1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2019 Intel Corporation
4  *
5  */
6 
7 #include "gem/i915_gem_internal.h"
8 
9 #include "i915_drv.h"
10 #include "i915_reg.h"
11 #include "intel_de.h"
12 #include "intel_display_types.h"
13 #include "intel_dsb.h"
14 
15 struct i915_vma;
16 
17 enum dsb_id {
18 	INVALID_DSB = -1,
19 	DSB1,
20 	DSB2,
21 	DSB3,
22 	MAX_DSB_PER_PIPE
23 };
24 
25 struct intel_dsb {
26 	enum dsb_id id;
27 
28 	u32 *cmd_buf;
29 	struct i915_vma *vma;
30 	struct intel_crtc *crtc;
31 
32 	/*
33 	 * maximum number of dwords the buffer will hold.
34 	 */
35 	unsigned int size;
36 
37 	/*
38 	 * free_pos will point the first free dword and
39 	 * help in calculating tail of command buffer.
40 	 */
41 	unsigned int free_pos;
42 
43 	/*
44 	 * ins_start_offset will help to store start dword of the dsb
45 	 * instuction and help in identifying the batch of auto-increment
46 	 * register.
47 	 */
48 	unsigned int ins_start_offset;
49 };
50 
51 /**
52  * DOC: DSB
53  *
54  * A DSB (Display State Buffer) is a queue of MMIO instructions in the memory
55  * which can be offloaded to DSB HW in Display Controller. DSB HW is a DMA
56  * engine that can be programmed to download the DSB from memory.
57  * It allows driver to batch submit display HW programming. This helps to
58  * reduce loading time and CPU activity, thereby making the context switch
59  * faster. DSB Support added from Gen12 Intel graphics based platform.
60  *
61  * DSB's can access only the pipe, plane, and transcoder Data Island Packet
62  * registers.
63  *
64  * DSB HW can support only register writes (both indexed and direct MMIO
65  * writes). There are no registers reads possible with DSB HW engine.
66  */
67 
68 /* DSB opcodes. */
69 #define DSB_OPCODE_SHIFT		24
70 #define DSB_OPCODE_NOOP			0x0
71 #define DSB_OPCODE_MMIO_WRITE		0x1
72 #define DSB_OPCODE_WAIT_USEC		0x2
73 #define DSB_OPCODE_WAIT_LINES		0x3
74 #define DSB_OPCODE_WAIT_VBLANKS		0x4
75 #define DSB_OPCODE_WAIT_DSL_IN		0x5
76 #define DSB_OPCODE_WAIT_DSL_OUT		0x6
77 #define DSB_OPCODE_INTERRUPT		0x7
78 #define DSB_OPCODE_INDEXED_WRITE	0x9
79 #define DSB_OPCODE_POLL			0xA
80 #define DSB_BYTE_EN			0xF
81 #define DSB_BYTE_EN_SHIFT		20
82 #define DSB_REG_VALUE_MASK		0xfffff
83 
84 static bool assert_dsb_has_room(struct intel_dsb *dsb)
85 {
86 	struct intel_crtc *crtc = dsb->crtc;
87 	struct drm_i915_private *i915 = to_i915(crtc->base.dev);
88 
89 	/* each instruction is 2 dwords */
90 	return !drm_WARN(&i915->drm, dsb->free_pos > dsb->size - 2,
91 			 "DSB buffer overflow\n");
92 }
93 
94 static bool is_dsb_busy(struct drm_i915_private *i915, enum pipe pipe,
95 			enum dsb_id id)
96 {
97 	return intel_de_read(i915, DSB_CTRL(pipe, id)) & DSB_STATUS_BUSY;
98 }
99 
100 static void intel_dsb_emit(struct intel_dsb *dsb, u32 ldw, u32 udw)
101 {
102 	u32 *buf = dsb->cmd_buf;
103 
104 	if (!assert_dsb_has_room(dsb))
105 		return;
106 
107 	/* Every instruction should be 8 byte aligned. */
108 	dsb->free_pos = ALIGN(dsb->free_pos, 2);
109 
110 	dsb->ins_start_offset = dsb->free_pos;
111 
112 	buf[dsb->free_pos++] = ldw;
113 	buf[dsb->free_pos++] = udw;
114 }
115 
116 static bool intel_dsb_prev_ins_is_write(struct intel_dsb *dsb,
117 					u32 opcode, i915_reg_t reg)
118 {
119 	const u32 *buf = dsb->cmd_buf;
120 	u32 prev_opcode, prev_reg;
121 
122 	prev_opcode = buf[dsb->ins_start_offset + 1] >> DSB_OPCODE_SHIFT;
123 	prev_reg = buf[dsb->ins_start_offset + 1] & DSB_REG_VALUE_MASK;
124 
125 	return prev_opcode == opcode && prev_reg == i915_mmio_reg_offset(reg);
126 }
127 
128 static bool intel_dsb_prev_ins_is_mmio_write(struct intel_dsb *dsb, i915_reg_t reg)
129 {
130 	return intel_dsb_prev_ins_is_write(dsb, DSB_OPCODE_MMIO_WRITE, reg);
131 }
132 
133 static bool intel_dsb_prev_ins_is_indexed_write(struct intel_dsb *dsb, i915_reg_t reg)
134 {
135 	return intel_dsb_prev_ins_is_write(dsb, DSB_OPCODE_INDEXED_WRITE, reg);
136 }
137 
138 /**
139  * intel_dsb_reg_write() - Emit register wriite to the DSB context
140  * @dsb: DSB context
141  * @reg: register address.
142  * @val: value.
143  *
144  * This function is used for writing register-value pair in command
145  * buffer of DSB.
146  */
147 void intel_dsb_reg_write(struct intel_dsb *dsb,
148 			 i915_reg_t reg, u32 val)
149 {
150 	/*
151 	 * For example the buffer will look like below for 3 dwords for auto
152 	 * increment register:
153 	 * +--------------------------------------------------------+
154 	 * | size = 3 | offset &| value1 | value2 | value3 | zero   |
155 	 * |          | opcode  |        |        |        |        |
156 	 * +--------------------------------------------------------+
157 	 * +          +         +        +        +        +        +
158 	 * 0          4         8        12       16       20       24
159 	 * Byte
160 	 *
161 	 * As every instruction is 8 byte aligned the index of dsb instruction
162 	 * will start always from even number while dealing with u32 array. If
163 	 * we are writing odd no of dwords, Zeros will be added in the end for
164 	 * padding.
165 	 */
166 	if (!intel_dsb_prev_ins_is_mmio_write(dsb, reg) &&
167 	    !intel_dsb_prev_ins_is_indexed_write(dsb, reg)) {
168 		intel_dsb_emit(dsb, val,
169 			       (DSB_OPCODE_MMIO_WRITE << DSB_OPCODE_SHIFT) |
170 			       (DSB_BYTE_EN << DSB_BYTE_EN_SHIFT) |
171 			       i915_mmio_reg_offset(reg));
172 	} else {
173 		u32 *buf = dsb->cmd_buf;
174 
175 		if (!assert_dsb_has_room(dsb))
176 			return;
177 
178 		/* convert to indexed write? */
179 		if (intel_dsb_prev_ins_is_mmio_write(dsb, reg)) {
180 			u32 prev_val = buf[dsb->ins_start_offset + 0];
181 
182 			buf[dsb->ins_start_offset + 0] = 1; /* count */
183 			buf[dsb->ins_start_offset + 1] =
184 				(DSB_OPCODE_INDEXED_WRITE << DSB_OPCODE_SHIFT) |
185 				i915_mmio_reg_offset(reg);
186 			buf[dsb->ins_start_offset + 2] = prev_val;
187 
188 			dsb->free_pos++;
189 		}
190 
191 		buf[dsb->free_pos++] = val;
192 		/* Update the count */
193 		buf[dsb->ins_start_offset]++;
194 
195 		/* if number of data words is odd, then the last dword should be 0.*/
196 		if (dsb->free_pos & 0x1)
197 			buf[dsb->free_pos] = 0;
198 	}
199 }
200 
201 static u32 intel_dsb_align_tail(struct intel_dsb *dsb)
202 {
203 	u32 aligned_tail, tail;
204 
205 	tail = dsb->free_pos * 4;
206 	aligned_tail = ALIGN(tail, CACHELINE_BYTES);
207 
208 	if (aligned_tail > tail)
209 		memset(&dsb->cmd_buf[dsb->free_pos], 0,
210 		       aligned_tail - tail);
211 
212 	dsb->free_pos = aligned_tail / 4;
213 
214 	return aligned_tail;
215 }
216 
217 /**
218  * intel_dsb_commit() - Trigger workload execution of DSB.
219  * @dsb: DSB context
220  *
221  * This function is used to do actual write to hardware using DSB.
222  */
223 void intel_dsb_commit(struct intel_dsb *dsb)
224 {
225 	struct intel_crtc *crtc = dsb->crtc;
226 	struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
227 	enum pipe pipe = crtc->pipe;
228 	u32 tail;
229 
230 	tail = intel_dsb_align_tail(dsb);
231 	if (tail == 0)
232 		return;
233 
234 	if (is_dsb_busy(dev_priv, pipe, dsb->id)) {
235 		drm_err(&dev_priv->drm, "DSB engine is busy.\n");
236 		goto reset;
237 	}
238 
239 	intel_de_write(dev_priv, DSB_CTRL(pipe, dsb->id),
240 		       DSB_ENABLE);
241 	intel_de_write(dev_priv, DSB_HEAD(pipe, dsb->id),
242 		       i915_ggtt_offset(dsb->vma));
243 	intel_de_write(dev_priv, DSB_TAIL(pipe, dsb->id),
244 		       i915_ggtt_offset(dsb->vma) + tail);
245 
246 	drm_dbg_kms(&dev_priv->drm,
247 		    "DSB execution started - head 0x%x, tail 0x%x\n",
248 		    i915_ggtt_offset(dsb->vma),
249 		    i915_ggtt_offset(dsb->vma) + tail);
250 
251 	if (wait_for(!is_dsb_busy(dev_priv, pipe, dsb->id), 1))
252 		drm_err(&dev_priv->drm,
253 			"Timed out waiting for DSB workload completion.\n");
254 
255 reset:
256 	dsb->free_pos = 0;
257 	dsb->ins_start_offset = 0;
258 	intel_de_write(dev_priv, DSB_CTRL(pipe, dsb->id), 0);
259 }
260 
261 /**
262  * intel_dsb_prepare() - Allocate, pin and map the DSB command buffer.
263  * @crtc: the CRTC
264  * @max_cmds: number of commands we need to fit into command buffer
265  *
266  * This function prepare the command buffer which is used to store dsb
267  * instructions with data.
268  *
269  * Returns:
270  * DSB context, NULL on failure
271  */
272 struct intel_dsb *intel_dsb_prepare(struct intel_crtc *crtc,
273 				    unsigned int max_cmds)
274 {
275 	struct drm_i915_private *i915 = to_i915(crtc->base.dev);
276 	struct drm_i915_gem_object *obj;
277 	intel_wakeref_t wakeref;
278 	struct intel_dsb *dsb;
279 	struct i915_vma *vma;
280 	unsigned int size;
281 	u32 *buf;
282 
283 	if (!HAS_DSB(i915))
284 		return NULL;
285 
286 	dsb = kzalloc(sizeof(*dsb), GFP_KERNEL);
287 	if (!dsb)
288 		goto out;
289 
290 	wakeref = intel_runtime_pm_get(&i915->runtime_pm);
291 
292 	/* ~1 qword per instruction, full cachelines */
293 	size = ALIGN(max_cmds * 8, CACHELINE_BYTES);
294 
295 	obj = i915_gem_object_create_internal(i915, PAGE_ALIGN(size));
296 	if (IS_ERR(obj))
297 		goto out_put_rpm;
298 
299 	vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, 0);
300 	if (IS_ERR(vma)) {
301 		i915_gem_object_put(obj);
302 		goto out_put_rpm;
303 	}
304 
305 	buf = i915_gem_object_pin_map_unlocked(vma->obj, I915_MAP_WC);
306 	if (IS_ERR(buf)) {
307 		i915_vma_unpin_and_release(&vma, I915_VMA_RELEASE_MAP);
308 		goto out_put_rpm;
309 	}
310 
311 	intel_runtime_pm_put(&i915->runtime_pm, wakeref);
312 
313 	dsb->id = DSB1;
314 	dsb->vma = vma;
315 	dsb->crtc = crtc;
316 	dsb->cmd_buf = buf;
317 	dsb->size = size / 4; /* in dwords */
318 	dsb->free_pos = 0;
319 	dsb->ins_start_offset = 0;
320 
321 	return dsb;
322 
323 out_put_rpm:
324 	intel_runtime_pm_put(&i915->runtime_pm, wakeref);
325 	kfree(dsb);
326 out:
327 	drm_info_once(&i915->drm,
328 		      "DSB queue setup failed, will fallback to MMIO for display HW programming\n");
329 
330 	return NULL;
331 }
332 
333 /**
334  * intel_dsb_cleanup() - To cleanup DSB context.
335  * @dsb: DSB context
336  *
337  * This function cleanup the DSB context by unpinning and releasing
338  * the VMA object associated with it.
339  */
340 void intel_dsb_cleanup(struct intel_dsb *dsb)
341 {
342 	i915_vma_unpin_and_release(&dsb->vma, I915_VMA_RELEASE_MAP);
343 	kfree(dsb);
344 }
345