projects / linux-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
[media] v4l: ti-vpe: Make sure in job_ready that we have the needed number of dst_bufs
[linux-block.git] / drivers / media / platform / ti-vpe / vpe.c
CommitLineData
45719127
AT		 1/*
2 * TI VPE mem2mem driver, based on the virtual v4l2-mem2mem example driver
3 *
4 * Copyright (c) 2013 Texas Instruments Inc.
5 * David Griego, <dagriego@biglakesoftware.com>
6 * Dale Farnsworth, <dale@farnsworth.org>
7 * Archit Taneja, <archit@ti.com>
8 *
9 * Copyright (c) 2009-2010 Samsung Electronics Co., Ltd.
10 * Pawel Osciak, <pawel@osciak.com>
11 * Marek Szyprowski, <m.szyprowski@samsung.com>
12 *
13 * Based on the virtual v4l2-mem2mem example device
14 *
15 * This program is free software; you can redistribute it and/or modify it
16 * under the terms of the GNU General Public License version 2 as published by
17 * the Free Software Foundation
18 */
19
20#include <linux/delay.h>
21#include <linux/dma-mapping.h>
22#include <linux/err.h>
23#include <linux/fs.h>
24#include <linux/interrupt.h>
25#include <linux/io.h>
26#include <linux/ioctl.h>
27#include <linux/module.h>
28#include <linux/platform_device.h>
29#include <linux/pm_runtime.h>
30#include <linux/sched.h>
31#include <linux/slab.h>
32#include <linux/videodev2.h>
a51cd8f5 33#include <linux/log2.h>
45719127
AT		 34
35#include <media/v4l2-common.h>
36#include <media/v4l2-ctrls.h>
37#include <media/v4l2-device.h>
38#include <media/v4l2-event.h>
39#include <media/v4l2-ioctl.h>
40#include <media/v4l2-mem2mem.h>
41#include <media/videobuf2-core.h>
42#include <media/videobuf2-dma-contig.h>
43
44#include "vpdma.h"
45#include "vpe_regs.h"
44687b2e 46#include "sc.h"
6948082d 47#include "csc.h"
45719127
AT		 48
49#define VPE_MODULE_NAME "vpe"
50
51/* minimum and maximum frame sizes */
52#define MIN_W 128
53#define MIN_H 128
54#define MAX_W 1920
55#define MAX_H 1080
56
57/* required alignments */
58#define S_ALIGN 0 /* multiple of 1 */
59#define H_ALIGN 1 /* multiple of 2 */
45719127
AT		 60
61/* flags that indicate a format can be used for capture/output */
62#define VPE_FMT_TYPE_CAPTURE (1 << 0)
63#define VPE_FMT_TYPE_OUTPUT (1 << 1)
64
65/* used as plane indices */
66#define VPE_MAX_PLANES 2
67#define VPE_LUMA 0
68#define VPE_CHROMA 1
69
70/* per m2m context info */
585e6f01
AT		 71#define VPE_MAX_SRC_BUFS 3 /* need 3 src fields to de-interlace */
72
45719127
AT		 73#define VPE_DEF_BUFS_PER_JOB 1 /* default one buffer per batch job */
74
75/*
76 * each VPE context can need up to 3 config desciptors, 7 input descriptors,
77 * 3 output descriptors, and 10 control descriptors
78 */
79#define VPE_DESC_LIST_SIZE (10 * VPDMA_DTD_DESC_SIZE + \
80 13 * VPDMA_CFD_CTD_DESC_SIZE)
81
82#define vpe_dbg(vpedev, fmt, arg...) \
83 dev_dbg((vpedev)->v4l2_dev.dev, fmt, ##arg)
84#define vpe_err(vpedev, fmt, arg...) \
85 dev_err((vpedev)->v4l2_dev.dev, fmt, ##arg)
86
87struct vpe_us_coeffs {
88 unsigned short anchor_fid0_c0;
89 unsigned short anchor_fid0_c1;
90 unsigned short anchor_fid0_c2;
91 unsigned short anchor_fid0_c3;
92 unsigned short interp_fid0_c0;
93 unsigned short interp_fid0_c1;
94 unsigned short interp_fid0_c2;
95 unsigned short interp_fid0_c3;
96 unsigned short anchor_fid1_c0;
97 unsigned short anchor_fid1_c1;
98 unsigned short anchor_fid1_c2;
99 unsigned short anchor_fid1_c3;
100 unsigned short interp_fid1_c0;
101 unsigned short interp_fid1_c1;
102 unsigned short interp_fid1_c2;
103 unsigned short interp_fid1_c3;
104};
105
106/*
107 * Default upsampler coefficients
108 */
109static const struct vpe_us_coeffs us_coeffs[] = {
110 {
111 /* Coefficients for progressive input */
112 0x00C8, 0x0348, 0x0018, 0x3FD8, 0x3FB8, 0x0378, 0x00E8, 0x3FE8,
113 0x00C8, 0x0348, 0x0018, 0x3FD8, 0x3FB8, 0x0378, 0x00E8, 0x3FE8,
114 },
585e6f01
AT		 115 {
116 /* Coefficients for Top Field Interlaced input */
117 0x0051, 0x03D5, 0x3FE3, 0x3FF7, 0x3FB5, 0x02E9, 0x018F, 0x3FD3,
118 /* Coefficients for Bottom Field Interlaced input */
119 0x016B, 0x0247, 0x00B1, 0x3F9D, 0x3FCF, 0x03DB, 0x005D, 0x3FF9,
120 },
121};
122
123/*
124 * the following registers are for configuring some of the parameters of the
125 * motion and edge detection blocks inside DEI, these generally remain the same,
126 * these could be passed later via userspace if some one needs to tweak these.
127 */
128struct vpe_dei_regs {
129 unsigned long mdt_spacial_freq_thr_reg; /* VPE_DEI_REG2 */
130 unsigned long edi_config_reg; /* VPE_DEI_REG3 */
131 unsigned long edi_lut_reg0; /* VPE_DEI_REG4 */
132 unsigned long edi_lut_reg1; /* VPE_DEI_REG5 */
133 unsigned long edi_lut_reg2; /* VPE_DEI_REG6 */
134 unsigned long edi_lut_reg3; /* VPE_DEI_REG7 */
135};
136
137/*
138 * default expert DEI register values, unlikely to be modified.
139 */
140static const struct vpe_dei_regs dei_regs = {
141 0x020C0804u,
142 0x0118100Fu,
143 0x08040200u,
144 0x1010100Cu,
145 0x10101010u,
146 0x10101010u,
45719127
AT		 147};
148
149/*
150 * The port_data structure contains per-port data.
151 */
152struct vpe_port_data {
153 enum vpdma_channel channel; /* VPDMA channel */
585e6f01 154 u8 vb_index; /* input frame f, f-1, f-2 index */
45719127
AT		 155 u8 vb_part; /* plane index for co-panar formats */
156};
157
158/*
159 * Define indices into the port_data tables
160 */
161#define VPE_PORT_LUMA1_IN 0
162#define VPE_PORT_CHROMA1_IN 1
585e6f01
AT		 163#define VPE_PORT_LUMA2_IN 2
164#define VPE_PORT_CHROMA2_IN 3
165#define VPE_PORT_LUMA3_IN 4
166#define VPE_PORT_CHROMA3_IN 5
167#define VPE_PORT_MV_IN 6
168#define VPE_PORT_MV_OUT 7
45719127
AT		 169#define VPE_PORT_LUMA_OUT 8
170#define VPE_PORT_CHROMA_OUT 9
171#define VPE_PORT_RGB_OUT 10
172
173static const struct vpe_port_data port_data[11] = {
174 [VPE_PORT_LUMA1_IN] = {
175 .channel = VPE_CHAN_LUMA1_IN,
585e6f01 176 .vb_index = 0,
45719127
AT		 177 .vb_part = VPE_LUMA,
178 },
179 [VPE_PORT_CHROMA1_IN] = {
180 .channel = VPE_CHAN_CHROMA1_IN,
585e6f01
AT		 181 .vb_index = 0,
182 .vb_part = VPE_CHROMA,
183 },
184 [VPE_PORT_LUMA2_IN] = {
185 .channel = VPE_CHAN_LUMA2_IN,
186 .vb_index = 1,
187 .vb_part = VPE_LUMA,
188 },
189 [VPE_PORT_CHROMA2_IN] = {
190 .channel = VPE_CHAN_CHROMA2_IN,
191 .vb_index = 1,
192 .vb_part = VPE_CHROMA,
193 },
194 [VPE_PORT_LUMA3_IN] = {
195 .channel = VPE_CHAN_LUMA3_IN,
196 .vb_index = 2,
197 .vb_part = VPE_LUMA,
198 },
199 [VPE_PORT_CHROMA3_IN] = {
200 .channel = VPE_CHAN_CHROMA3_IN,
201 .vb_index = 2,
45719127
AT		 202 .vb_part = VPE_CHROMA,
203 },
585e6f01
AT		 204 [VPE_PORT_MV_IN] = {
205 .channel = VPE_CHAN_MV_IN,
206 },
207 [VPE_PORT_MV_OUT] = {
208 .channel = VPE_CHAN_MV_OUT,
209 },
45719127
AT		 210 [VPE_PORT_LUMA_OUT] = {
211 .channel = VPE_CHAN_LUMA_OUT,
212 .vb_part = VPE_LUMA,
213 },
214 [VPE_PORT_CHROMA_OUT] = {
215 .channel = VPE_CHAN_CHROMA_OUT,
216 .vb_part = VPE_CHROMA,
217 },
218 [VPE_PORT_RGB_OUT] = {
219 .channel = VPE_CHAN_RGB_OUT,
220 .vb_part = VPE_LUMA,
221 },
222};
223
224
225/* driver info for each of the supported video formats */
226struct vpe_fmt {
227 char *name; /* human-readable name */
228 u32 fourcc; /* standard format identifier */
229 u8 types; /* CAPTURE and/or OUTPUT */
230 u8 coplanar; /* set for unpacked Luma and Chroma */
231 /* vpdma format info for each plane */
232 struct vpdma_data_format const *vpdma_fmt[VPE_MAX_PLANES];
233};
234
235static struct vpe_fmt vpe_formats[] = {
236 {
237 .name = "YUV 422 co-planar",
238 .fourcc = V4L2_PIX_FMT_NV16,
239 .types = VPE_FMT_TYPE_CAPTURE | VPE_FMT_TYPE_OUTPUT,
240 .coplanar = 1,
241 .vpdma_fmt = { &vpdma_yuv_fmts[VPDMA_DATA_FMT_Y444],
242 &vpdma_yuv_fmts[VPDMA_DATA_FMT_C444],
243 },
244 },
245 {
246 .name = "YUV 420 co-planar",
247 .fourcc = V4L2_PIX_FMT_NV12,
248 .types = VPE_FMT_TYPE_CAPTURE | VPE_FMT_TYPE_OUTPUT,
249 .coplanar = 1,
250 .vpdma_fmt = { &vpdma_yuv_fmts[VPDMA_DATA_FMT_Y420],
251 &vpdma_yuv_fmts[VPDMA_DATA_FMT_C420],
252 },
253 },
254 {
255 .name = "YUYV 422 packed",
256 .fourcc = V4L2_PIX_FMT_YUYV,
257 .types = VPE_FMT_TYPE_CAPTURE | VPE_FMT_TYPE_OUTPUT,
258 .coplanar = 0,
259 .vpdma_fmt = { &vpdma_yuv_fmts[VPDMA_DATA_FMT_YC422],
260 },
261 },
262 {
263 .name = "UYVY 422 packed",
264 .fourcc = V4L2_PIX_FMT_UYVY,
265 .types = VPE_FMT_TYPE_CAPTURE | VPE_FMT_TYPE_OUTPUT,
266 .coplanar = 0,
267 .vpdma_fmt = { &vpdma_yuv_fmts[VPDMA_DATA_FMT_CY422],
268 },
269 },
30496799
AT		 270 {
271 .name = "RGB888 packed",
272 .fourcc = V4L2_PIX_FMT_RGB24,
273 .types = VPE_FMT_TYPE_CAPTURE,
274 .coplanar = 0,
275 .vpdma_fmt = { &vpdma_rgb_fmts[VPDMA_DATA_FMT_RGB24],
276 },
277 },
278 {
279 .name = "ARGB32",
280 .fourcc = V4L2_PIX_FMT_RGB32,
281 .types = VPE_FMT_TYPE_CAPTURE,
282 .coplanar = 0,
283 .vpdma_fmt = { &vpdma_rgb_fmts[VPDMA_DATA_FMT_ARGB32],
284 },
285 },
286 {
287 .name = "BGR888 packed",
288 .fourcc = V4L2_PIX_FMT_BGR24,
289 .types = VPE_FMT_TYPE_CAPTURE,
290 .coplanar = 0,
291 .vpdma_fmt = { &vpdma_rgb_fmts[VPDMA_DATA_FMT_BGR24],
292 },
293 },
294 {
295 .name = "ABGR32",
296 .fourcc = V4L2_PIX_FMT_BGR32,
297 .types = VPE_FMT_TYPE_CAPTURE,
298 .coplanar = 0,
299 .vpdma_fmt = { &vpdma_rgb_fmts[VPDMA_DATA_FMT_ABGR32],
300 },
301 },
45719127
AT		 302};
303
304/*
305 * per-queue, driver-specific private data.
306 * there is one source queue and one destination queue for each m2m context.
307 */
308struct vpe_q_data {
309 unsigned int width; /* frame width */
310 unsigned int height; /* frame height */
311 unsigned int bytesperline[VPE_MAX_PLANES]; /* bytes per line in memory */
312 enum v4l2_colorspace colorspace;
585e6f01 313 enum v4l2_field field; /* supported field value */
45719127
AT		 314 unsigned int flags;
315 unsigned int sizeimage[VPE_MAX_PLANES]; /* image size in memory */
316 struct v4l2_rect c_rect; /* crop/compose rectangle */
317 struct vpe_fmt *fmt; /* format info */
318};
319
320/* vpe_q_data flag bits */
321#define Q_DATA_FRAME_1D (1 << 0)
322#define Q_DATA_MODE_TILED (1 << 1)
585e6f01 323#define Q_DATA_INTERLACED (1 << 2)
45719127
AT		 324
325enum {
326 Q_DATA_SRC = 0,
327 Q_DATA_DST = 1,
328};
329
330/* find our format description corresponding to the passed v4l2_format */
331static struct vpe_fmt *find_format(struct v4l2_format *f)
332{
333 struct vpe_fmt *fmt;
334 unsigned int k;
335
336 for (k = 0; k < ARRAY_SIZE(vpe_formats); k++) {
337 fmt = &vpe_formats[k];
338 if (fmt->fourcc == f->fmt.pix.pixelformat)
339 return fmt;
340 }
341
342 return NULL;
343}
344
345/*
346 * there is one vpe_dev structure in the driver, it is shared by
347 * all instances.
348 */
349struct vpe_dev {
350 struct v4l2_device v4l2_dev;
351 struct video_device vfd;
352 struct v4l2_m2m_dev *m2m_dev;
353
354 atomic_t num_instances; /* count of driver instances */
355 dma_addr_t loaded_mmrs; /* shadow mmrs in device */
356 struct mutex dev_mutex;
357 spinlock_t lock;
358
359 int irq;
360 void __iomem *base;
44687b2e 361 struct resource *res;
45719127
AT		 362
363 struct vb2_alloc_ctx *alloc_ctx;
364 struct vpdma_data *vpdma; /* vpdma data handle */
44687b2e 365 struct sc_data *sc; /* scaler data handle */
6948082d 366 struct csc_data *csc; /* csc data handle */
45719127
AT		 367};
368
369/*
370 * There is one vpe_ctx structure for each m2m context.
371 */
372struct vpe_ctx {
373 struct v4l2_fh fh;
374 struct vpe_dev *dev;
375 struct v4l2_m2m_ctx *m2m_ctx;
376 struct v4l2_ctrl_handler hdl;
377
585e6f01 378 unsigned int field; /* current field */
45719127
AT		 379 unsigned int sequence; /* current frame/field seq */
380 unsigned int aborting; /* abort after next irq */
381
382 unsigned int bufs_per_job; /* input buffers per batch */
383 unsigned int bufs_completed; /* bufs done in this batch */
384
385 struct vpe_q_data q_data[2]; /* src & dst queue data */
585e6f01 386 struct vb2_buffer *src_vbs[VPE_MAX_SRC_BUFS];
45719127
AT		 387 struct vb2_buffer *dst_vb;
388
585e6f01
AT		 389 dma_addr_t mv_buf_dma[2]; /* dma addrs of motion vector in/out bufs */
390 void *mv_buf[2]; /* virtual addrs of motion vector bufs */
391 size_t mv_buf_size; /* current motion vector buffer size */
45719127 392 struct vpdma_buf mmr_adb; /* shadow reg addr/data block */
773f0657
AT		 393 struct vpdma_buf sc_coeff_h; /* h coeff buffer */
394 struct vpdma_buf sc_coeff_v; /* v coeff buffer */
45719127
AT		 395 struct vpdma_desc_list desc_list; /* DMA descriptor list */
396
585e6f01 397 bool deinterlacing; /* using de-interlacer */
45719127 398 bool load_mmrs; /* have new shadow reg values */
585e6f01
AT		 399
400 unsigned int src_mv_buf_selector;
45719127
AT		 401};
402
403
404/*
405 * M2M devices get 2 queues.
406 * Return the queue given the type.
407 */
408static struct vpe_q_data *get_q_data(struct vpe_ctx *ctx,
409 enum v4l2_buf_type type)
410{
411 switch (type) {
412 case V4L2_BUF_TYPE_VIDEO_OUTPUT_MPLANE:
413 return &ctx->q_data[Q_DATA_SRC];
414 case V4L2_BUF_TYPE_VIDEO_CAPTURE_MPLANE:
415 return &ctx->q_data[Q_DATA_DST];
416 default:
417 BUG();
418 }
419 return NULL;
420}
421
422static u32 read_reg(struct vpe_dev *dev, int offset)
423{
424 return ioread32(dev->base + offset);
425}
426
427static void write_reg(struct vpe_dev *dev, int offset, u32 value)
428{
429 iowrite32(value, dev->base + offset);
430}
431
432/* register field read/write helpers */
433static int get_field(u32 value, u32 mask, int shift)
434{
435 return (value & (mask << shift)) >> shift;
436}
437
438static int read_field_reg(struct vpe_dev *dev, int offset, u32 mask, int shift)
439{
440 return get_field(read_reg(dev, offset), mask, shift);
441}
442
443static void write_field(u32 *valp, u32 field, u32 mask, int shift)
444{
445 u32 val = *valp;
446
447 val &= ~(mask << shift);
448 val |= (field & mask) << shift;
449 *valp = val;
450}
451
452static void write_field_reg(struct vpe_dev *dev, int offset, u32 field,
453 u32 mask, int shift)
454{
455 u32 val = read_reg(dev, offset);
456
457 write_field(&val, field, mask, shift);
458
459 write_reg(dev, offset, val);
460}
461
462/*
463 * DMA address/data block for the shadow registers
464 */
465struct vpe_mmr_adb {
466 struct vpdma_adb_hdr out_fmt_hdr;
467 u32 out_fmt_reg[1];
468 u32 out_fmt_pad[3];
469 struct vpdma_adb_hdr us1_hdr;
470 u32 us1_regs[8];
471 struct vpdma_adb_hdr us2_hdr;
472 u32 us2_regs[8];
473 struct vpdma_adb_hdr us3_hdr;
474 u32 us3_regs[8];
475 struct vpdma_adb_hdr dei_hdr;
585e6f01 476 u32 dei_regs[8];
bbee8b39
AT		 477 struct vpdma_adb_hdr sc_hdr0;
478 u32 sc_regs0[7];
479 u32 sc_pad0[1];
480 struct vpdma_adb_hdr sc_hdr8;
481 u32 sc_regs8[6];
482 u32 sc_pad8[2];
483 struct vpdma_adb_hdr sc_hdr17;
484 u32 sc_regs17[9];
485 u32 sc_pad17[3];
45719127
AT		 486 struct vpdma_adb_hdr csc_hdr;
487 u32 csc_regs[6];
488 u32 csc_pad[2];
489};
490
44687b2e
AT		 491#define GET_OFFSET_TOP(ctx, obj, reg) \
492 ((obj)->res->start - ctx->dev->res->start + reg)
493
45719127
AT		 494#define VPE_SET_MMR_ADB_HDR(ctx, hdr, regs, offset_a) \
495 VPDMA_SET_MMR_ADB_HDR(ctx->mmr_adb, vpe_mmr_adb, hdr, regs, offset_a)
496/*
497 * Set the headers for all of the address/data block structures.
498 */
499static void init_adb_hdrs(struct vpe_ctx *ctx)
500{
501 VPE_SET_MMR_ADB_HDR(ctx, out_fmt_hdr, out_fmt_reg, VPE_CLK_FORMAT_SELECT);
502 VPE_SET_MMR_ADB_HDR(ctx, us1_hdr, us1_regs, VPE_US1_R0);
503 VPE_SET_MMR_ADB_HDR(ctx, us2_hdr, us2_regs, VPE_US2_R0);
504 VPE_SET_MMR_ADB_HDR(ctx, us3_hdr, us3_regs, VPE_US3_R0);
505 VPE_SET_MMR_ADB_HDR(ctx, dei_hdr, dei_regs, VPE_DEI_FRAME_SIZE);
bbee8b39 506 VPE_SET_MMR_ADB_HDR(ctx, sc_hdr0, sc_regs0,
44687b2e 507 GET_OFFSET_TOP(ctx, ctx->dev->sc, CFG_SC0));
bbee8b39
AT		 508 VPE_SET_MMR_ADB_HDR(ctx, sc_hdr8, sc_regs8,
509 GET_OFFSET_TOP(ctx, ctx->dev->sc, CFG_SC8));
510 VPE_SET_MMR_ADB_HDR(ctx, sc_hdr17, sc_regs17,
511 GET_OFFSET_TOP(ctx, ctx->dev->sc, CFG_SC17));
6948082d
AT		 512 VPE_SET_MMR_ADB_HDR(ctx, csc_hdr, csc_regs,
513 GET_OFFSET_TOP(ctx, ctx->dev->csc, CSC_CSC00));
45719127
AT		 514};
515
585e6f01
AT		 516/*
517 * Allocate or re-allocate the motion vector DMA buffers
518 * There are two buffers, one for input and one for output.
519 * However, the roles are reversed after each field is processed.
520 * In other words, after each field is processed, the previous
521 * output (dst) MV buffer becomes the new input (src) MV buffer.
522 */
523static int realloc_mv_buffers(struct vpe_ctx *ctx, size_t size)
524{
525 struct device *dev = ctx->dev->v4l2_dev.dev;
526
527 if (ctx->mv_buf_size == size)
528 return 0;
529
530 if (ctx->mv_buf[0])
531 dma_free_coherent(dev, ctx->mv_buf_size, ctx->mv_buf[0],
532 ctx->mv_buf_dma[0]);
533
534 if (ctx->mv_buf[1])
535 dma_free_coherent(dev, ctx->mv_buf_size, ctx->mv_buf[1],
536 ctx->mv_buf_dma[1]);
537
538 if (size == 0)
539 return 0;
540
541 ctx->mv_buf[0] = dma_alloc_coherent(dev, size, &ctx->mv_buf_dma[0],
542 GFP_KERNEL);
543 if (!ctx->mv_buf[0]) {
544 vpe_err(ctx->dev, "failed to allocate motion vector buffer\n");
545 return -ENOMEM;
546 }
547
548 ctx->mv_buf[1] = dma_alloc_coherent(dev, size, &ctx->mv_buf_dma[1],
549 GFP_KERNEL);
550 if (!ctx->mv_buf[1]) {
551 vpe_err(ctx->dev, "failed to allocate motion vector buffer\n");
552 dma_free_coherent(dev, size, ctx->mv_buf[0],
553 ctx->mv_buf_dma[0]);
554
555 return -ENOMEM;
556 }
557
558 ctx->mv_buf_size = size;
559 ctx->src_mv_buf_selector = 0;
560
561 return 0;
562}
563
564static void free_mv_buffers(struct vpe_ctx *ctx)
565{
566 realloc_mv_buffers(ctx, 0);
567}
568
569/*
570 * While de-interlacing, we keep the two most recent input buffers
571 * around. This function frees those two buffers when we have
572 * finished processing the current stream.
573 */
574static void free_vbs(struct vpe_ctx *ctx)
575{
576 struct vpe_dev *dev = ctx->dev;
577 unsigned long flags;
578
579 if (ctx->src_vbs[2] == NULL)
580 return;
581
582 spin_lock_irqsave(&dev->lock, flags);
583 if (ctx->src_vbs[2]) {
584 v4l2_m2m_buf_done(ctx->src_vbs[2], VB2_BUF_STATE_DONE);
585 v4l2_m2m_buf_done(ctx->src_vbs[1], VB2_BUF_STATE_DONE);
586 }
587 spin_unlock_irqrestore(&dev->lock, flags);
588}
589
45719127
AT		 590/*
591 * Enable or disable the VPE clocks
592 */
593static void vpe_set_clock_enable(struct vpe_dev *dev, bool on)
594{
595 u32 val = 0;
596
597 if (on)
598 val = VPE_DATA_PATH_CLK_ENABLE | VPE_VPEDMA_CLK_ENABLE;
599 write_reg(dev, VPE_CLK_ENABLE, val);
600}
601
602static void vpe_top_reset(struct vpe_dev *dev)
603{
604
605 write_field_reg(dev, VPE_CLK_RESET, 1, VPE_DATA_PATH_CLK_RESET_MASK,
606 VPE_DATA_PATH_CLK_RESET_SHIFT);
607
608 usleep_range(100, 150);
609
610 write_field_reg(dev, VPE_CLK_RESET, 0, VPE_DATA_PATH_CLK_RESET_MASK,
611 VPE_DATA_PATH_CLK_RESET_SHIFT);
612}
613
614static void vpe_top_vpdma_reset(struct vpe_dev *dev)
615{
616 write_field_reg(dev, VPE_CLK_RESET, 1, VPE_VPDMA_CLK_RESET_MASK,
617 VPE_VPDMA_CLK_RESET_SHIFT);
618
619 usleep_range(100, 150);
620
621 write_field_reg(dev, VPE_CLK_RESET, 0, VPE_VPDMA_CLK_RESET_MASK,
622 VPE_VPDMA_CLK_RESET_SHIFT);
623}
624
625/*
626 * Load the correct of upsampler coefficients into the shadow MMRs
627 */
628static void set_us_coefficients(struct vpe_ctx *ctx)
629{
630 struct vpe_mmr_adb *mmr_adb = ctx->mmr_adb.addr;
585e6f01 631 struct vpe_q_data *s_q_data = &ctx->q_data[Q_DATA_SRC];
45719127
AT		 632 u32 *us1_reg = &mmr_adb->us1_regs[0];
633 u32 *us2_reg = &mmr_adb->us2_regs[0];
634 u32 *us3_reg = &mmr_adb->us3_regs[0];
635 const unsigned short *cp, *end_cp;
636
637 cp = &us_coeffs[0].anchor_fid0_c0;
638
585e6f01
AT		 639 if (s_q_data->flags & Q_DATA_INTERLACED) /* interlaced */
640 cp += sizeof(us_coeffs[0]) / sizeof(*cp);
641
45719127
AT		 642 end_cp = cp + sizeof(us_coeffs[0]) / sizeof(*cp);
643
644 while (cp < end_cp) {
645 write_field(us1_reg, *cp++, VPE_US_C0_MASK, VPE_US_C0_SHIFT);
646 write_field(us1_reg, *cp++, VPE_US_C1_MASK, VPE_US_C1_SHIFT);
647 *us2_reg++ = *us1_reg;
648 *us3_reg++ = *us1_reg++;
649 }
650 ctx->load_mmrs = true;
651}
652
653/*
654 * Set the upsampler config mode and the VPDMA line mode in the shadow MMRs.
655 */
656static void set_cfg_and_line_modes(struct vpe_ctx *ctx)
657{
658 struct vpe_fmt *fmt = ctx->q_data[Q_DATA_SRC].fmt;
659 struct vpe_mmr_adb *mmr_adb = ctx->mmr_adb.addr;
660 u32 *us1_reg0 = &mmr_adb->us1_regs[0];
661 u32 *us2_reg0 = &mmr_adb->us2_regs[0];
662 u32 *us3_reg0 = &mmr_adb->us3_regs[0];
663 int line_mode = 1;
664 int cfg_mode = 1;
665
666 /*
667 * Cfg Mode 0: YUV420 source, enable upsampler, DEI is de-interlacing.
668 * Cfg Mode 1: YUV422 source, disable upsampler, DEI is de-interlacing.
669 */
670
671 if (fmt->fourcc == V4L2_PIX_FMT_NV12) {
672 cfg_mode = 0;
673 line_mode = 0; /* double lines to line buffer */
674 }
675
676 write_field(us1_reg0, cfg_mode, VPE_US_MODE_MASK, VPE_US_MODE_SHIFT);
677 write_field(us2_reg0, cfg_mode, VPE_US_MODE_MASK, VPE_US_MODE_SHIFT);
678 write_field(us3_reg0, cfg_mode, VPE_US_MODE_MASK, VPE_US_MODE_SHIFT);
679
680 /* regs for now */
681 vpdma_set_line_mode(ctx->dev->vpdma, line_mode, VPE_CHAN_CHROMA1_IN);
585e6f01
AT		 682 vpdma_set_line_mode(ctx->dev->vpdma, line_mode, VPE_CHAN_CHROMA2_IN);
683 vpdma_set_line_mode(ctx->dev->vpdma, line_mode, VPE_CHAN_CHROMA3_IN);
45719127
AT		 684
685 /* frame start for input luma */
686 vpdma_set_frame_start_event(ctx->dev->vpdma, VPDMA_FSEVENT_CHANNEL_ACTIVE,
687 VPE_CHAN_LUMA1_IN);
585e6f01
AT		 688 vpdma_set_frame_start_event(ctx->dev->vpdma, VPDMA_FSEVENT_CHANNEL_ACTIVE,
689 VPE_CHAN_LUMA2_IN);
690 vpdma_set_frame_start_event(ctx->dev->vpdma, VPDMA_FSEVENT_CHANNEL_ACTIVE,
691 VPE_CHAN_LUMA3_IN);
45719127
AT		 692
693 /* frame start for input chroma */
694 vpdma_set_frame_start_event(ctx->dev->vpdma, VPDMA_FSEVENT_CHANNEL_ACTIVE,
695 VPE_CHAN_CHROMA1_IN);
585e6f01
AT		 696 vpdma_set_frame_start_event(ctx->dev->vpdma, VPDMA_FSEVENT_CHANNEL_ACTIVE,
697 VPE_CHAN_CHROMA2_IN);
698 vpdma_set_frame_start_event(ctx->dev->vpdma, VPDMA_FSEVENT_CHANNEL_ACTIVE,
699 VPE_CHAN_CHROMA3_IN);
700
701 /* frame start for MV in client */
702 vpdma_set_frame_start_event(ctx->dev->vpdma, VPDMA_FSEVENT_CHANNEL_ACTIVE,
703 VPE_CHAN_MV_IN);
45719127
AT		 704
705 ctx->load_mmrs = true;
706}
707
708/*
709 * Set the shadow registers that are modified when the source
710 * format changes.
711 */
712static void set_src_registers(struct vpe_ctx *ctx)
713{
714 set_us_coefficients(ctx);
715}
716
717/*
718 * Set the shadow registers that are modified when the destination
719 * format changes.
720 */
721static void set_dst_registers(struct vpe_ctx *ctx)
722{
723 struct vpe_mmr_adb *mmr_adb = ctx->mmr_adb.addr;
30496799 724 enum v4l2_colorspace clrspc = ctx->q_data[Q_DATA_DST].colorspace;
45719127
AT		 725 struct vpe_fmt *fmt = ctx->q_data[Q_DATA_DST].fmt;
726 u32 val = 0;
727
30496799
AT		 728 if (clrspc == V4L2_COLORSPACE_SRGB)
729 val |= VPE_RGB_OUT_SELECT;
45719127
AT		 730 else if (fmt->fourcc == V4L2_PIX_FMT_NV16)
731 val |= VPE_COLOR_SEPARATE_422;
732
30496799
AT		 733 /*
734 * the source of CHR_DS and CSC is always the scaler, irrespective of
735 * whether it's used or not
736 */
737 val |= VPE_DS_SRC_DEI_SCALER | VPE_CSC_SRC_DEI_SCALER;
45719127
AT		 738
739 if (fmt->fourcc != V4L2_PIX_FMT_NV12)
740 val |= VPE_DS_BYPASS;
741
742 mmr_adb->out_fmt_reg[0] = val;
743
744 ctx->load_mmrs = true;
745}
746
747/*
748 * Set the de-interlacer shadow register values
749 */
585e6f01 750static void set_dei_regs(struct vpe_ctx *ctx)
45719127
AT		 751{
752 struct vpe_mmr_adb *mmr_adb = ctx->mmr_adb.addr;
753 struct vpe_q_data *s_q_data = &ctx->q_data[Q_DATA_SRC];
754 unsigned int src_h = s_q_data->c_rect.height;
755 unsigned int src_w = s_q_data->c_rect.width;
756 u32 *dei_mmr0 = &mmr_adb->dei_regs[0];
585e6f01 757 bool deinterlace = true;
45719127
AT		 758 u32 val = 0;
759
760 /*
761 * according to TRM, we should set DEI in progressive bypass mode when
762 * the input content is progressive, however, DEI is bypassed correctly
763 * for both progressive and interlace content in interlace bypass mode.
764 * It has been recommended not to use progressive bypass mode.
765 */
585e6f01
AT		 766 if ((!ctx->deinterlacing && (s_q_data->flags & Q_DATA_INTERLACED)) ||
767 !(s_q_data->flags & Q_DATA_INTERLACED)) {
768 deinterlace = false;
769 val = VPE_DEI_INTERLACE_BYPASS;
770 }
771
772 src_h = deinterlace ? src_h * 2 : src_h;
45719127
AT		 773
774 val |= (src_h << VPE_DEI_HEIGHT_SHIFT) |
775 (src_w << VPE_DEI_WIDTH_SHIFT) |
776 VPE_DEI_FIELD_FLUSH;
777
778 *dei_mmr0 = val;
779
780 ctx->load_mmrs = true;
781}
782
585e6f01
AT		 783static void set_dei_shadow_registers(struct vpe_ctx *ctx)
784{
785 struct vpe_mmr_adb *mmr_adb = ctx->mmr_adb.addr;
786 u32 *dei_mmr = &mmr_adb->dei_regs[0];
787 const struct vpe_dei_regs *cur = &dei_regs;
788
789 dei_mmr[2] = cur->mdt_spacial_freq_thr_reg;
790 dei_mmr[3] = cur->edi_config_reg;
791 dei_mmr[4] = cur->edi_lut_reg0;
792 dei_mmr[5] = cur->edi_lut_reg1;
793 dei_mmr[6] = cur->edi_lut_reg2;
794 dei_mmr[7] = cur->edi_lut_reg3;
795
796 ctx->load_mmrs = true;
797}
798
45719127
AT		 799/*
800 * Set the shadow registers whose values are modified when either the
801 * source or destination format is changed.
802 */
803static int set_srcdst_params(struct vpe_ctx *ctx)
804{
585e6f01
AT		 805 struct vpe_q_data *s_q_data = &ctx->q_data[Q_DATA_SRC];
806 struct vpe_q_data *d_q_data = &ctx->q_data[Q_DATA_DST];
44687b2e 807 struct vpe_mmr_adb *mmr_adb = ctx->mmr_adb.addr;
773f0657
AT		 808 unsigned int src_w = s_q_data->c_rect.width;
809 unsigned int src_h = s_q_data->c_rect.height;
810 unsigned int dst_w = d_q_data->c_rect.width;
811 unsigned int dst_h = d_q_data->c_rect.height;
585e6f01
AT		 812 size_t mv_buf_size;
813 int ret;
814
45719127 815 ctx->sequence = 0;
585e6f01
AT		 816 ctx->field = V4L2_FIELD_TOP;
817
818 if ((s_q_data->flags & Q_DATA_INTERLACED) &&
819 !(d_q_data->flags & Q_DATA_INTERLACED)) {
a51cd8f5 820 int bytes_per_line;
585e6f01
AT		 821 const struct vpdma_data_format *mv =
822 &vpdma_misc_fmts[VPDMA_DATA_FMT_MV];
823
a51cd8f5
AT		 824 /*
825 * we make sure that the source image has a 16 byte aligned
826 * stride, we need to do the same for the motion vector buffer
827 * by aligning it's stride to the next 16 byte boundry. this
828 * extra space will not be used by the de-interlacer, but will
829 * ensure that vpdma operates correctly
830 */
831 bytes_per_line = ALIGN((s_q_data->width * mv->depth) >> 3,
832 VPDMA_STRIDE_ALIGN);
833 mv_buf_size = bytes_per_line * s_q_data->height;
773f0657
AT		 834
835 ctx->deinterlacing = 1;
836 src_h <<= 1;
585e6f01
AT		 837 } else {
838 ctx->deinterlacing = 0;
839 mv_buf_size = 0;
840 }
841
842 free_vbs(ctx);
843
844 ret = realloc_mv_buffers(ctx, mv_buf_size);
845 if (ret)
846 return ret;
45719127
AT		 847
848 set_cfg_and_line_modes(ctx);
585e6f01 849 set_dei_regs(ctx);
6948082d 850
30496799
AT		 851 csc_set_coeff(ctx->dev->csc, &mmr_adb->csc_regs[0],
852 s_q_data->colorspace, d_q_data->colorspace);
bbee8b39 853
773f0657
AT		 854 sc_set_hs_coeffs(ctx->dev->sc, ctx->sc_coeff_h.addr, src_w, dst_w);
855 sc_set_vs_coeffs(ctx->dev->sc, ctx->sc_coeff_v.addr, src_h, dst_h);
bbee8b39
AT		 856
857 sc_config_scaler(ctx->dev->sc, &mmr_adb->sc_regs0[0],
858 &mmr_adb->sc_regs8[0], &mmr_adb->sc_regs17[0],
859 src_w, src_h, dst_w, dst_h);
45719127
AT		 860
861 return 0;
862}
863
864/*
865 * Return the vpe_ctx structure for a given struct file
866 */
867static struct vpe_ctx *file2ctx(struct file *file)
868{
869 return container_of(file->private_data, struct vpe_ctx, fh);
870}
871
872/*
873 * mem2mem callbacks
874 */
875
876/**
877 * job_ready() - check whether an instance is ready to be scheduled to run
878 */
879static int job_ready(void *priv)
880{
881 struct vpe_ctx *ctx = priv;
882 int needed = ctx->bufs_per_job;
883
585e6f01
AT		 884 if (ctx->deinterlacing && ctx->src_vbs[2] == NULL)
885 needed += 2; /* need additional two most recent fields */
886
45719127
AT		 887 if (v4l2_m2m_num_src_bufs_ready(ctx->m2m_ctx) < needed)
888 return 0;
889
db476163
AT		 890 if (v4l2_m2m_num_dst_bufs_ready(ctx->m2m_ctx) < needed)
891 return 0;
892
45719127
AT		 893 return 1;
894}
895
896static void job_abort(void *priv)
897{
898 struct vpe_ctx *ctx = priv;
899
900 /* Will cancel the transaction in the next interrupt handler */
901 ctx->aborting = 1;
902}
903
904/*
905 * Lock access to the device
906 */
907static void vpe_lock(void *priv)
908{
909 struct vpe_ctx *ctx = priv;
910 struct vpe_dev *dev = ctx->dev;
911 mutex_lock(&dev->dev_mutex);
912}
913
914static void vpe_unlock(void *priv)
915{
916 struct vpe_ctx *ctx = priv;
917 struct vpe_dev *dev = ctx->dev;
918 mutex_unlock(&dev->dev_mutex);
919}
920
921static void vpe_dump_regs(struct vpe_dev *dev)
922{
923#define DUMPREG(r) vpe_dbg(dev, "%-35s %08x\n", #r, read_reg(dev, VPE_##r))
924
925 vpe_dbg(dev, "VPE Registers:\n");
926
927 DUMPREG(PID);
928 DUMPREG(SYSCONFIG);
929 DUMPREG(INT0_STATUS0_RAW);
930 DUMPREG(INT0_STATUS0);
931 DUMPREG(INT0_ENABLE0);
932 DUMPREG(INT0_STATUS1_RAW);
933 DUMPREG(INT0_STATUS1);
934 DUMPREG(INT0_ENABLE1);
935 DUMPREG(CLK_ENABLE);
936 DUMPREG(CLK_RESET);
937 DUMPREG(CLK_FORMAT_SELECT);
938 DUMPREG(CLK_RANGE_MAP);
939 DUMPREG(US1_R0);
940 DUMPREG(US1_R1);
941 DUMPREG(US1_R2);
942 DUMPREG(US1_R3);
943 DUMPREG(US1_R4);
944 DUMPREG(US1_R5);
945 DUMPREG(US1_R6);
946 DUMPREG(US1_R7);
947 DUMPREG(US2_R0);
948 DUMPREG(US2_R1);
949 DUMPREG(US2_R2);
950 DUMPREG(US2_R3);
951 DUMPREG(US2_R4);
952 DUMPREG(US2_R5);
953 DUMPREG(US2_R6);
954 DUMPREG(US2_R7);
955 DUMPREG(US3_R0);
956 DUMPREG(US3_R1);
957 DUMPREG(US3_R2);
958 DUMPREG(US3_R3);
959 DUMPREG(US3_R4);
960 DUMPREG(US3_R5);
961 DUMPREG(US3_R6);
962 DUMPREG(US3_R7);
963 DUMPREG(DEI_FRAME_SIZE);
964 DUMPREG(MDT_BYPASS);
965 DUMPREG(MDT_SF_THRESHOLD);
966 DUMPREG(EDI_CONFIG);
967 DUMPREG(DEI_EDI_LUT_R0);
968 DUMPREG(DEI_EDI_LUT_R1);
969 DUMPREG(DEI_EDI_LUT_R2);
970 DUMPREG(DEI_EDI_LUT_R3);
971 DUMPREG(DEI_FMD_WINDOW_R0);
972 DUMPREG(DEI_FMD_WINDOW_R1);
973 DUMPREG(DEI_FMD_CONTROL_R0);
974 DUMPREG(DEI_FMD_CONTROL_R1);
975 DUMPREG(DEI_FMD_STATUS_R0);
976 DUMPREG(DEI_FMD_STATUS_R1);
977 DUMPREG(DEI_FMD_STATUS_R2);
45719127 978#undef DUMPREG
44687b2e
AT		 979
980 sc_dump_regs(dev->sc);
6948082d 981 csc_dump_regs(dev->csc);
45719127
AT		 982}
983
984static void add_out_dtd(struct vpe_ctx *ctx, int port)
985{
986 struct vpe_q_data *q_data = &ctx->q_data[Q_DATA_DST];
987 const struct vpe_port_data *p_data = &port_data[port];
988 struct vb2_buffer *vb = ctx->dst_vb;
989 struct v4l2_rect *c_rect = &q_data->c_rect;
990 struct vpe_fmt *fmt = q_data->fmt;
991 const struct vpdma_data_format *vpdma_fmt;
585e6f01 992 int mv_buf_selector = !ctx->src_mv_buf_selector;
45719127
AT		 993 dma_addr_t dma_addr;
994 u32 flags = 0;
995
585e6f01
AT		 996 if (port == VPE_PORT_MV_OUT) {
997 vpdma_fmt = &vpdma_misc_fmts[VPDMA_DATA_FMT_MV];
998 dma_addr = ctx->mv_buf_dma[mv_buf_selector];
999 } else {
1000 /* to incorporate interleaved formats */
1001 int plane = fmt->coplanar ? p_data->vb_part : 0;
1002
1003 vpdma_fmt = fmt->vpdma_fmt[plane];
1004 dma_addr = vb2_dma_contig_plane_dma_addr(vb, plane);
1005 if (!dma_addr) {
1006 vpe_err(ctx->dev,
1007 "acquiring output buffer(%d) dma_addr failed\n",
1008 port);
1009 return;
1010 }
45719127
AT		 1011 }
1012
1013 if (q_data->flags & Q_DATA_FRAME_1D)
1014 flags |= VPDMA_DATA_FRAME_1D;
1015 if (q_data->flags & Q_DATA_MODE_TILED)
1016 flags |= VPDMA_DATA_MODE_TILED;
1017
1018 vpdma_add_out_dtd(&ctx->desc_list, c_rect, vpdma_fmt, dma_addr,
1019 p_data->channel, flags);
1020}
1021
1022static void add_in_dtd(struct vpe_ctx *ctx, int port)
1023{
1024 struct vpe_q_data *q_data = &ctx->q_data[Q_DATA_SRC];
1025 const struct vpe_port_data *p_data = &port_data[port];
585e6f01 1026 struct vb2_buffer *vb = ctx->src_vbs[p_data->vb_index];
45719127
AT		 1027 struct v4l2_rect *c_rect = &q_data->c_rect;
1028 struct vpe_fmt *fmt = q_data->fmt;
1029 const struct vpdma_data_format *vpdma_fmt;
585e6f01
AT		 1030 int mv_buf_selector = ctx->src_mv_buf_selector;
1031 int field = vb->v4l2_buf.field == V4L2_FIELD_BOTTOM;
45719127
AT		 1032 dma_addr_t dma_addr;
1033 u32 flags = 0;
1034
585e6f01
AT		 1035 if (port == VPE_PORT_MV_IN) {
1036 vpdma_fmt = &vpdma_misc_fmts[VPDMA_DATA_FMT_MV];
1037 dma_addr = ctx->mv_buf_dma[mv_buf_selector];
1038 } else {
1039 /* to incorporate interleaved formats */
1040 int plane = fmt->coplanar ? p_data->vb_part : 0;
45719127 1041
585e6f01
AT		 1042 vpdma_fmt = fmt->vpdma_fmt[plane];
1043
1044 dma_addr = vb2_dma_contig_plane_dma_addr(vb, plane);
1045 if (!dma_addr) {
1046 vpe_err(ctx->dev,
1047 "acquiring input buffer(%d) dma_addr failed\n",
1048 port);
1049 return;
1050 }
45719127
AT		 1051 }
1052
1053 if (q_data->flags & Q_DATA_FRAME_1D)
1054 flags |= VPDMA_DATA_FRAME_1D;
1055 if (q_data->flags & Q_DATA_MODE_TILED)
1056 flags |= VPDMA_DATA_MODE_TILED;
1057
1058 vpdma_add_in_dtd(&ctx->desc_list, q_data->width, q_data->height,
1059 c_rect, vpdma_fmt, dma_addr, p_data->channel, field, flags);
1060}
1061
1062/*
1063 * Enable the expected IRQ sources
1064 */
1065static void enable_irqs(struct vpe_ctx *ctx)
1066{
1067 write_reg(ctx->dev, VPE_INT0_ENABLE0_SET, VPE_INT0_LIST0_COMPLETE);
585e6f01
AT		 1068 write_reg(ctx->dev, VPE_INT0_ENABLE1_SET, VPE_DEI_ERROR_INT |
1069 VPE_DS1_UV_ERROR_INT);
45719127
AT		 1070
1071 vpdma_enable_list_complete_irq(ctx->dev->vpdma, 0, true);
1072}
1073
1074static void disable_irqs(struct vpe_ctx *ctx)
1075{
1076 write_reg(ctx->dev, VPE_INT0_ENABLE0_CLR, 0xffffffff);
1077 write_reg(ctx->dev, VPE_INT0_ENABLE1_CLR, 0xffffffff);
1078
1079 vpdma_enable_list_complete_irq(ctx->dev->vpdma, 0, false);
1080}
1081
1082/* device_run() - prepares and starts the device
1083 *
1084 * This function is only called when both the source and destination
1085 * buffers are in place.
1086 */
1087static void device_run(void *priv)
1088{
1089 struct vpe_ctx *ctx = priv;
773f0657 1090 struct sc_data *sc = ctx->dev->sc;
45719127
AT		 1091 struct vpe_q_data *d_q_data = &ctx->q_data[Q_DATA_DST];
1092
585e6f01
AT		 1093 if (ctx->deinterlacing && ctx->src_vbs[2] == NULL) {
1094 ctx->src_vbs[2] = v4l2_m2m_src_buf_remove(ctx->m2m_ctx);
1095 WARN_ON(ctx->src_vbs[2] == NULL);
1096 ctx->src_vbs[1] = v4l2_m2m_src_buf_remove(ctx->m2m_ctx);
1097 WARN_ON(ctx->src_vbs[1] == NULL);
1098 }
1099
1100 ctx->src_vbs[0] = v4l2_m2m_src_buf_remove(ctx->m2m_ctx);
1101 WARN_ON(ctx->src_vbs[0] == NULL);
45719127
AT		 1102 ctx->dst_vb = v4l2_m2m_dst_buf_remove(ctx->m2m_ctx);
1103 WARN_ON(ctx->dst_vb == NULL);
1104
1105 /* config descriptors */
1106 if (ctx->dev->loaded_mmrs != ctx->mmr_adb.dma_addr || ctx->load_mmrs) {
1107 vpdma_map_desc_buf(ctx->dev->vpdma, &ctx->mmr_adb);
1108 vpdma_add_cfd_adb(&ctx->desc_list, CFD_MMR_CLIENT, &ctx->mmr_adb);
1109 ctx->dev->loaded_mmrs = ctx->mmr_adb.dma_addr;
1110 ctx->load_mmrs = false;
1111 }
1112
773f0657
AT		 1113 if (sc->loaded_coeff_h != ctx->sc_coeff_h.dma_addr ||
1114 sc->load_coeff_h) {
1115 vpdma_map_desc_buf(ctx->dev->vpdma, &ctx->sc_coeff_h);
1116 vpdma_add_cfd_block(&ctx->desc_list, CFD_SC_CLIENT,
1117 &ctx->sc_coeff_h, 0);
1118
1119 sc->loaded_coeff_h = ctx->sc_coeff_h.dma_addr;
1120 sc->load_coeff_h = false;
1121 }
1122
1123 if (sc->loaded_coeff_v != ctx->sc_coeff_v.dma_addr ||
1124 sc->load_coeff_v) {
1125 vpdma_map_desc_buf(ctx->dev->vpdma, &ctx->sc_coeff_v);
1126 vpdma_add_cfd_block(&ctx->desc_list, CFD_SC_CLIENT,
1127 &ctx->sc_coeff_v, SC_COEF_SRAM_SIZE >> 4);
1128
1129 sc->loaded_coeff_v = ctx->sc_coeff_v.dma_addr;
1130 sc->load_coeff_v = false;
1131 }
1132
585e6f01
AT		 1133 /* output data descriptors */
1134 if (ctx->deinterlacing)
1135 add_out_dtd(ctx, VPE_PORT_MV_OUT);
1136
30496799
AT		 1137 if (d_q_data->colorspace == V4L2_COLORSPACE_SRGB) {
1138 add_out_dtd(ctx, VPE_PORT_RGB_OUT);
1139 } else {
1140 add_out_dtd(ctx, VPE_PORT_LUMA_OUT);
1141 if (d_q_data->fmt->coplanar)
1142 add_out_dtd(ctx, VPE_PORT_CHROMA_OUT);
1143 }
45719127 1144
585e6f01
AT		 1145 /* input data descriptors */
1146 if (ctx->deinterlacing) {
1147 add_in_dtd(ctx, VPE_PORT_LUMA3_IN);
1148 add_in_dtd(ctx, VPE_PORT_CHROMA3_IN);
1149
1150 add_in_dtd(ctx, VPE_PORT_LUMA2_IN);
1151 add_in_dtd(ctx, VPE_PORT_CHROMA2_IN);
1152 }
1153
45719127
AT		 1154 add_in_dtd(ctx, VPE_PORT_LUMA1_IN);
1155 add_in_dtd(ctx, VPE_PORT_CHROMA1_IN);
1156
585e6f01
AT		 1157 if (ctx->deinterlacing)
1158 add_in_dtd(ctx, VPE_PORT_MV_IN);
1159
45719127
AT		 1160 /* sync on channel control descriptors for input ports */
1161 vpdma_add_sync_on_channel_ctd(&ctx->desc_list, VPE_CHAN_LUMA1_IN);
1162 vpdma_add_sync_on_channel_ctd(&ctx->desc_list, VPE_CHAN_CHROMA1_IN);
1163
585e6f01
AT		 1164 if (ctx->deinterlacing) {
1165 vpdma_add_sync_on_channel_ctd(&ctx->desc_list,
1166 VPE_CHAN_LUMA2_IN);
1167 vpdma_add_sync_on_channel_ctd(&ctx->desc_list,
1168 VPE_CHAN_CHROMA2_IN);
1169
1170 vpdma_add_sync_on_channel_ctd(&ctx->desc_list,
1171 VPE_CHAN_LUMA3_IN);
1172 vpdma_add_sync_on_channel_ctd(&ctx->desc_list,
1173 VPE_CHAN_CHROMA3_IN);
1174
1175 vpdma_add_sync_on_channel_ctd(&ctx->desc_list, VPE_CHAN_MV_IN);
1176 }
1177
45719127 1178 /* sync on channel control descriptors for output ports */
30496799
AT		 1179 if (d_q_data->colorspace == V4L2_COLORSPACE_SRGB) {
1180 vpdma_add_sync_on_channel_ctd(&ctx->desc_list,
1181 VPE_CHAN_RGB_OUT);
1182 } else {
1183 vpdma_add_sync_on_channel_ctd(&ctx->desc_list,
1184 VPE_CHAN_LUMA_OUT);
1185 if (d_q_data->fmt->coplanar)
1186 vpdma_add_sync_on_channel_ctd(&ctx->desc_list,
1187 VPE_CHAN_CHROMA_OUT);
1188 }
45719127 1189
585e6f01
AT		 1190 if (ctx->deinterlacing)
1191 vpdma_add_sync_on_channel_ctd(&ctx->desc_list, VPE_CHAN_MV_OUT);
1192
45719127
AT		 1193 enable_irqs(ctx);
1194
1195 vpdma_map_desc_buf(ctx->dev->vpdma, &ctx->desc_list.buf);
1196 vpdma_submit_descs(ctx->dev->vpdma, &ctx->desc_list);
1197}
1198
585e6f01
AT		 1199static void dei_error(struct vpe_ctx *ctx)
1200{
1201 dev_warn(ctx->dev->v4l2_dev.dev,
1202 "received DEI error interrupt\n");
1203}
1204
45719127
AT		 1205static void ds1_uv_error(struct vpe_ctx *ctx)
1206{
1207 dev_warn(ctx->dev->v4l2_dev.dev,
1208 "received downsampler error interrupt\n");
1209}
1210
1211static irqreturn_t vpe_irq(int irq_vpe, void *data)
1212{
1213 struct vpe_dev *dev = (struct vpe_dev *)data;
1214 struct vpe_ctx *ctx;
585e6f01 1215 struct vpe_q_data *d_q_data;
45719127
AT		 1216 struct vb2_buffer *s_vb, *d_vb;
1217 struct v4l2_buffer *s_buf, *d_buf;
1218 unsigned long flags;
1219 u32 irqst0, irqst1;
1220
1221 irqst0 = read_reg(dev, VPE_INT0_STATUS0);
1222 if (irqst0) {
1223 write_reg(dev, VPE_INT0_STATUS0_CLR, irqst0);
1224 vpe_dbg(dev, "INT0_STATUS0 = 0x%08x\n", irqst0);
1225 }
1226
1227 irqst1 = read_reg(dev, VPE_INT0_STATUS1);
1228 if (irqst1) {
1229 write_reg(dev, VPE_INT0_STATUS1_CLR, irqst1);
1230 vpe_dbg(dev, "INT0_STATUS1 = 0x%08x\n", irqst1);
1231 }
1232
1233 ctx = v4l2_m2m_get_curr_priv(dev->m2m_dev);
1234 if (!ctx) {
1235 vpe_err(dev, "instance released before end of transaction\n");
1236 goto handled;
1237 }
1238
585e6f01
AT		 1239 if (irqst1) {
1240 if (irqst1 & VPE_DEI_ERROR_INT) {
1241 irqst1 &= ~VPE_DEI_ERROR_INT;
1242 dei_error(ctx);
1243 }
1244 if (irqst1 & VPE_DS1_UV_ERROR_INT) {
1245 irqst1 &= ~VPE_DS1_UV_ERROR_INT;
1246 ds1_uv_error(ctx);
1247 }
45719127
AT		 1248 }
1249
1250 if (irqst0) {
1251 if (irqst0 & VPE_INT0_LIST0_COMPLETE)
1252 vpdma_clear_list_stat(ctx->dev->vpdma);
1253
1254 irqst0 &= ~(VPE_INT0_LIST0_COMPLETE);
1255 }
1256
1257 if (irqst0 | irqst1) {
1258 dev_warn(dev->v4l2_dev.dev, "Unexpected interrupt: "
1259 "INT0_STATUS0 = 0x%08x, INT0_STATUS1 = 0x%08x\n",
1260 irqst0, irqst1);
1261 }
1262
1263 disable_irqs(ctx);
1264
1265 vpdma_unmap_desc_buf(dev->vpdma, &ctx->desc_list.buf);
1266 vpdma_unmap_desc_buf(dev->vpdma, &ctx->mmr_adb);
773f0657
AT		 1267 vpdma_unmap_desc_buf(dev->vpdma, &ctx->sc_coeff_h);
1268 vpdma_unmap_desc_buf(dev->vpdma, &ctx->sc_coeff_v);
45719127
AT		 1269
1270 vpdma_reset_desc_list(&ctx->desc_list);
1271
585e6f01
AT		 1272 /* the previous dst mv buffer becomes the next src mv buffer */
1273 ctx->src_mv_buf_selector = !ctx->src_mv_buf_selector;
1274
45719127
AT		 1275 if (ctx->aborting)
1276 goto finished;
1277
585e6f01 1278 s_vb = ctx->src_vbs[0];
45719127
AT		 1279 d_vb = ctx->dst_vb;
1280 s_buf = &s_vb->v4l2_buf;
1281 d_buf = &d_vb->v4l2_buf;
1282
1283 d_buf->timestamp = s_buf->timestamp;
309f4d62
SA		 1284 d_buf->flags &= ~V4L2_BUF_FLAG_TSTAMP_SRC_MASK;
1285 d_buf->flags |= s_buf->flags & V4L2_BUF_FLAG_TSTAMP_SRC_MASK;
45719127
AT		 1286 if (s_buf->flags & V4L2_BUF_FLAG_TIMECODE) {
1287 d_buf->flags |= V4L2_BUF_FLAG_TIMECODE;
1288 d_buf->timecode = s_buf->timecode;
1289 }
45719127 1290 d_buf->sequence = ctx->sequence;
585e6f01
AT		 1291 d_buf->field = ctx->field;
1292
1293 d_q_data = &ctx->q_data[Q_DATA_DST];
1294 if (d_q_data->flags & Q_DATA_INTERLACED) {
1295 if (ctx->field == V4L2_FIELD_BOTTOM) {
1296 ctx->sequence++;
1297 ctx->field = V4L2_FIELD_TOP;
1298 } else {
1299 WARN_ON(ctx->field != V4L2_FIELD_TOP);
1300 ctx->field = V4L2_FIELD_BOTTOM;
1301 }
1302 } else {
1303 ctx->sequence++;
1304 }
45719127 1305
585e6f01
AT		 1306 if (ctx->deinterlacing)
1307 s_vb = ctx->src_vbs[2];
45719127
AT		 1308
1309 spin_lock_irqsave(&dev->lock, flags);
1310 v4l2_m2m_buf_done(s_vb, VB2_BUF_STATE_DONE);
1311 v4l2_m2m_buf_done(d_vb, VB2_BUF_STATE_DONE);
1312 spin_unlock_irqrestore(&dev->lock, flags);
1313
585e6f01
AT		 1314 if (ctx->deinterlacing) {
1315 ctx->src_vbs[2] = ctx->src_vbs[1];
1316 ctx->src_vbs[1] = ctx->src_vbs[0];
1317 }
1318
45719127
AT		 1319 ctx->bufs_completed++;
1320 if (ctx->bufs_completed < ctx->bufs_per_job) {
1321 device_run(ctx);
1322 goto handled;
1323 }
1324
1325finished:
1326 vpe_dbg(ctx->dev, "finishing transaction\n");
1327 ctx->bufs_completed = 0;
1328 v4l2_m2m_job_finish(dev->m2m_dev, ctx->m2m_ctx);
1329handled:
1330 return IRQ_HANDLED;
1331}
1332
1333/*
1334 * video ioctls
1335 */
1336static int vpe_querycap(struct file *file, void *priv,
1337 struct v4l2_capability *cap)
1338{
1339 strncpy(cap->driver, VPE_MODULE_NAME, sizeof(cap->driver) - 1);
1340 strncpy(cap->card, VPE_MODULE_NAME, sizeof(cap->card) - 1);
1341 strlcpy(cap->bus_info, VPE_MODULE_NAME, sizeof(cap->bus_info));
1342 cap->device_caps = V4L2_CAP_VIDEO_M2M | V4L2_CAP_STREAMING;
1343 cap->capabilities = cap->device_caps | V4L2_CAP_DEVICE_CAPS;
1344 return 0;
1345}
1346
1347static int __enum_fmt(struct v4l2_fmtdesc *f, u32 type)
1348{
1349 int i, index;
1350 struct vpe_fmt *fmt = NULL;
1351
1352 index = 0;
1353 for (i = 0; i < ARRAY_SIZE(vpe_formats); ++i) {
1354 if (vpe_formats[i].types & type) {
1355 if (index == f->index) {
1356 fmt = &vpe_formats[i];
1357 break;
1358 }
1359 index++;
1360 }
1361 }
1362
1363 if (!fmt)
1364 return -EINVAL;
1365
1366 strncpy(f->description, fmt->name, sizeof(f->description) - 1);
1367 f->pixelformat = fmt->fourcc;
1368 return 0;
1369}
1370
1371static int vpe_enum_fmt(struct file *file, void *priv,
1372 struct v4l2_fmtdesc *f)
1373{
1374 if (V4L2_TYPE_IS_OUTPUT(f->type))
1375 return __enum_fmt(f, VPE_FMT_TYPE_OUTPUT);
1376
1377 return __enum_fmt(f, VPE_FMT_TYPE_CAPTURE);
1378}
1379
1380static int vpe_g_fmt(struct file *file, void *priv, struct v4l2_format *f)
1381{
1382 struct v4l2_pix_format_mplane *pix = &f->fmt.pix_mp;
1383 struct vpe_ctx *ctx = file2ctx(file);
1384 struct vb2_queue *vq;
1385 struct vpe_q_data *q_data;
1386 int i;
1387
1388 vq = v4l2_m2m_get_vq(ctx->m2m_ctx, f->type);
1389 if (!vq)
1390 return -EINVAL;
1391
1392 q_data = get_q_data(ctx, f->type);
1393
1394 pix->width = q_data->width;
1395 pix->height = q_data->height;
1396 pix->pixelformat = q_data->fmt->fourcc;
585e6f01 1397 pix->field = q_data->field;
45719127
AT		 1398
1399 if (V4L2_TYPE_IS_OUTPUT(f->type)) {
1400 pix->colorspace = q_data->colorspace;
1401 } else {
1402 struct vpe_q_data *s_q_data;
1403
1404 /* get colorspace from the source queue */
1405 s_q_data = get_q_data(ctx, V4L2_BUF_TYPE_VIDEO_OUTPUT_MPLANE);
1406
1407 pix->colorspace = s_q_data->colorspace;
1408 }
1409
1410 pix->num_planes = q_data->fmt->coplanar ? 2 : 1;
1411
1412 for (i = 0; i < pix->num_planes; i++) {
1413 pix->plane_fmt[i].bytesperline = q_data->bytesperline[i];
1414 pix->plane_fmt[i].sizeimage = q_data->sizeimage[i];
1415 }
1416
1417 return 0;
1418}
1419
1420static int __vpe_try_fmt(struct vpe_ctx *ctx, struct v4l2_format *f,
1421 struct vpe_fmt *fmt, int type)
1422{
1423 struct v4l2_pix_format_mplane *pix = &f->fmt.pix_mp;
1424 struct v4l2_plane_pix_format *plane_fmt;
a51cd8f5
AT		 1425 unsigned int w_align;
1426 int i, depth, depth_bytes;
45719127
AT		 1427
1428 if (!fmt || !(fmt->types & type)) {
1429 vpe_err(ctx->dev, "Fourcc format (0x%08x) invalid.\n",
1430 pix->pixelformat);
1431 return -EINVAL;
1432 }
1433
585e6f01
AT		 1434 if (pix->field != V4L2_FIELD_NONE && pix->field != V4L2_FIELD_ALTERNATE)
1435 pix->field = V4L2_FIELD_NONE;
45719127 1436
a51cd8f5
AT		 1437 depth = fmt->vpdma_fmt[VPE_LUMA]->depth;
1438
1439 /*
1440 * the line stride should 16 byte aligned for VPDMA to work, based on
1441 * the bytes per pixel, figure out how much the width should be aligned
1442 * to make sure line stride is 16 byte aligned
1443 */
1444 depth_bytes = depth >> 3;
1445
1446 if (depth_bytes == 3)
1447 /*
1448 * if bpp is 3(as in some RGB formats), the pixel width doesn't
1449 * really help in ensuring line stride is 16 byte aligned
1450 */
1451 w_align = 4;
1452 else
1453 /*
1454 * for the remainder bpp(4, 2 and 1), the pixel width alignment
1455 * can ensure a line stride alignment of 16 bytes. For example,
1456 * if bpp is 2, then the line stride can be 16 byte aligned if
1457 * the width is 8 byte aligned
1458 */
1459 w_align = order_base_2(VPDMA_DESC_ALIGN / depth_bytes);
1460
1461 v4l_bound_align_image(&pix->width, MIN_W, MAX_W, w_align,
45719127
AT		 1462 &pix->height, MIN_H, MAX_H, H_ALIGN,
1463 S_ALIGN);
1464
1465 pix->num_planes = fmt->coplanar ? 2 : 1;
1466 pix->pixelformat = fmt->fourcc;
1467
30496799
AT		 1468 if (!pix->colorspace) {
1469 if (fmt->fourcc == V4L2_PIX_FMT_RGB24 ||
1470 fmt->fourcc == V4L2_PIX_FMT_BGR24 ||
1471 fmt->fourcc == V4L2_PIX_FMT_RGB32 ||
1472 fmt->fourcc == V4L2_PIX_FMT_BGR32) {
1473 pix->colorspace = V4L2_COLORSPACE_SRGB;
1474 } else {
1475 if (pix->height > 1280) /* HD */
1476 pix->colorspace = V4L2_COLORSPACE_REC709;
1477 else /* SD */
1478 pix->colorspace = V4L2_COLORSPACE_SMPTE170M;
1479 }
45719127
AT		 1480 }
1481
1482 for (i = 0; i < pix->num_planes; i++) {
45719127
AT		 1483 plane_fmt = &pix->plane_fmt[i];
1484 depth = fmt->vpdma_fmt[i]->depth;
1485
1486 if (i == VPE_LUMA)
a51cd8f5 1487 plane_fmt->bytesperline = (pix->width * depth) >> 3;
45719127
AT		 1488 else
1489 plane_fmt->bytesperline = pix->width;
1490
1491 plane_fmt->sizeimage =
1492 (pix->height * pix->width * depth) >> 3;
1493 }
1494
1495 return 0;
1496}
1497
1498static int vpe_try_fmt(struct file *file, void *priv, struct v4l2_format *f)
1499{
1500 struct vpe_ctx *ctx = file2ctx(file);
1501 struct vpe_fmt *fmt = find_format(f);
1502
1503 if (V4L2_TYPE_IS_OUTPUT(f->type))
1504 return __vpe_try_fmt(ctx, f, fmt, VPE_FMT_TYPE_OUTPUT);
1505 else
1506 return __vpe_try_fmt(ctx, f, fmt, VPE_FMT_TYPE_CAPTURE);
1507}
1508
1509static int __vpe_s_fmt(struct vpe_ctx *ctx, struct v4l2_format *f)
1510{
1511 struct v4l2_pix_format_mplane *pix = &f->fmt.pix_mp;
1512 struct v4l2_plane_pix_format *plane_fmt;
1513 struct vpe_q_data *q_data;
1514 struct vb2_queue *vq;
1515 int i;
1516
1517 vq = v4l2_m2m_get_vq(ctx->m2m_ctx, f->type);
1518 if (!vq)
1519 return -EINVAL;
1520
1521 if (vb2_is_busy(vq)) {
1522 vpe_err(ctx->dev, "queue busy\n");
1523 return -EBUSY;
1524 }
1525
1526 q_data = get_q_data(ctx, f->type);
1527 if (!q_data)
1528 return -EINVAL;
1529
1530 q_data->fmt = find_format(f);
1531 q_data->width = pix->width;
1532 q_data->height = pix->height;
1533 q_data->colorspace = pix->colorspace;
585e6f01 1534 q_data->field = pix->field;
45719127
AT		 1535
1536 for (i = 0; i < pix->num_planes; i++) {
1537 plane_fmt = &pix->plane_fmt[i];
1538
1539 q_data->bytesperline[i] = plane_fmt->bytesperline;
1540 q_data->sizeimage[i] = plane_fmt->sizeimage;
1541 }
1542
1543 q_data->c_rect.left = 0;
1544 q_data->c_rect.top = 0;
1545 q_data->c_rect.width = q_data->width;
1546 q_data->c_rect.height = q_data->height;
1547
585e6f01
AT		 1548 if (q_data->field == V4L2_FIELD_ALTERNATE)
1549 q_data->flags |= Q_DATA_INTERLACED;
1550 else
1551 q_data->flags &= ~Q_DATA_INTERLACED;
1552
45719127
AT		 1553 vpe_dbg(ctx->dev, "Setting format for type %d, wxh: %dx%d, fmt: %d bpl_y %d",
1554 f->type, q_data->width, q_data->height, q_data->fmt->fourcc,
1555 q_data->bytesperline[VPE_LUMA]);
1556 if (q_data->fmt->coplanar)
1557 vpe_dbg(ctx->dev, " bpl_uv %d\n",
1558 q_data->bytesperline[VPE_CHROMA]);
1559
1560 return 0;
1561}
1562
1563static int vpe_s_fmt(struct file *file, void *priv, struct v4l2_format *f)
1564{
1565 int ret;
1566 struct vpe_ctx *ctx = file2ctx(file);
1567
1568 ret = vpe_try_fmt(file, priv, f);
1569 if (ret)
1570 return ret;
1571
1572 ret = __vpe_s_fmt(ctx, f);
1573 if (ret)
1574 return ret;
1575
1576 if (V4L2_TYPE_IS_OUTPUT(f->type))
1577 set_src_registers(ctx);
1578 else
1579 set_dst_registers(ctx);
1580
1581 return set_srcdst_params(ctx);
1582}
1583
1584static int vpe_reqbufs(struct file *file, void *priv,
1585 struct v4l2_requestbuffers *reqbufs)
1586{
1587 struct vpe_ctx *ctx = file2ctx(file);
1588
1589 return v4l2_m2m_reqbufs(file, ctx->m2m_ctx, reqbufs);
1590}
1591
1592static int vpe_querybuf(struct file *file, void *priv, struct v4l2_buffer *buf)
1593{
1594 struct vpe_ctx *ctx = file2ctx(file);
1595
1596 return v4l2_m2m_querybuf(file, ctx->m2m_ctx, buf);
1597}
1598
1599static int vpe_qbuf(struct file *file, void *priv, struct v4l2_buffer *buf)
1600{
1601 struct vpe_ctx *ctx = file2ctx(file);
1602
1603 return v4l2_m2m_qbuf(file, ctx->m2m_ctx, buf);
1604}
1605
1606static int vpe_dqbuf(struct file *file, void *priv, struct v4l2_buffer *buf)
1607{
1608 struct vpe_ctx *ctx = file2ctx(file);
1609
1610 return v4l2_m2m_dqbuf(file, ctx->m2m_ctx, buf);
1611}
1612
1613static int vpe_streamon(struct file *file, void *priv, enum v4l2_buf_type type)
1614{
1615 struct vpe_ctx *ctx = file2ctx(file);
1616
1617 return v4l2_m2m_streamon(file, ctx->m2m_ctx, type);
1618}
1619
1620static int vpe_streamoff(struct file *file, void *priv, enum v4l2_buf_type type)
1621{
1622 struct vpe_ctx *ctx = file2ctx(file);
1623
1624 vpe_dump_regs(ctx->dev);
1625 vpdma_dump_regs(ctx->dev->vpdma);
1626
1627 return v4l2_m2m_streamoff(file, ctx->m2m_ctx, type);
1628}
1629
1630/*
1631 * defines number of buffers/frames a context can process with VPE before
1632 * switching to a different context. default value is 1 buffer per context
1633 */
1634#define V4L2_CID_VPE_BUFS_PER_JOB (V4L2_CID_USER_TI_VPE_BASE + 0)
1635
1636static int vpe_s_ctrl(struct v4l2_ctrl *ctrl)
1637{
1638 struct vpe_ctx *ctx =
1639 container_of(ctrl->handler, struct vpe_ctx, hdl);
1640
1641 switch (ctrl->id) {
1642 case V4L2_CID_VPE_BUFS_PER_JOB:
1643 ctx->bufs_per_job = ctrl->val;
1644 break;
1645
1646 default:
1647 vpe_err(ctx->dev, "Invalid control\n");
1648 return -EINVAL;
1649 }
1650
1651 return 0;
1652}
1653
1654static const struct v4l2_ctrl_ops vpe_ctrl_ops = {
1655 .s_ctrl = vpe_s_ctrl,
1656};
1657
1658static const struct v4l2_ioctl_ops vpe_ioctl_ops = {
1659 .vidioc_querycap = vpe_querycap,
1660
1661 .vidioc_enum_fmt_vid_cap_mplane = vpe_enum_fmt,
1662 .vidioc_g_fmt_vid_cap_mplane = vpe_g_fmt,
1663 .vidioc_try_fmt_vid_cap_mplane = vpe_try_fmt,
1664 .vidioc_s_fmt_vid_cap_mplane = vpe_s_fmt,
1665
1666 .vidioc_enum_fmt_vid_out_mplane = vpe_enum_fmt,
1667 .vidioc_g_fmt_vid_out_mplane = vpe_g_fmt,
1668 .vidioc_try_fmt_vid_out_mplane = vpe_try_fmt,
1669 .vidioc_s_fmt_vid_out_mplane = vpe_s_fmt,
1670
1671 .vidioc_reqbufs = vpe_reqbufs,
1672 .vidioc_querybuf = vpe_querybuf,
1673
1674 .vidioc_qbuf = vpe_qbuf,
1675 .vidioc_dqbuf = vpe_dqbuf,
1676
1677 .vidioc_streamon = vpe_streamon,
1678 .vidioc_streamoff = vpe_streamoff,
1679 .vidioc_subscribe_event = v4l2_ctrl_subscribe_event,
1680 .vidioc_unsubscribe_event = v4l2_event_unsubscribe,
1681};
1682
1683/*
1684 * Queue operations
1685 */
1686static int vpe_queue_setup(struct vb2_queue *vq,
1687 const struct v4l2_format *fmt,
1688 unsigned int *nbuffers, unsigned int *nplanes,
1689 unsigned int sizes[], void *alloc_ctxs[])
1690{
1691 int i;
1692 struct vpe_ctx *ctx = vb2_get_drv_priv(vq);
1693 struct vpe_q_data *q_data;
1694
1695 q_data = get_q_data(ctx, vq->type);
1696
1697 *nplanes = q_data->fmt->coplanar ? 2 : 1;
1698
1699 for (i = 0; i < *nplanes; i++) {
1700 sizes[i] = q_data->sizeimage[i];
1701 alloc_ctxs[i] = ctx->dev->alloc_ctx;
1702 }
1703
1704 vpe_dbg(ctx->dev, "get %d buffer(s) of size %d", *nbuffers,
1705 sizes[VPE_LUMA]);
1706 if (q_data->fmt->coplanar)
1707 vpe_dbg(ctx->dev, " and %d\n", sizes[VPE_CHROMA]);
1708
1709 return 0;
1710}
1711
1712static int vpe_buf_prepare(struct vb2_buffer *vb)
1713{
1714 struct vpe_ctx *ctx = vb2_get_drv_priv(vb->vb2_queue);
1715 struct vpe_q_data *q_data;
1716 int i, num_planes;
1717
1718 vpe_dbg(ctx->dev, "type: %d\n", vb->vb2_queue->type);
1719
1720 q_data = get_q_data(ctx, vb->vb2_queue->type);
1721 num_planes = q_data->fmt->coplanar ? 2 : 1;
1722
1723 for (i = 0; i < num_planes; i++) {
1724 if (vb2_plane_size(vb, i) < q_data->sizeimage[i]) {
1725 vpe_err(ctx->dev,
1726 "data will not fit into plane (%lu < %lu)\n",
1727 vb2_plane_size(vb, i),
1728 (long) q_data->sizeimage[i]);
1729 return -EINVAL;
1730 }
1731 }
1732
1733 for (i = 0; i < num_planes; i++)
1734 vb2_set_plane_payload(vb, i, q_data->sizeimage[i]);
1735
1736 return 0;
1737}
1738
1739static void vpe_buf_queue(struct vb2_buffer *vb)
1740{
1741 struct vpe_ctx *ctx = vb2_get_drv_priv(vb->vb2_queue);
1742 v4l2_m2m_buf_queue(ctx->m2m_ctx, vb);
1743}
1744
1745static void vpe_wait_prepare(struct vb2_queue *q)
1746{
1747 struct vpe_ctx *ctx = vb2_get_drv_priv(q);
1748 vpe_unlock(ctx);
1749}
1750
1751static void vpe_wait_finish(struct vb2_queue *q)
1752{
1753 struct vpe_ctx *ctx = vb2_get_drv_priv(q);
1754 vpe_lock(ctx);
1755}
1756
1757static struct vb2_ops vpe_qops = {
1758 .queue_setup = vpe_queue_setup,
1759 .buf_prepare = vpe_buf_prepare,
1760 .buf_queue = vpe_buf_queue,
1761 .wait_prepare = vpe_wait_prepare,
1762 .wait_finish = vpe_wait_finish,
1763};
1764
1765static int queue_init(void *priv, struct vb2_queue *src_vq,
1766 struct vb2_queue *dst_vq)
1767{
1768 struct vpe_ctx *ctx = priv;
1769 int ret;
1770
1771 memset(src_vq, 0, sizeof(*src_vq));
1772 src_vq->type = V4L2_BUF_TYPE_VIDEO_OUTPUT_MPLANE;
1773 src_vq->io_modes = VB2_MMAP;
1774 src_vq->drv_priv = ctx;
1775 src_vq->buf_struct_size = sizeof(struct v4l2_m2m_buffer);
1776 src_vq->ops = &vpe_qops;
1777 src_vq->mem_ops = &vb2_dma_contig_memops;
ade48681 1778 src_vq->timestamp_flags = V4L2_BUF_FLAG_TIMESTAMP_COPY;
45719127
AT		 1779
1780 ret = vb2_queue_init(src_vq);
1781 if (ret)
1782 return ret;
1783
1784 memset(dst_vq, 0, sizeof(*dst_vq));
1785 dst_vq->type = V4L2_BUF_TYPE_VIDEO_CAPTURE_MPLANE;
1786 dst_vq->io_modes = VB2_MMAP;
1787 dst_vq->drv_priv = ctx;
1788 dst_vq->buf_struct_size = sizeof(struct v4l2_m2m_buffer);
1789 dst_vq->ops = &vpe_qops;
1790 dst_vq->mem_ops = &vb2_dma_contig_memops;
ade48681 1791 dst_vq->timestamp_flags = V4L2_BUF_FLAG_TIMESTAMP_COPY;
45719127
AT		 1792
1793 return vb2_queue_init(dst_vq);
1794}
1795
1796static const struct v4l2_ctrl_config vpe_bufs_per_job = {
1797 .ops = &vpe_ctrl_ops,
1798 .id = V4L2_CID_VPE_BUFS_PER_JOB,
1799 .name = "Buffers Per Transaction",
1800 .type = V4L2_CTRL_TYPE_INTEGER,
1801 .def = VPE_DEF_BUFS_PER_JOB,
1802 .min = 1,
1803 .max = VIDEO_MAX_FRAME,
1804 .step = 1,
1805};
1806
1807/*
1808 * File operations
1809 */
1810static int vpe_open(struct file *file)
1811{
1812 struct vpe_dev *dev = video_drvdata(file);
1813 struct vpe_ctx *ctx = NULL;
1814 struct vpe_q_data *s_q_data;
1815 struct v4l2_ctrl_handler *hdl;
1816 int ret;
1817
1818 vpe_dbg(dev, "vpe_open\n");
1819
1820 if (!dev->vpdma->ready) {
1821 vpe_err(dev, "vpdma firmware not loaded\n");
1822 return -ENODEV;
1823 }
1824
1825 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1826 if (!ctx)
1827 return -ENOMEM;
1828
1829 ctx->dev = dev;
1830
1831 if (mutex_lock_interruptible(&dev->dev_mutex)) {
1832 ret = -ERESTARTSYS;
1833 goto free_ctx;
1834 }
1835
1836 ret = vpdma_create_desc_list(&ctx->desc_list, VPE_DESC_LIST_SIZE,
1837 VPDMA_LIST_TYPE_NORMAL);
1838 if (ret != 0)
1839 goto unlock;
1840
1841 ret = vpdma_alloc_desc_buf(&ctx->mmr_adb, sizeof(struct vpe_mmr_adb));
1842 if (ret != 0)
1843 goto free_desc_list;
1844
773f0657
AT		 1845 ret = vpdma_alloc_desc_buf(&ctx->sc_coeff_h, SC_COEF_SRAM_SIZE);
1846 if (ret != 0)
1847 goto free_mmr_adb;
1848
1849 ret = vpdma_alloc_desc_buf(&ctx->sc_coeff_v, SC_COEF_SRAM_SIZE);
1850 if (ret != 0)
1851 goto free_sc_h;
1852
45719127
AT		 1853 init_adb_hdrs(ctx);
1854
1855 v4l2_fh_init(&ctx->fh, video_devdata(file));
1856 file->private_data = &ctx->fh;
1857
1858 hdl = &ctx->hdl;
1859 v4l2_ctrl_handler_init(hdl, 1);
1860 v4l2_ctrl_new_custom(hdl, &vpe_bufs_per_job, NULL);
1861 if (hdl->error) {
1862 ret = hdl->error;
1863 goto exit_fh;
1864 }
1865 ctx->fh.ctrl_handler = hdl;
1866 v4l2_ctrl_handler_setup(hdl);
1867
1868 s_q_data = &ctx->q_data[Q_DATA_SRC];
1869 s_q_data->fmt = &vpe_formats[2];
1870 s_q_data->width = 1920;
1871 s_q_data->height = 1080;
1872 s_q_data->sizeimage[VPE_LUMA] = (s_q_data->width * s_q_data->height *
1873 s_q_data->fmt->vpdma_fmt[VPE_LUMA]->depth) >> 3;
30496799 1874 s_q_data->colorspace = V4L2_COLORSPACE_SMPTE170M;
585e6f01 1875 s_q_data->field = V4L2_FIELD_NONE;
45719127
AT		 1876 s_q_data->c_rect.left = 0;
1877 s_q_data->c_rect.top = 0;
1878 s_q_data->c_rect.width = s_q_data->width;
1879 s_q_data->c_rect.height = s_q_data->height;
1880 s_q_data->flags = 0;
1881
1882 ctx->q_data[Q_DATA_DST] = *s_q_data;
1883
585e6f01 1884 set_dei_shadow_registers(ctx);
45719127
AT		 1885 set_src_registers(ctx);
1886 set_dst_registers(ctx);
1887 ret = set_srcdst_params(ctx);
1888 if (ret)
1889 goto exit_fh;
1890
1891 ctx->m2m_ctx = v4l2_m2m_ctx_init(dev->m2m_dev, ctx, &queue_init);
1892
1893 if (IS_ERR(ctx->m2m_ctx)) {
1894 ret = PTR_ERR(ctx->m2m_ctx);
1895 goto exit_fh;
1896 }
1897
1898 v4l2_fh_add(&ctx->fh);
1899
1900 /*
1901 * for now, just report the creation of the first instance, we can later
1902 * optimize the driver to enable or disable clocks when the first
1903 * instance is created or the last instance released
1904 */
1905 if (atomic_inc_return(&dev->num_instances) == 1)
1906 vpe_dbg(dev, "first instance created\n");
1907
1908 ctx->bufs_per_job = VPE_DEF_BUFS_PER_JOB;
1909
1910 ctx->load_mmrs = true;
1911
1912 vpe_dbg(dev, "created instance %p, m2m_ctx: %p\n",
1913 ctx, ctx->m2m_ctx);
1914
1915 mutex_unlock(&dev->dev_mutex);
1916
1917 return 0;
1918exit_fh:
1919 v4l2_ctrl_handler_free(hdl);
1920 v4l2_fh_exit(&ctx->fh);
773f0657
AT		 1921 vpdma_free_desc_buf(&ctx->sc_coeff_v);
1922free_sc_h:
1923 vpdma_free_desc_buf(&ctx->sc_coeff_h);
1924free_mmr_adb:
45719127
AT		 1925 vpdma_free_desc_buf(&ctx->mmr_adb);
1926free_desc_list:
1927 vpdma_free_desc_list(&ctx->desc_list);
1928unlock:
1929 mutex_unlock(&dev->dev_mutex);
1930free_ctx:
1931 kfree(ctx);
1932 return ret;
1933}
1934
1935static int vpe_release(struct file *file)
1936{
1937 struct vpe_dev *dev = video_drvdata(file);
1938 struct vpe_ctx *ctx = file2ctx(file);
1939
1940 vpe_dbg(dev, "releasing instance %p\n", ctx);
1941
1942 mutex_lock(&dev->dev_mutex);
585e6f01
AT		 1943 free_vbs(ctx);
1944 free_mv_buffers(ctx);
45719127
AT		 1945 vpdma_free_desc_list(&ctx->desc_list);
1946 vpdma_free_desc_buf(&ctx->mmr_adb);
1947
1948 v4l2_fh_del(&ctx->fh);
1949 v4l2_fh_exit(&ctx->fh);
1950 v4l2_ctrl_handler_free(&ctx->hdl);
1951 v4l2_m2m_ctx_release(ctx->m2m_ctx);
1952
1953 kfree(ctx);
1954
1955 /*
1956 * for now, just report the release of the last instance, we can later
1957 * optimize the driver to enable or disable clocks when the first
1958 * instance is created or the last instance released
1959 */
1960 if (atomic_dec_return(&dev->num_instances) == 0)
1961 vpe_dbg(dev, "last instance released\n");
1962
1963 mutex_unlock(&dev->dev_mutex);
1964
1965 return 0;
1966}
1967
1968static unsigned int vpe_poll(struct file *file,
1969 struct poll_table_struct *wait)
1970{
1971 struct vpe_ctx *ctx = file2ctx(file);
1972 struct vpe_dev *dev = ctx->dev;
1973 int ret;
1974
1975 mutex_lock(&dev->dev_mutex);
1976 ret = v4l2_m2m_poll(file, ctx->m2m_ctx, wait);
1977 mutex_unlock(&dev->dev_mutex);
1978 return ret;
1979}
1980
1981static int vpe_mmap(struct file *file, struct vm_area_struct *vma)
1982{
1983 struct vpe_ctx *ctx = file2ctx(file);
1984 struct vpe_dev *dev = ctx->dev;
1985 int ret;
1986
1987 if (mutex_lock_interruptible(&dev->dev_mutex))
1988 return -ERESTARTSYS;
1989 ret = v4l2_m2m_mmap(file, ctx->m2m_ctx, vma);
1990 mutex_unlock(&dev->dev_mutex);
1991 return ret;
1992}
1993
1994static const struct v4l2_file_operations vpe_fops = {
1995 .owner = THIS_MODULE,
1996 .open = vpe_open,
1997 .release = vpe_release,
1998 .poll = vpe_poll,
1999 .unlocked_ioctl = video_ioctl2,
2000 .mmap = vpe_mmap,
2001};
2002
2003static struct video_device vpe_videodev = {
2004 .name = VPE_MODULE_NAME,
2005 .fops = &vpe_fops,
2006 .ioctl_ops = &vpe_ioctl_ops,
2007 .minor = -1,
2008 .release = video_device_release,
2009 .vfl_dir = VFL_DIR_M2M,
2010};
2011
2012static struct v4l2_m2m_ops m2m_ops = {
2013 .device_run = device_run,
2014 .job_ready = job_ready,
2015 .job_abort = job_abort,
2016 .lock = vpe_lock,
2017 .unlock = vpe_unlock,
2018};
2019
2020static int vpe_runtime_get(struct platform_device *pdev)
2021{
2022 int r;
2023
2024 dev_dbg(&pdev->dev, "vpe_runtime_get\n");
2025
2026 r = pm_runtime_get_sync(&pdev->dev);
2027 WARN_ON(r < 0);
2028 return r < 0 ? r : 0;
2029}
2030
2031static void vpe_runtime_put(struct platform_device *pdev)
2032{
2033
2034 int r;
2035
2036 dev_dbg(&pdev->dev, "vpe_runtime_put\n");
2037
2038 r = pm_runtime_put_sync(&pdev->dev);
2039 WARN_ON(r < 0 && r != -ENOSYS);
2040}
2041
2042static int vpe_probe(struct platform_device *pdev)
2043{
2044 struct vpe_dev *dev;
2045 struct video_device *vfd;
45719127
AT		 2046 int ret, irq, func;
2047
2048 dev = devm_kzalloc(&pdev->dev, sizeof(*dev), GFP_KERNEL);
b68231a1
WY		 2049 if (!dev)
2050 return -ENOMEM;
45719127
AT		 2051
2052 spin_lock_init(&dev->lock);
2053
2054 ret = v4l2_device_register(&pdev->dev, &dev->v4l2_dev);
2055 if (ret)
2056 return ret;
2057
2058 atomic_set(&dev->num_instances, 0);
2059 mutex_init(&dev->dev_mutex);
2060
44687b2e
AT		 2061 dev->res = platform_get_resource_byname(pdev, IORESOURCE_MEM,
2062 "vpe_top");
45719127
AT		 2063 /*
2064 * HACK: we get resource info from device tree in the form of a list of
2065 * VPE sub blocks, the driver currently uses only the base of vpe_top
2066 * for register access, the driver should be changed later to access
2067 * registers based on the sub block base addresses
2068 */
44687b2e 2069 dev->base = devm_ioremap(&pdev->dev, dev->res->start, SZ_32K);
b68231a1
WY		 2070 if (!dev->base) {
2071 ret = -ENOMEM;
45719127
AT		 2072 goto v4l2_dev_unreg;
2073 }
2074
2075 irq = platform_get_irq(pdev, 0);
2076 ret = devm_request_irq(&pdev->dev, irq, vpe_irq, 0, VPE_MODULE_NAME,
2077 dev);
2078 if (ret)
2079 goto v4l2_dev_unreg;
2080
2081 platform_set_drvdata(pdev, dev);
2082
2083 dev->alloc_ctx = vb2_dma_contig_init_ctx(&pdev->dev);
2084 if (IS_ERR(dev->alloc_ctx)) {
2085 vpe_err(dev, "Failed to alloc vb2 context\n");
2086 ret = PTR_ERR(dev->alloc_ctx);
2087 goto v4l2_dev_unreg;
2088 }
2089
2090 dev->m2m_dev = v4l2_m2m_init(&m2m_ops);
2091 if (IS_ERR(dev->m2m_dev)) {
2092 vpe_err(dev, "Failed to init mem2mem device\n");
2093 ret = PTR_ERR(dev->m2m_dev);
2094 goto rel_ctx;
2095 }
2096
2097 pm_runtime_enable(&pdev->dev);
2098
2099 ret = vpe_runtime_get(pdev);
2100 if (ret)
2101 goto rel_m2m;
2102
2103 /* Perform clk enable followed by reset */
2104 vpe_set_clock_enable(dev, 1);
2105
2106 vpe_top_reset(dev);
2107
2108 func = read_field_reg(dev, VPE_PID, VPE_PID_FUNC_MASK,
2109 VPE_PID_FUNC_SHIFT);
2110 vpe_dbg(dev, "VPE PID function %x\n", func);
2111
2112 vpe_top_vpdma_reset(dev);
2113
44687b2e
AT		 2114 dev->sc = sc_create(pdev);
2115 if (IS_ERR(dev->sc)) {
2116 ret = PTR_ERR(dev->sc);
2117 goto runtime_put;
2118 }
2119
6948082d
AT		 2120 dev->csc = csc_create(pdev);
2121 if (IS_ERR(dev->csc)) {
2122 ret = PTR_ERR(dev->csc);
2123 goto runtime_put;
2124 }
2125
45719127 2126 dev->vpdma = vpdma_create(pdev);
6676cafe
WY		 2127 if (IS_ERR(dev->vpdma)) {
2128 ret = PTR_ERR(dev->vpdma);
45719127 2129 goto runtime_put;
6676cafe 2130 }
45719127
AT		 2131
2132 vfd = &dev->vfd;
2133 *vfd = vpe_videodev;
2134 vfd->lock = &dev->dev_mutex;
2135 vfd->v4l2_dev = &dev->v4l2_dev;
2136
2137 ret = video_register_device(vfd, VFL_TYPE_GRABBER, 0);
2138 if (ret) {
2139 vpe_err(dev, "Failed to register video device\n");
2140 goto runtime_put;
2141 }
2142
2143 video_set_drvdata(vfd, dev);
2144 snprintf(vfd->name, sizeof(vfd->name), "%s", vpe_videodev.name);
2145 dev_info(dev->v4l2_dev.dev, "Device registered as /dev/video%d\n",
2146 vfd->num);
2147
2148 return 0;
2149
2150runtime_put:
2151 vpe_runtime_put(pdev);
2152rel_m2m:
2153 pm_runtime_disable(&pdev->dev);
2154 v4l2_m2m_release(dev->m2m_dev);
2155rel_ctx:
2156 vb2_dma_contig_cleanup_ctx(dev->alloc_ctx);
2157v4l2_dev_unreg:
2158 v4l2_device_unregister(&dev->v4l2_dev);
2159
2160 return ret;
2161}
2162
2163static int vpe_remove(struct platform_device *pdev)
2164{
2165 struct vpe_dev *dev =
2166 (struct vpe_dev *) platform_get_drvdata(pdev);
2167
2168 v4l2_info(&dev->v4l2_dev, "Removing " VPE_MODULE_NAME);
2169
2170 v4l2_m2m_release(dev->m2m_dev);
2171 video_unregister_device(&dev->vfd);
2172 v4l2_device_unregister(&dev->v4l2_dev);
2173 vb2_dma_contig_cleanup_ctx(dev->alloc_ctx);
2174
2175 vpe_set_clock_enable(dev, 0);
2176 vpe_runtime_put(pdev);
2177 pm_runtime_disable(&pdev->dev);
2178
2179 return 0;
2180}
2181
2182#if defined(CONFIG_OF)
2183static const struct of_device_id vpe_of_match[] = {
2184 {
2185 .compatible = "ti,vpe",
2186 },
2187 {},
2188};
2189#else
2190#define vpe_of_match NULL
2191#endif
2192
2193static struct platform_driver vpe_pdrv = {
2194 .probe = vpe_probe,
2195 .remove = vpe_remove,
2196 .driver = {
2197 .name = VPE_MODULE_NAME,
2198 .owner = THIS_MODULE,
2199 .of_match_table = vpe_of_match,
2200 },
2201};
2202
903cbb83 2203module_platform_driver(vpe_pdrv);
45719127
AT		 2204
2205MODULE_DESCRIPTION("TI VPE driver");
2206MODULE_AUTHOR("Dale Farnsworth, <dale@farnsworth.org>");
2207MODULE_LICENSE("GPL");
Linux 6.x block layer and io_uring tree(s)