct intel_ring_buffer *ring = &dev_priv->ring[RCS];
02e792fb 393 int ret;
02e792fb 394
5cd68c98
CW
395 /* Only wait if there is actually an old frame to release to
396 * guarantee forward progress.
397 */
03f77ea5
DV
398 if (!overlay->old_vid_bo)
399 return 0;
400
5cd68c98
CW
401 if (I915_READ(ISR) & I915_OVERLAY_PLANE_FLIP_PENDING_INTERRUPT) {
402 /* synchronous slowpath */
6d90c952 403 ret = intel_ring_begin(ring, 2);
acb868d3 404 if (ret)
e1f99ce6 405 return ret;
e1f99ce6 406
6d90c952
DV
407 intel_ring_emit(ring, MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP);
408 intel_ring_emit(ring, MI_NOOP);
409 intel_ring_advance(ring);
5cd68c98 410
acb868d3 411 ret = intel_overlay_do_wait_request(overlay,
b303cf95 412 intel_overlay_release_old_vid_tail);
5cd68c98
CW
413 if (ret)
414 return ret;
415 }
02e792fb 416
5cd68c98 417 intel_overlay_release_old_vid_tail(overlay);
02e792fb
DV
418 return 0;
419}
420
421struct put_image_params {
422 int format;
423 short dst_x;
424 short dst_y;
425 short dst_w;
426 short dst_h;
427 short src_w;
428 short src_scan_h;
429 short src_scan_w;
430 short src_h;
431 short stride_Y;
432 short stride_UV;
433 int offset_Y;
434 int offset_U;
435 int offset_V;
436};
437
438static int packed_depth_bytes(u32 format)
439{
440 switch (format & I915_OVERLAY_DEPTH_MASK) {
722506f0
CW
441 case I915_OVERLAY_YUV422:
442 return 4;
443 case I915_OVERLAY_YUV411:
444 /* return 6; not implemented */
445 default:
446 return -EINVAL;
02e792fb
DV
447 }
448}
449
450static int packed_width_bytes(u32 format, short width)
451{
452 switch (format & I915_OVERLAY_DEPTH_MASK) {
722506f0
CW
453 case I915_OVERLAY_YUV422:
454 return width << 1;
455 default:
456 return -EINVAL;
02e792fb
DV
457 }
458}
459
460static int uv_hsubsampling(u32 format)
461{
462 switch (format & I915_OVERLAY_DEPTH_MASK) {
722506f0
CW
463 case I915_OVERLAY_YUV422:
464 case I915_OVERLAY_YUV420:
465 return 2;
466 case I915_OVERLAY_YUV411:
467 case I915_OVERLAY_YUV410:
468 return 4;
469 default:
470 return -EINVAL;
02e792fb
DV
471 }
472}
473
474static int uv_vsubsampling(u32 format)
475{
476 switch (format & I915_OVERLAY_DEPTH_MASK) {
722506f0
CW
477 case I915_OVERLAY_YUV420:
478 case I915_OVERLAY_YUV410:
479 return 2;
480 case I915_OVERLAY_YUV422:
481 case I915_OVERLAY_YUV411:
482 return 1;
483 default:
484 return -EINVAL;
02e792fb
DV
485 }
486}
487
488static u32 calc_swidthsw(struct drm_device *dev, u32 offset, u32 width)
489{
490 u32 mask, shift, ret;
a6c45cf0 491 if (IS_GEN2(dev)) {
02e792fb
DV
492 mask = 0x1f;
493 shift = 5;
a6c45cf0
CW
494 } else {
495 mask = 0x3f;
496 shift = 6;
02e792fb
DV
497 }
498 ret = ((offset + width + mask) >> shift) - (offset >> shift);
a6c45cf0 499 if (!IS_GEN2(dev))
02e792fb 500 ret <<= 1;
0206e353 501 ret -= 1;
02e792fb
DV
502 return ret << 2;
503}
504
505static const u16 y_static_hcoeffs[N_HORIZ_Y_TAPS * N_PHASES] = {
506 0x3000, 0xb4a0, 0x1930, 0x1920, 0xb4a0,
507 0x3000, 0xb500, 0x19d0, 0x1880, 0xb440,
508 0x3000, 0xb540, 0x1a88, 0x2f80, 0xb3e0,
509 0x3000, 0xb580, 0x1b30, 0x2e20, 0xb380,
510 0x3000, 0xb5c0, 0x1bd8, 0x2cc0, 0xb320,
511 0x3020, 0xb5e0, 0x1c60, 0x2b80, 0xb2c0,
512 0x3020, 0xb5e0, 0x1cf8, 0x2a20, 0xb260,
513 0x3020, 0xb5e0, 0x1d80, 0x28e0, 0xb200,
514 0x3020, 0xb5c0, 0x1e08, 0x3f40, 0xb1c0,
515 0x3020, 0xb580, 0x1e78, 0x3ce0, 0xb160,
516 0x3040, 0xb520, 0x1ed8, 0x3aa0, 0xb120,
517 0x3040, 0xb4a0, 0x1f30, 0x3880, 0xb0e0,
518 0x3040, 0xb400, 0x1f78, 0x3680, 0xb0a0,
519 0x3020, 0xb340, 0x1fb8, 0x34a0, 0xb060,
520 0x3020, 0xb240, 0x1fe0, 0x32e0, 0xb040,
521 0x3020, 0xb140, 0x1ff8, 0x3160, 0xb020,
722506f0
CW
522 0xb000, 0x3000, 0x0800, 0x3000, 0xb000
523};
524
02e792fb
DV
525static const u16 uv_static_hcoeffs[N_HORIZ_UV_TAPS * N_PHASES] = {
526 0x3000, 0x1800, 0x1800, 0xb000, 0x18d0, 0x2e60,
527 0xb000, 0x1990, 0x2ce0, 0xb020, 0x1a68, 0x2b40,
528 0xb040, 0x1b20, 0x29e0, 0xb060, 0x1bd8, 0x2880,
529 0xb080, 0x1c88, 0x3e60, 0xb0a0, 0x1d28, 0x3c00,
530 0xb0c0, 0x1db8, 0x39e0, 0xb0e0, 0x1e40, 0x37e0,
531 0xb100, 0x1eb8, 0x3620, 0xb100, 0x1f18, 0x34a0,
532 0xb100, 0x1f68, 0x3360, 0xb0e0, 0x1fa8, 0x3240,
533 0xb0c0, 0x1fe0, 0x3140, 0xb060, 0x1ff0, 0x30a0,
722506f0
CW
534 0x3000, 0x0800, 0x3000
535};
02e792fb 536
75020bc1 537static void update_polyphase_filter(struct overlay_registers __iomem *regs)
02e792fb 538{
75020bc1
BW
539 memcpy_toio(regs->Y_HCOEFS, y_static_hcoeffs, sizeof(y_static_hcoeffs));
540 memcpy_toio(regs->UV_HCOEFS, uv_static_hcoeffs,
541 sizeof(uv_static_hcoeffs));
02e792fb
DV
542}
543
544static bool update_scaling_factors(struct intel_overlay *overlay,
75020bc1 545 struct overlay_registers __iomem *regs,
02e792fb
DV
546 struct put_image_params *params)
547{
548 /* fixed point with a 12 bit shift */
549 u32 xscale, yscale, xscale_UV, yscale_UV;
550#define FP_SHIFT 12
551#define FRACT_MASK 0xfff
552 bool scale_changed = false;
553 int uv_hscale = uv_hsubsampling(params->format);
554 int uv_vscale = uv_vsubsampling(params->format);
555
556 if (params->dst_w > 1)
557 xscale = ((params->src_scan_w - 1) << FP_SHIFT)
558 /(params->dst_w);
559 else
560 xscale = 1 << FP_SHIFT;
561
562 if (params->dst_h > 1)
563 yscale = ((params->src_scan_h - 1) << FP_SHIFT)
564 /(params->dst_h);
565 else
566 yscale = 1 << FP_SHIFT;
567
568 /*if (params->format & I915_OVERLAY_YUV_PLANAR) {*/
722506f0
CW
569 xscale_UV = xscale/uv_hscale;
570 yscale_UV = yscale/uv_vscale;
571 /* make the Y scale to UV scale ratio an exact multiply */
572 xscale = xscale_UV * uv_hscale;
573 yscale = yscale_UV * uv_vscale;
02e792fb 574 /*} else {
722506f0
CW
575 xscale_UV = 0;
576 yscale_UV = 0;
577 }*/
02e792fb
DV
578
579 if (xscale != overlay->old_xscale || yscale != overlay->old_yscale)
580 scale_changed = true;
581 overlay->old_xscale = xscale;
582 overlay->old_yscale = yscale;
583
75020bc1
BW
584 iowrite32(((yscale & FRACT_MASK) << 20) |
585 ((xscale >> FP_SHIFT) << 16) |
586 ((xscale & FRACT_MASK) << 3),
587 &regs->YRGBSCALE);
722506f0 588
75020bc1
BW
589 iowrite32(((yscale_UV & FRACT_MASK) << 20) |
590 ((xscale_UV >> FP_SHIFT) << 16) |
591 ((xscale_UV & FRACT_MASK) << 3),
592 &regs->UVSCALE);
722506f0 593
75020bc1
BW
594 iowrite32((((yscale >> FP_SHIFT) << 16) |
595 ((yscale_UV >> FP_SHIFT) << 0)),
596 &regs->UVSCALEV);
02e792fb
DV
597
598 if (scale_changed)
599 update_polyphase_filter(regs);
600
601 return scale_changed;
602}
603
604static void update_colorkey(struct intel_overlay *overlay,
75020bc1 605 struct overlay_registers __iomem *regs)
02e792fb
DV
606{
607 u32 key = overlay->color_key;
6ba3ddd9 608
02e792fb 609 switch (overlay->crtc->base.fb->bits_per_pixel) {
722506f0 610 case 8:
75020bc1
BW
611 iowrite32(0, &regs->DCLRKV);
612 iowrite32(CLK_RGB8I_MASK | DST_KEY_ENABLE, &regs->DCLRKM);
6ba3ddd9
CW
613 break;
614
722506f0
CW
615 case 16:
616 if (overlay->crtc->base.fb->depth == 15) {
75020bc1
BW
617 iowrite32(RGB15_TO_COLORKEY(key), &regs->DCLRKV);
618 iowrite32(CLK_RGB15_MASK | DST_KEY_ENABLE,
619 &regs->DCLRKM);
722506f0 620 } else {
75020bc1
BW
621 iowrite32(RGB16_TO_COLORKEY(key), &regs->DCLRKV);
622 iowrite32(CLK_RGB16_MASK | DST_KEY_ENABLE,
623 &regs->DCLRKM);
722506f0 624 }
6ba3ddd9
CW
625 break;
626
722506f0
CW
627 case 24:
628 case 32:
75020bc1
BW
629 iowrite32(key, &regs->DCLRKV);
630 iowrite32(CLK_RGB24_MASK | DST_KEY_ENABLE, &regs->DCLRKM);
6ba3ddd9 631 break;
02e792fb
DV
632 }
633}
634
635static u32 overlay_cmd_reg(struct put_image_params *params)
636{
637 u32 cmd = OCMD_ENABLE | OCMD_BUF_TYPE_FRAME | OCMD_BUFFER0;
638
639 if (params->format & I915_OVERLAY_YUV_PLANAR) {
640 switch (params->format & I915_OVERLAY_DEPTH_MASK) {
722506f0
CW
641 case I915_OVERLAY_YUV422:
642 cmd |= OCMD_YUV_422_PLANAR;
643 break;
644 case I915_OVERLAY_YUV420:
645 cmd |= OCMD_YUV_420_PLANAR;
646 break;
647 case I915_OVERLAY_YUV411:
648 case I915_OVERLAY_YUV410:
649 cmd |= OCMD_YUV_410_PLANAR;
650 break;
02e792fb
DV
651 }
652 } else { /* YUV packed */
653 switch (params->format & I915_OVERLAY_DEPTH_MASK) {
722506f0
CW
654 case I915_OVERLAY_YUV422:
655 cmd |= OCMD_YUV_422_PACKED;
656 break;
657 case I915_OVERLAY_YUV411:
658 cmd |= OCMD_YUV_411_PACKED;
659 break;
02e792fb
DV
660 }
661
662 switch (params->format & I915_OVERLAY_SWAP_MASK) {
722506f0
CW
663 case I915_OVERLAY_NO_SWAP:
664 break;
665 case I915_OVERLAY_UV_SWAP:
666 cmd |= OCMD_UV_SWAP;
667 break;
668 case I915_OVERLAY_Y_SWAP:
669 cmd |= OCMD_Y_SWAP;
670 break;
671 case I915_OVERLAY_Y_AND_UV_SWAP:
672 cmd |= OCMD_Y_AND_UV_SWAP;
673 break;
02e792fb
DV
674 }
675 }
676
677 return cmd;
678}
679
5fe82c5e 680static int intel_overlay_do_put_image(struct intel_overlay *overlay,
05394f39 681 struct drm_i915_gem_object *new_bo,
5fe82c5e 682 struct put_image_params *params)
02e792fb
DV
683{
684 int ret, tmp_width;
75020bc1 685 struct overlay_registers __iomem *regs;
02e792fb 686 bool scale_changed = false;
02e792fb 687 struct drm_device *dev = overlay->dev;
75020bc1 688 u32 swidth, swidthsw, sheight, ostride;
02e792fb
DV
689
690 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
691 BUG_ON(!mutex_is_locked(&dev->mode_config.mutex));
692 BUG_ON(!overlay);
693
02e792fb
DV
694 ret = intel_overlay_release_old_vid(overlay);
695 if (ret != 0)
696 return ret;
697
2da3b9b9 698 ret = i915_gem_object_pin_to_display_plane(new_bo, 0, NULL);
02e792fb
DV
699 if (ret != 0)
700 return ret;
701
d9e86c0e
CW
702 ret = i915_gem_object_put_fence(new_bo);
703 if (ret)
704 goto out_unpin;
705
02e792fb 706 if (!overlay->active) {
75020bc1 707 u32 oconfig;
8d74f656 708 regs = intel_overlay_map_regs(overlay);
02e792fb
DV
709 if (!regs) {
710 ret = -ENOMEM;
711 goto out_unpin;
712 }
75020bc1 713 oconfig = OCONF_CC_OUT_8BIT;
a6c45cf0 714 if (IS_GEN4(overlay->dev))
75020bc1
BW
715 oconfig |= OCONF_CSC_MODE_BT709;
716 oconfig |= overlay->crtc->pipe == 0 ?
02e792fb 717 OCONF_PIPE_A : OCONF_PIPE_B;
75020bc1 718 iowrite32(oconfig, &regs->OCONFIG);
9bb2ff73 719 intel_overlay_unmap_regs(overlay, regs);
02e792fb
DV
720
721 ret = intel_overlay_on(overlay);
722 if (ret != 0)
723 goto out_unpin;
724 }
725
8d74f656 726 regs = intel_overlay_map_regs(overlay);
02e792fb
DV
727 if (!regs) {
728 ret = -ENOMEM;
729 goto out_unpin;
730 }
731
75020bc1
BW
732 iowrite32((params->dst_y << 16) | params->dst_x, &regs->DWINPOS);
733 iowrite32((params->dst_h << 16) | params->dst_w, &regs->DWINSZ);
02e792fb
DV
734
735 if (params->format & I915_OVERLAY_YUV_PACKED)
736 tmp_width = packed_width_bytes(params->format, params->src_w);
737 else
738 tmp_width = params->src_w;
739
75020bc1
BW
740 swidth = params->src_w;
741 swidthsw = calc_swidthsw(overlay->dev, params->offset_Y, tmp_width);
742 sheight = params->src_h;
f343c5f6 743 iowrite32(i915_gem_obj_ggtt_offset(new_bo) + params->offset_Y, &regs->OBUF_0Y);
75020bc1 744 ostride = params->stride_Y;
02e792fb
DV
745
746 if (params->format & I915_OVERLAY_YUV_PLANAR) {
747 int uv_hscale = uv_hsubsampling(params->format);
748 int uv_vscale = uv_vsubsampling(params->format);
749 u32 tmp_U, tmp_V;
75020bc1 750 swidth |= (params->src_w/uv_hscale) << 16;
02e792fb 751 tmp_U = calc_swidthsw(overlay->dev, params->offset_U,
722506f0 752 params->src_w/uv_hscale);
02e792fb 753 tmp_V = calc_swidthsw(overlay->dev, params->offset_V,
722506f0 754 params->src_w/uv_hscale);
75020bc1
BW
755 swidthsw |= max_t(u32, tmp_U, tmp_V) << 16;
756 sheight |= (params->src_h/uv_vscale) << 16;
f343c5f6
BW
757 iowrite32(i915_gem_obj_ggtt_offset(new_bo) + params->offset_U, &regs->OBUF_0U);
758 iowrite32(i915_gem_obj_ggtt_offset(new_bo) + params->offset_V, &regs->OBUF_0V);
75020bc1 759 ostride |= params->stride_UV << 16;
02e792fb
DV
760 }
761
75020bc1
BW
762 iowrite32(swidth, &regs->SWIDTH);
763 iowrite32(swidthsw, &regs->SWIDTHSW);
764 iowrite32(sheight, &regs->SHEIGHT);
765 iowrite32(ostride, &regs->OSTRIDE);
766
02e792fb
DV
767 scale_changed = update_scaling_factors(overlay, regs, params);
768
769 update_colorkey(overlay, regs);
770
75020bc1 771 iowrite32(overlay_cmd_reg(params), &regs->OCMD);
02e792fb 772
9bb2ff73 773 intel_overlay_unmap_regs(overlay, regs);
02e792fb 774
8dc5d147
CW
775 ret = intel_overlay_continue(overlay, scale_changed);
776 if (ret)
777 goto out_unpin;
02e792fb
DV
778
779 overlay->old_vid_bo = overlay->vid_bo;
05394f39 780 overlay->vid_bo = new_bo;
02e792fb
DV
781
782 return 0;
783
784out_unpin:
785 i915_gem_object_unpin(new_bo);
786 return ret;
787}
788
ce453d81 789int intel_overlay_switch_off(struct intel_overlay *overlay)
02e792fb 790{
75020bc1 791 struct overlay_registers __iomem *regs;
02e792fb 792 struct drm_device *dev = overlay->dev;
5dcdbcb0 793 int ret;
02e792fb
DV
794
795 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
796 BUG_ON(!mutex_is_locked(&dev->mode_config.mutex));
797
ce453d81 798 ret = intel_overlay_recover_from_interrupt(overlay);
b303cf95
CW
799 if (ret != 0)
800 return ret;
9bedb974 801
02e792fb
DV
802 if (!overlay->active)
803 return 0;
804
02e792fb
DV
805 ret = intel_overlay_release_old_vid(overlay);
806 if (ret != 0)
807 return ret;
808
8d74f656 809 regs = intel_overlay_map_regs(overlay);
75020bc1 810 iowrite32(0, &regs->OCMD);
9bb2ff73 811 intel_overlay_unmap_regs(overlay, regs);
02e792fb 812
ce453d81 813 ret = intel_overlay_off(overlay);
03f77ea5
DV
814 if (ret != 0)
815 return ret;
816
12ca45fe 817 intel_overlay_off_tail(overlay);
02e792fb
DV
818 return 0;
819}
820
821static int check_overlay_possible_on_crtc(struct intel_overlay *overlay,
822 struct intel_crtc *crtc)
823{
f7abfe8b 824 if (!crtc->active)
02e792fb
DV
825 return -EINVAL;
826
02e792fb 827 /* can't use the overlay with double wide pipe */
4926cb76 828 if (crtc->config.double_wide)
02e792fb
DV
829 return -EINVAL;
830
831 return 0;
832}
833
834static void update_pfit_vscale_ratio(struct intel_overlay *overlay)
835{
836 struct drm_device *dev = overlay->dev;
722506f0 837 drm_i915_private_t *dev_priv = dev->dev_private;
02e792fb 838 u32 pfit_control = I915_READ(PFIT_CONTROL);
446d2183 839 u32 ratio;
02e792fb
DV
840
841 /* XXX: This is not the same logic as in the xorg driver, but more in
446d2183
CW
842 * line with the intel documentation for the i965
843 */
a6c45cf0 844 if (INTEL_INFO(dev)->gen >= 4) {
0206e353 845 /* on i965 use the PGM reg to read out the autoscaler values */
a6c45cf0
CW
846 ratio = I915_READ(PFIT_PGM_RATIOS) >> PFIT_VERT_SCALE_SHIFT_965;
847 } else {
446d2183
CW
848 if (pfit_control & VERT_AUTO_SCALE)
849 ratio = I915_READ(PFIT_AUTO_RATIOS);
02e792fb 850 else
446d2183
CW
851 ratio = I915_READ(PFIT_PGM_RATIOS);
852 ratio >>= PFIT_VERT_SCALE_SHIFT;
02e792fb
DV
853 }
854
855 overlay->pfit_vscale_ratio = ratio;
856}
857
858static int check_overlay_dst(struct intel_overlay *overlay,
859 struct drm_intel_overlay_put_image *rec)
860{
861 struct drm_display_mode *mode = &overlay->crtc->base.mode;
862
75c13993
DV
863 if (rec->dst_x < mode->hdisplay &&
864 rec->dst_x + rec->dst_width <= mode->hdisplay &&
865 rec->dst_y < mode->vdisplay &&
866 rec->dst_y + rec->dst_height <= mode->vdisplay)
02e792fb
DV
867 return 0;
868 else
869 return -EINVAL;
870}
871
872static int check_overlay_scaling(struct put_image_params *rec)
873{
874 u32 tmp;
875
876 /* downscaling limit is 8.0 */
877 tmp = ((rec->src_scan_h << 16) / rec->dst_h) >> 16;
878 if (tmp > 7)
879 return -EINVAL;
880 tmp = ((rec->src_scan_w << 16) / rec->dst_w) >> 16;
881 if (tmp > 7)
882 return -EINVAL;
883
884 return 0;
885}
886
887static int check_overlay_src(struct drm_device *dev,
888 struct drm_intel_overlay_put_image *rec,
05394f39 889 struct drm_i915_gem_object *new_bo)
02e792fb 890{
02e792fb
DV
891 int uv_hscale = uv_hsubsampling(rec->flags);
892 int uv_vscale = uv_vsubsampling(rec->flags);
8f28f54a
DC
893 u32 stride_mask;
894 int depth;
895 u32 tmp;
02e792fb
DV
896
897 /* check src dimensions */
898 if (IS_845G(dev) || IS_I830(dev)) {
722506f0 899 if (rec->src_height > IMAGE_MAX_HEIGHT_LEGACY ||
9f7c3f44 900 rec->src_width > IMAGE_MAX_WIDTH_LEGACY)
02e792fb
DV
901 return -EINVAL;
902 } else {
722506f0 903 if (rec->src_height > IMAGE_MAX_HEIGHT ||
9f7c3f44 904 rec->src_width > IMAGE_MAX_WIDTH)
02e792fb
DV
905 return -EINVAL;
906 }
9f7c3f44 907
02e792fb 908 /* better safe than sorry, use 4 as the maximal subsampling ratio */
722506f0 909 if (rec->src_height < N_VERT_Y_TAPS*4 ||
9f7c3f44 910 rec->src_width < N_HORIZ_Y_TAPS*4)
02e792fb
DV
911 return -EINVAL;
912
a1efd14a 913 /* check alignment constraints */
02e792fb 914 switch (rec->flags & I915_OVERLAY_TYPE_MASK) {
722506f0
CW
915 case I915_OVERLAY_RGB:
916 /* not implemented */
917 return -EINVAL;
9f7c3f44 918
722506f0 919 case I915_OVERLAY_YUV_PACKED:
722506f0 920 if (uv_vscale != 1)
02e792fb 921 return -EINVAL;
9f7c3f44
CW
922
923 depth = packed_depth_bytes(rec->flags);
722506f0
CW
924 if (depth < 0)
925 return depth;
9f7c3f44 926
722506f0
CW
927 /* ignore UV planes */
928 rec->stride_UV = 0;
929 rec->offset_U = 0;
930 rec->offset_V = 0;
931 /* check pixel alignment */
932 if (rec->offset_Y % depth)
933 return -EINVAL;
934 break;
9f7c3f44 935
722506f0
CW
936 case I915_OVERLAY_YUV_PLANAR:
937 if (uv_vscale < 0 || uv_hscale < 0)
02e792fb 938 return -EINVAL;
722506f0
CW
939 /* no offset restrictions for planar formats */
940 break;
9f7c3f44 941
722506f0
CW
942 default:
943 return -EINVAL;
02e792fb
DV
944 }
945
946 if (rec->src_width % uv_hscale)
947 return -EINVAL;
948
949 /* stride checking */
a1efd14a
CW
950 if (IS_I830(dev) || IS_845G(dev))
951 stride_mask = 255;
952 else
953 stride_mask = 63;
02e792fb
DV
954
955 if (rec->stride_Y & stride_mask || rec->stride_UV & stride_mask)
956 return -EINVAL;
a6c45cf0 957 if (IS_GEN4(dev) && rec->stride_Y < 512)
02e792fb
DV
958 return -EINVAL;
959
960 tmp = (rec->flags & I915_OVERLAY_TYPE_MASK) == I915_OVERLAY_YUV_PLANAR ?
9f7c3f44
CW
961 4096 : 8192;
962 if (rec->stride_Y > tmp || rec->stride_UV > 2*1024)
02e792fb
DV
963 return -EINVAL;
964
965 /* check buffer dimensions */
966 switch (rec->flags & I915_OVERLAY_TYPE_MASK) {
722506f0
CW
967 case I915_OVERLAY_RGB:
968 case I915_OVERLAY_YUV_PACKED:
969 /* always 4 Y values per depth pixels */
970 if (packed_width_bytes(rec->flags, rec->src_width) > rec->stride_Y)
971 return -EINVAL;
972
973 tmp = rec->stride_Y*rec->src_height;
05394f39 974 if (rec->offset_Y + tmp > new_bo->base.size)
722506f0
CW
975 return -EINVAL;
976 break;
977
978 case I915_OVERLAY_YUV_PLANAR:
979 if (rec->src_width > rec->stride_Y)
980 return -EINVAL;
981 if (rec->src_width/uv_hscale > rec->stride_UV)
982 return -EINVAL;
983
9f7c3f44 984 tmp = rec->stride_Y * rec->src_height;
05394f39 985 if (rec->offset_Y + tmp > new_bo->base.size)
722506f0 986 return -EINVAL;
9f7c3f44
CW
987
988 tmp = rec->stride_UV * (rec->src_height / uv_vscale);
05394f39
CW
989 if (rec->offset_U + tmp > new_bo->base.size ||
990 rec->offset_V + tmp > new_bo->base.size)
722506f0
CW
991 return -EINVAL;
992 break;
02e792fb
DV
993 }
994
995 return 0;
996}
997
e9e331a8
CW
998/**
999 * Return the pipe currently connected to the panel fitter,
1000 * or -1 if the panel fitter is not present or not in use
1001 */
1002static int intel_panel_fitter_pipe(struct drm_device *dev)
1003{
1004 struct drm_i915_private *dev_priv = dev->dev_private;
1005 u32 pfit_control;
1006
1007 /* i830 doesn't have a panel fitter */
1008 if (IS_I830(dev))
1009 return -1;
1010
1011 pfit_control = I915_READ(PFIT_CONTROL);
1012
1013 /* See if the panel fitter is in use */
1014 if ((pfit_control & PFIT_ENABLE) == 0)
1015 return -1;
1016
1017 /* 965 can place panel fitter on either pipe */
a6c45cf0 1018 if (IS_GEN4(dev))
e9e331a8
CW
1019 return (pfit_control >> 29) & 0x3;
1020
1021 /* older chips can only use pipe 1 */
1022 return 1;
1023}
1024
02e792fb 1025int intel_overlay_put_image(struct drm_device *dev, void *data,
0206e353 1026 struct drm_file *file_priv)
02e792fb
DV
1027{
1028 struct drm_intel_overlay_put_image *put_image_rec = data;
1029 drm_i915_private_t *dev_priv = dev->dev_private;
1030 struct intel_overlay *overlay;
1031 struct drm_mode_object *drmmode_obj;
1032 struct intel_crtc *crtc;
05394f39 1033 struct drm_i915_gem_object *new_bo;
02e792fb
DV
1034 struct put_image_params *params;
1035 int ret;
1036
1cff8f6b 1037 /* No need to check for DRIVER_MODESET - we don't set it up then. */
02e792fb
DV
1038 overlay = dev_priv->overlay;
1039 if (!overlay) {
1040 DRM_DEBUG("userspace bug: no overlay\n");
1041 return -ENODEV;
1042 }
1043
1044 if (!(put_image_rec->flags & I915_OVERLAY_ENABLE)) {
a0e99e68 1045 drm_modeset_lock_all(dev);
02e792fb
DV
1046 mutex_lock(&dev->struct_mutex);
1047
ce453d81 1048 ret = intel_overlay_switch_off(overlay);
02e792fb
DV
1049
1050 mutex_unlock(&dev->struct_mutex);
a0e99e68 1051 drm_modeset_unlock_all(dev);
02e792fb
DV
1052
1053 return ret;
1054 }
1055
b14c5679 1056 params = kmalloc(sizeof(*params), GFP_KERNEL);
02e792fb
DV
1057 if (!params)
1058 return -ENOMEM;
1059
1060 drmmode_obj = drm_mode_object_find(dev, put_image_rec->crtc_id,
722506f0 1061 DRM_MODE_OBJECT_CRTC);
915a428e
DC
1062 if (!drmmode_obj) {
1063 ret = -ENOENT;
1064 goto out_free;
1065 }
02e792fb
DV
1066 crtc = to_intel_crtc(obj_to_crtc(drmmode_obj));
1067
05394f39
CW
1068 new_bo = to_intel_bo(drm_gem_object_lookup(dev, file_priv,
1069 put_image_rec->bo_handle));
c8725226 1070 if (&new_bo->base == NULL) {
915a428e
DC
1071 ret = -ENOENT;
1072 goto out_free;
1073 }
02e792fb 1074
a0e99e68 1075 drm_modeset_lock_all(dev);
02e792fb
DV
1076 mutex_lock(&dev->struct_mutex);
1077
d9e86c0e
CW
1078 if (new_bo->tiling_mode) {
1079 DRM_ERROR("buffer used for overlay image can not be tiled\n");
1080 ret = -EINVAL;
1081 goto out_unlock;
1082 }
1083
ce453d81 1084 ret = intel_overlay_recover_from_interrupt(overlay);
b303cf95
CW
1085 if (ret != 0)
1086 goto out_unlock;
03f77ea5 1087
02e792fb
DV
1088 if (overlay->crtc != crtc) {
1089 struct drm_display_mode *mode = &crtc->base.mode;
ce453d81 1090 ret = intel_overlay_switch_off(overlay);
02e792fb
DV
1091 if (ret != 0)
1092 goto out_unlock;
1093
1094 ret = check_overlay_possible_on_crtc(overlay, crtc);
1095 if (ret != 0)
1096 goto out_unlock;
1097
1098 overlay->crtc = crtc;
1099 crtc->overlay = overlay;
1100
e9e331a8
CW
1101 /* line too wide, i.e. one-line-mode */
1102 if (mode->hdisplay > 1024 &&
1103 intel_panel_fitter_pipe(dev) == crtc->pipe) {
02e792fb
DV
1104 overlay->pfit_active = 1;
1105 update_pfit_vscale_ratio(overlay);
1106 } else
1107 overlay->pfit_active = 0;
1108 }
1109
1110 ret = check_overlay_dst(overlay, put_image_rec);
1111 if (ret != 0)
1112 goto out_unlock;
1113
1114 if (overlay->pfit_active) {
1115 params->dst_y = ((((u32)put_image_rec->dst_y) << 12) /
722506f0 1116 overlay->pfit_vscale_ratio);
02e792fb
DV
1117 /* shifting right rounds downwards, so add 1 */
1118 params->dst_h = ((((u32)put_image_rec->dst_height) << 12) /
722506f0 1119 overlay->pfit_vscale_ratio) + 1;
02e792fb
DV
1120 } else {
1121 params->dst_y = put_image_rec->dst_y;
1122 params->dst_h = put_image_rec->dst_height;
1123 }
1124 params->dst_x = put_image_rec->dst_x;
1125 params->dst_w = put_image_rec->dst_width;
1126
1127 params->src_w = put_image_rec->src_width;
1128 params->src_h = put_image_rec->src_height;
1129 params->src_scan_w = put_image_rec->src_scan_width;
1130 params->src_scan_h = put_image_rec->src_scan_height;
722506f0
CW
1131 if (params->src_scan_h > params->src_h ||
1132 params->src_scan_w > params->src_w) {
02e792fb
DV
1133 ret = -EINVAL;
1134 goto out_unlock;
1135 }
1136
1137 ret = check_overlay_src(dev, put_image_rec, new_bo);
1138 if (ret != 0)
1139 goto out_unlock;
1140 params->format = put_image_rec->flags & ~I915_OVERLAY_FLAGS_MASK;
1141 params->stride_Y = put_image_rec->stride_Y;
1142 params->stride_UV = put_image_rec->stride_UV;
1143 params->offset_Y = put_image_rec->offset_Y;
1144 params->offset_U = put_image_rec->offset_U;
1145 params->offset_V = put_image_rec->offset_V;
1146
1147 /* Check scaling after src size to prevent a divide-by-zero. */
1148 ret = check_overlay_scaling(params);
1149 if (ret != 0)
1150 goto out_unlock;
1151
1152 ret = intel_overlay_do_put_image(overlay, new_bo, params);
1153 if (ret != 0)
1154 goto out_unlock;
1155
1156 mutex_unlock(&dev->struct_mutex);
a0e99e68 1157 drm_modeset_unlock_all(dev);
02e792fb
DV
1158
1159 kfree(params);
1160
1161 return 0;
1162
1163out_unlock:
1164 mutex_unlock(&dev->struct_mutex);
a0e99e68 1165 drm_modeset_unlock_all(dev);
05394f39 1166 drm_gem_object_unreference_unlocked(&new_bo->base);
915a428e 1167out_free:
02e792fb
DV
1168 kfree(params);
1169
1170 return ret;
1171}
1172
1173static void update_reg_attrs(struct intel_overlay *overlay,
75020bc1 1174 struct overlay_registers __iomem *regs)
02e792fb 1175{
75020bc1
BW
1176 iowrite32((overlay->contrast << 18) | (overlay->brightness & 0xff),
1177 &regs->OCLRC0);
1178 iowrite32(overlay->saturation, &regs->OCLRC1);
02e792fb
DV
1179}
1180
1181static bool check_gamma_bounds(u32 gamma1, u32 gamma2)
1182{
1183 int i;
1184
1185 if (gamma1 & 0xff000000 || gamma2 & 0xff000000)
1186 return false;
1187
1188 for (i = 0; i < 3; i++) {
722506f0 1189 if (((gamma1 >> i*8) & 0xff) >= ((gamma2 >> i*8) & 0xff))
02e792fb
DV
1190 return false;
1191 }
1192
1193 return true;
1194}
1195
1196static bool check_gamma5_errata(u32 gamma5)
1197{
1198 int i;
1199
1200 for (i = 0; i < 3; i++) {
1201 if (((gamma5 >> i*8) & 0xff) == 0x80)
1202 return false;
1203 }
1204
1205 return true;
1206}
1207
1208static int check_gamma(struct drm_intel_overlay_attrs *attrs)
1209{
722506f0
CW
1210 if (!check_gamma_bounds(0, attrs->gamma0) ||
1211 !check_gamma_bounds(attrs->gamma0, attrs->gamma1) ||
1212 !check_gamma_bounds(attrs->gamma1, attrs->gamma2) ||
1213 !check_gamma_bounds(attrs->gamma2, attrs->gamma3) ||
1214 !check_gamma_bounds(attrs->gamma3, attrs->gamma4) ||
1215 !check_gamma_bounds(attrs->gamma4, attrs->gamma5) ||
1216 !check_gamma_bounds(attrs->gamma5, 0x00ffffff))
02e792fb 1217 return -EINVAL;
722506f0 1218
02e792fb
DV
1219 if (!check_gamma5_errata(attrs->gamma5))
1220 return -EINVAL;
722506f0 1221
02e792fb
DV
1222 return 0;
1223}
1224
1225int intel_overlay_attrs(struct drm_device *dev, void *data,
0206e353 1226 struct drm_file *file_priv)
02e792fb
DV
1227{
1228 struct drm_intel_overlay_attrs *attrs = data;
0206e353 1229 drm_i915_private_t *dev_priv = dev->dev_private;
02e792fb 1230 struct intel_overlay *overlay;
75020bc1 1231 struct overlay_registers __iomem *regs;
02e792fb
DV
1232 int ret;
1233
1cff8f6b 1234 /* No need to check for DRIVER_MODESET - we don't set it up then. */
02e792fb
DV
1235 overlay = dev_priv->overlay;
1236 if (!overlay) {
1237 DRM_DEBUG("userspace bug: no overlay\n");
1238 return -ENODEV;
1239 }
1240
a0e99e68 1241 drm_modeset_lock_all(dev);
02e792fb
DV
1242 mutex_lock(&dev->struct_mutex);
1243
60fc332c 1244 ret = -EINVAL;
02e792fb 1245 if (!(attrs->flags & I915_OVERLAY_UPDATE_ATTRS)) {
60fc332c 1246 attrs->color_key = overlay->color_key;
02e792fb 1247 attrs->brightness = overlay->brightness;
60fc332c 1248 attrs->contrast = overlay->contrast;
02e792fb
DV
1249 attrs->saturation = overlay->saturation;
1250
a6c45cf0 1251 if (!IS_GEN2(dev)) {
02e792fb
DV
1252 attrs->gamma0 = I915_READ(OGAMC0);
1253 attrs->gamma1 = I915_READ(OGAMC1);
1254 attrs->gamma2 = I915_READ(OGAMC2);
1255 attrs->gamma3 = I915_READ(OGAMC3);
1256 attrs->gamma4 = I915_READ(OGAMC4);
1257 attrs->gamma5 = I915_READ(OGAMC5);
1258 }
02e792fb 1259 } else {
60fc332c 1260 if (attrs->brightness < -128 || attrs->brightness > 127)
02e792fb 1261 goto out_unlock;
60fc332c 1262 if (attrs->contrast > 255)
02e792fb 1263 goto out_unlock;
60fc332c 1264 if (attrs->saturation > 1023)
02e792fb 1265 goto out_unlock;
02e792fb 1266
60fc332c
CW
1267 overlay->color_key = attrs->color_key;
1268 overlay->brightness = attrs->brightness;
1269 overlay->contrast = attrs->contrast;
1270 overlay->saturation = attrs->saturation;
02e792fb 1271
8d74f656 1272 regs = intel_overlay_map_regs(overlay);
02e792fb
DV
1273 if (!regs) {
1274 ret = -ENOMEM;
1275 goto out_unlock;
1276 }
1277
1278 update_reg_attrs(overlay, regs);
1279
9bb2ff73 1280 intel_overlay_unmap_regs(overlay, regs);
02e792fb
DV
1281
1282 if (attrs->flags & I915_OVERLAY_UPDATE_GAMMA) {
a6c45cf0 1283 if (IS_GEN2(dev))
02e792fb 1284 goto out_unlock;
02e792fb
DV
1285
1286 if (overlay->active) {
1287 ret = -EBUSY;
1288 goto out_unlock;
1289 }
1290
1291 ret = check_gamma(attrs);
60fc332c 1292 if (ret)
02e792fb
DV
1293 goto out_unlock;
1294
1295 I915_WRITE(OGAMC0, attrs->gamma0);
1296 I915_WRITE(OGAMC1, attrs->gamma1);
1297 I915_WRITE(OGAMC2, attrs->gamma2);
1298 I915_WRITE(OGAMC3, attrs->gamma3);
1299 I915_WRITE(OGAMC4, attrs->gamma4);
1300 I915_WRITE(OGAMC5, attrs->gamma5);
1301 }
02e792fb
DV
1302 }
1303
60fc332c 1304 ret = 0;
02e792fb
DV
1305out_unlock:
1306 mutex_unlock(&dev->struct_mutex);
a0e99e68 1307 drm_modeset_unlock_all(dev);
02e792fb
DV
1308
1309 return ret;
1310}
1311
1312void intel_setup_overlay(struct drm_device *dev)
1313{
0206e353 1314 drm_i915_private_t *dev_priv = dev->dev_private;
02e792fb 1315 struct intel_overlay *overlay;
05394f39 1316 struct drm_i915_gem_object *reg_bo;
75020bc1 1317 struct overlay_registers __iomem *regs;
02e792fb
DV
1318 int ret;
1319
31578148 1320 if (!HAS_OVERLAY(dev))
02e792fb
DV
1321 return;
1322
b14c5679 1323 overlay = kzalloc(sizeof(*overlay), GFP_KERNEL);
02e792fb
DV
1324 if (!overlay)
1325 return;
79d24273
CW
1326
1327 mutex_lock(&dev->struct_mutex);
1328 if (WARN_ON(dev_priv->overlay))
1329 goto out_free;
1330
02e792fb
DV
1331 overlay->dev = dev;
1332
f63a484c
DV
1333 reg_bo = NULL;
1334 if (!OVERLAY_NEEDS_PHYSICAL(dev))
1335 reg_bo = i915_gem_object_create_stolen(dev, PAGE_SIZE);
80405138
CW
1336 if (reg_bo == NULL)
1337 reg_bo = i915_gem_alloc_object(dev, PAGE_SIZE);
1338 if (reg_bo == NULL)
02e792fb 1339 goto out_free;
05394f39 1340 overlay->reg_bo = reg_bo;
02e792fb 1341
31578148 1342 if (OVERLAY_NEEDS_PHYSICAL(dev)) {
02e792fb 1343 ret = i915_gem_attach_phys_object(dev, reg_bo,
6eeefaf3 1344 I915_GEM_PHYS_OVERLAY_REGS,
a2930128 1345 PAGE_SIZE);
0206e353
AJ
1346 if (ret) {
1347 DRM_ERROR("failed to attach phys overlay regs\n");
1348 goto out_free_bo;
1349 }
05394f39 1350 overlay->flip_addr = reg_bo->phys_obj->handle->busaddr;
31578148 1351 } else {
c37e2204 1352 ret = i915_gem_obj_ggtt_pin(reg_bo, PAGE_SIZE, true, false);
02e792fb 1353 if (ret) {
0206e353
AJ
1354 DRM_ERROR("failed to pin overlay register bo\n");
1355 goto out_free_bo;
1356 }
f343c5f6 1357 overlay->flip_addr = i915_gem_obj_ggtt_offset(reg_bo);
0ddc1289
CW
1358
1359 ret = i915_gem_object_set_to_gtt_domain(reg_bo, true);
1360 if (ret) {
0206e353
AJ
1361 DRM_ERROR("failed to move overlay register bo into the GTT\n");
1362 goto out_unpin_bo;
1363 }
02e792fb
DV
1364 }
1365
1366 /* init all values */
1367 overlay->color_key = 0x0101fe;
1368 overlay->brightness = -19;
1369 overlay->contrast = 75;
1370 overlay->saturation = 146;
1371
8d74f656 1372 regs = intel_overlay_map_regs(overlay);
02e792fb 1373 if (!regs)
79d24273 1374 goto out_unpin_bo;
02e792fb 1375
75020bc1 1376 memset_io(regs, 0, sizeof(struct overlay_registers));
02e792fb 1377 update_polyphase_filter(regs);
02e792fb
DV
1378 update_reg_attrs(overlay, regs);
1379
9bb2ff73 1380 intel_overlay_unmap_regs(overlay, regs);
02e792fb
DV
1381
1382 dev_priv->overlay = overlay;
79d24273 1383 mutex_unlock(&dev->struct_mutex);
02e792fb
DV
1384 DRM_INFO("initialized overlay support\n");
1385 return;
1386
0ddc1289 1387out_unpin_bo:
79d24273
CW
1388 if (!OVERLAY_NEEDS_PHYSICAL(dev))
1389 i915_gem_object_unpin(reg_bo);
02e792fb 1390out_free_bo:
05394f39 1391 drm_gem_object_unreference(&reg_bo->base);
02e792fb 1392out_free:
79d24273 1393 mutex_unlock(&dev->struct_mutex);
02e792fb
DV
1394 kfree(overlay);
1395 return;
1396}
1397
1398void intel_cleanup_overlay(struct drm_device *dev)
1399{
722506f0 1400 drm_i915_private_t *dev_priv = dev->dev_private;
02e792fb 1401
62cf4e6f
CW
1402 if (!dev_priv->overlay)
1403 return;
02e792fb 1404
62cf4e6f
CW
1405 /* The bo's should be free'd by the generic code already.
1406 * Furthermore modesetting teardown happens beforehand so the
1407 * hardware should be off already */
1408 BUG_ON(dev_priv->overlay->active);
1409
1410 drm_gem_object_unreference_unlocked(&dev_priv->overlay->reg_bo->base);
1411 kfree(dev_priv->overlay);
02e792fb 1412}
6ef3d427
CW
1413
1414struct intel_overlay_error_state {
1415 struct overlay_registers regs;
1416 unsigned long base;
1417 u32 dovsta;
1418 u32 isr;
1419};
1420
75020bc1 1421static struct overlay_registers __iomem *
c48c43e4 1422intel_overlay_map_regs_atomic(struct intel_overlay *overlay)
3bd3c932 1423{
c48c43e4 1424 drm_i915_private_t *dev_priv = overlay->dev->dev_private;
75020bc1 1425 struct overlay_registers __iomem *regs;
3bd3c932
CW
1426
1427 if (OVERLAY_NEEDS_PHYSICAL(overlay->dev))
75020bc1
BW
1428 /* Cast to make sparse happy, but it's wc memory anyway, so
1429 * equivalent to the wc io mapping on X86. */
1430 regs = (struct overlay_registers __iomem *)
1431 overlay->reg_bo->phys_obj->handle->vaddr;
3bd3c932 1432 else
5d4545ae 1433 regs = io_mapping_map_atomic_wc(dev_priv->gtt.mappable,
f343c5f6 1434 i915_gem_obj_ggtt_offset(overlay->reg_bo));
3bd3c932
CW
1435
1436 return regs;
1437}
1438
1439static void intel_overlay_unmap_regs_atomic(struct intel_overlay *overlay,
75020bc1 1440 struct overlay_registers __iomem *regs)
3bd3c932
CW
1441{
1442 if (!OVERLAY_NEEDS_PHYSICAL(overlay->dev))
c48c43e4 1443 io_mapping_unmap_atomic(regs);
3bd3c932
CW
1444}
1445
1446
6ef3d427
CW
1447struct intel_overlay_error_state *
1448intel_overlay_capture_error_state(struct drm_device *dev)
1449{
0206e353 1450 drm_i915_private_t *dev_priv = dev->dev_private;
6ef3d427
CW
1451 struct intel_overlay *overlay = dev_priv->overlay;
1452 struct intel_overlay_error_state *error;
1453 struct overlay_registers __iomem *regs;
1454
1455 if (!overlay || !overlay->active)
1456 return NULL;
1457
1458 error = kmalloc(sizeof(*error), GFP_ATOMIC);
1459 if (error == NULL)
1460 return NULL;
1461
1462 error->dovsta = I915_READ(DOVSTA);
1463 error->isr = I915_READ(ISR);
31578148 1464 if (OVERLAY_NEEDS_PHYSICAL(overlay->dev))
75020bc1 1465 error->base = (__force long)overlay->reg_bo->phys_obj->handle->vaddr;
31578148 1466 else
f343c5f6 1467 error->base = i915_gem_obj_ggtt_offset(overlay->reg_bo);
6ef3d427
CW
1468
1469 regs = intel_overlay_map_regs_atomic(overlay);
1470 if (!regs)
1471 goto err;
1472
1473 memcpy_fromio(&error->regs, regs, sizeof(struct overlay_registers));
c48c43e4 1474 intel_overlay_unmap_regs_atomic(overlay, regs);
6ef3d427
CW
1475
1476 return error;
1477
1478err:
1479 kfree(error);
1480 return NULL;
1481}
1482
1483void
edc3d884
MK
1484intel_overlay_print_error_state(struct drm_i915_error_state_buf *m,
1485 struct intel_overlay_error_state *error)
6ef3d427 1486{
edc3d884
MK
1487 i915_error_printf(m, "Overlay, status: 0x%08x, interrupt: 0x%08x\n",
1488 error->dovsta, error->isr);
1489 i915_error_printf(m, " Register file at 0x%08lx:\n",
1490 error->base);
6ef3d427 1491
edc3d884 1492#define P(x) i915_error_printf(m, " " #x ": 0x%08x\n", error->regs.x)
6ef3d427
CW
1493 P(OBUF_0Y);
1494 P(OBUF_1Y);
1495 P(OBUF_0U);
1496 P(OBUF_0V);
1497 P(OBUF_1U);
1498 P(OBUF_1V);
1499 P(OSTRIDE);
1500 P(YRGB_VPH);
1501 P(UV_VPH);
1502 P(HORZ_PH);
1503 P(INIT_PHS);
1504 P(DWINPOS);
1505 P(DWINSZ);
1506 P(SWIDTH);
1507 P(SWIDTHSW);
1508 P(SHEIGHT);
1509 P(YRGBSCALE);
1510 P(UVSCALE);
1511 P(OCLRC0);
1512 P(OCLRC1);
1513 P(DCLRKV);
1514 P(DCLRKM);
1515 P(SCLRKVH);
1516 P(SCLRKVL);
1517 P(SCLRKEN);
1518 P(OCONFIG);
1519 P(OCMD);
1520 P(OSTART_0Y);
1521 P(OSTART_1Y);
1522 P(OSTART_0U);
1523 P(OSTART_0V);
1524 P(OSTART_1U);
1525 P(OSTART_1V);
1526 P(OTILEOFF_0Y);
1527 P(OTILEOFF_1Y);
1528 P(OTILEOFF_0U);
1529 P(OTILEOFF_0V);
1530 P(OTILEOFF_1U);
1531 P(OTILEOFF_1V);
1532 P(FASTHSCALE);
1533 P(UVSCALEV);
1534#undef P
1535}