xref: /openbmc/linux/drivers/gpu/drm/gud/gud_pipe.c (revision dd21bfa4)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright 2020 Noralf Trønnes
4  */
5 
6 #include <linux/lz4.h>
7 #include <linux/usb.h>
8 #include <linux/workqueue.h>
9 
10 #include <drm/drm_atomic.h>
11 #include <drm/drm_connector.h>
12 #include <drm/drm_damage_helper.h>
13 #include <drm/drm_drv.h>
14 #include <drm/drm_format_helper.h>
15 #include <drm/drm_fourcc.h>
16 #include <drm/drm_framebuffer.h>
17 #include <drm/drm_gem.h>
18 #include <drm/drm_gem_framebuffer_helper.h>
19 #include <drm/drm_print.h>
20 #include <drm/drm_rect.h>
21 #include <drm/drm_simple_kms_helper.h>
22 #include <drm/gud.h>
23 
24 #include "gud_internal.h"
25 
26 /*
27  * Some userspace rendering loops runs all displays in the same loop.
28  * This means that a fast display will have to wait for a slow one.
29  * For this reason gud does flushing asynchronous by default.
30  * The down side is that in e.g. a single display setup userspace thinks
31  * the display is insanely fast since the driver reports back immediately
32  * that the flush/pageflip is done. This wastes CPU and power.
33  * Such users might want to set this module parameter to false.
34  */
35 static bool gud_async_flush = true;
36 module_param_named(async_flush, gud_async_flush, bool, 0644);
37 MODULE_PARM_DESC(async_flush, "Enable asynchronous flushing [default=true]");
38 
39 /*
40  * FIXME: The driver is probably broken on Big Endian machines.
41  * See discussion:
42  * https://lore.kernel.org/dri-devel/CAKb7UvihLX0hgBOP3VBG7O+atwZcUVCPVuBdfmDMpg0NjXe-cQ@mail.gmail.com/
43  */
44 
45 static bool gud_is_big_endian(void)
46 {
47 #if defined(__BIG_ENDIAN)
48 	return true;
49 #else
50 	return false;
51 #endif
52 }
53 
54 static size_t gud_xrgb8888_to_r124(u8 *dst, const struct drm_format_info *format,
55 				   void *src, struct drm_framebuffer *fb,
56 				   struct drm_rect *rect)
57 {
58 	unsigned int block_width = drm_format_info_block_width(format, 0);
59 	unsigned int bits_per_pixel = 8 / block_width;
60 	unsigned int x, y, width, height;
61 	u8 pix, *pix8, *block = dst; /* Assign to silence compiler warning */
62 	size_t len;
63 	void *buf;
64 
65 	WARN_ON_ONCE(format->char_per_block[0] != 1);
66 
67 	/* Start on a byte boundary */
68 	rect->x1 = ALIGN_DOWN(rect->x1, block_width);
69 	width = drm_rect_width(rect);
70 	height = drm_rect_height(rect);
71 	len = drm_format_info_min_pitch(format, 0, width) * height;
72 
73 	buf = kmalloc(width * height, GFP_KERNEL);
74 	if (!buf)
75 		return 0;
76 
77 	drm_fb_xrgb8888_to_gray8(buf, 0, src, fb, rect);
78 	pix8 = buf;
79 
80 	for (y = 0; y < height; y++) {
81 		for (x = 0; x < width; x++) {
82 			unsigned int pixpos = x % block_width; /* within byte from the left */
83 			unsigned int pixshift = (block_width - pixpos - 1) * bits_per_pixel;
84 
85 			if (!pixpos) {
86 				block = dst++;
87 				*block = 0;
88 			}
89 
90 			pix = (*pix8++) >> (8 - bits_per_pixel);
91 			*block |= pix << pixshift;
92 		}
93 	}
94 
95 	kfree(buf);
96 
97 	return len;
98 }
99 
100 static size_t gud_xrgb8888_to_color(u8 *dst, const struct drm_format_info *format,
101 				    void *src, struct drm_framebuffer *fb,
102 				    struct drm_rect *rect)
103 {
104 	unsigned int block_width = drm_format_info_block_width(format, 0);
105 	unsigned int bits_per_pixel = 8 / block_width;
106 	u8 r, g, b, pix, *block = dst; /* Assign to silence compiler warning */
107 	unsigned int x, y, width;
108 	u32 *pix32;
109 	size_t len;
110 
111 	/* Start on a byte boundary */
112 	rect->x1 = ALIGN_DOWN(rect->x1, block_width);
113 	width = drm_rect_width(rect);
114 	len = drm_format_info_min_pitch(format, 0, width) * drm_rect_height(rect);
115 
116 	for (y = rect->y1; y < rect->y2; y++) {
117 		pix32 = src + (y * fb->pitches[0]);
118 		pix32 += rect->x1;
119 
120 		for (x = 0; x < width; x++) {
121 			unsigned int pixpos = x % block_width; /* within byte from the left */
122 			unsigned int pixshift = (block_width - pixpos - 1) * bits_per_pixel;
123 
124 			if (!pixpos) {
125 				block = dst++;
126 				*block = 0;
127 			}
128 
129 			r = *pix32 >> 16;
130 			g = *pix32 >> 8;
131 			b = *pix32++;
132 
133 			switch (format->format) {
134 			case GUD_DRM_FORMAT_XRGB1111:
135 				pix = ((r >> 7) << 2) | ((g >> 7) << 1) | (b >> 7);
136 				break;
137 			default:
138 				WARN_ON_ONCE(1);
139 				return len;
140 			}
141 
142 			*block |= pix << pixshift;
143 		}
144 	}
145 
146 	return len;
147 }
148 
149 static int gud_prep_flush(struct gud_device *gdrm, struct drm_framebuffer *fb,
150 			  const struct drm_format_info *format, struct drm_rect *rect,
151 			  struct gud_set_buffer_req *req)
152 {
153 	struct dma_buf_attachment *import_attach = fb->obj[0]->import_attach;
154 	u8 compression = gdrm->compression;
155 	struct dma_buf_map map[DRM_FORMAT_MAX_PLANES];
156 	struct dma_buf_map map_data[DRM_FORMAT_MAX_PLANES];
157 	void *vaddr, *buf;
158 	size_t pitch, len;
159 	int ret = 0;
160 
161 	pitch = drm_format_info_min_pitch(format, 0, drm_rect_width(rect));
162 	len = pitch * drm_rect_height(rect);
163 	if (len > gdrm->bulk_len)
164 		return -E2BIG;
165 
166 	ret = drm_gem_fb_vmap(fb, map, map_data);
167 	if (ret)
168 		return ret;
169 
170 	vaddr = map_data[0].vaddr;
171 
172 	ret = drm_gem_fb_begin_cpu_access(fb, DMA_FROM_DEVICE);
173 	if (ret)
174 		goto vunmap;
175 retry:
176 	if (compression)
177 		buf = gdrm->compress_buf;
178 	else
179 		buf = gdrm->bulk_buf;
180 
181 	/*
182 	 * Imported buffers are assumed to be write-combined and thus uncached
183 	 * with slow reads (at least on ARM).
184 	 */
185 	if (format != fb->format) {
186 		if (format->format == GUD_DRM_FORMAT_R1) {
187 			len = gud_xrgb8888_to_r124(buf, format, vaddr, fb, rect);
188 			if (!len) {
189 				ret = -ENOMEM;
190 				goto end_cpu_access;
191 			}
192 		} else if (format->format == DRM_FORMAT_R8) {
193 			drm_fb_xrgb8888_to_gray8(buf, 0, vaddr, fb, rect);
194 		} else if (format->format == DRM_FORMAT_RGB332) {
195 			drm_fb_xrgb8888_to_rgb332(buf, 0, vaddr, fb, rect);
196 		} else if (format->format == DRM_FORMAT_RGB565) {
197 			drm_fb_xrgb8888_to_rgb565(buf, 0, vaddr, fb, rect, gud_is_big_endian());
198 		} else if (format->format == DRM_FORMAT_RGB888) {
199 			drm_fb_xrgb8888_to_rgb888(buf, 0, vaddr, fb, rect);
200 		} else {
201 			len = gud_xrgb8888_to_color(buf, format, vaddr, fb, rect);
202 		}
203 	} else if (gud_is_big_endian() && format->cpp[0] > 1) {
204 		drm_fb_swab(buf, 0, vaddr, fb, rect, !import_attach);
205 	} else if (compression && !import_attach && pitch == fb->pitches[0]) {
206 		/* can compress directly from the framebuffer */
207 		buf = vaddr + rect->y1 * pitch;
208 	} else {
209 		drm_fb_memcpy(buf, 0, vaddr, fb, rect);
210 	}
211 
212 	memset(req, 0, sizeof(*req));
213 	req->x = cpu_to_le32(rect->x1);
214 	req->y = cpu_to_le32(rect->y1);
215 	req->width = cpu_to_le32(drm_rect_width(rect));
216 	req->height = cpu_to_le32(drm_rect_height(rect));
217 	req->length = cpu_to_le32(len);
218 
219 	if (compression & GUD_COMPRESSION_LZ4) {
220 		int complen;
221 
222 		complen = LZ4_compress_default(buf, gdrm->bulk_buf, len, len, gdrm->lz4_comp_mem);
223 		if (complen <= 0) {
224 			compression = 0;
225 			goto retry;
226 		}
227 
228 		req->compression = GUD_COMPRESSION_LZ4;
229 		req->compressed_length = cpu_to_le32(complen);
230 	}
231 
232 end_cpu_access:
233 	drm_gem_fb_end_cpu_access(fb, DMA_FROM_DEVICE);
234 vunmap:
235 	drm_gem_fb_vunmap(fb, map);
236 
237 	return ret;
238 }
239 
240 struct gud_usb_bulk_context {
241 	struct timer_list timer;
242 	struct usb_sg_request sgr;
243 };
244 
245 static void gud_usb_bulk_timeout(struct timer_list *t)
246 {
247 	struct gud_usb_bulk_context *ctx = from_timer(ctx, t, timer);
248 
249 	usb_sg_cancel(&ctx->sgr);
250 }
251 
252 static int gud_usb_bulk(struct gud_device *gdrm, size_t len)
253 {
254 	struct gud_usb_bulk_context ctx;
255 	int ret;
256 
257 	ret = usb_sg_init(&ctx.sgr, gud_to_usb_device(gdrm), gdrm->bulk_pipe, 0,
258 			  gdrm->bulk_sgt.sgl, gdrm->bulk_sgt.nents, len, GFP_KERNEL);
259 	if (ret)
260 		return ret;
261 
262 	timer_setup_on_stack(&ctx.timer, gud_usb_bulk_timeout, 0);
263 	mod_timer(&ctx.timer, jiffies + msecs_to_jiffies(3000));
264 
265 	usb_sg_wait(&ctx.sgr);
266 
267 	if (!del_timer_sync(&ctx.timer))
268 		ret = -ETIMEDOUT;
269 	else if (ctx.sgr.status < 0)
270 		ret = ctx.sgr.status;
271 	else if (ctx.sgr.bytes != len)
272 		ret = -EIO;
273 
274 	destroy_timer_on_stack(&ctx.timer);
275 
276 	return ret;
277 }
278 
279 static int gud_flush_rect(struct gud_device *gdrm, struct drm_framebuffer *fb,
280 			  const struct drm_format_info *format, struct drm_rect *rect)
281 {
282 	struct gud_set_buffer_req req;
283 	size_t len, trlen;
284 	int ret;
285 
286 	drm_dbg(&gdrm->drm, "Flushing [FB:%d] " DRM_RECT_FMT "\n", fb->base.id, DRM_RECT_ARG(rect));
287 
288 	ret = gud_prep_flush(gdrm, fb, format, rect, &req);
289 	if (ret)
290 		return ret;
291 
292 	len = le32_to_cpu(req.length);
293 
294 	if (req.compression)
295 		trlen = le32_to_cpu(req.compressed_length);
296 	else
297 		trlen = len;
298 
299 	gdrm->stats_length += len;
300 	/* Did it wrap around? */
301 	if (gdrm->stats_length <= len && gdrm->stats_actual_length) {
302 		gdrm->stats_length = len;
303 		gdrm->stats_actual_length = 0;
304 	}
305 	gdrm->stats_actual_length += trlen;
306 
307 	if (!(gdrm->flags & GUD_DISPLAY_FLAG_FULL_UPDATE) || gdrm->prev_flush_failed) {
308 		ret = gud_usb_set(gdrm, GUD_REQ_SET_BUFFER, 0, &req, sizeof(req));
309 		if (ret)
310 			return ret;
311 	}
312 
313 	ret = gud_usb_bulk(gdrm, trlen);
314 	if (ret)
315 		gdrm->stats_num_errors++;
316 
317 	return ret;
318 }
319 
320 void gud_clear_damage(struct gud_device *gdrm)
321 {
322 	gdrm->damage.x1 = INT_MAX;
323 	gdrm->damage.y1 = INT_MAX;
324 	gdrm->damage.x2 = 0;
325 	gdrm->damage.y2 = 0;
326 }
327 
328 static void gud_add_damage(struct gud_device *gdrm, struct drm_rect *damage)
329 {
330 	gdrm->damage.x1 = min(gdrm->damage.x1, damage->x1);
331 	gdrm->damage.y1 = min(gdrm->damage.y1, damage->y1);
332 	gdrm->damage.x2 = max(gdrm->damage.x2, damage->x2);
333 	gdrm->damage.y2 = max(gdrm->damage.y2, damage->y2);
334 }
335 
336 static void gud_retry_failed_flush(struct gud_device *gdrm, struct drm_framebuffer *fb,
337 				   struct drm_rect *damage)
338 {
339 	/*
340 	 * pipe_update waits for the worker when the display mode is going to change.
341 	 * This ensures that the width and height is still the same making it safe to
342 	 * add back the damage.
343 	 */
344 
345 	mutex_lock(&gdrm->damage_lock);
346 	if (!gdrm->fb) {
347 		drm_framebuffer_get(fb);
348 		gdrm->fb = fb;
349 	}
350 	gud_add_damage(gdrm, damage);
351 	mutex_unlock(&gdrm->damage_lock);
352 
353 	/* Retry only once to avoid a possible storm in case of continues errors. */
354 	if (!gdrm->prev_flush_failed)
355 		queue_work(system_long_wq, &gdrm->work);
356 	gdrm->prev_flush_failed = true;
357 }
358 
359 void gud_flush_work(struct work_struct *work)
360 {
361 	struct gud_device *gdrm = container_of(work, struct gud_device, work);
362 	const struct drm_format_info *format;
363 	struct drm_framebuffer *fb;
364 	struct drm_rect damage;
365 	unsigned int i, lines;
366 	int idx, ret = 0;
367 	size_t pitch;
368 
369 	if (!drm_dev_enter(&gdrm->drm, &idx))
370 		return;
371 
372 	mutex_lock(&gdrm->damage_lock);
373 	fb = gdrm->fb;
374 	gdrm->fb = NULL;
375 	damage = gdrm->damage;
376 	gud_clear_damage(gdrm);
377 	mutex_unlock(&gdrm->damage_lock);
378 
379 	if (!fb)
380 		goto out;
381 
382 	format = fb->format;
383 	if (format->format == DRM_FORMAT_XRGB8888 && gdrm->xrgb8888_emulation_format)
384 		format = gdrm->xrgb8888_emulation_format;
385 
386 	/* Split update if it's too big */
387 	pitch = drm_format_info_min_pitch(format, 0, drm_rect_width(&damage));
388 	lines = drm_rect_height(&damage);
389 
390 	if (gdrm->bulk_len < lines * pitch)
391 		lines = gdrm->bulk_len / pitch;
392 
393 	for (i = 0; i < DIV_ROUND_UP(drm_rect_height(&damage), lines); i++) {
394 		struct drm_rect rect = damage;
395 
396 		rect.y1 += i * lines;
397 		rect.y2 = min_t(u32, rect.y1 + lines, damage.y2);
398 
399 		ret = gud_flush_rect(gdrm, fb, format, &rect);
400 		if (ret) {
401 			if (ret != -ENODEV && ret != -ECONNRESET &&
402 			    ret != -ESHUTDOWN && ret != -EPROTO) {
403 				bool prev_flush_failed = gdrm->prev_flush_failed;
404 
405 				gud_retry_failed_flush(gdrm, fb, &damage);
406 				if (!prev_flush_failed)
407 					dev_err_ratelimited(fb->dev->dev,
408 							    "Failed to flush framebuffer: error=%d\n", ret);
409 			}
410 			break;
411 		}
412 
413 		gdrm->prev_flush_failed = false;
414 	}
415 
416 	drm_framebuffer_put(fb);
417 out:
418 	drm_dev_exit(idx);
419 }
420 
421 static void gud_fb_queue_damage(struct gud_device *gdrm, struct drm_framebuffer *fb,
422 				struct drm_rect *damage)
423 {
424 	struct drm_framebuffer *old_fb = NULL;
425 
426 	mutex_lock(&gdrm->damage_lock);
427 
428 	if (fb != gdrm->fb) {
429 		old_fb = gdrm->fb;
430 		drm_framebuffer_get(fb);
431 		gdrm->fb = fb;
432 	}
433 
434 	gud_add_damage(gdrm, damage);
435 
436 	mutex_unlock(&gdrm->damage_lock);
437 
438 	queue_work(system_long_wq, &gdrm->work);
439 
440 	if (old_fb)
441 		drm_framebuffer_put(old_fb);
442 }
443 
444 int gud_pipe_check(struct drm_simple_display_pipe *pipe,
445 		   struct drm_plane_state *new_plane_state,
446 		   struct drm_crtc_state *new_crtc_state)
447 {
448 	struct gud_device *gdrm = to_gud_device(pipe->crtc.dev);
449 	struct drm_plane_state *old_plane_state = pipe->plane.state;
450 	const struct drm_display_mode *mode = &new_crtc_state->mode;
451 	struct drm_atomic_state *state = new_plane_state->state;
452 	struct drm_framebuffer *old_fb = old_plane_state->fb;
453 	struct drm_connector_state *connector_state = NULL;
454 	struct drm_framebuffer *fb = new_plane_state->fb;
455 	const struct drm_format_info *format = fb->format;
456 	struct drm_connector *connector;
457 	unsigned int i, num_properties;
458 	struct gud_state_req *req;
459 	int idx, ret;
460 	size_t len;
461 
462 	if (WARN_ON_ONCE(!fb))
463 		return -EINVAL;
464 
465 	if (old_plane_state->rotation != new_plane_state->rotation)
466 		new_crtc_state->mode_changed = true;
467 
468 	if (old_fb && old_fb->format != format)
469 		new_crtc_state->mode_changed = true;
470 
471 	if (!new_crtc_state->mode_changed && !new_crtc_state->connectors_changed)
472 		return 0;
473 
474 	/* Only one connector is supported */
475 	if (hweight32(new_crtc_state->connector_mask) != 1)
476 		return -EINVAL;
477 
478 	if (format->format == DRM_FORMAT_XRGB8888 && gdrm->xrgb8888_emulation_format)
479 		format = gdrm->xrgb8888_emulation_format;
480 
481 	for_each_new_connector_in_state(state, connector, connector_state, i) {
482 		if (connector_state->crtc)
483 			break;
484 	}
485 
486 	/*
487 	 * DRM_IOCTL_MODE_OBJ_SETPROPERTY on the rotation property will not have
488 	 * the connector included in the state.
489 	 */
490 	if (!connector_state) {
491 		struct drm_connector_list_iter conn_iter;
492 
493 		drm_connector_list_iter_begin(pipe->crtc.dev, &conn_iter);
494 		drm_for_each_connector_iter(connector, &conn_iter) {
495 			if (connector->state->crtc) {
496 				connector_state = connector->state;
497 				break;
498 			}
499 		}
500 		drm_connector_list_iter_end(&conn_iter);
501 	}
502 
503 	if (WARN_ON_ONCE(!connector_state))
504 		return -ENOENT;
505 
506 	len = struct_size(req, properties,
507 			  GUD_PROPERTIES_MAX_NUM + GUD_CONNECTOR_PROPERTIES_MAX_NUM);
508 	req = kzalloc(len, GFP_KERNEL);
509 	if (!req)
510 		return -ENOMEM;
511 
512 	gud_from_display_mode(&req->mode, mode);
513 
514 	req->format = gud_from_fourcc(format->format);
515 	if (WARN_ON_ONCE(!req->format)) {
516 		ret = -EINVAL;
517 		goto out;
518 	}
519 
520 	req->connector = drm_connector_index(connector_state->connector);
521 
522 	ret = gud_connector_fill_properties(connector_state, req->properties);
523 	if (ret < 0)
524 		goto out;
525 
526 	num_properties = ret;
527 	for (i = 0; i < gdrm->num_properties; i++) {
528 		u16 prop = gdrm->properties[i];
529 		u64 val;
530 
531 		switch (prop) {
532 		case GUD_PROPERTY_ROTATION:
533 			/* DRM UAPI matches the protocol so use value directly */
534 			val = new_plane_state->rotation;
535 			break;
536 		default:
537 			WARN_ON_ONCE(1);
538 			ret = -EINVAL;
539 			goto out;
540 		}
541 
542 		req->properties[num_properties + i].prop = cpu_to_le16(prop);
543 		req->properties[num_properties + i].val = cpu_to_le64(val);
544 		num_properties++;
545 	}
546 
547 	if (drm_dev_enter(fb->dev, &idx)) {
548 		len = struct_size(req, properties, num_properties);
549 		ret = gud_usb_set(gdrm, GUD_REQ_SET_STATE_CHECK, 0, req, len);
550 		drm_dev_exit(idx);
551 	}  else {
552 		ret = -ENODEV;
553 	}
554 out:
555 	kfree(req);
556 
557 	return ret;
558 }
559 
560 void gud_pipe_update(struct drm_simple_display_pipe *pipe,
561 		     struct drm_plane_state *old_state)
562 {
563 	struct drm_device *drm = pipe->crtc.dev;
564 	struct gud_device *gdrm = to_gud_device(drm);
565 	struct drm_plane_state *state = pipe->plane.state;
566 	struct drm_framebuffer *fb = state->fb;
567 	struct drm_crtc *crtc = &pipe->crtc;
568 	struct drm_rect damage;
569 	int idx;
570 
571 	if (crtc->state->mode_changed || !crtc->state->enable) {
572 		cancel_work_sync(&gdrm->work);
573 		mutex_lock(&gdrm->damage_lock);
574 		if (gdrm->fb) {
575 			drm_framebuffer_put(gdrm->fb);
576 			gdrm->fb = NULL;
577 		}
578 		gud_clear_damage(gdrm);
579 		mutex_unlock(&gdrm->damage_lock);
580 	}
581 
582 	if (!drm_dev_enter(drm, &idx))
583 		return;
584 
585 	if (!old_state->fb)
586 		gud_usb_set_u8(gdrm, GUD_REQ_SET_CONTROLLER_ENABLE, 1);
587 
588 	if (fb && (crtc->state->mode_changed || crtc->state->connectors_changed))
589 		gud_usb_set(gdrm, GUD_REQ_SET_STATE_COMMIT, 0, NULL, 0);
590 
591 	if (crtc->state->active_changed)
592 		gud_usb_set_u8(gdrm, GUD_REQ_SET_DISPLAY_ENABLE, crtc->state->active);
593 
594 	if (drm_atomic_helper_damage_merged(old_state, state, &damage)) {
595 		if (gdrm->flags & GUD_DISPLAY_FLAG_FULL_UPDATE)
596 			drm_rect_init(&damage, 0, 0, fb->width, fb->height);
597 		gud_fb_queue_damage(gdrm, fb, &damage);
598 		if (!gud_async_flush)
599 			flush_work(&gdrm->work);
600 	}
601 
602 	if (!crtc->state->enable)
603 		gud_usb_set_u8(gdrm, GUD_REQ_SET_CONTROLLER_ENABLE, 0);
604 
605 	drm_dev_exit(idx);
606 }
607