]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blame - drivers/gpu/drm/i915/intel_overlay.c
drm/i915: Rename drm_gem_object_unreference in preparation for lockless free
[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 170struct intel_overlay {
1ee8da6d 171 struct drm_i915_private *i915;
23f09ce3
CW
172 struct intel_crtc *crtc;
173 struct drm_i915_gem_object *vid_bo;
174 struct drm_i915_gem_object *old_vid_bo;
209c2a5e
VS
175 bool active;
176 bool pfit_active;
23f09ce3 177 u32 pfit_vscale_ratio; /* shifted-point number, (1<<12) == 1.0 */
ea9da4e4
CW
178 u32 color_key:24;
179 u32 color_key_enabled:1;
23f09ce3
CW
180 u32 brightness, contrast, saturation;
181 u32 old_xscale, old_yscale;
182 /* register access */
183 u32 flip_addr;
184 struct drm_i915_gem_object *reg_bo;
185 /* flip handling */
9bfc01a2 186 struct drm_i915_gem_request *last_flip_req;
b303cf95 187 void (*flip_tail)(struct intel_overlay *);
23f09ce3 188};
02e792fb 189
75020bc1 190static struct overlay_registers __iomem *
8d74f656 191intel_overlay_map_regs(struct intel_overlay *overlay)
02e792fb 192{
1ee8da6d 193 struct drm_i915_private *dev_priv = overlay->i915;
75020bc1 194 struct overlay_registers __iomem *regs;
02e792fb 195
1ee8da6d 196 if (OVERLAY_NEEDS_PHYSICAL(dev_priv))
00731155 197 regs = (struct overlay_registers __iomem *)overlay->reg_bo->phys_handle->vaddr;
9bb2ff73 198 else
1ee8da6d 199 regs = io_mapping_map_wc(dev_priv->ggtt.mappable,
d8dab00d
CW
200 overlay->flip_addr,
201 PAGE_SIZE);
02e792fb 202
9bb2ff73 203 return regs;
8d74f656 204}
02e792fb 205
9bb2ff73 206static void intel_overlay_unmap_regs(struct intel_overlay *overlay,
75020bc1 207 struct overlay_registers __iomem *regs)
8d74f656 208{
1ee8da6d 209 if (!OVERLAY_NEEDS_PHYSICAL(overlay->i915))
9bb2ff73 210 io_mapping_unmap(regs);
02e792fb
DV
211}
212
b6c028e0 213static int intel_overlay_do_wait_request(struct intel_overlay *overlay,
dad540ce 214 struct drm_i915_gem_request *req,
b303cf95 215 void (*tail)(struct intel_overlay *))
02e792fb 216{
b6c028e0 217 int ret;
02e792fb 218
77589f56 219 WARN_ON(overlay->last_flip_req);
dad540ce 220 i915_gem_request_assign(&overlay->last_flip_req, req);
75289874 221 i915_add_request(req);
acb868d3 222
b303cf95 223 overlay->flip_tail = tail;
a4b3a571 224 ret = i915_wait_request(overlay->last_flip_req);
b6c028e0 225 if (ret)
03f77ea5 226 return ret;
02e792fb 227
9bfc01a2 228 i915_gem_request_assign(&overlay->last_flip_req, NULL);
02e792fb 229 return 0;
02e792fb
DV
230}
231
02e792fb
DV
232/* overlay needs to be disable in OCMD reg */
233static int intel_overlay_on(struct intel_overlay *overlay)
234{
1ee8da6d 235 struct drm_i915_private *dev_priv = overlay->i915;
4a570db5 236 struct intel_engine_cs *engine = &dev_priv->engine[RCS];
dad540ce 237 struct drm_i915_gem_request *req;
02e792fb 238 int ret;
02e792fb 239
77589f56 240 WARN_ON(overlay->active);
1ee8da6d 241 WARN_ON(IS_I830(dev_priv) && !(dev_priv->quirks & QUIRK_PIPEA_FORCE));
106dadac 242
e2f80391 243 req = i915_gem_request_alloc(engine, NULL);
26827088
DG
244 if (IS_ERR(req))
245 return PTR_ERR(req);
e1f99ce6 246
5fb9de1a 247 ret = intel_ring_begin(req, 4);
dad540ce 248 if (ret) {
aa9b7810 249 i915_add_request_no_flush(req);
dad540ce
JH
250 return ret;
251 }
252
1c7c4301
VS
253 overlay->active = true;
254
e2f80391
TU
255 intel_ring_emit(engine, MI_OVERLAY_FLIP | MI_OVERLAY_ON);
256 intel_ring_emit(engine, overlay->flip_addr | OFC_UPDATE);
257 intel_ring_emit(engine, MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP);
258 intel_ring_emit(engine, MI_NOOP);
259 intel_ring_advance(engine);
02e792fb 260
dad540ce 261 return intel_overlay_do_wait_request(overlay, req, NULL);
02e792fb
DV
262}
263
264/* overlay needs to be enabled in OCMD reg */
8dc5d147
CW
265static int intel_overlay_continue(struct intel_overlay *overlay,
266 bool load_polyphase_filter)
02e792fb 267{
1ee8da6d 268 struct drm_i915_private *dev_priv = overlay->i915;
4a570db5 269 struct intel_engine_cs *engine = &dev_priv->engine[RCS];
dad540ce 270 struct drm_i915_gem_request *req;
02e792fb
DV
271 u32 flip_addr = overlay->flip_addr;
272 u32 tmp;
e1f99ce6 273 int ret;
02e792fb 274
77589f56 275 WARN_ON(!overlay->active);
02e792fb
DV
276
277 if (load_polyphase_filter)
278 flip_addr |= OFC_UPDATE;
279
280 /* check for underruns */
281 tmp = I915_READ(DOVSTA);
282 if (tmp & (1 << 17))
283 DRM_DEBUG("overlay underrun, DOVSTA: %x\n", tmp);
284
e2f80391 285 req = i915_gem_request_alloc(engine, NULL);
26827088
DG
286 if (IS_ERR(req))
287 return PTR_ERR(req);
acb868d3 288
5fb9de1a 289 ret = intel_ring_begin(req, 2);
dad540ce 290 if (ret) {
aa9b7810 291 i915_add_request_no_flush(req);
dad540ce
JH
292 return ret;
293 }
294
e2f80391
TU
295 intel_ring_emit(engine, MI_OVERLAY_FLIP | MI_OVERLAY_CONTINUE);
296 intel_ring_emit(engine, flip_addr);
297 intel_ring_advance(engine);
5a5a0c64 298
9bfc01a2 299 WARN_ON(overlay->last_flip_req);
dad540ce 300 i915_gem_request_assign(&overlay->last_flip_req, req);
75289874 301 i915_add_request(req);
bf7dc5b7
JH
302
303 return 0;
5a5a0c64
DV
304}
305
b303cf95 306static void intel_overlay_release_old_vid_tail(struct intel_overlay *overlay)
5a5a0c64 307{
05394f39 308 struct drm_i915_gem_object *obj = overlay->old_vid_bo;
5a5a0c64 309
d7f46fc4 310 i915_gem_object_ggtt_unpin(obj);
f8c417cd 311 i915_gem_object_put(obj);
5a5a0c64 312
b303cf95
CW
313 overlay->old_vid_bo = NULL;
314}
03f77ea5 315
b303cf95
CW
316static void intel_overlay_off_tail(struct intel_overlay *overlay)
317{
05394f39 318 struct drm_i915_gem_object *obj = overlay->vid_bo;
02e792fb 319
b303cf95 320 /* never have the overlay hw on without showing a frame */
77589f56
VS
321 if (WARN_ON(!obj))
322 return;
02e792fb 323
d7f46fc4 324 i915_gem_object_ggtt_unpin(obj);
f8c417cd 325 i915_gem_object_put(obj);
b303cf95 326 overlay->vid_bo = NULL;
03f77ea5 327
b303cf95
CW
328 overlay->crtc->overlay = NULL;
329 overlay->crtc = NULL;
209c2a5e 330 overlay->active = false;
02e792fb
DV
331}
332
333/* overlay needs to be disabled in OCMD reg */
ce453d81 334static int intel_overlay_off(struct intel_overlay *overlay)
02e792fb 335{
1ee8da6d 336 struct drm_i915_private *dev_priv = overlay->i915;
4a570db5 337 struct intel_engine_cs *engine = &dev_priv->engine[RCS];
dad540ce 338 struct drm_i915_gem_request *req;
8dc5d147 339 u32 flip_addr = overlay->flip_addr;
e1f99ce6 340 int ret;
02e792fb 341
77589f56 342 WARN_ON(!overlay->active);
02e792fb
DV
343
344 /* According to intel docs the overlay hw may hang (when switching
345 * off) without loading the filter coeffs. It is however unclear whether
346 * this applies to the disabling of the overlay or to the switching off
347 * of the hw. Do it in both cases */
348 flip_addr |= OFC_UPDATE;
349
e2f80391 350 req = i915_gem_request_alloc(engine, NULL);
26827088
DG
351 if (IS_ERR(req))
352 return PTR_ERR(req);
acb868d3 353
5fb9de1a 354 ret = intel_ring_begin(req, 6);
dad540ce 355 if (ret) {
aa9b7810 356 i915_add_request_no_flush(req);
dad540ce
JH
357 return ret;
358 }
359
02e792fb 360 /* wait for overlay to go idle */
e2f80391
TU
361 intel_ring_emit(engine, MI_OVERLAY_FLIP | MI_OVERLAY_CONTINUE);
362 intel_ring_emit(engine, flip_addr);
363 intel_ring_emit(engine, MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP);
02e792fb 364 /* turn overlay off */
1ee8da6d 365 if (IS_I830(dev_priv)) {
a9193983
DV
366 /* Workaround: Don't disable the overlay fully, since otherwise
367 * it dies on the next OVERLAY_ON cmd. */
e2f80391
TU
368 intel_ring_emit(engine, MI_NOOP);
369 intel_ring_emit(engine, MI_NOOP);
370 intel_ring_emit(engine, MI_NOOP);
a9193983 371 } else {
e2f80391
TU
372 intel_ring_emit(engine, MI_OVERLAY_FLIP | MI_OVERLAY_OFF);
373 intel_ring_emit(engine, flip_addr);
374 intel_ring_emit(engine,
375 MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP);
a9193983 376 }
e2f80391 377 intel_ring_advance(engine);
02e792fb 378
dad540ce 379 return intel_overlay_do_wait_request(overlay, req, intel_overlay_off_tail);
12ca45fe
DV
380}
381
03f77ea5
DV
382/* recover from an interruption due to a signal
383 * We have to be careful not to repeat work forever an make forward progess. */
ce453d81 384static int intel_overlay_recover_from_interrupt(struct intel_overlay *overlay)
03f77ea5 385{
03f77ea5 386 int ret;
03f77ea5 387
9bfc01a2 388 if (overlay->last_flip_req == NULL)
b303cf95 389 return 0;
03f77ea5 390
a4b3a571 391 ret = i915_wait_request(overlay->last_flip_req);
b6c028e0 392 if (ret)
03f77ea5
DV
393 return ret;
394
b303cf95
CW
395 if (overlay->flip_tail)
396 overlay->flip_tail(overlay);
03f77ea5 397
9bfc01a2 398 i915_gem_request_assign(&overlay->last_flip_req, NULL);
03f77ea5
DV
399 return 0;
400}
401
5a5a0c64
DV
402/* Wait for pending overlay flip and release old frame.
403 * Needs to be called before the overlay register are changed
8d74f656
CW
404 * via intel_overlay_(un)map_regs
405 */
02e792fb
DV
406static int intel_overlay_release_old_vid(struct intel_overlay *overlay)
407{
1ee8da6d 408 struct drm_i915_private *dev_priv = overlay->i915;
4a570db5 409 struct intel_engine_cs *engine = &dev_priv->engine[RCS];
02e792fb 410 int ret;
02e792fb 411
91c8a326 412 lockdep_assert_held(&dev_priv->drm.struct_mutex);
1362b776 413
5cd68c98
CW
414 /* Only wait if there is actually an old frame to release to
415 * guarantee forward progress.
416 */
03f77ea5
DV
417 if (!overlay->old_vid_bo)
418 return 0;
419
5cd68c98
CW
420 if (I915_READ(ISR) & I915_OVERLAY_PLANE_FLIP_PENDING_INTERRUPT) {
421 /* synchronous slowpath */
dad540ce
JH
422 struct drm_i915_gem_request *req;
423
e2f80391 424 req = i915_gem_request_alloc(engine, NULL);
26827088
DG
425 if (IS_ERR(req))
426 return PTR_ERR(req);
e1f99ce6 427
5fb9de1a 428 ret = intel_ring_begin(req, 2);
dad540ce 429 if (ret) {
aa9b7810 430 i915_add_request_no_flush(req);
dad540ce
JH
431 return ret;
432 }
433
e2f80391
TU
434 intel_ring_emit(engine,
435 MI_WAIT_FOR_EVENT | MI_WAIT_FOR_OVERLAY_FLIP);
436 intel_ring_emit(engine, MI_NOOP);
437 intel_ring_advance(engine);
5cd68c98 438
dad540ce 439 ret = intel_overlay_do_wait_request(overlay, req,
b303cf95 440 intel_overlay_release_old_vid_tail);
5cd68c98
CW
441 if (ret)
442 return ret;
443 }
02e792fb 444
5cd68c98 445 intel_overlay_release_old_vid_tail(overlay);
a071fa00
DV
446
447
448 i915_gem_track_fb(overlay->old_vid_bo, NULL,
449 INTEL_FRONTBUFFER_OVERLAY(overlay->crtc->pipe));
02e792fb
DV
450 return 0;
451}
452
1362b776
VS
453void intel_overlay_reset(struct drm_i915_private *dev_priv)
454{
455 struct intel_overlay *overlay = dev_priv->overlay;
456
457 if (!overlay)
458 return;
459
460 intel_overlay_release_old_vid(overlay);
461
462 overlay->last_flip_req = NULL;
463 overlay->old_xscale = 0;
464 overlay->old_yscale = 0;
465 overlay->crtc = NULL;
466 overlay->active = false;
467}
468
02e792fb
DV
469struct put_image_params {
470 int format;
471 short dst_x;
472 short dst_y;
473 short dst_w;
474 short dst_h;
475 short src_w;
476 short src_scan_h;
477 short src_scan_w;
478 short src_h;
479 short stride_Y;
480 short stride_UV;
481 int offset_Y;
482 int offset_U;
483 int offset_V;
484};
485
486static int packed_depth_bytes(u32 format)
487{
488 switch (format & I915_OVERLAY_DEPTH_MASK) {
722506f0
CW
489 case I915_OVERLAY_YUV422:
490 return 4;
491 case I915_OVERLAY_YUV411:
492 /* return 6; not implemented */
493 default:
494 return -EINVAL;
02e792fb
DV
495 }
496}
497
498static int packed_width_bytes(u32 format, short width)
499{
500 switch (format & I915_OVERLAY_DEPTH_MASK) {
722506f0
CW
501 case I915_OVERLAY_YUV422:
502 return width << 1;
503 default:
504 return -EINVAL;
02e792fb
DV
505 }
506}
507
508static int uv_hsubsampling(u32 format)
509{
510 switch (format & I915_OVERLAY_DEPTH_MASK) {
722506f0
CW
511 case I915_OVERLAY_YUV422:
512 case I915_OVERLAY_YUV420:
513 return 2;
514 case I915_OVERLAY_YUV411:
515 case I915_OVERLAY_YUV410:
516 return 4;
517 default:
518 return -EINVAL;
02e792fb
DV
519 }
520}
521
522static int uv_vsubsampling(u32 format)
523{
524 switch (format & I915_OVERLAY_DEPTH_MASK) {
722506f0
CW
525 case I915_OVERLAY_YUV420:
526 case I915_OVERLAY_YUV410:
527 return 2;
528 case I915_OVERLAY_YUV422:
529 case I915_OVERLAY_YUV411:
530 return 1;
531 default:
532 return -EINVAL;
02e792fb
DV
533 }
534}
535
1ee8da6d 536static u32 calc_swidthsw(struct drm_i915_private *dev_priv, u32 offset, u32 width)
02e792fb
DV
537{
538 u32 mask, shift, ret;
1ee8da6d 539 if (IS_GEN2(dev_priv)) {
02e792fb
DV
540 mask = 0x1f;
541 shift = 5;
a6c45cf0
CW
542 } else {
543 mask = 0x3f;
544 shift = 6;
02e792fb
DV
545 }
546 ret = ((offset + width + mask) >> shift) - (offset >> shift);
1ee8da6d 547 if (!IS_GEN2(dev_priv))
02e792fb 548 ret <<= 1;
0206e353 549 ret -= 1;
02e792fb
DV
550 return ret << 2;
551}
552
553static const u16 y_static_hcoeffs[N_HORIZ_Y_TAPS * N_PHASES] = {
554 0x3000, 0xb4a0, 0x1930, 0x1920, 0xb4a0,
555 0x3000, 0xb500, 0x19d0, 0x1880, 0xb440,
556 0x3000, 0xb540, 0x1a88, 0x2f80, 0xb3e0,
557 0x3000, 0xb580, 0x1b30, 0x2e20, 0xb380,
558 0x3000, 0xb5c0, 0x1bd8, 0x2cc0, 0xb320,
559 0x3020, 0xb5e0, 0x1c60, 0x2b80, 0xb2c0,
560 0x3020, 0xb5e0, 0x1cf8, 0x2a20, 0xb260,
561 0x3020, 0xb5e0, 0x1d80, 0x28e0, 0xb200,
562 0x3020, 0xb5c0, 0x1e08, 0x3f40, 0xb1c0,
563 0x3020, 0xb580, 0x1e78, 0x3ce0, 0xb160,
564 0x3040, 0xb520, 0x1ed8, 0x3aa0, 0xb120,
565 0x3040, 0xb4a0, 0x1f30, 0x3880, 0xb0e0,
566 0x3040, 0xb400, 0x1f78, 0x3680, 0xb0a0,
567 0x3020, 0xb340, 0x1fb8, 0x34a0, 0xb060,
568 0x3020, 0xb240, 0x1fe0, 0x32e0, 0xb040,
569 0x3020, 0xb140, 0x1ff8, 0x3160, 0xb020,
722506f0
CW
570 0xb000, 0x3000, 0x0800, 0x3000, 0xb000
571};
572
02e792fb
DV
573static const u16 uv_static_hcoeffs[N_HORIZ_UV_TAPS * N_PHASES] = {
574 0x3000, 0x1800, 0x1800, 0xb000, 0x18d0, 0x2e60,
575 0xb000, 0x1990, 0x2ce0, 0xb020, 0x1a68, 0x2b40,
576 0xb040, 0x1b20, 0x29e0, 0xb060, 0x1bd8, 0x2880,
577 0xb080, 0x1c88, 0x3e60, 0xb0a0, 0x1d28, 0x3c00,
578 0xb0c0, 0x1db8, 0x39e0, 0xb0e0, 0x1e40, 0x37e0,
579 0xb100, 0x1eb8, 0x3620, 0xb100, 0x1f18, 0x34a0,
580 0xb100, 0x1f68, 0x3360, 0xb0e0, 0x1fa8, 0x3240,
581 0xb0c0, 0x1fe0, 0x3140, 0xb060, 0x1ff0, 0x30a0,
722506f0
CW
582 0x3000, 0x0800, 0x3000
583};
02e792fb 584
75020bc1 585static void update_polyphase_filter(struct overlay_registers __iomem *regs)
02e792fb 586{
75020bc1
BW
587 memcpy_toio(regs->Y_HCOEFS, y_static_hcoeffs, sizeof(y_static_hcoeffs));
588 memcpy_toio(regs->UV_HCOEFS, uv_static_hcoeffs,
589 sizeof(uv_static_hcoeffs));
02e792fb
DV
590}
591
592static bool update_scaling_factors(struct intel_overlay *overlay,
75020bc1 593 struct overlay_registers __iomem *regs,
02e792fb
DV
594 struct put_image_params *params)
595{
596 /* fixed point with a 12 bit shift */
597 u32 xscale, yscale, xscale_UV, yscale_UV;
598#define FP_SHIFT 12
599#define FRACT_MASK 0xfff
600 bool scale_changed = false;
601 int uv_hscale = uv_hsubsampling(params->format);
602 int uv_vscale = uv_vsubsampling(params->format);
603
604 if (params->dst_w > 1)
605 xscale = ((params->src_scan_w - 1) << FP_SHIFT)
606 /(params->dst_w);
607 else
608 xscale = 1 << FP_SHIFT;
609
610 if (params->dst_h > 1)
611 yscale = ((params->src_scan_h - 1) << FP_SHIFT)
612 /(params->dst_h);
613 else
614 yscale = 1 << FP_SHIFT;
615
616 /*if (params->format & I915_OVERLAY_YUV_PLANAR) {*/
722506f0
CW
617 xscale_UV = xscale/uv_hscale;
618 yscale_UV = yscale/uv_vscale;
619 /* make the Y scale to UV scale ratio an exact multiply */
620 xscale = xscale_UV * uv_hscale;
621 yscale = yscale_UV * uv_vscale;
02e792fb 622 /*} else {
722506f0
CW
623 xscale_UV = 0;
624 yscale_UV = 0;
625 }*/
02e792fb
DV
626
627 if (xscale != overlay->old_xscale || yscale != overlay->old_yscale)
628 scale_changed = true;
629 overlay->old_xscale = xscale;
630 overlay->old_yscale = yscale;
631
75020bc1
BW
632 iowrite32(((yscale & FRACT_MASK) << 20) |
633 ((xscale >> FP_SHIFT) << 16) |
634 ((xscale & FRACT_MASK) << 3),
635 &regs->YRGBSCALE);
722506f0 636
75020bc1
BW
637 iowrite32(((yscale_UV & FRACT_MASK) << 20) |
638 ((xscale_UV >> FP_SHIFT) << 16) |
639 ((xscale_UV & FRACT_MASK) << 3),
640 &regs->UVSCALE);
722506f0 641
75020bc1
BW
642 iowrite32((((yscale >> FP_SHIFT) << 16) |
643 ((yscale_UV >> FP_SHIFT) << 0)),
644 &regs->UVSCALEV);
02e792fb
DV
645
646 if (scale_changed)
647 update_polyphase_filter(regs);
648
649 return scale_changed;
650}
651
652static void update_colorkey(struct intel_overlay *overlay,
75020bc1 653 struct overlay_registers __iomem *regs)
02e792fb
DV
654{
655 u32 key = overlay->color_key;
ea9da4e4
CW
656 u32 flags;
657
658 flags = 0;
659 if (overlay->color_key_enabled)
660 flags |= DST_KEY_ENABLE;
6ba3ddd9 661
f4510a27 662 switch (overlay->crtc->base.primary->fb->bits_per_pixel) {
722506f0 663 case 8:
ea9da4e4
CW
664 key = 0;
665 flags |= CLK_RGB8I_MASK;
6ba3ddd9
CW
666 break;
667
722506f0 668 case 16:
f4510a27 669 if (overlay->crtc->base.primary->fb->depth == 15) {
ea9da4e4
CW
670 key = RGB15_TO_COLORKEY(key);
671 flags |= CLK_RGB15_MASK;
722506f0 672 } else {
ea9da4e4
CW
673 key = RGB16_TO_COLORKEY(key);
674 flags |= CLK_RGB16_MASK;
722506f0 675 }
6ba3ddd9
CW
676 break;
677
722506f0
CW
678 case 24:
679 case 32:
ea9da4e4 680 flags |= CLK_RGB24_MASK;
6ba3ddd9 681 break;
02e792fb 682 }
ea9da4e4
CW
683
684 iowrite32(key, &regs->DCLRKV);
685 iowrite32(flags, &regs->DCLRKM);
02e792fb
DV
686}
687
688static u32 overlay_cmd_reg(struct put_image_params *params)
689{
690 u32 cmd = OCMD_ENABLE | OCMD_BUF_TYPE_FRAME | OCMD_BUFFER0;
691
692 if (params->format & I915_OVERLAY_YUV_PLANAR) {
693 switch (params->format & I915_OVERLAY_DEPTH_MASK) {
722506f0
CW
694 case I915_OVERLAY_YUV422:
695 cmd |= OCMD_YUV_422_PLANAR;
696 break;
697 case I915_OVERLAY_YUV420:
698 cmd |= OCMD_YUV_420_PLANAR;
699 break;
700 case I915_OVERLAY_YUV411:
701 case I915_OVERLAY_YUV410:
702 cmd |= OCMD_YUV_410_PLANAR;
703 break;
02e792fb
DV
704 }
705 } else { /* YUV packed */
706 switch (params->format & I915_OVERLAY_DEPTH_MASK) {
722506f0
CW
707 case I915_OVERLAY_YUV422:
708 cmd |= OCMD_YUV_422_PACKED;
709 break;
710 case I915_OVERLAY_YUV411:
711 cmd |= OCMD_YUV_411_PACKED;
712 break;
02e792fb
DV
713 }
714
715 switch (params->format & I915_OVERLAY_SWAP_MASK) {
722506f0
CW
716 case I915_OVERLAY_NO_SWAP:
717 break;
718 case I915_OVERLAY_UV_SWAP:
719 cmd |= OCMD_UV_SWAP;
720 break;
721 case I915_OVERLAY_Y_SWAP:
722 cmd |= OCMD_Y_SWAP;
723 break;
724 case I915_OVERLAY_Y_AND_UV_SWAP:
725 cmd |= OCMD_Y_AND_UV_SWAP;
726 break;
02e792fb
DV
727 }
728 }
729
730 return cmd;
731}
732
5fe82c5e 733static int intel_overlay_do_put_image(struct intel_overlay *overlay,
05394f39 734 struct drm_i915_gem_object *new_bo,
5fe82c5e 735 struct put_image_params *params)
02e792fb
DV
736{
737 int ret, tmp_width;
75020bc1 738 struct overlay_registers __iomem *regs;
02e792fb 739 bool scale_changed = false;
1ee8da6d 740 struct drm_i915_private *dev_priv = overlay->i915;
75020bc1 741 u32 swidth, swidthsw, sheight, ostride;
a071fa00 742 enum pipe pipe = overlay->crtc->pipe;
02e792fb 743
91c8a326
CW
744 lockdep_assert_held(&dev_priv->drm.struct_mutex);
745 WARN_ON(!drm_modeset_is_locked(&dev_priv->drm.mode_config.connection_mutex));
02e792fb 746
02e792fb
DV
747 ret = intel_overlay_release_old_vid(overlay);
748 if (ret != 0)
749 return ret;
750
7580d774 751 ret = i915_gem_object_pin_to_display_plane(new_bo, 0,
e6617330 752 &i915_ggtt_view_normal);
02e792fb
DV
753 if (ret != 0)
754 return ret;
755
d9e86c0e
CW
756 ret = i915_gem_object_put_fence(new_bo);
757 if (ret)
758 goto out_unpin;
759
02e792fb 760 if (!overlay->active) {
75020bc1 761 u32 oconfig;
8d74f656 762 regs = intel_overlay_map_regs(overlay);
02e792fb
DV
763 if (!regs) {
764 ret = -ENOMEM;
765 goto out_unpin;
766 }
75020bc1 767 oconfig = OCONF_CC_OUT_8BIT;
1ee8da6d 768 if (IS_GEN4(dev_priv))
75020bc1 769 oconfig |= OCONF_CSC_MODE_BT709;
a071fa00 770 oconfig |= pipe == 0 ?
02e792fb 771 OCONF_PIPE_A : OCONF_PIPE_B;
75020bc1 772 iowrite32(oconfig, &regs->OCONFIG);
9bb2ff73 773 intel_overlay_unmap_regs(overlay, regs);
02e792fb
DV
774
775 ret = intel_overlay_on(overlay);
776 if (ret != 0)
777 goto out_unpin;
778 }
779
8d74f656 780 regs = intel_overlay_map_regs(overlay);
02e792fb
DV
781 if (!regs) {
782 ret = -ENOMEM;
783 goto out_unpin;
784 }
785
75020bc1
BW
786 iowrite32((params->dst_y << 16) | params->dst_x, &regs->DWINPOS);
787 iowrite32((params->dst_h << 16) | params->dst_w, &regs->DWINSZ);
02e792fb
DV
788
789 if (params->format & I915_OVERLAY_YUV_PACKED)
790 tmp_width = packed_width_bytes(params->format, params->src_w);
791 else
792 tmp_width = params->src_w;
793
75020bc1 794 swidth = params->src_w;
1ee8da6d 795 swidthsw = calc_swidthsw(dev_priv, params->offset_Y, tmp_width);
75020bc1 796 sheight = params->src_h;
f343c5f6 797 iowrite32(i915_gem_obj_ggtt_offset(new_bo) + params->offset_Y, &regs->OBUF_0Y);
75020bc1 798 ostride = params->stride_Y;
02e792fb
DV
799
800 if (params->format & I915_OVERLAY_YUV_PLANAR) {
801 int uv_hscale = uv_hsubsampling(params->format);
802 int uv_vscale = uv_vsubsampling(params->format);
803 u32 tmp_U, tmp_V;
75020bc1 804 swidth |= (params->src_w/uv_hscale) << 16;
1ee8da6d 805 tmp_U = calc_swidthsw(dev_priv, params->offset_U,
722506f0 806 params->src_w/uv_hscale);
1ee8da6d 807 tmp_V = calc_swidthsw(dev_priv, params->offset_V,
722506f0 808 params->src_w/uv_hscale);
75020bc1
BW
809 swidthsw |= max_t(u32, tmp_U, tmp_V) << 16;
810 sheight |= (params->src_h/uv_vscale) << 16;
f343c5f6
BW
811 iowrite32(i915_gem_obj_ggtt_offset(new_bo) + params->offset_U, &regs->OBUF_0U);
812 iowrite32(i915_gem_obj_ggtt_offset(new_bo) + params->offset_V, &regs->OBUF_0V);
75020bc1 813 ostride |= params->stride_UV << 16;
02e792fb
DV
814 }
815
75020bc1
BW
816 iowrite32(swidth, &regs->SWIDTH);
817 iowrite32(swidthsw, &regs->SWIDTHSW);
818 iowrite32(sheight, &regs->SHEIGHT);
819 iowrite32(ostride, &regs->OSTRIDE);
820
02e792fb
DV
821 scale_changed = update_scaling_factors(overlay, regs, params);
822
823 update_colorkey(overlay, regs);
824
75020bc1 825 iowrite32(overlay_cmd_reg(params), &regs->OCMD);
02e792fb 826
9bb2ff73 827 intel_overlay_unmap_regs(overlay, regs);
02e792fb 828
8dc5d147
CW
829 ret = intel_overlay_continue(overlay, scale_changed);
830 if (ret)
831 goto out_unpin;
02e792fb 832
a071fa00
DV
833 i915_gem_track_fb(overlay->vid_bo, new_bo,
834 INTEL_FRONTBUFFER_OVERLAY(pipe));
835
02e792fb 836 overlay->old_vid_bo = overlay->vid_bo;
05394f39 837 overlay->vid_bo = new_bo;
02e792fb 838
91c8a326
CW
839 intel_frontbuffer_flip(&dev_priv->drm,
840 INTEL_FRONTBUFFER_OVERLAY(pipe));
f99d7069 841
02e792fb
DV
842 return 0;
843
844out_unpin:
d7f46fc4 845 i915_gem_object_ggtt_unpin(new_bo);
02e792fb
DV
846 return ret;
847}
848
ce453d81 849int intel_overlay_switch_off(struct intel_overlay *overlay)
02e792fb 850{
1ee8da6d 851 struct drm_i915_private *dev_priv = overlay->i915;
75020bc1 852 struct overlay_registers __iomem *regs;
5dcdbcb0 853 int ret;
02e792fb 854
91c8a326
CW
855 lockdep_assert_held(&dev_priv->drm.struct_mutex);
856 WARN_ON(!drm_modeset_is_locked(&dev_priv->drm.mode_config.connection_mutex));
02e792fb 857
ce453d81 858 ret = intel_overlay_recover_from_interrupt(overlay);
b303cf95
CW
859 if (ret != 0)
860 return ret;
9bedb974 861
02e792fb
DV
862 if (!overlay->active)
863 return 0;
864
02e792fb
DV
865 ret = intel_overlay_release_old_vid(overlay);
866 if (ret != 0)
867 return ret;
868
8d74f656 869 regs = intel_overlay_map_regs(overlay);
75020bc1 870 iowrite32(0, &regs->OCMD);
9bb2ff73 871 intel_overlay_unmap_regs(overlay, regs);
02e792fb 872
ce453d81 873 ret = intel_overlay_off(overlay);
03f77ea5
DV
874 if (ret != 0)
875 return ret;
876
12ca45fe 877 intel_overlay_off_tail(overlay);
02e792fb
DV
878 return 0;
879}
880
881static int check_overlay_possible_on_crtc(struct intel_overlay *overlay,
882 struct intel_crtc *crtc)
883{
f7abfe8b 884 if (!crtc->active)
02e792fb
DV
885 return -EINVAL;
886
02e792fb 887 /* can't use the overlay with double wide pipe */
6e3c9717 888 if (crtc->config->double_wide)
02e792fb
DV
889 return -EINVAL;
890
891 return 0;
892}
893
894static void update_pfit_vscale_ratio(struct intel_overlay *overlay)
895{
1ee8da6d 896 struct drm_i915_private *dev_priv = overlay->i915;
02e792fb 897 u32 pfit_control = I915_READ(PFIT_CONTROL);
446d2183 898 u32 ratio;
02e792fb
DV
899
900 /* XXX: This is not the same logic as in the xorg driver, but more in
446d2183
CW
901 * line with the intel documentation for the i965
902 */
1ee8da6d 903 if (INTEL_GEN(dev_priv) >= 4) {
0206e353 904 /* on i965 use the PGM reg to read out the autoscaler values */
a6c45cf0
CW
905 ratio = I915_READ(PFIT_PGM_RATIOS) >> PFIT_VERT_SCALE_SHIFT_965;
906 } else {
446d2183
CW
907 if (pfit_control & VERT_AUTO_SCALE)
908 ratio = I915_READ(PFIT_AUTO_RATIOS);
02e792fb 909 else
446d2183
CW
910 ratio = I915_READ(PFIT_PGM_RATIOS);
911 ratio >>= PFIT_VERT_SCALE_SHIFT;
02e792fb
DV
912 }
913
914 overlay->pfit_vscale_ratio = ratio;
915}
916
917static int check_overlay_dst(struct intel_overlay *overlay,
918 struct drm_intel_overlay_put_image *rec)
919{
920 struct drm_display_mode *mode = &overlay->crtc->base.mode;
921
75c13993
DV
922 if (rec->dst_x < mode->hdisplay &&
923 rec->dst_x + rec->dst_width <= mode->hdisplay &&
924 rec->dst_y < mode->vdisplay &&
925 rec->dst_y + rec->dst_height <= mode->vdisplay)
02e792fb
DV
926 return 0;
927 else
928 return -EINVAL;
929}
930
931static int check_overlay_scaling(struct put_image_params *rec)
932{
933 u32 tmp;
934
935 /* downscaling limit is 8.0 */
936 tmp = ((rec->src_scan_h << 16) / rec->dst_h) >> 16;
937 if (tmp > 7)
938 return -EINVAL;
939 tmp = ((rec->src_scan_w << 16) / rec->dst_w) >> 16;
940 if (tmp > 7)
941 return -EINVAL;
942
943 return 0;
944}
945
1ee8da6d 946static int check_overlay_src(struct drm_i915_private *dev_priv,
02e792fb 947 struct drm_intel_overlay_put_image *rec,
05394f39 948 struct drm_i915_gem_object *new_bo)
02e792fb 949{
02e792fb
DV
950 int uv_hscale = uv_hsubsampling(rec->flags);
951 int uv_vscale = uv_vsubsampling(rec->flags);
8f28f54a
DC
952 u32 stride_mask;
953 int depth;
954 u32 tmp;
02e792fb
DV
955
956 /* check src dimensions */
1ee8da6d 957 if (IS_845G(dev_priv) || IS_I830(dev_priv)) {
722506f0 958 if (rec->src_height > IMAGE_MAX_HEIGHT_LEGACY ||
9f7c3f44 959 rec->src_width > IMAGE_MAX_WIDTH_LEGACY)
02e792fb
DV
960 return -EINVAL;
961 } else {
722506f0 962 if (rec->src_height > IMAGE_MAX_HEIGHT ||
9f7c3f44 963 rec->src_width > IMAGE_MAX_WIDTH)
02e792fb
DV
964 return -EINVAL;
965 }
9f7c3f44 966
02e792fb 967 /* better safe than sorry, use 4 as the maximal subsampling ratio */
722506f0 968 if (rec->src_height < N_VERT_Y_TAPS*4 ||
9f7c3f44 969 rec->src_width < N_HORIZ_Y_TAPS*4)
02e792fb
DV
970 return -EINVAL;
971
a1efd14a 972 /* check alignment constraints */
02e792fb 973 switch (rec->flags & I915_OVERLAY_TYPE_MASK) {
722506f0
CW
974 case I915_OVERLAY_RGB:
975 /* not implemented */
976 return -EINVAL;
9f7c3f44 977
722506f0 978 case I915_OVERLAY_YUV_PACKED:
722506f0 979 if (uv_vscale != 1)
02e792fb 980 return -EINVAL;
9f7c3f44
CW
981
982 depth = packed_depth_bytes(rec->flags);
722506f0
CW
983 if (depth < 0)
984 return depth;
9f7c3f44 985
722506f0
CW
986 /* ignore UV planes */
987 rec->stride_UV = 0;
988 rec->offset_U = 0;
989 rec->offset_V = 0;
990 /* check pixel alignment */
991 if (rec->offset_Y % depth)
992 return -EINVAL;
993 break;
9f7c3f44 994
722506f0
CW
995 case I915_OVERLAY_YUV_PLANAR:
996 if (uv_vscale < 0 || uv_hscale < 0)
02e792fb 997 return -EINVAL;
722506f0
CW
998 /* no offset restrictions for planar formats */
999 break;
9f7c3f44 1000
722506f0
CW
1001 default:
1002 return -EINVAL;
02e792fb
DV
1003 }
1004
1005 if (rec->src_width % uv_hscale)
1006 return -EINVAL;
1007
1008 /* stride checking */
1ee8da6d 1009 if (IS_I830(dev_priv) || IS_845G(dev_priv))
a1efd14a
CW
1010 stride_mask = 255;
1011 else
1012 stride_mask = 63;
02e792fb
DV
1013
1014 if (rec->stride_Y & stride_mask || rec->stride_UV & stride_mask)
1015 return -EINVAL;
1ee8da6d 1016 if (IS_GEN4(dev_priv) && rec->stride_Y < 512)
02e792fb
DV
1017 return -EINVAL;
1018
1019 tmp = (rec->flags & I915_OVERLAY_TYPE_MASK) == I915_OVERLAY_YUV_PLANAR ?
9f7c3f44
CW
1020 4096 : 8192;
1021 if (rec->stride_Y > tmp || rec->stride_UV > 2*1024)
02e792fb
DV
1022 return -EINVAL;
1023
1024 /* check buffer dimensions */
1025 switch (rec->flags & I915_OVERLAY_TYPE_MASK) {
722506f0
CW
1026 case I915_OVERLAY_RGB:
1027 case I915_OVERLAY_YUV_PACKED:
1028 /* always 4 Y values per depth pixels */
1029 if (packed_width_bytes(rec->flags, rec->src_width) > rec->stride_Y)
1030 return -EINVAL;
1031
1032 tmp = rec->stride_Y*rec->src_height;
05394f39 1033 if (rec->offset_Y + tmp > new_bo->base.size)
722506f0
CW
1034 return -EINVAL;
1035 break;
1036
1037 case I915_OVERLAY_YUV_PLANAR:
1038 if (rec->src_width > rec->stride_Y)
1039 return -EINVAL;
1040 if (rec->src_width/uv_hscale > rec->stride_UV)
1041 return -EINVAL;
1042
9f7c3f44 1043 tmp = rec->stride_Y * rec->src_height;
05394f39 1044 if (rec->offset_Y + tmp > new_bo->base.size)
722506f0 1045 return -EINVAL;
9f7c3f44
CW
1046
1047 tmp = rec->stride_UV * (rec->src_height / uv_vscale);
05394f39
CW
1048 if (rec->offset_U + tmp > new_bo->base.size ||
1049 rec->offset_V + tmp > new_bo->base.size)
722506f0
CW
1050 return -EINVAL;
1051 break;
02e792fb
DV
1052 }
1053
1054 return 0;
1055}
1056
e9e331a8
CW
1057/**
1058 * Return the pipe currently connected to the panel fitter,
1059 * or -1 if the panel fitter is not present or not in use
1060 */
1ee8da6d 1061static int intel_panel_fitter_pipe(struct drm_i915_private *dev_priv)
e9e331a8 1062{
e9e331a8
CW
1063 u32 pfit_control;
1064
1065 /* i830 doesn't have a panel fitter */
1ee8da6d
CW
1066 if (INTEL_GEN(dev_priv) <= 3 &&
1067 (IS_I830(dev_priv) || !IS_MOBILE(dev_priv)))
e9e331a8
CW
1068 return -1;
1069
1070 pfit_control = I915_READ(PFIT_CONTROL);
1071
1072 /* See if the panel fitter is in use */
1073 if ((pfit_control & PFIT_ENABLE) == 0)
1074 return -1;
1075
1076 /* 965 can place panel fitter on either pipe */
1ee8da6d 1077 if (IS_GEN4(dev_priv))
e9e331a8
CW
1078 return (pfit_control >> 29) & 0x3;
1079
1080 /* older chips can only use pipe 1 */
1081 return 1;
1082}
1083
1ee8da6d
CW
1084int intel_overlay_put_image_ioctl(struct drm_device *dev, void *data,
1085 struct drm_file *file_priv)
02e792fb
DV
1086{
1087 struct drm_intel_overlay_put_image *put_image_rec = data;
fac5e23e 1088 struct drm_i915_private *dev_priv = to_i915(dev);
02e792fb 1089 struct intel_overlay *overlay;
7707e653 1090 struct drm_crtc *drmmode_crtc;
02e792fb 1091 struct intel_crtc *crtc;
05394f39 1092 struct drm_i915_gem_object *new_bo;
02e792fb
DV
1093 struct put_image_params *params;
1094 int ret;
1095
02e792fb
DV
1096 overlay = dev_priv->overlay;
1097 if (!overlay) {
1098 DRM_DEBUG("userspace bug: no overlay\n");
1099 return -ENODEV;
1100 }
1101
1102 if (!(put_image_rec->flags & I915_OVERLAY_ENABLE)) {
a0e99e68 1103 drm_modeset_lock_all(dev);
02e792fb
DV
1104 mutex_lock(&dev->struct_mutex);
1105
ce453d81 1106 ret = intel_overlay_switch_off(overlay);
02e792fb
DV
1107
1108 mutex_unlock(&dev->struct_mutex);
a0e99e68 1109 drm_modeset_unlock_all(dev);
02e792fb
DV
1110
1111 return ret;
1112 }
1113
b14c5679 1114 params = kmalloc(sizeof(*params), GFP_KERNEL);
02e792fb
DV
1115 if (!params)
1116 return -ENOMEM;
1117
7707e653
RC
1118 drmmode_crtc = drm_crtc_find(dev, put_image_rec->crtc_id);
1119 if (!drmmode_crtc) {
915a428e
DC
1120 ret = -ENOENT;
1121 goto out_free;
1122 }
7707e653 1123 crtc = to_intel_crtc(drmmode_crtc);
02e792fb 1124
03ac0642
CW
1125 new_bo = i915_gem_object_lookup(file_priv, put_image_rec->bo_handle);
1126 if (!new_bo) {
915a428e
DC
1127 ret = -ENOENT;
1128 goto out_free;
1129 }
02e792fb 1130
a0e99e68 1131 drm_modeset_lock_all(dev);
02e792fb
DV
1132 mutex_lock(&dev->struct_mutex);
1133
d9e86c0e 1134 if (new_bo->tiling_mode) {
3b25b31f 1135 DRM_DEBUG_KMS("buffer used for overlay image can not be tiled\n");
d9e86c0e
CW
1136 ret = -EINVAL;
1137 goto out_unlock;
1138 }
1139
ce453d81 1140 ret = intel_overlay_recover_from_interrupt(overlay);
b303cf95
CW
1141 if (ret != 0)
1142 goto out_unlock;
03f77ea5 1143
02e792fb
DV
1144 if (overlay->crtc != crtc) {
1145 struct drm_display_mode *mode = &crtc->base.mode;
ce453d81 1146 ret = intel_overlay_switch_off(overlay);
02e792fb
DV
1147 if (ret != 0)
1148 goto out_unlock;
1149
1150 ret = check_overlay_possible_on_crtc(overlay, crtc);
1151 if (ret != 0)
1152 goto out_unlock;
1153
1154 overlay->crtc = crtc;
1155 crtc->overlay = overlay;
1156
e9e331a8
CW
1157 /* line too wide, i.e. one-line-mode */
1158 if (mode->hdisplay > 1024 &&
1ee8da6d 1159 intel_panel_fitter_pipe(dev_priv) == crtc->pipe) {
209c2a5e 1160 overlay->pfit_active = true;
02e792fb
DV
1161 update_pfit_vscale_ratio(overlay);
1162 } else
209c2a5e 1163 overlay->pfit_active = false;
02e792fb
DV
1164 }
1165
1166 ret = check_overlay_dst(overlay, put_image_rec);
1167 if (ret != 0)
1168 goto out_unlock;
1169
1170 if (overlay->pfit_active) {
1171 params->dst_y = ((((u32)put_image_rec->dst_y) << 12) /
722506f0 1172 overlay->pfit_vscale_ratio);
02e792fb
DV
1173 /* shifting right rounds downwards, so add 1 */
1174 params->dst_h = ((((u32)put_image_rec->dst_height) << 12) /
722506f0 1175 overlay->pfit_vscale_ratio) + 1;
02e792fb
DV
1176 } else {
1177 params->dst_y = put_image_rec->dst_y;
1178 params->dst_h = put_image_rec->dst_height;
1179 }
1180 params->dst_x = put_image_rec->dst_x;
1181 params->dst_w = put_image_rec->dst_width;
1182
1183 params->src_w = put_image_rec->src_width;
1184 params->src_h = put_image_rec->src_height;
1185 params->src_scan_w = put_image_rec->src_scan_width;
1186 params->src_scan_h = put_image_rec->src_scan_height;
722506f0
CW
1187 if (params->src_scan_h > params->src_h ||
1188 params->src_scan_w > params->src_w) {
02e792fb
DV
1189 ret = -EINVAL;
1190 goto out_unlock;
1191 }
1192
1ee8da6d 1193 ret = check_overlay_src(dev_priv, put_image_rec, new_bo);
02e792fb
DV
1194 if (ret != 0)
1195 goto out_unlock;
1196 params->format = put_image_rec->flags & ~I915_OVERLAY_FLAGS_MASK;
1197 params->stride_Y = put_image_rec->stride_Y;
1198 params->stride_UV = put_image_rec->stride_UV;
1199 params->offset_Y = put_image_rec->offset_Y;
1200 params->offset_U = put_image_rec->offset_U;
1201 params->offset_V = put_image_rec->offset_V;
1202
1203 /* Check scaling after src size to prevent a divide-by-zero. */
1204 ret = check_overlay_scaling(params);
1205 if (ret != 0)
1206 goto out_unlock;
1207
1208 ret = intel_overlay_do_put_image(overlay, new_bo, params);
1209 if (ret != 0)
1210 goto out_unlock;
1211
1212 mutex_unlock(&dev->struct_mutex);
a0e99e68 1213 drm_modeset_unlock_all(dev);
02e792fb
DV
1214
1215 kfree(params);
1216
1217 return 0;
1218
1219out_unlock:
1220 mutex_unlock(&dev->struct_mutex);
a0e99e68 1221 drm_modeset_unlock_all(dev);
f8c417cd 1222 i915_gem_object_put(new_bo);
915a428e 1223out_free:
02e792fb
DV
1224 kfree(params);
1225
1226 return ret;
1227}
1228
1229static void update_reg_attrs(struct intel_overlay *overlay,
75020bc1 1230 struct overlay_registers __iomem *regs)
02e792fb 1231{
75020bc1
BW
1232 iowrite32((overlay->contrast << 18) | (overlay->brightness & 0xff),
1233 &regs->OCLRC0);
1234 iowrite32(overlay->saturation, &regs->OCLRC1);
02e792fb
DV
1235}
1236
1237static bool check_gamma_bounds(u32 gamma1, u32 gamma2)
1238{
1239 int i;
1240
1241 if (gamma1 & 0xff000000 || gamma2 & 0xff000000)
1242 return false;
1243
1244 for (i = 0; i < 3; i++) {
722506f0 1245 if (((gamma1 >> i*8) & 0xff) >= ((gamma2 >> i*8) & 0xff))
02e792fb
DV
1246 return false;
1247 }
1248
1249 return true;
1250}
1251
1252static bool check_gamma5_errata(u32 gamma5)
1253{
1254 int i;
1255
1256 for (i = 0; i < 3; i++) {
1257 if (((gamma5 >> i*8) & 0xff) == 0x80)
1258 return false;
1259 }
1260
1261 return true;
1262}
1263
1264static int check_gamma(struct drm_intel_overlay_attrs *attrs)
1265{
722506f0
CW
1266 if (!check_gamma_bounds(0, attrs->gamma0) ||
1267 !check_gamma_bounds(attrs->gamma0, attrs->gamma1) ||
1268 !check_gamma_bounds(attrs->gamma1, attrs->gamma2) ||
1269 !check_gamma_bounds(attrs->gamma2, attrs->gamma3) ||
1270 !check_gamma_bounds(attrs->gamma3, attrs->gamma4) ||
1271 !check_gamma_bounds(attrs->gamma4, attrs->gamma5) ||
1272 !check_gamma_bounds(attrs->gamma5, 0x00ffffff))
02e792fb 1273 return -EINVAL;
722506f0 1274
02e792fb
DV
1275 if (!check_gamma5_errata(attrs->gamma5))
1276 return -EINVAL;
722506f0 1277
02e792fb
DV
1278 return 0;
1279}
1280
1ee8da6d
CW
1281int intel_overlay_attrs_ioctl(struct drm_device *dev, void *data,
1282 struct drm_file *file_priv)
02e792fb
DV
1283{
1284 struct drm_intel_overlay_attrs *attrs = data;
fac5e23e 1285 struct drm_i915_private *dev_priv = to_i915(dev);
02e792fb 1286 struct intel_overlay *overlay;
75020bc1 1287 struct overlay_registers __iomem *regs;
02e792fb
DV
1288 int ret;
1289
02e792fb
DV
1290 overlay = dev_priv->overlay;
1291 if (!overlay) {
1292 DRM_DEBUG("userspace bug: no overlay\n");
1293 return -ENODEV;
1294 }
1295
a0e99e68 1296 drm_modeset_lock_all(dev);
02e792fb
DV
1297 mutex_lock(&dev->struct_mutex);
1298
60fc332c 1299 ret = -EINVAL;
02e792fb 1300 if (!(attrs->flags & I915_OVERLAY_UPDATE_ATTRS)) {
60fc332c 1301 attrs->color_key = overlay->color_key;
02e792fb 1302 attrs->brightness = overlay->brightness;
60fc332c 1303 attrs->contrast = overlay->contrast;
02e792fb
DV
1304 attrs->saturation = overlay->saturation;
1305
1ee8da6d 1306 if (!IS_GEN2(dev_priv)) {
02e792fb
DV
1307 attrs->gamma0 = I915_READ(OGAMC0);
1308 attrs->gamma1 = I915_READ(OGAMC1);
1309 attrs->gamma2 = I915_READ(OGAMC2);
1310 attrs->gamma3 = I915_READ(OGAMC3);
1311 attrs->gamma4 = I915_READ(OGAMC4);
1312 attrs->gamma5 = I915_READ(OGAMC5);
1313 }
02e792fb 1314 } else {
60fc332c 1315 if (attrs->brightness < -128 || attrs->brightness > 127)
02e792fb 1316 goto out_unlock;
60fc332c 1317 if (attrs->contrast > 255)
02e792fb 1318 goto out_unlock;
60fc332c 1319 if (attrs->saturation > 1023)
02e792fb 1320 goto out_unlock;
02e792fb 1321
60fc332c
CW
1322 overlay->color_key = attrs->color_key;
1323 overlay->brightness = attrs->brightness;
1324 overlay->contrast = attrs->contrast;
1325 overlay->saturation = attrs->saturation;
02e792fb 1326
8d74f656 1327 regs = intel_overlay_map_regs(overlay);
02e792fb
DV
1328 if (!regs) {
1329 ret = -ENOMEM;
1330 goto out_unlock;
1331 }
1332
1333 update_reg_attrs(overlay, regs);
1334
9bb2ff73 1335 intel_overlay_unmap_regs(overlay, regs);
02e792fb
DV
1336
1337 if (attrs->flags & I915_OVERLAY_UPDATE_GAMMA) {
1ee8da6d 1338 if (IS_GEN2(dev_priv))
02e792fb 1339 goto out_unlock;
02e792fb
DV
1340
1341 if (overlay->active) {
1342 ret = -EBUSY;
1343 goto out_unlock;
1344 }
1345
1346 ret = check_gamma(attrs);
60fc332c 1347 if (ret)
02e792fb
DV
1348 goto out_unlock;
1349
1350 I915_WRITE(OGAMC0, attrs->gamma0);
1351 I915_WRITE(OGAMC1, attrs->gamma1);
1352 I915_WRITE(OGAMC2, attrs->gamma2);
1353 I915_WRITE(OGAMC3, attrs->gamma3);
1354 I915_WRITE(OGAMC4, attrs->gamma4);
1355 I915_WRITE(OGAMC5, attrs->gamma5);
1356 }
02e792fb 1357 }
ea9da4e4 1358 overlay->color_key_enabled = (attrs->flags & I915_OVERLAY_DISABLE_DEST_COLORKEY) == 0;
02e792fb 1359
60fc332c 1360 ret = 0;
02e792fb
DV
1361out_unlock:
1362 mutex_unlock(&dev->struct_mutex);
a0e99e68 1363 drm_modeset_unlock_all(dev);
02e792fb
DV
1364
1365 return ret;
1366}
1367
1ee8da6d 1368void intel_setup_overlay(struct drm_i915_private *dev_priv)
02e792fb 1369{
02e792fb 1370 struct intel_overlay *overlay;
05394f39 1371 struct drm_i915_gem_object *reg_bo;
75020bc1 1372 struct overlay_registers __iomem *regs;
02e792fb
DV
1373 int ret;
1374
1ee8da6d 1375 if (!HAS_OVERLAY(dev_priv))
02e792fb
DV
1376 return;
1377
b14c5679 1378 overlay = kzalloc(sizeof(*overlay), GFP_KERNEL);
02e792fb
DV
1379 if (!overlay)
1380 return;
79d24273 1381
91c8a326 1382 mutex_lock(&dev_priv->drm.struct_mutex);
79d24273
CW
1383 if (WARN_ON(dev_priv->overlay))
1384 goto out_free;
1385
1ee8da6d 1386 overlay->i915 = dev_priv;
02e792fb 1387
f63a484c 1388 reg_bo = NULL;
1ee8da6d 1389 if (!OVERLAY_NEEDS_PHYSICAL(dev_priv))
91c8a326
CW
1390 reg_bo = i915_gem_object_create_stolen(&dev_priv->drm,
1391 PAGE_SIZE);
80405138 1392 if (reg_bo == NULL)
91c8a326 1393 reg_bo = i915_gem_object_create(&dev_priv->drm, PAGE_SIZE);
fe3db79b 1394 if (IS_ERR(reg_bo))
02e792fb 1395 goto out_free;
05394f39 1396 overlay->reg_bo = reg_bo;
02e792fb 1397
1ee8da6d 1398 if (OVERLAY_NEEDS_PHYSICAL(dev_priv)) {
00731155 1399 ret = i915_gem_object_attach_phys(reg_bo, PAGE_SIZE);
0206e353
AJ
1400 if (ret) {
1401 DRM_ERROR("failed to attach phys overlay regs\n");
1402 goto out_free_bo;
1403 }
00731155 1404 overlay->flip_addr = reg_bo->phys_handle->busaddr;
31578148 1405 } else {
1ec9e26d 1406 ret = i915_gem_obj_ggtt_pin(reg_bo, PAGE_SIZE, PIN_MAPPABLE);
02e792fb 1407 if (ret) {
0206e353
AJ
1408 DRM_ERROR("failed to pin overlay register bo\n");
1409 goto out_free_bo;
1410 }
f343c5f6 1411 overlay->flip_addr = i915_gem_obj_ggtt_offset(reg_bo);
0ddc1289
CW
1412
1413 ret = i915_gem_object_set_to_gtt_domain(reg_bo, true);
1414 if (ret) {
0206e353
AJ
1415 DRM_ERROR("failed to move overlay register bo into the GTT\n");
1416 goto out_unpin_bo;
1417 }
02e792fb
DV
1418 }
1419
1420 /* init all values */
1421 overlay->color_key = 0x0101fe;
ea9da4e4 1422 overlay->color_key_enabled = true;
02e792fb
DV
1423 overlay->brightness = -19;
1424 overlay->contrast = 75;
1425 overlay->saturation = 146;
1426
8d74f656 1427 regs = intel_overlay_map_regs(overlay);
02e792fb 1428 if (!regs)
79d24273 1429 goto out_unpin_bo;
02e792fb 1430
75020bc1 1431 memset_io(regs, 0, sizeof(struct overlay_registers));
02e792fb 1432 update_polyphase_filter(regs);
02e792fb
DV
1433 update_reg_attrs(overlay, regs);
1434
9bb2ff73 1435 intel_overlay_unmap_regs(overlay, regs);
02e792fb
DV
1436
1437 dev_priv->overlay = overlay;
91c8a326 1438 mutex_unlock(&dev_priv->drm.struct_mutex);
02e792fb
DV
1439 DRM_INFO("initialized overlay support\n");
1440 return;
1441
0ddc1289 1442out_unpin_bo:
1ee8da6d 1443 if (!OVERLAY_NEEDS_PHYSICAL(dev_priv))
d7f46fc4 1444 i915_gem_object_ggtt_unpin(reg_bo);
02e792fb 1445out_free_bo:
f8c417cd 1446 i915_gem_object_put(reg_bo);
02e792fb 1447out_free:
91c8a326 1448 mutex_unlock(&dev_priv->drm.struct_mutex);
02e792fb
DV
1449 kfree(overlay);
1450 return;
1451}
1452
1ee8da6d 1453void intel_cleanup_overlay(struct drm_i915_private *dev_priv)
02e792fb 1454{
62cf4e6f
CW
1455 if (!dev_priv->overlay)
1456 return;
02e792fb 1457
62cf4e6f
CW
1458 /* The bo's should be free'd by the generic code already.
1459 * Furthermore modesetting teardown happens beforehand so the
1460 * hardware should be off already */
77589f56 1461 WARN_ON(dev_priv->overlay->active);
62cf4e6f
CW
1462
1463 drm_gem_object_unreference_unlocked(&dev_priv->overlay->reg_bo->base);
1464 kfree(dev_priv->overlay);
02e792fb 1465}
6ef3d427
CW
1466
1467struct intel_overlay_error_state {
1468 struct overlay_registers regs;
1469 unsigned long base;
1470 u32 dovsta;
1471 u32 isr;
1472};
1473
75020bc1 1474static struct overlay_registers __iomem *
c48c43e4 1475intel_overlay_map_regs_atomic(struct intel_overlay *overlay)
3bd3c932 1476{
1ee8da6d 1477 struct drm_i915_private *dev_priv = overlay->i915;
75020bc1 1478 struct overlay_registers __iomem *regs;
3bd3c932 1479
1ee8da6d 1480 if (OVERLAY_NEEDS_PHYSICAL(dev_priv))
75020bc1
BW
1481 /* Cast to make sparse happy, but it's wc memory anyway, so
1482 * equivalent to the wc io mapping on X86. */
1483 regs = (struct overlay_registers __iomem *)
00731155 1484 overlay->reg_bo->phys_handle->vaddr;
3bd3c932 1485 else
1ee8da6d 1486 regs = io_mapping_map_atomic_wc(dev_priv->ggtt.mappable,
da6ca034 1487 overlay->flip_addr);
3bd3c932
CW
1488
1489 return regs;
1490}
1491
1492static void intel_overlay_unmap_regs_atomic(struct intel_overlay *overlay,
75020bc1 1493 struct overlay_registers __iomem *regs)
3bd3c932 1494{
1ee8da6d 1495 if (!OVERLAY_NEEDS_PHYSICAL(overlay->i915))
c48c43e4 1496 io_mapping_unmap_atomic(regs);
3bd3c932
CW
1497}
1498
6ef3d427 1499struct intel_overlay_error_state *
c033666a 1500intel_overlay_capture_error_state(struct drm_i915_private *dev_priv)
6ef3d427 1501{
6ef3d427
CW
1502 struct intel_overlay *overlay = dev_priv->overlay;
1503 struct intel_overlay_error_state *error;
1504 struct overlay_registers __iomem *regs;
1505
1506 if (!overlay || !overlay->active)
1507 return NULL;
1508
1509 error = kmalloc(sizeof(*error), GFP_ATOMIC);
1510 if (error == NULL)
1511 return NULL;
1512
1513 error->dovsta = I915_READ(DOVSTA);
1514 error->isr = I915_READ(ISR);
da6ca034 1515 error->base = overlay->flip_addr;
6ef3d427
CW
1516
1517 regs = intel_overlay_map_regs_atomic(overlay);
1518 if (!regs)
1519 goto err;
1520
1521 memcpy_fromio(&error->regs, regs, sizeof(struct overlay_registers));
c48c43e4 1522 intel_overlay_unmap_regs_atomic(overlay, regs);
6ef3d427
CW
1523
1524 return error;
1525
1526err:
1527 kfree(error);
1528 return NULL;
1529}
1530
1531void
edc3d884
MK
1532intel_overlay_print_error_state(struct drm_i915_error_state_buf *m,
1533 struct intel_overlay_error_state *error)
6ef3d427 1534{
edc3d884
MK
1535 i915_error_printf(m, "Overlay, status: 0x%08x, interrupt: 0x%08x\n",
1536 error->dovsta, error->isr);
1537 i915_error_printf(m, " Register file at 0x%08lx:\n",
1538 error->base);
6ef3d427 1539
edc3d884 1540#define P(x) i915_error_printf(m, " " #x ": 0x%08x\n", error->regs.x)
6ef3d427
CW
1541 P(OBUF_0Y);
1542 P(OBUF_1Y);
1543 P(OBUF_0U);
1544 P(OBUF_0V);
1545 P(OBUF_1U);
1546 P(OBUF_1V);
1547 P(OSTRIDE);
1548 P(YRGB_VPH);
1549 P(UV_VPH);
1550 P(HORZ_PH);
1551 P(INIT_PHS);
1552 P(DWINPOS);
1553 P(DWINSZ);
1554 P(SWIDTH);
1555 P(SWIDTHSW);
1556 P(SHEIGHT);
1557 P(YRGBSCALE);
1558 P(UVSCALE);
1559 P(OCLRC0);
1560 P(OCLRC1);
1561 P(DCLRKV);
1562 P(DCLRKM);
1563 P(SCLRKVH);
1564 P(SCLRKVL);
1565 P(SCLRKEN);
1566 P(OCONFIG);
1567 P(OCMD);
1568 P(OSTART_0Y);
1569 P(OSTART_1Y);
1570 P(OSTART_0U);
1571 P(OSTART_0V);
1572 P(OSTART_1U);
1573 P(OSTART_1V);
1574 P(OTILEOFF_0Y);
1575 P(OTILEOFF_1Y);
1576 P(OTILEOFF_0U);
1577 P(OTILEOFF_0V);
1578 P(OTILEOFF_1U);
1579 P(OTILEOFF_1V);
1580 P(FASTHSCALE);
1581 P(UVSCALEV);
1582#undef P
1583}