1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (c) 2014, The Linux Foundation. All rights reserved.
4  * Copyright (C) 2013 Red Hat
5  * Author: Rob Clark <robdclark@gmail.com>
6  */
7 
8 #include <drm/drm_util.h>
9 
10 #include "mdp5_kms.h"
11 #include "mdp5_smp.h"
12 
13 
14 struct mdp5_smp {
15 	struct drm_device *dev;
16 
17 	uint8_t reserved[MAX_CLIENTS]; /* fixed MMBs allocation per client */
18 
19 	int blk_cnt;
20 	int blk_size;
21 
22 	/* register cache */
23 	u32 alloc_w[22];
24 	u32 alloc_r[22];
25 	u32 pipe_reqprio_fifo_wm0[SSPP_MAX];
26 	u32 pipe_reqprio_fifo_wm1[SSPP_MAX];
27 	u32 pipe_reqprio_fifo_wm2[SSPP_MAX];
28 };
29 
30 static inline
31 struct mdp5_kms *get_kms(struct mdp5_smp *smp)
32 {
33 	struct msm_drm_private *priv = smp->dev->dev_private;
34 
35 	return to_mdp5_kms(to_mdp_kms(priv->kms));
36 }
37 
38 static inline u32 pipe2client(enum mdp5_pipe pipe, int plane)
39 {
40 #define CID_UNUSED	0
41 
42 	if (WARN_ON(plane >= pipe2nclients(pipe)))
43 		return CID_UNUSED;
44 
45 	/*
46 	 * Note on SMP clients:
47 	 * For ViG pipes, fetch Y/Cr/Cb-components clients are always
48 	 * consecutive, and in that order.
49 	 *
50 	 * e.g.:
51 	 * if mdp5_cfg->smp.clients[SSPP_VIG0] = N,
52 	 *	Y  plane's client ID is N
53 	 *	Cr plane's client ID is N + 1
54 	 *	Cb plane's client ID is N + 2
55 	 */
56 
57 	return mdp5_cfg->smp.clients[pipe] + plane;
58 }
59 
60 /* allocate blocks for the specified request: */
61 static int smp_request_block(struct mdp5_smp *smp,
62 		struct mdp5_smp_state *state,
63 		u32 cid, int nblks)
64 {
65 	void *cs = state->client_state[cid];
66 	int i, avail, cnt = smp->blk_cnt;
67 	uint8_t reserved;
68 
69 	/* we shouldn't be requesting blocks for an in-use client: */
70 	WARN_ON(bitmap_weight(cs, cnt) > 0);
71 
72 	reserved = smp->reserved[cid];
73 
74 	if (reserved) {
75 		nblks = max(0, nblks - reserved);
76 		DBG("%d MMBs allocated (%d reserved)", nblks, reserved);
77 	}
78 
79 	avail = cnt - bitmap_weight(state->state, cnt);
80 	if (nblks > avail) {
81 		DRM_DEV_ERROR(smp->dev->dev, "out of blks (req=%d > avail=%d)\n",
82 				nblks, avail);
83 		return -ENOSPC;
84 	}
85 
86 	for (i = 0; i < nblks; i++) {
87 		int blk = find_first_zero_bit(state->state, cnt);
88 		set_bit(blk, cs);
89 		set_bit(blk, state->state);
90 	}
91 
92 	return 0;
93 }
94 
95 static void set_fifo_thresholds(struct mdp5_smp *smp,
96 		enum mdp5_pipe pipe, int nblks)
97 {
98 	u32 smp_entries_per_blk = smp->blk_size / (128 / BITS_PER_BYTE);
99 	u32 val;
100 
101 	/* 1/4 of SMP pool that is being fetched */
102 	val = (nblks * smp_entries_per_blk) / 4;
103 
104 	smp->pipe_reqprio_fifo_wm0[pipe] = val * 1;
105 	smp->pipe_reqprio_fifo_wm1[pipe] = val * 2;
106 	smp->pipe_reqprio_fifo_wm2[pipe] = val * 3;
107 }
108 
109 /*
110  * NOTE: looks like if horizontal decimation is used (if we supported that)
111  * then the width used to calculate SMP block requirements is the post-
112  * decimated width.  Ie. SMP buffering sits downstream of decimation (which
113  * presumably happens during the dma from scanout buffer).
114  */
115 uint32_t mdp5_smp_calculate(struct mdp5_smp *smp,
116 		const struct mdp_format *format,
117 		u32 width, bool hdecim)
118 {
119 	struct mdp5_kms *mdp5_kms = get_kms(smp);
120 	int rev = mdp5_cfg_get_hw_rev(mdp5_kms->cfg);
121 	int i, hsub, nplanes, nlines;
122 	u32 fmt = format->base.pixel_format;
123 	uint32_t blkcfg = 0;
124 
125 	nplanes = drm_format_num_planes(fmt);
126 	hsub = drm_format_horz_chroma_subsampling(fmt);
127 
128 	/* different if BWC (compressed framebuffer?) enabled: */
129 	nlines = 2;
130 
131 	/* Newer MDPs have split/packing logic, which fetches sub-sampled
132 	 * U and V components (splits them from Y if necessary) and packs
133 	 * them together, writes to SMP using a single client.
134 	 */
135 	if ((rev > 0) && (format->chroma_sample > CHROMA_FULL)) {
136 		fmt = DRM_FORMAT_NV24;
137 		nplanes = 2;
138 
139 		/* if decimation is enabled, HW decimates less on the
140 		 * sub sampled chroma components
141 		 */
142 		if (hdecim && (hsub > 1))
143 			hsub = 1;
144 	}
145 
146 	for (i = 0; i < nplanes; i++) {
147 		int n, fetch_stride, cpp;
148 
149 		cpp = drm_format_plane_cpp(fmt, i);
150 		fetch_stride = width * cpp / (i ? hsub : 1);
151 
152 		n = DIV_ROUND_UP(fetch_stride * nlines, smp->blk_size);
153 
154 		/* for hw rev v1.00 */
155 		if (rev == 0)
156 			n = roundup_pow_of_two(n);
157 
158 		blkcfg |= (n << (8 * i));
159 	}
160 
161 	return blkcfg;
162 }
163 
164 int mdp5_smp_assign(struct mdp5_smp *smp, struct mdp5_smp_state *state,
165 		enum mdp5_pipe pipe, uint32_t blkcfg)
166 {
167 	struct mdp5_kms *mdp5_kms = get_kms(smp);
168 	struct drm_device *dev = mdp5_kms->dev;
169 	int i, ret;
170 
171 	for (i = 0; i < pipe2nclients(pipe); i++) {
172 		u32 cid = pipe2client(pipe, i);
173 		int n = blkcfg & 0xff;
174 
175 		if (!n)
176 			continue;
177 
178 		DBG("%s[%d]: request %d SMP blocks", pipe2name(pipe), i, n);
179 		ret = smp_request_block(smp, state, cid, n);
180 		if (ret) {
181 			DRM_DEV_ERROR(dev->dev, "Cannot allocate %d SMP blocks: %d\n",
182 					n, ret);
183 			return ret;
184 		}
185 
186 		blkcfg >>= 8;
187 	}
188 
189 	state->assigned |= (1 << pipe);
190 
191 	return 0;
192 }
193 
194 /* Release SMP blocks for all clients of the pipe */
195 void mdp5_smp_release(struct mdp5_smp *smp, struct mdp5_smp_state *state,
196 		enum mdp5_pipe pipe)
197 {
198 	int i;
199 	int cnt = smp->blk_cnt;
200 
201 	for (i = 0; i < pipe2nclients(pipe); i++) {
202 		u32 cid = pipe2client(pipe, i);
203 		void *cs = state->client_state[cid];
204 
205 		/* update global state: */
206 		bitmap_andnot(state->state, state->state, cs, cnt);
207 
208 		/* clear client's state */
209 		bitmap_zero(cs, cnt);
210 	}
211 
212 	state->released |= (1 << pipe);
213 }
214 
215 /* NOTE: SMP_ALLOC_* regs are *not* double buffered, so release has to
216  * happen after scanout completes.
217  */
218 static unsigned update_smp_state(struct mdp5_smp *smp,
219 		u32 cid, mdp5_smp_state_t *assigned)
220 {
221 	int cnt = smp->blk_cnt;
222 	unsigned nblks = 0;
223 	u32 blk, val;
224 
225 	for_each_set_bit(blk, *assigned, cnt) {
226 		int idx = blk / 3;
227 		int fld = blk % 3;
228 
229 		val = smp->alloc_w[idx];
230 
231 		switch (fld) {
232 		case 0:
233 			val &= ~MDP5_SMP_ALLOC_W_REG_CLIENT0__MASK;
234 			val |= MDP5_SMP_ALLOC_W_REG_CLIENT0(cid);
235 			break;
236 		case 1:
237 			val &= ~MDP5_SMP_ALLOC_W_REG_CLIENT1__MASK;
238 			val |= MDP5_SMP_ALLOC_W_REG_CLIENT1(cid);
239 			break;
240 		case 2:
241 			val &= ~MDP5_SMP_ALLOC_W_REG_CLIENT2__MASK;
242 			val |= MDP5_SMP_ALLOC_W_REG_CLIENT2(cid);
243 			break;
244 		}
245 
246 		smp->alloc_w[idx] = val;
247 		smp->alloc_r[idx] = val;
248 
249 		nblks++;
250 	}
251 
252 	return nblks;
253 }
254 
255 static void write_smp_alloc_regs(struct mdp5_smp *smp)
256 {
257 	struct mdp5_kms *mdp5_kms = get_kms(smp);
258 	int i, num_regs;
259 
260 	num_regs = smp->blk_cnt / 3 + 1;
261 
262 	for (i = 0; i < num_regs; i++) {
263 		mdp5_write(mdp5_kms, REG_MDP5_SMP_ALLOC_W_REG(i),
264 			   smp->alloc_w[i]);
265 		mdp5_write(mdp5_kms, REG_MDP5_SMP_ALLOC_R_REG(i),
266 			   smp->alloc_r[i]);
267 	}
268 }
269 
270 static void write_smp_fifo_regs(struct mdp5_smp *smp)
271 {
272 	struct mdp5_kms *mdp5_kms = get_kms(smp);
273 	int i;
274 
275 	for (i = 0; i < mdp5_kms->num_hwpipes; i++) {
276 		struct mdp5_hw_pipe *hwpipe = mdp5_kms->hwpipes[i];
277 		enum mdp5_pipe pipe = hwpipe->pipe;
278 
279 		mdp5_write(mdp5_kms, REG_MDP5_PIPE_REQPRIO_FIFO_WM_0(pipe),
280 			   smp->pipe_reqprio_fifo_wm0[pipe]);
281 		mdp5_write(mdp5_kms, REG_MDP5_PIPE_REQPRIO_FIFO_WM_1(pipe),
282 			   smp->pipe_reqprio_fifo_wm1[pipe]);
283 		mdp5_write(mdp5_kms, REG_MDP5_PIPE_REQPRIO_FIFO_WM_2(pipe),
284 			   smp->pipe_reqprio_fifo_wm2[pipe]);
285 	}
286 }
287 
288 void mdp5_smp_prepare_commit(struct mdp5_smp *smp, struct mdp5_smp_state *state)
289 {
290 	enum mdp5_pipe pipe;
291 
292 	for_each_set_bit(pipe, &state->assigned, sizeof(state->assigned) * 8) {
293 		unsigned i, nblks = 0;
294 
295 		for (i = 0; i < pipe2nclients(pipe); i++) {
296 			u32 cid = pipe2client(pipe, i);
297 			void *cs = state->client_state[cid];
298 
299 			nblks += update_smp_state(smp, cid, cs);
300 
301 			DBG("assign %s:%u, %u blks",
302 				pipe2name(pipe), i, nblks);
303 		}
304 
305 		set_fifo_thresholds(smp, pipe, nblks);
306 	}
307 
308 	write_smp_alloc_regs(smp);
309 	write_smp_fifo_regs(smp);
310 
311 	state->assigned = 0;
312 }
313 
314 void mdp5_smp_complete_commit(struct mdp5_smp *smp, struct mdp5_smp_state *state)
315 {
316 	enum mdp5_pipe pipe;
317 
318 	for_each_set_bit(pipe, &state->released, sizeof(state->released) * 8) {
319 		DBG("release %s", pipe2name(pipe));
320 		set_fifo_thresholds(smp, pipe, 0);
321 	}
322 
323 	write_smp_fifo_regs(smp);
324 
325 	state->released = 0;
326 }
327 
328 void mdp5_smp_dump(struct mdp5_smp *smp, struct drm_printer *p)
329 {
330 	struct mdp5_kms *mdp5_kms = get_kms(smp);
331 	struct mdp5_hw_pipe_state *hwpstate;
332 	struct mdp5_smp_state *state;
333 	struct mdp5_global_state *global_state;
334 	int total = 0, i, j;
335 
336 	drm_printf(p, "name\tinuse\tplane\n");
337 	drm_printf(p, "----\t-----\t-----\n");
338 
339 	if (drm_can_sleep())
340 		drm_modeset_lock(&mdp5_kms->glob_state_lock, NULL);
341 
342 	global_state = mdp5_get_existing_global_state(mdp5_kms);
343 
344 	/* grab these *after* we hold the state_lock */
345 	hwpstate = &global_state->hwpipe;
346 	state = &global_state->smp;
347 
348 	for (i = 0; i < mdp5_kms->num_hwpipes; i++) {
349 		struct mdp5_hw_pipe *hwpipe = mdp5_kms->hwpipes[i];
350 		struct drm_plane *plane = hwpstate->hwpipe_to_plane[hwpipe->idx];
351 		enum mdp5_pipe pipe = hwpipe->pipe;
352 		for (j = 0; j < pipe2nclients(pipe); j++) {
353 			u32 cid = pipe2client(pipe, j);
354 			void *cs = state->client_state[cid];
355 			int inuse = bitmap_weight(cs, smp->blk_cnt);
356 
357 			drm_printf(p, "%s:%d\t%d\t%s\n",
358 				pipe2name(pipe), j, inuse,
359 				plane ? plane->name : NULL);
360 
361 			total += inuse;
362 		}
363 	}
364 
365 	drm_printf(p, "TOTAL:\t%d\t(of %d)\n", total, smp->blk_cnt);
366 	drm_printf(p, "AVAIL:\t%d\n", smp->blk_cnt -
367 			bitmap_weight(state->state, smp->blk_cnt));
368 
369 	if (drm_can_sleep())
370 		drm_modeset_unlock(&mdp5_kms->glob_state_lock);
371 }
372 
373 void mdp5_smp_destroy(struct mdp5_smp *smp)
374 {
375 	kfree(smp);
376 }
377 
378 struct mdp5_smp *mdp5_smp_init(struct mdp5_kms *mdp5_kms, const struct mdp5_smp_block *cfg)
379 {
380 	struct mdp5_smp_state *state;
381 	struct mdp5_global_state *global_state;
382 	struct mdp5_smp *smp = NULL;
383 	int ret;
384 
385 	smp = kzalloc(sizeof(*smp), GFP_KERNEL);
386 	if (unlikely(!smp)) {
387 		ret = -ENOMEM;
388 		goto fail;
389 	}
390 
391 	smp->dev = mdp5_kms->dev;
392 	smp->blk_cnt = cfg->mmb_count;
393 	smp->blk_size = cfg->mmb_size;
394 
395 	global_state = mdp5_get_existing_global_state(mdp5_kms);
396 	state = &global_state->smp;
397 
398 	/* statically tied MMBs cannot be re-allocated: */
399 	bitmap_copy(state->state, cfg->reserved_state, smp->blk_cnt);
400 	memcpy(smp->reserved, cfg->reserved, sizeof(smp->reserved));
401 
402 	return smp;
403 fail:
404 	if (smp)
405 		mdp5_smp_destroy(smp);
406 
407 	return ERR_PTR(ret);
408 }
409